diff --git a/CMakeLists.txt b/CMakeLists.txt index df00e977ebb547980e69ee421779c57717d771a9..6aa2e1715b92d73aa4e5e97d5e52ffac51451d80 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -127,6 +127,9 @@ set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING "A path setting fluid shared and static libraries") +set(FLUID_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_inference_install_dir" CACHE STRING + "A path setting fluid inference shared and static libraries") + if (WITH_C_API AND WITH_PYTHON) message(WARNING "It is suggest not embedded a python interpreter in Paddle " "when using C-API. It will give an unpredictable behavior when using a " diff --git a/README.md b/README.md index 46fdef5e376d3f5bf49ef10c62f5b3a6637913c1..8ee67f66423df8bce27f70015be8752457cd9784 100644 --- a/README.md +++ b/README.md @@ -2,8 +2,8 @@ [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle) -[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html) -[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html) +[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.0/getstarted/index_en.html) +[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) @@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle. Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle. -### Latest PaddlePaddle Release: [Fluid 0.15.0](https://github.com/PaddlePaddle/Paddle/tree/v0.15.0) +### Latest PaddlePaddle Release: [Fluid 1.0.1](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0) ### Install Latest Stable Release: ``` # Linux CPU @@ -27,9 +27,9 @@ pip install paddlepaddle # Linux GPU cuda9cudnn7 pip install paddlepaddle-gpu # Linux GPU cuda8cudnn7 -pip install paddlepaddle-gpu==0.15.0.post87 +pip install paddlepaddle-gpu==1.0.1.post87 # Linux GPU cuda8cudnn5 -pip install paddlepaddle-gpu==0.15.0.post85 +pip install paddlepaddle-gpu==1.0.1.post85 # For installation on other platform, refer to http://paddlepaddle.org/ ``` @@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==0.15.0.post85 ## Installation -It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/install/install_doc.html) on our website. +It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) on our website. ## Documentation -We provide [English](http://paddlepaddle.org/documentation/docs/en/0.15.0/getstarted/index_en.html) and -[Chinese](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/index.html) documentation. +We provide [English](http://paddlepaddle.org/documentation/docs/en/1.0.0/getstarted/index_en.html) and +[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) documentation. - [Deep Learning 101](https://github.com/PaddlePaddle/book) You might want to start from this online interactive book that can run in a Jupyter Notebook. -- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/user_guides/howto/training/cluster_howto.html) +- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.0/user_guides/howto/training/cluster_howto.html) You can run distributed training jobs on MPI clusters. -- [Python API](http://paddlepaddle.org/documentation/api/zh/0.15.0/fluid.html) +- [Python API](http://paddlepaddle.org/documentation/api/zh/1.0/fluid.html) Our new API enables much shorter programs. -- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/advanced_usage/development/contribute_to_paddle.html) +- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.0/advanced_usage/development/contribute_to_paddle.html) We appreciate your contributions! diff --git a/benchmark/fluid/run.sh b/benchmark/fluid/run.sh old mode 100644 new mode 100755 diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 5bf82b4ddf10a646ca540ac4ee2cfd3d3bc6cf58..34581e43e86631a556f03ef08fc424698b4a99dc 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -311,6 +311,8 @@ function(cc_test TARGET_NAME) set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true) set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true) set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true) + # No unit test should exceed 10 minutes. + set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600) endif() endfunction(cc_test) @@ -629,6 +631,8 @@ function(py_test TARGET_NAME) PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS} ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) + # No unit test should exceed 10 minutes. + set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600) endif() endfunction() diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index 077072f6eadb0c48f4ae32f94828613d89ed01c9..67cca09b64c1ed7a503a886e78347d786eae0de7 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -18,7 +18,7 @@ function(copy TARGET) set(oneValueArgs "") set(multiValueArgs SRCS DSTS DEPS) cmake_parse_arguments(copy_lib "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - set(inference_lib_dist_dep ${TARGET} ${inference_lib_dist_dep} PARENT_SCOPE) + set(fluid_lib_dist_dep ${TARGET} ${fluid_lib_dist_dep} PARENT_SCOPE) list(LENGTH copy_lib_SRCS copy_lib_SRCS_len) list(LENGTH copy_lib_DSTS copy_lib_DSTS_len) @@ -150,16 +150,16 @@ if (WITH_ANAKIN AND WITH_MKL) SRCS ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api ${ANAKIN_INSTALL_DIR} # anakin release - DSTS ${dst_dir}/inference/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin) + DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin) list(APPEND inference_deps anakin_inference_lib) endif() set(module "inference") copy(inference_lib DEPS ${inference_deps} SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.* - ${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci + ${src_dir}/${module}/api/paddle_inference_api.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h - DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} + DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ) set(module "platform") @@ -185,20 +185,41 @@ copy(cmake_cache SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt DSTS ${FLUID_INSTALL_DIR}) -add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) +# This command generates a complete fluid library for both train and inference +add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep}) + +# Following commands generate a inference-only fluid library +# third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR} +copy(third_party DEPS fluid_lib_dist + SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt + DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR} +) + +# only need libpaddle_fluid.so/a and paddle_inference_api.h for inference-only library +copy(inference_api_lib DEPS fluid_lib_dist + SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.* + ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_inference_api.h + DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include +) + +add_custom_target(inference_lib_dist DEPENDS third_party inference_api_lib) # paddle fluid version -execute_process( - COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1 - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} - OUTPUT_VARIABLE PADDLE_GIT_COMMIT) -set(version_file ${FLUID_INSTALL_DIR}/version.txt) -file(WRITE ${version_file} - "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n" - "WITH_MKL: ${WITH_MKL}\n" - "WITH_GPU: ${WITH_GPU}\n") -if(WITH_GPU) - file(APPEND ${version_file} - "CUDA version: ${CUDA_VERSION}\n" - "CUDNN version: v${CUDNN_MAJOR_VERSION}\n") -endif() +function(version version_file) + execute_process( + COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1 + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} + OUTPUT_VARIABLE PADDLE_GIT_COMMIT) + file(WRITE ${version_file} + "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n" + "WITH_MKL: ${WITH_MKL}\n" + "WITH_MKLDNN: ${WITH_MKLDNN}\n" + "WITH_GPU: ${WITH_GPU}\n") + if(WITH_GPU) + file(APPEND ${version_file} + "CUDA version: ${CUDA_VERSION}\n" + "CUDNN version: v${CUDNN_MAJOR_VERSION}\n") + endif() +endfunction() +version(${FLUID_INSTALL_DIR}/version.txt) +version(${FLUID_INFERENCE_INSTALL_DIR}/version.txt) diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index d0ae80274690dcf79b0585948cf9bf89e943c383..19ef23cdfa90912ff6fbd050a685d10861d909d2 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -61,12 +61,12 @@ paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100)) paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None)) +paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None)) paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None)) paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None)) paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'], varargs=None, keywords=None, defaults=None) -paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn'], varargs=None, keywords=None, defaults=(None, None, False)) -paddle.fluid.layers.softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None)) +paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None)) +paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None)) paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None)) paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None)) paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False)) @@ -75,7 +75,8 @@ paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'outp paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None)) paddle.fluid.layers.sequence_expand_as ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen', 'name'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.sequence_unpad ArgSpec(args=['x', 'length', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None)) paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) @@ -84,6 +85,7 @@ paddle.fluid.layers.reduce_min ArgSpec(args=['input', 'dim', 'keep_dim', 'name'] paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None)) paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None)) paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,)) @@ -95,8 +97,8 @@ paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_ti paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None)) -paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples'], varargs=None, keywords=None, defaults=(None, None, None, None)) -paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None)) +paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None)) paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None) @@ -114,6 +116,7 @@ paddle.fluid.layers.pad ArgSpec(args=['x', 'paddings', 'pad_value', 'name'], var paddle.fluid.layers.pad_constant_like ArgSpec(args=['x', 'y', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None)) paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 0.1, 'float32', None)) paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)) +paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None)) paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)) paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR')) paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)) @@ -127,6 +130,7 @@ paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None paddle.fluid.layers.log ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.margin_rank_loss ArgSpec(args=['label', 'left', 'right', 'margin', 'name'], varargs=None, keywords=None, defaults=(0.1, None)) paddle.fluid.layers.elu ArgSpec(args=['x', 'alpha', 'name'], varargs=None, keywords=None, defaults=(1.0, None)) paddle.fluid.layers.relu6 ArgSpec(args=['x', 'threshold', 'name'], varargs=None, keywords=None, defaults=(6.0, None)) paddle.fluid.layers.pow ArgSpec(args=['x', 'factor', 'name'], varargs=None, keywords=None, defaults=(1.0, None)) @@ -170,6 +174,7 @@ paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None)) paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None)) paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)) paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None)) paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None) @@ -178,7 +183,6 @@ paddle.fluid.layers.batch ArgSpec(args=['reader', 'batch_size'], varargs=None, k paddle.fluid.layers.double_buffer ArgSpec(args=['reader', 'place', 'name'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.layers.random_data_generator ArgSpec(args=['low', 'high', 'shapes', 'lod_levels', 'for_parallel'], varargs=None, keywords=None, defaults=(True,)) paddle.fluid.layers.py_reader ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)) -paddle.fluid.layers.create_py_reader_by_data ArgSpec(args=['capacity', 'feed_list', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, True)) paddle.fluid.layers.Preprocessor.__init__ ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.Preprocessor.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.layers.Preprocessor.inputs ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index 9fbefabc841e3f6940860f60d959fee97495e4c9..d09b94a3fd32952985a37cf4246c7640d2db4f56 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -64,7 +64,8 @@ class OpHandleBase { virtual bool IsMultiDeviceTransfer() { return false; } const platform::DeviceContext *DeviceContext(platform::Place place) { - return dev_ctxes_[place]; + auto it = dev_ctxes_.find(place); + return it != dev_ctxes_.end() ? it->second : nullptr; } void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) { diff --git a/paddle/fluid/framework/details/var_handle.h b/paddle/fluid/framework/details/var_handle.h index d8c2bc40b9458a1d5a7dd8a32277d04f69295f09..a1f458c660ce9f73bc9ac2ed194091ad0b8f8400 100644 --- a/paddle/fluid/framework/details/var_handle.h +++ b/paddle/fluid/framework/details/var_handle.h @@ -49,6 +49,8 @@ struct VarHandleBase { void AddOutput(OpHandleBase* out, ir::Node* node) { if (pending_ops_.find(out) == pending_ops_.end()) { + PADDLE_ENFORCE(out != nullptr, "The output of %s should not be nullptr", + this->Node()->Name()); pending_ops_.insert(out); node_->outputs.push_back(node); } diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 8f6a5e88583829f6fd5a174c40af73b8d20bc09c..9095119efce49a5a789f70fb1e677a8d02e5b5fb 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -46,6 +46,41 @@ ExecutorPrepareContext::~ExecutorPrepareContext() { VLOG(5) << "destroy ExecutorPrepareContext"; } +template +static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op, + GarbageCollector* gc, + RefCntMap* ref_cnts) { + std::unordered_set erase_tensors; + + auto handler = [&](const VariableNameMap& name_map) { + for (auto& name_pair : name_map) { + for (auto& name : name_pair.second) { + auto it = ref_cnts->find(name); + if (it == ref_cnts->end()) continue; + if ((it->second)-- == 1) { + auto* var = scope.FindVar(name); + if (var != nullptr) { + VLOG(10) << "Erase tensor \'" << name << "\'"; + if (var->IsType()) { + erase_tensors.insert(var->GetMutable()); + } else if (var->IsType()) { + erase_tensors.insert( + var->GetMutable()->mutable_value()); + } + } + } + } + } + }; + + handler(op->Inputs()); + handler(op->Outputs()); + + if (!erase_tensors.empty()) { + gc->Add(erase_tensors); + } +} + Executor::Executor(const platform::Place& place) : place_(place) {} void Executor::Close() { @@ -66,7 +101,7 @@ void InitializeVariable(Variable* var, proto::VarType::Type var_type) { } else if (var_type == proto::VarType::FETCH_LIST) { var->GetMutable(); } else if (var_type == proto::VarType::STEP_SCOPES) { - var->GetMutable>(); + var->GetMutable>(); } else if (var_type == proto::VarType::LOD_RANK_TABLE) { var->GetMutable(); } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) { @@ -331,9 +366,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, } int64_t max_memory_size = GetEagerDeletionThreshold(); - std::unique_ptr> gc; - if (max_memory_size >= 0) { + // WhileOp would set keep_kids to false + // WhileGradOp would need the scopes created in WhileOp + // Perhaps, we should not perform eager deletion in WhileOp + // The scopes and variables created by WhileOp would be deleted + // in WhileGradOp. + if (max_memory_size >= 0 && !keep_kids) { ctx->ResetReferenceCount(); #ifdef PADDLE_WITH_CUDA if (platform::is_gpu_place(place_)) { @@ -352,45 +391,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, op->Run(*local_scope, place_); if (gc != nullptr) { - std::vector erase_vars; - for (auto& input : op->Inputs()) { - for (auto& input_name : input.second) { - auto it = ctx->cur_ref_cnts_.find(input_name); - if (it == ctx->cur_ref_cnts_.end()) continue; - if (it->second == 1) { // should delete it - erase_vars.emplace_back(input_name); - ctx->cur_ref_cnts_.erase(input_name); - } else { - --(it->second); - } - } - } - - for (auto& output : op->Outputs()) { - for (auto& output_name : output.second) { - auto it = ctx->cur_ref_cnts_.find(output_name); - if (it == ctx->cur_ref_cnts_.end()) continue; - if (it->second == 1) { - erase_vars.emplace_back(output_name); - ctx->cur_ref_cnts_.erase(output_name); - } else { - --(it->second); - } - } - } - - if (!erase_vars.empty()) { - std::vector erase_tensors; - for (auto& name : erase_vars) { - auto* var = local_scope->FindVar(name); - if (var == nullptr) continue; - if (var->IsType()) { - auto* tensor = var->GetMutable(); - erase_tensors.push_back(tensor); - } - } - if (!erase_tensors.empty()) gc->Add(erase_tensors); - } + DeleteUnusedTensors(*local_scope, op.get(), gc.get(), + &(ctx->cur_ref_cnts_)); } } diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index f0cc1338a8af50030a70a9797cbcd1b0567272b5..36b36d49c2728dbef93042158dffa26d8f56d529 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -32,38 +32,32 @@ template std::unordered_map GetNonPersistableReferenceCount( const ProgramDesc& prog, size_t block_id) { auto& block = prog.Block(block_id); - std::unordered_set ignored_vars; std::unordered_map ref_cnts; - for (auto var_desc : block.AllVars()) { - auto type = var_desc->Proto()->type().type(); - if (type != proto::VarType::LOD_TENSOR || var_desc->Persistable()) { - ignored_vars.insert(var_desc->Name()); // ignore persistable vars - } - } - - for (auto op_desc : block.AllOps()) { - for (auto& input : op_desc->Inputs()) { - for (auto& input_name : input.second) { - if (!ignored_vars.count(input_name)) { - if (ref_cnts.count(input_name)) - ++ref_cnts[input_name]; - else - ref_cnts[input_name] = 1; + auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) { + for (auto& name_pair : name_map) { + for (auto& name : name_pair.second) { + auto* var_desc = block.FindVar(name); + if (var_desc == nullptr || var_desc->Persistable()) continue; + auto type = var_desc->Proto()->type().type(); + if (type != proto::VarType::LOD_TENSOR && + type != proto::VarType::SELECTED_ROWS) { + continue; } - } - } - for (auto& output : op_desc->Outputs()) { - for (auto output_name : output.second) { - if (!ignored_vars.count(output_name)) { - if (ref_cnts.count(output_name)) - ++ref_cnts[output_name]; - else - ref_cnts[output_name] = 1; + auto it = ref_cnts.find(name); + if (it != ref_cnts.end()) { + ++it->second; + } else { + ref_cnts[name] = 1; } } } + }; + + for (auto op_desc : block.AllOps()) { + update_ref_cnts(op_desc, op_desc->Inputs()); + update_ref_cnts(op_desc, op_desc->Outputs()); } return ref_cnts; } diff --git a/paddle/fluid/framework/feed_fetch_method.cc b/paddle/fluid/framework/feed_fetch_method.cc index 8e1f93c5ebd448903d70f9668539e077875836e4..3e9353f5cf67d8de62c5551f12ea786e49190549 100644 --- a/paddle/fluid/framework/feed_fetch_method.cc +++ b/paddle/fluid/framework/feed_fetch_method.cc @@ -27,8 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input, // be created. VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index; Variable* g_feed_value = scope->Var(var_name); - auto& feed_inputs = - *(g_feed_value->GetMutable>()); + auto& feed_inputs = *(g_feed_value->GetMutable()); if (index >= feed_inputs.size()) { feed_inputs.resize(index + 1); } diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index 0076a8bece31f9a977b375717c25688fc0c95819..3aa2c7b9ea013dd977ef0051700df54e26a81307 100644 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -10,7 +10,7 @@ function(pass_library TARGET DEST) set(oneValueArgs "") set(multiValueArgs SRCS DEPS) cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass ${op_library_DEPS}) + cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS}) # add more DEST here, such as train, dist and collect USE_PASS into a file automatically. if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference") message(STATUS "add pass ${TARGET} ${DEST}") @@ -25,21 +25,29 @@ cc_library(graph_helper SRCS graph_helper.cc DEPS graph) cc_library(pass SRCS pass.cc DEPS graph node graph_helper) cc_library(graph_traits SRCS graph_traits.cc DEPS graph) cc_library(graph_pattern_detector SRCS graph_pattern_detector.cc DEPS graph graph_helper graph_traits) +cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass) pass_library(graph_to_program_pass base) pass_library(graph_viz_pass base) pass_library(fc_fuse_pass inference) -if (WITH_MKLDNN) - pass_library(conv_relu_mkldnn_fuse_pass inference) -endif () pass_library(attention_lstm_fuse_pass inference) pass_library(infer_clean_graph_pass inference) pass_library(fc_lstm_fuse_pass inference) pass_library(embedding_fc_lstm_fuse_pass inference) pass_library(fc_gru_fuse_pass inference) pass_library(seq_concat_fc_fuse_pass inference) +pass_library(conv_bn_fuse_pass inference) +pass_library(seqconv_eltadd_relu_fuse_pass inference) +if(WITH_MKLDNN) + pass_library(mkldnn_placement_pass base) + pass_library(conv_bias_mkldnn_fuse_pass inference) + pass_library(conv_relu_mkldnn_fuse_pass inference) +endif() cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector ) +if(WITH_MKLDNN) + pass_library(conv_elementwise_add_mkldnn_fuse_pass inference) +endif() set(GLOB_PASS_LIB ${PASS_LIBRARY} CACHE INTERNAL "Global PASS library") @@ -53,4 +61,5 @@ cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS g cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto) if (WITH_MKLDNN) cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass) + cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass) endif () diff --git a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc index 1c75cb5a82029b6a542a3a2f031a353f5e40f4ea..6090f1fe76a49dddad0640123b1fa4db8c489634 100644 --- a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc +++ b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc @@ -262,7 +262,7 @@ std::unique_ptr AttentionLSTMFusePass::ApplyImpl( std::unordered_set specified_vars({"data_lod_attention", "cell_init", "hidden_init", "data", "week", "minute"}); - int count = 0; + size_t count = 0; for (auto* node : graph->Nodes()) { if (node->IsVar() && specified_vars.count(node->Name())) { ++count; diff --git a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..449cc78be15bcd2575ce2e6846b41e475f8921f6 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc @@ -0,0 +1,137 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h" +#include +#include +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace framework { +namespace ir { + +template +LoDTensor tensor_apply_eltwise(const LoDTensor& vec_a, const LoDTensor& vec_b, + BinaryOperation f) { + PADDLE_ENFORCE_EQ(vec_a.dims(), vec_b.dims()); + LoDTensor vec_y; + vec_y.Resize(vec_a.dims()); + const float* a = vec_a.data(); + const float* b = vec_b.data(); + float* y = vec_y.mutable_data(platform::CPUPlace()); + for (int i = 0; i < vec_a.numel(); i++) { + y[i] = f(a[i], b[i]); + } + return vec_y; +} + +std::unique_ptr ConvBiasFusePass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init(name_scope_, graph.get()); + + auto* scope = param_scope(); + PADDLE_ENFORCE(scope); + + GraphPatternDetector gpd; + auto* conv_input = + gpd.mutable_pattern() + ->NewNode(patterns::PDNodeName(name_scope_, "conv_input")) + ->AsInput() + ->assert_is_op_input("conv2d", "Input"); + patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_); + conv_bias_pattern(conv_input); + int found_conv_bias_count = 0; + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(4) << "handle ConvBias fuse"; + GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight, + conv_bias_pattern); // Filter + GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_bias_pattern); // tmp + GET_IR_NODE_FROM_SUBGRAPH(conv, conv, conv_bias_pattern); // CONV op + // bias + GET_IR_NODE_FROM_SUBGRAPH(eltwise_bias, eltwise_bias, conv_bias_pattern); + // output + GET_IR_NODE_FROM_SUBGRAPH(eltwise_out, eltwise_out, conv_bias_pattern); + // elementwise_add op + GET_IR_NODE_FROM_SUBGRAPH(eltwise, eltwise, conv_bias_pattern); + + PADDLE_ENFORCE(subgraph.count(conv_input)); + + // check if fuse can be done and if MKL-DNN should be used + FuseOptions fuse_option = FindFuseOption(*conv, *eltwise); + if (fuse_option == DO_NOT_FUSE || fuse_option == FUSE_NATIVE) { + VLOG(3) << "do not perform conv+bias fuse"; + return; + } + + auto* eltwise_bias_tensor = + scope->FindVar(eltwise_bias->Name())->GetMutable(); + + auto input_names = conv->Op()->InputNames(); + bool has_bias = std::find(input_names.begin(), input_names.end(), "Bias") != + input_names.end(); + if (has_bias && conv->Op()->Input("Bias").size() > 0) { + auto conv_bias_names = conv->Op()->Input("Bias"); + // add eltwise bias to existing conv bias + PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1); + auto* conv_bias_var = scope->FindVar(conv_bias_names[0]); + auto* conv_bias_tensor = conv_bias_var->GetMutable(); + PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(), eltwise_bias_tensor->dims()); + *conv_bias_tensor = tensor_apply_eltwise( + *conv_bias_tensor, *eltwise_bias_tensor, std::plus()); + + conv->Op()->SetOutput("Output", + std::vector({eltwise_out->Name()})); + + GraphSafeRemoveNodes(graph.get(), {eltwise, conv_out}); + + IR_NODE_LINK_TO(conv, eltwise_out); + } else { + // take eltwise bias as conv bias + OpDesc desc; + + desc.SetInput( + "Input", std::vector({subgraph.at(conv_input)->Name()})); + desc.SetInput("Filter", std::vector({conv_weight->Name()})); + desc.SetInput("Bias", std::vector({eltwise_bias->Name()})); + desc.SetOutput("Output", std::vector({eltwise_out->Name()})); + desc.SetType("conv2d"); + + for (auto& attr : conv->Op()->GetAttrMap()) { + desc.SetAttr(attr.first, attr.second); + } + auto conv_bias_node = g->CreateOpNode(&desc); + + IR_NODE_LINK_TO(subgraph.at(conv_input), conv_bias_node); + IR_NODE_LINK_TO(conv_weight, conv_bias_node); + IR_NODE_LINK_TO(eltwise_bias, conv_bias_node); + IR_NODE_LINK_TO(conv_bias_node, eltwise_out); + + GraphSafeRemoveNodes(graph.get(), {conv, eltwise, conv_out}); + } + + found_conv_bias_count++; + }; + gpd(graph.get(), handler); + AddStatis(found_conv_bias_count); + return graph; +} +} // namespace ir +} // namespace framework +} // namespace paddle +REGISTER_PASS(conv_bias_mkldnn_fuse_pass, + paddle::framework::ir::ConvBiasFusePass); diff --git a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..5775b83b88730ec298c421a15f5c0b83c27b0750 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h @@ -0,0 +1,36 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#pragma once +#include +#include "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" +#include "paddle/fluid/framework/ir/pass.h" +namespace paddle { +namespace framework { +namespace ir { +/* +* Fuse the Conv and Elementwise_add to a ConvBiasOp. +*/ +class ConvBiasFusePass : public FusePassBase { + public: + virtual ~ConvBiasFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + const std::string name_scope_{"conv_bias_mkldnn_fuse"}; +}; +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..846a14e365e6bd7f056d409130a3b246371931da --- /dev/null +++ b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc @@ -0,0 +1,298 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/conv_bn_fuse_pass.h" +#include +#include +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/operators/math/cpu_vec.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace framework { +namespace ir { + +#define GET_CONV_BN_NODES(pattern_name) \ + /* OPERATORS */ \ + GET_IR_NODE_FROM_SUBGRAPH(conv, conv, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(batch_norm, batch_norm, pattern_name); \ + /* CONV inputs */ \ + GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight, pattern_name); \ + /* CONV outputs */ \ + GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, pattern_name); \ + /* BN inputs */ \ + GET_IR_NODE_FROM_SUBGRAPH(bn_scale, bn_scale, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_bias, bn_bias, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_mean, bn_mean, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_variance, bn_variance, pattern_name); \ + /* BN outputs */ \ + GET_IR_NODE_FROM_SUBGRAPH(bn_out, bn_out, pattern_name); /* Out */ \ + GET_IR_NODE_FROM_SUBGRAPH(bn_mean_out, bn_mean_out, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_variance_out, bn_variance_out, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_saved_mean, bn_saved_mean, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_saved_variance, bn_saved_variance, pattern_name) + +void recompute_bias_and_weights(const Scope* scope, + ir::Node* conv_weight, // + const ir::Node& bn_scale, // + const LoDTensor& bn_bias_tensor, // + const ir::Node& bn_mean, // + const ir::Node& bn_variance, // + LoDTensor* eltwise_y_in_tensor, // + float epsilon) { + using EigenVectorArrayMap = + Eigen::Map>; + using ConstEigenVectorArrayMap = + Eigen::Map>; + using EigenMatrixArrayMap = Eigen::Map< + Eigen::Array>; + + // Re-compute bias of conv2d from BN + PADDLE_ENFORCE_EQ(eltwise_y_in_tensor->dims(), bn_bias_tensor.dims()); + + auto* scale_tensor = scope->FindVar(bn_scale.Name())->GetMutable(); + auto* variance_tensor = + scope->FindVar(bn_variance.Name())->GetMutable(); + auto* mean_tensor = scope->FindVar(bn_mean.Name())->GetMutable(); + + ConstEigenVectorArrayMap scale_array(scale_tensor->data(), + scale_tensor->numel(), 1); + EigenVectorArrayMap variance_array( + variance_tensor->mutable_data(platform::CPUPlace()), + variance_tensor->numel(), 1); + ConstEigenVectorArrayMap mean_array(mean_tensor->data(), + mean_tensor->numel(), 1); + ConstEigenVectorArrayMap bn_bias_array(bn_bias_tensor.data(), + bn_bias_tensor.numel(), 1); + + // variance will not be used anymore, so make it std_array and then tmp_array + variance_array += epsilon; + variance_array = variance_array.sqrt(); + variance_array = scale_array / variance_array; + + EigenVectorArrayMap eltwise_y_in_array( + eltwise_y_in_tensor->mutable_data(platform::CPUPlace()), + eltwise_y_in_tensor->numel(), 1); + + eltwise_y_in_array = + ((eltwise_y_in_array - mean_array) * variance_array) + bn_bias_array; + + // Re-compute weight of conv2d from BN + auto* weights = scope->FindVar(conv_weight->Name())->GetMutable(); + auto weights_shape = weights->dims(); + auto weights_shape_2d = flatten_to_2d(weights_shape, 1); + + EigenMatrixArrayMap weights_array_2d( + weights->mutable_data(platform::CPUPlace()), weights_shape_2d[0], + weights_shape_2d[1]); + + weights_array_2d.colwise() *= variance_array; +} + +std::unique_ptr ConvBNFusePass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init(name_scope_, graph.get()); + + auto* scope = param_scope(); + PADDLE_ENFORCE(scope); + + GraphPatternDetector gpd; + auto* conv_input = + gpd.mutable_pattern() + ->NewNode(patterns::PDNodeName(name_scope_, "conv_input")) + ->AsInput() + ->assert_is_op_input("conv2d", "Input"); + patterns::ConvBN conv_bn_pattern(gpd.mutable_pattern(), name_scope_); + conv_bn_pattern(conv_input, false /*with_eltwise_add*/); + + int found_conv_bn_count = 0; + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(4) << "handle ConvBN fuse"; + + // conv, batch_norm, + // conv_weight, conv_out, + // bn_scale, bn_bias, bn_mean, bn_variance, + // bn_out, bn_mean_out, bn_variance_out, bn_saved_mean, + // bn_saved_variance + GET_CONV_BN_NODES(conv_bn_pattern); + + // check if fuse can be done and if MKL-DNN should be used + FuseOptions fuse_option = FindFuseOption(*conv, *batch_norm); + if (fuse_option == DO_NOT_FUSE) { + VLOG(3) << "do not perform conv+bn fuse"; + return; + } + + // Create eltwise_y (conv bias) variable + VarDesc eltwise_y_in_desc( + patterns::PDNodeName(name_scope_, "eltwise_y_in")); + eltwise_y_in_desc.SetPersistable(true); + auto* eltwise_y_in_node = g->CreateVarNode(&eltwise_y_in_desc); + auto* eltwise_y_in_tensor = + scope->Var(eltwise_y_in_node->Name())->GetMutable(); + + // Get batch norm bias + auto* bn_bias_tensor = + scope->FindVar(bn_bias->Name())->GetMutable(); + + // Initialize eltwise_y + eltwise_y_in_tensor->Resize(bn_bias_tensor->dims()); + std::fill_n(eltwise_y_in_tensor->mutable_data(platform::CPUPlace()), + eltwise_y_in_tensor->numel(), 0.0f); + + // update weights and biases + float epsilon = boost::get(batch_norm->Op()->GetAttr("epsilon")); + recompute_bias_and_weights(scope, conv_weight, *bn_scale, *bn_bias_tensor, + *bn_mean, *bn_variance, eltwise_y_in_tensor, + epsilon); + + // with MKL-DNN fuse conv+bn into conv with bias + // without MKL-DNN fuse conv+bn into conv+elementwise_add + if (fuse_option == FUSE_MKLDNN) { + auto input_names = conv->Op()->InputNames(); + bool has_bias = std::find(input_names.begin(), input_names.end(), + "Bias") != input_names.end(); + if (has_bias && conv->Op()->Input("Bias").size() > 0) { + // reuse existing conv bias node + auto conv_bias_names = conv->Op()->Input("Bias"); + PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1); + auto* conv_bias_var = scope->FindVar(conv_bias_names[0]); + auto* conv_bias_tensor = conv_bias_var->GetMutable(); + PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(), + eltwise_y_in_tensor->dims()); + + auto eigen_conv_bias = EigenVector::From(*conv_bias_tensor); + eigen_conv_bias += EigenVector::From(*eltwise_y_in_tensor); + } else { + // add new conv_bias node + conv->Op()->SetInput( + "Bias", std::vector({eltwise_y_in_node->Name()})); + IR_NODE_LINK_TO(eltwise_y_in_node, conv); + } + conv->Op()->SetOutput("Output", + std::vector({bn_out->Name()})); + + GraphSafeRemoveNodes( + graph.get(), + {conv_out, bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, + bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance}); + + IR_NODE_LINK_TO(conv, bn_out); + found_conv_bn_count++; + } else { // fuse_option == FUSE_NATIVE + // create an elementwise add node. + OpDesc desc; + desc.SetInput("X", std::vector({conv_out->Name()})); + desc.SetInput("Y", std::vector({eltwise_y_in_node->Name()})); + desc.SetOutput("Out", std::vector({bn_out->Name()})); + desc.SetType("elementwise_add"); + desc.SetAttr("axis", 1); + auto eltwise_op = g->CreateOpNode(&desc); // OpDesc will be copied. + + GraphSafeRemoveNodes( + graph.get(), + {bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out, + bn_variance_out, bn_saved_mean, bn_saved_variance}); + + IR_NODE_LINK_TO(conv_out, eltwise_op); + IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op); + IR_NODE_LINK_TO(eltwise_op, bn_out); + found_conv_bn_count++; + } + }; + + gpd(graph.get(), handler); + + AddStatis(found_conv_bn_count); + return graph; +} + +std::unique_ptr ConvEltwiseAddBNFusePass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init(name_scope_, graph.get()); + + auto* scope = param_scope(); + PADDLE_ENFORCE(scope); + + GraphPatternDetector gpd; + auto* conv_input = + gpd.mutable_pattern() + ->NewNode(patterns::PDNodeName(name_scope_, "conv_input")) + ->AsInput() + ->assert_is_op_input("conv2d", "Input"); + patterns::ConvBN conv_bn_pattern(gpd.mutable_pattern(), name_scope_); + conv_bn_pattern(conv_input, true /*with_eltwise_add*/); + + int found_conv_bn_count = 0; + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(4) << "handle ConvBN fuse"; + + // conv, batch_norm, + // conv_weight, conv_out, + // bn_scale, bn_bias, bn_mean, bn_variance, + // bn_out, bn_mean_out, bn_variance_out, bn_saved_mean,bn_saved_variance + GET_CONV_BN_NODES(conv_bn_pattern); + // OPERATORS + GET_IR_NODE_FROM_SUBGRAPH(eltwise, eltwise, conv_bn_pattern); + // BIAS inputs + GET_IR_NODE_FROM_SUBGRAPH(eltwise_y_in, eltwise_y_in, conv_bn_pattern); + // BIAS outputs + GET_IR_NODE_FROM_SUBGRAPH(eltwise_out, eltwise_out, conv_bn_pattern); + + // Get eltwise_y (conv bias) variable + auto* eltwise_y_in_tensor = + scope->FindVar(eltwise_y_in->Name())->GetMutable(); + + // Get batch norm bias + auto* bn_bias_tensor = + scope->FindVar(bn_bias->Name())->GetMutable(); + + // update weights and biases + float epsilon = boost::get(batch_norm->Op()->GetAttr("epsilon")); + recompute_bias_and_weights(scope, conv_weight, *bn_scale, *bn_bias_tensor, + *bn_mean, *bn_variance, eltwise_y_in_tensor, + epsilon); + + // Update the elementwise_add node + eltwise->Op()->SetAttr("axis", 1); + eltwise->Op()->SetOutput("Out", std::vector({bn_out->Name()})); + + GraphSafeRemoveNodes( + graph.get(), + {bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out, + bn_variance_out, bn_saved_mean, bn_saved_variance, eltwise_out}); + + IR_NODE_LINK_TO(eltwise, bn_out); + + found_conv_bn_count++; + }; + + gpd(graph.get(), handler); + + AddStatis(found_conv_bn_count); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(conv_bn_fuse_pass, paddle::framework::ir::ConvBNFusePass); +REGISTER_PASS(conv_eltwiseadd_bn_fuse_pass, + paddle::framework::ir::ConvEltwiseAddBNFusePass); diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass.h b/paddle/fluid/framework/ir/conv_bn_fuse_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..2c9eb574fe8e054e0ae221f08f664b91f05d95c9 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_bn_fuse_pass.h @@ -0,0 +1,49 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +/* + * Fuse the Conv and BatchNorm to a ConvBNMKLDNNOp. + */ +class ConvBNFusePass : public FusePassBase { + public: + virtual ~ConvBNFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + const std::string name_scope_{"conv_bn_fuse"}; +}; + +class ConvEltwiseAddBNFusePass : public FusePassBase { + public: + virtual ~ConvEltwiseAddBNFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + const std::string name_scope_{"conv_eltwiseadd_bn_fuse"}; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..8d0035ae98b093979eb8bbcc0a8d6ae5356d951f --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc @@ -0,0 +1,154 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h" +#include +#include + +#include "paddle/fluid/framework/ir/graph_traits.h" + +namespace paddle { +namespace framework { +namespace ir { +namespace { + +// The function keeps the graph consistent by replacing +// a node 'from' in the set of inputs nodes +// of the visited node by a node 'to'. +void CorrectGraphEdges(Graph* graph, Node* from, Node* to) { + for (auto& node : GraphTraits::DFS(*graph)) { + auto from_in_inputs = + std::find(std::begin(node.inputs), std::end(node.inputs), from); + + if (from_in_inputs != std::end(node.inputs)) { + IR_NODE_LINK_TO(to, (&node)); + + auto inputs = node.Op()->Inputs(); + + using input_type = VariableNameMap::value_type; + + std::for_each(std::begin(inputs), std::end(inputs), + [from, to, &node](const input_type& i) -> void { + auto param_names = i.second; + auto pi = std::find(std::begin(param_names), + std::end(param_names), from->Name()); + + if (pi != std::end(param_names)) { + node.Op()->SetInput(i.first, {to->Name()}); + } + }); + } + } +} +} // namespace +using graph_ptr = std::unique_ptr; + +graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const { + FusePassBase::Init(name_scope_, graph.get()); + + GraphPatternDetector gpd; + auto pattern = gpd.mutable_pattern(); + + patterns::Conv conv_pattern{pattern, name_scope_}; + auto conv_output = conv_pattern(); + + patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope_}; + elementwise_add_pattern(conv_output); + + conv_output->AsIntermediate(); + + auto conv_op_has_bias = [](const Node& conv_op) -> std::pair { + auto bias_input_names = conv_op.Op()->Inputs(); + auto bias_it = bias_input_names.find("Bias"); + + if (bias_it != std::end(bias_input_names)) { + bool has_bias = !bias_it->second.empty(); + + if (has_bias) { + auto conv_bias_names = bias_it->second; + auto conv_bias_names_it = + std::find_if(std::begin(conv_op.inputs), std::end(conv_op.inputs), + [&conv_bias_names](Node* n) -> bool { + return n->Name() == conv_bias_names[0]; + }); + return std::make_pair(has_bias, *conv_bias_names_it); + } + } + + return std::make_pair(false, nullptr); + }; + + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern); + GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern); + GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern); + GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern); + GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op, + elementwise_add_pattern); + GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x, + elementwise_add_pattern); + GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out, + elementwise_add_pattern); + + if (FindFuseOption(*conv_op, *elementwise_add_op) != FUSE_MKLDNN) return; + + OpDesc op_desc; + op_desc.SetType("conv2d"); + + op_desc.SetInput("Input", {conv_input->Name()}); + op_desc.SetInput("Filter", {conv_filter->Name()}); + op_desc.SetInput("ResidualData", {elementwise_add_x->Name()}); + op_desc.SetOutput("Output", {conv_output->Name()}); + + bool has_bias; + Node* conv_bias; + + std::tie(has_bias, conv_bias) = conv_op_has_bias(*conv_op); + + if (has_bias) { + op_desc.SetInput("Bias", {conv_bias->Name()}); + } + + for (const auto& attr : conv_op->Op()->GetAttrMap()) { + op_desc.SetAttr(attr.first, attr.second); + } + + op_desc.SetAttr("fuse_residual_connection", true); + + auto fused_conv_op = g->CreateOpNode(&op_desc); + + IR_NODE_LINK_TO(conv_input, fused_conv_op); + IR_NODE_LINK_TO(conv_filter, fused_conv_op); + IR_NODE_LINK_TO(elementwise_add_x, fused_conv_op); + IR_NODE_LINK_TO(fused_conv_op, conv_output); + + if (has_bias) { + IR_NODE_LINK_TO(conv_bias, fused_conv_op); + } + + CorrectGraphEdges(g, elementwise_add_out, conv_output); + GraphSafeRemoveNodes(g, {elementwise_add_out, conv_op, elementwise_add_op}); + }; + + gpd(graph.get(), handler); + + return graph; +} +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(conv_elementwise_add_mkldnn_fuse_pass, + paddle::framework::ir::ConvElementwiseAddMKLDNNFusePass); diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..f4a899f1adb5e993895a40a8cfb846a67b41bb22 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h @@ -0,0 +1,38 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +class ConvElementwiseAddMKLDNNFusePass : public FusePassBase { + public: + virtual ~ConvElementwiseAddMKLDNNFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + + const std::string name_scope_{"residual_connections_fuse_pass"}; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..348a3dfc5da78e860742595a60a0b7a8b2d92243 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc @@ -0,0 +1,247 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include + +#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h" +#include "paddle/fluid/framework/ir/graph_traits.h" + +namespace paddle { +namespace framework { +namespace ir { + +namespace { +constexpr int nodes_removed = 3; +constexpr int nodes_added = 1; + +void SetOp(ProgramDesc* prog, const std::string& type, + const std::vector>& inputs, + const std::pair& output) { + auto op = prog->MutableBlock(0)->AppendOp(); + op->SetType(type); + op->SetAttr("use_mkldnn", true); + + for (const auto& input : inputs) { + op->SetInput(input.first, {input.second}); + } + + op->SetOutput(output.first, {output.second}); +} + +struct IsReachable { + using func = std::function; + + auto operator()(const std::unique_ptr& graph) -> func { + auto find_node = [](const std::unique_ptr& graph, + const std::string& name) -> Node* { + for (auto& node : GraphTraits::DFS(*graph)) { + if (name == node.Name()) { + return &node; + } + } + + return nullptr; + }; + + return [&](std::string from, const std::string to) -> bool { + if (from == to) return true; + + std::map visited; + + for (auto& node : GraphTraits::DFS(*graph)) { + visited[node.Name()] = false; + } + + visited[from] = true; + + std::list queue; + queue.push_back(from); + + while (!queue.empty()) { + auto cur = find_node(graph, queue.front()); + queue.pop_front(); + + if (cur == nullptr) return false; + + for (auto n : cur->outputs) { + if (n->Name() == to) return true; + + if (!visited[n->Name()]) { + visited[n->Name()] = true; + queue.push_back(n->Name()); + } + } + } + return false; + }; + } +}; + +void AssertOpsCount(const std::unique_ptr& graph) { + int conv_count = 0; + int elementwise_add_count = 0; + + for (auto* node : graph->Nodes()) { + if (node->IsOp() && node->Op()->Type() == "conv2d") { + ++conv_count; + } + if (node->IsOp() && node->Op()->Type() == "elementwise_add") { + ++elementwise_add_count; + } + } + EXPECT_EQ(conv_count, 1); + EXPECT_EQ(elementwise_add_count, 0); +} + +ProgramDesc BuildProgramDesc(const std::vector& transient_vars, + const std::vector& persistent_vars) { + ProgramDesc prog; + + auto add_var_to_prog = [&prog](const std::string& var_name) -> VarDesc* { + auto var = prog.MutableBlock(0)->Var(var_name); + var->SetType(proto::VarType::LOD_TENSOR); + + return var; + }; + + for (const auto& v : transient_vars) { + add_var_to_prog(v); + } + + for (const auto& v : persistent_vars) { + auto var = add_var_to_prog(v); + var->SetPersistable(true); + } + + return prog; +} +} // namespace + +TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) { + auto prog = + BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"}); + + SetOp(&prog, "conv2d", + {{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}}, + {"Output", "b"}); + SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"}); + SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"}); + + std::unique_ptr graph(new ir::Graph(prog)); + + IsReachable is_reachable; + EXPECT_TRUE(is_reachable(graph)("a", "relu")); + + auto pass = + PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass"); + int original_nodes_num = graph->Nodes().size(); + graph = pass->Apply(std::move(graph)); + int current_nodes_num = graph->Nodes().size(); + + EXPECT_TRUE(is_reachable(graph)("a", "relu")); + + EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added, + current_nodes_num); + + AssertOpsCount(graph); +} + +TEST(ConvElementwiseAddMKLDNNFusePass, + ConvolutionWithElementwiseAddReluNoBias) { + auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"}); + SetOp(&prog, "conv2d", {{"Input", "a"}, {"Filter", "weights"}}, + {"Output", "b"}); + SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"}); + SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"}); + + std::unique_ptr graph(new ir::Graph(prog)); + + IsReachable is_reachable; + + EXPECT_TRUE(is_reachable(graph)("a", "relu")); + + auto pass = + PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass"); + int original_nodes_num = graph->Nodes().size(); + graph = pass->Apply(std::move(graph)); + int current_nodes_num = graph->Nodes().size(); + + EXPECT_TRUE(is_reachable(graph)("a", "relu")); + + EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added, + current_nodes_num); + + AssertOpsCount(graph); +} + +TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionElementwiseAdd) { + auto prog = BuildProgramDesc({"a", "b", "c", "d"}, {"bias", "weights"}); + SetOp(&prog, "conv2d", + {{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}}, + {"Output", "b"}); + SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"}); + + std::unique_ptr graph(new ir::Graph(prog)); + + IsReachable is_reachable; + EXPECT_TRUE(is_reachable(graph)("a", "d")); + + auto pass = + PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass"); + int original_nodes_num = graph->Nodes().size(); + graph = pass->Apply(std::move(graph)); + int current_nodes_num = graph->Nodes().size(); + + EXPECT_FALSE(is_reachable(graph)("a", "d")); + + EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added, + current_nodes_num); + AssertOpsCount(graph); +} + +TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) { + auto prog = + BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"}); + SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"}); + SetOp(&prog, "conv2d", + {{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}}, + {"Output", "c"}); + SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "d"}}, {"Out", "e"}); + SetOp(&prog, "relu", {{"X", "e"}}, {"Out", "f"}); + + std::unique_ptr graph(new ir::Graph(prog)); + + IsReachable is_reachable; + + EXPECT_TRUE(is_reachable(graph)("a", "f")); + + auto pass = + PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass"); + int original_nodes_num = graph->Nodes().size(); + graph = pass->Apply(std::move(graph)); + int current_nodes_num = graph->Nodes().size(); + + EXPECT_TRUE(is_reachable(graph)("a", "f")); + + EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added, + current_nodes_num); + AssertOpsCount(graph); +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +USE_PASS(conv_elementwise_add_mkldnn_fuse_pass); diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc index d7df6389cfd595324e284e0da10f65213ccee80f..e359a3832ee8d549f8c58d63bc1cc6564ecadede 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc @@ -46,6 +46,12 @@ std::unique_ptr ConvReLUFusePass::ApplyImpl( GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, conv_relu_pattern); // Out GET_IR_NODE_FROM_SUBGRAPH(relu, relu, conv_relu_pattern); // ReLU op + FuseOptions fuse_option = FindFuseOption(*conv, *relu); + if (fuse_option == DO_NOT_FUSE) { + VLOG(3) << "do not perform conv+relu fuse"; + return; + } + // Transform Conv node into ConvReLU node. OpDesc* desc = conv->Op(); desc->SetOutput("Output", std::vector({relu_out->Name()})); diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc index 9dd780ec89ab991d6d99cb66fa2a9b683be2b9ca..8f4bab25ed4919881baf19a961a52aa229e06a8f 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc +++ b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc @@ -20,17 +20,19 @@ namespace paddle { namespace framework { namespace ir { -void SetOp(ProgramDesc* prog, const std::string& type, +void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name, const std::vector& inputs, - const std::vector& outputs) { + const std::vector& outputs, bool use_mkldnn = false) { auto* op = prog->MutableBlock(0)->AppendOp(); op->SetType(type); if (type == "conv2d") { - op->SetAttr("use_mkldnn", true); + op->SetAttr("use_mkldnn", use_mkldnn); + op->SetAttr("name", name); op->SetInput("Input", {inputs[0]}); op->SetInput("Filter", {inputs[1]}); op->SetInput("Bias", {inputs[2]}); } else if (type == "relu") { + op->SetAttr("use_mkldnn", use_mkldnn); op->SetInput("X", inputs); } op->SetOutput("Out", outputs); @@ -43,7 +45,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, ProgramDesc BuildProgramDesc() { ProgramDesc prog; for (auto& v : - std::vector({"a", "b", "c", "weights", "bias", "f", "g"})) { + std::vector({"a", "b", "c", "weights", "bias", "f", "g", + "h", "weights2", "bias2", "k", "l"})) { auto* var = prog.MutableBlock(0)->Var(v); var->SetType(proto::VarType::SELECTED_ROWS); if (v == "weights" || v == "bias") { @@ -51,14 +54,24 @@ ProgramDesc BuildProgramDesc() { } } - SetOp(&prog, "OP0", std::vector({"a"}), + SetOp(&prog, "OP0", "op0", std::vector({"a"}), std::vector({"b"})); - SetOp(&prog, "OP1", std::vector({"b"}), + SetOp(&prog, "OP1", "op1", std::vector({"b"}), std::vector({"c"})); - SetOp(&prog, "conv2d", std::vector({"c", "weights", "bias"}), - std::vector({"f"})); - SetOp(&prog, "relu", std::vector({"f"}), - std::vector({"g"})); + // conv+relu, both with MKL-DNN + SetOp(&prog, "conv2d", "conv1", + std::vector({"c", "weights", "bias"}), + std::vector({"f"}), true); + SetOp(&prog, "relu", "relu1", std::vector({"f"}), + std::vector({"g"}), true); + SetOp(&prog, "OP3", "op3", std::vector({"g"}), + std::vector({"h"})); + // conv+relu, only one with MKL-DNN + SetOp(&prog, "conv2d", "conv2", + std::vector({"h", "weights2", "bias2"}), + std::vector({"k"}), true); + SetOp(&prog, "relu", "relu2", std::vector({"k"}), + std::vector({"l"})); return prog; } @@ -88,10 +101,16 @@ TEST(ConvReLUFusePass, basic) { auto* op = node->Op(); ASSERT_TRUE(op->HasAttr("use_mkldnn")); EXPECT_TRUE(boost::get(op->GetAttr("use_mkldnn"))); - ASSERT_TRUE(op->HasAttr("fuse_relu")); - bool fuse_relu = boost::get(op->GetAttr("fuse_relu")); - if (fuse_relu) { - ++conv_relu_count; + // check if only "conv1" convolution is fused + auto op_name = boost::get(op->GetAttr("name")); + if (op_name == "conv1") { + ASSERT_TRUE(op->HasAttr("fuse_relu")); + bool fuse_relu = boost::get(op->GetAttr("fuse_relu")); + if (fuse_relu) { + ++conv_relu_count; + } + } else if (op_name == "conv2") { + ASSERT_FALSE(op->HasAttr("fuse_relu")); } } } diff --git a/paddle/fluid/framework/ir/fuse_pass_base.cc b/paddle/fluid/framework/ir/fuse_pass_base.cc new file mode 100644 index 0000000000000000000000000000000000000000..d70010089e4b4fbb4542ef7748b8e9ece48d3942 --- /dev/null +++ b/paddle/fluid/framework/ir/fuse_pass_base.cc @@ -0,0 +1,62 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/fuse_pass_base.h" + +namespace paddle { +namespace framework { +namespace ir { + +void FusePassBase::Init(const std::string& repr, Graph* graph) const { + repr_ = repr; + graph_ = graph; +} + +Scope* FusePassBase::param_scope() const { + PADDLE_ENFORCE(graph_->Has(kParamScopeAttr)); + return graph_->Get(kParamScopeAttr); +} + +void FusePassBase::AddStatis(int count_of_fused) const { + PADDLE_ENFORCE(graph_); + PADDLE_ENFORCE(!repr_.empty()); + if (!graph_->Has(kFuseStatisAttr)) { + graph_->Set(kFuseStatisAttr, new std::unordered_map); + } + auto& info = + graph_->Get>(kFuseStatisAttr); + info[repr_] = count_of_fused; +} + +FuseOptions FusePassBase::FindFuseOption(const Node& node1, + const Node& node2) const { +#ifdef PADDLE_WITH_MKLDNN + bool node1_mkldnn = node1.Op()->HasAttr("use_mkldnn") && + boost::get(node1.Op()->GetAttr("use_mkldnn")); + bool node2_mkldnn = node2.Op()->HasAttr("use_mkldnn") && + boost::get(node2.Op()->GetAttr("use_mkldnn")); + if (node1_mkldnn && node2_mkldnn) + return FUSE_MKLDNN; + else if (!node1_mkldnn && !node2_mkldnn) + return FUSE_NATIVE; + else + return DO_NOT_FUSE; +#else + return FUSE_NATIVE; +#endif +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/fuse_pass_base.h b/paddle/fluid/framework/ir/fuse_pass_base.h index 877bbeb502252cac77095981641d7ce283ca1eb7..c53b2a6186741d86f14faf1d21fa19aa09cec036 100644 --- a/paddle/fluid/framework/ir/fuse_pass_base.h +++ b/paddle/fluid/framework/ir/fuse_pass_base.h @@ -25,32 +25,24 @@ namespace ir { static const char kParamScopeAttr[] = "__param_scope__"; static const char kFuseStatisAttr[] = "__fuse_statis__"; +enum FuseOptions { + DO_NOT_FUSE, // fusing will not be done + FUSE_NATIVE, // fusing will be done without MKL-DNN + FUSE_MKLDNN // fusing will be done with MKL-DNN +}; + class FusePassBase : public Pass { public: - void Init(const std::string& repr, Graph* graph) const { - repr_ = repr; - graph_ = graph; - } - - Scope* param_scope() const { - PADDLE_ENFORCE(graph_->Has(kParamScopeAttr)); - return graph_->Get(kParamScopeAttr); - } - - void AddStatis(int count_of_fused) const { - PADDLE_ENFORCE(graph_); - PADDLE_ENFORCE(!repr_.empty()); - if (!graph_->Has(kFuseStatisAttr)) { - graph_->Set(kFuseStatisAttr, new std::unordered_map); - } - auto& info = - graph_->Get>(kFuseStatisAttr); - info[repr_] = count_of_fused; - } + void Init(const std::string& repr, Graph* graph) const; + Scope* param_scope() const; + void AddStatis(int count_of_fused) const; virtual ~FusePassBase() {} protected: + virtual FuseOptions FindFuseOption(const Node& node1, + const Node& node2) const; + mutable Graph* graph_; mutable std::string repr_; }; diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index 46c6a52c09e896596aa6d8e1e901955a68a4957d..29b604afbfcfc2bac67e447db8cd4c671c036dbe 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -259,6 +259,8 @@ GraphPatternDetector::DetectPatterns() { return result; } +// TODO(Superjomn) enhance the function as it marks unique unique as duplicates +// see https://github.com/PaddlePaddle/Paddle/issues/13550 void GraphPatternDetector::UniquePatterns( std::vector *subgraphs) { if (subgraphs->empty()) return; @@ -626,6 +628,112 @@ bool VarLinksFromOp(Node *node, const std::string &op_type) { return false; } +PDNode *patterns::ConvBN::operator()(paddle::framework::ir::PDNode *conv_input, + bool with_eltwise_add) { + // Create Operators + conv_input->assert_is_op_input("conv2d", "Input"); + auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d"); + + PDNode *eltwise_op = nullptr; + if (with_eltwise_add) { + eltwise_op = + pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add"); + } + auto *batch_norm_op = + pattern->NewNode(batch_norm_repr())->assert_is_op("batch_norm"); + // Create variables + // Conv Filter + auto *conv_weight_var = pattern->NewNode(conv_weight_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("conv2d", "Filter"); + + auto *conv_out_var = pattern->NewNode(conv_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("conv2d"); + + PDNode *eltwise_y_in_var = nullptr; + PDNode *eltwise_out_var = nullptr; + if (with_eltwise_add) { + // Conv output as Bias input + conv_out_var->assert_is_op_input("elementwise_add", "X"); + // Bias + eltwise_y_in_var = pattern->NewNode(eltwise_y_in_repr()) + ->assert_is_op_input("elementwise_add", "Y") + ->AsInput(); + eltwise_out_var = pattern->NewNode(eltwise_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("elementwise_add"); + } else { + // Conv output as BN input + conv_out_var->assert_is_op_input("batch_norm", "X"); + } + + // BN Scale + auto *bn_scale_var = pattern->NewNode(bn_scale_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("batch_norm", "Scale"); + // BN Bias + auto *bn_bias_var = pattern->NewNode(bn_bias_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("batch_norm", "Bias"); + // BN Mean + auto *bn_mean_var = pattern->NewNode(bn_mean_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("batch_norm", "Mean"); + // BN Variance + auto *bn_variance_var = pattern->NewNode(bn_variance_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("batch_norm", "Variance"); + + // BN output + auto *bn_out_var = pattern->NewNode(bn_out_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm"); + + auto *bn_mean_out_var = pattern->NewNode(bn_mean_out_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm", "MeanOut"); + + auto *bn_variance_out_var = + pattern->NewNode(bn_variance_out_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm", "VarianceOut"); + + auto *bn_saved_mean_var = + pattern->NewNode(bn_saved_mean_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm", "SavedMean"); + + auto *bn_saved_variance_var = + pattern->NewNode(bn_saved_variance_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm", "SavedVariance"); + + conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var}); + + if (with_eltwise_add) { + eltwise_op->LinksFrom({conv_out_var, eltwise_y_in_var}) + .LinksTo({eltwise_out_var}); + batch_norm_op + ->LinksFrom({eltwise_out_var, bn_scale_var, bn_bias_var, bn_mean_var, + bn_variance_var}) + .LinksTo({bn_out_var, bn_mean_out_var, bn_variance_out_var, + bn_saved_mean_var, bn_saved_variance_var}); + } else { + batch_norm_op + ->LinksFrom({conv_out_var, bn_scale_var, bn_bias_var, bn_mean_var, + bn_variance_var}) + .LinksTo({bn_out_var, bn_mean_out_var, bn_variance_out_var, + bn_saved_mean_var, bn_saved_variance_var}); + } + return bn_out_var; +} + PDNode *patterns::ConvReLU::operator()( paddle::framework::ir::PDNode *conv_input) { // Create Operators @@ -653,6 +761,51 @@ PDNode *patterns::ConvReLU::operator()( return relu_out_var; } +PDNode *patterns::SeqConvEltAddRelu::operator()( + paddle::framework::ir::PDNode *seqconv_input) { + // Create Operators + seqconv_input->assert_is_op_input("sequence_conv", "X"); + auto *seqconv_op = pattern->NewNode(seqconv_repr()) + ->assert_is_op("sequence_conv") + ->assert_op_attr("paddingTrainable", false) + ->assert_op_attr("contextStride", 1); + + auto *eltadd_op = + pattern->NewNode(eltadd_repr())->assert_is_op("elementwise_add"); + auto *relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu"); + // Create variables + // Filter + auto *seqconv_weight_var = + pattern->NewNode(seqconv_weight_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("sequence_conv", "Filter"); + // Bias + auto *eltadd_bias_var = pattern->NewNode(eltadd_bias_repr()) + ->AsInput() + ->assert_is_op_input("elementwise_add"); + // intermediate variable, will be removed in the IR after fuse. + auto *seqconv_out_var = pattern->NewNode(seqconv_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("sequence_conv") + ->assert_is_op_input("elementwise_add"); + auto *eltadd_out_var = pattern->NewNode(eltadd_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("elementwise_add") + ->assert_is_only_input_of_op("relu"); + // output + auto *relu_out_var = pattern->NewNode(relu_out_repr()) + ->AsOutput() + ->assert_is_op_output("relu"); + + seqconv_op->LinksFrom({seqconv_input, seqconv_weight_var}) + .LinksTo({seqconv_out_var}); + eltadd_op->LinksFrom({seqconv_out_var, eltadd_bias_var}) + .LinksTo({eltadd_out_var}); + relu_op->LinksFrom({eltadd_out_var}).LinksTo({relu_out_var}); + return relu_out_var; +} + PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x, bool with_bias) { // Create shared nodes. @@ -858,6 +1011,79 @@ PDNode *patterns::ElewiseAddActInplaceGrad::operator()( return ele_add_grad; } +PDNode *patterns::ConvBias::operator()( + paddle::framework::ir::PDNode *conv_input) { + // Create Operators + conv_input->assert_is_op_input("conv2d", "Input"); + auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d"); + auto *eltiwse_op = + pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add"); + // Create variables + // Filter + auto *conv_weight_var = pattern->NewNode(conv_weight_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("conv2d", "Filter"); + // intermediate variable, will be removed in the IR after fuse. + auto *conv_out_var = pattern->NewNode(conv_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("conv2d") + ->assert_is_op_input("elementwise_add"); + // Bias stored in elementwise_add + auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("elementwise_add", "Y"); + // output + auto *eltwise_out_var = pattern->NewNode(eltwise_out_repr()) + ->AsOutput() + ->assert_is_op_output("elementwise_add"); + conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var}); + eltiwse_op->LinksFrom({conv_out_var, eltwise_bias_var}) + .LinksTo({eltwise_out_var}); + return eltwise_out_var; +} + +PDNode *patterns::Conv::operator()() { + auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d"); + + auto input_var = pattern->NewNode(conv_input_repr()) + ->AsInput() + ->assert_is_op_input("conv2d", "Input"); + + auto filter_var = pattern->NewNode(conv_filter_repr()) + ->AsInput() + ->assert_is_op_input("conv2d", "Filter"); + + auto output_var = pattern->NewNode(conv_output_repr()) + ->AsOutput() + ->assert_is_op_output("conv2d", "Output"); + + conv_op->LinksFrom({input_var, filter_var}); + conv_op->LinksTo({output_var}); + + return output_var; +} + +PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var) { + auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr()) + ->assert_is_op("elementwise_add"); + + x_var->assert_is_op_input("elementwise_add", "X"); + + auto y_var = pattern->NewNode(elementwise_add_x_repr()) + ->AsInput() + ->assert_is_op_input("elementwise_add", "Y"); + + auto out_var = pattern->NewNode(elementwise_add_out_repr()) + ->AsOutput() + ->assert_is_op_output("elementwise_add", "Out"); + + elementwise_add_op->LinksFrom({x_var, y_var}); + elementwise_add_op->LinksTo({out_var}); + + return out_var; +} } // namespace ir } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h index 508113bf4fcab274394f2705c36eddbf4ba3c77a..9e462ac671ee931fc17a31f32a76049a0990341f 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.h +++ b/paddle/fluid/framework/ir/graph_pattern_detector.h @@ -128,6 +128,15 @@ struct PDNode { const std::unordered_set& op_types, const std::string& argument, int nth); + template + PDNode* assert_op_attr(const std::string& attr_name, const T& attr) { + asserts_.emplace_back([=](Node* x) { + return x && x->IsOp() && x->Op()->HasAttr(attr_name) && + boost::get(x->Op()->GetAttr(attr_name)) == attr; + }); + return this; + } + private: PDNode(PDPattern* pattern, const std::string& name = "", Type type = Type::kVar) @@ -375,6 +384,44 @@ struct PatternBase { size_t id_; }; +// Conv with batch norm +// op: conv + (elementwise_add +) batch_norm +// named nodes: +// conv_weight, conv_out, conv, +// bn_x, bn_scale, bn_bias, bn_mean, bn_variance, +// bn_batch_norm, bn_y, bn_mean_out, bn_variance_out, +// bn_saved_mean, bn_saved_variance +struct ConvBN : public PatternBase { + ConvBN(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "conv_bn") {} + + PDNode* operator()(PDNode* conv_input, bool with_eltwise_add); + + // declare operator node's name + PATTERN_DECL_NODE(conv); + PATTERN_DECL_NODE(batch_norm); + PATTERN_DECL_NODE(eltwise); // ELEMENTWISE_ADD + // CONV inputs + PATTERN_DECL_NODE(conv_weight); // Filter + // CONV outputs + PATTERN_DECL_NODE(conv_out); // tmp + // ELTWISE inputs + PATTERN_DECL_NODE(eltwise_y_in); + // ELTWISE outputs + PATTERN_DECL_NODE(eltwise_out); // tmp + // BN inputs + PATTERN_DECL_NODE(bn_scale); + PATTERN_DECL_NODE(bn_bias); + PATTERN_DECL_NODE(bn_mean); + PATTERN_DECL_NODE(bn_variance); + // BN outputs + PATTERN_DECL_NODE(bn_out); // Out + PATTERN_DECL_NODE(bn_mean_out); + PATTERN_DECL_NODE(bn_variance_out); + PATTERN_DECL_NODE(bn_saved_mean); + PATTERN_DECL_NODE(bn_saved_variance); +}; + // CONV with ReLU // op: conv + relu // named nodes: @@ -396,6 +443,31 @@ struct ConvReLU : public PatternBase { PATTERN_DECL_NODE(relu_out); }; +// SEQCONV with Elementwise_Add ReLU +// op: seqconv + elementwise_add + relu +// named nodes: +// seqconv_input, seqconv_weight, +// seqconv_out, seqconv, +// elementwise_add_bias, elementwise_add_out, elementwise_add +// relu_out, relu +struct SeqConvEltAddRelu : public PatternBase { + SeqConvEltAddRelu(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "seqconv_eltadd_relu") {} + + PDNode* operator()(PDNode* seqconv_input); + + // declare operator node's name + PATTERN_DECL_NODE(seqconv); + PATTERN_DECL_NODE(eltadd); + PATTERN_DECL_NODE(relu); + // declare variable node's name + PATTERN_DECL_NODE(seqconv_weight); + PATTERN_DECL_NODE(seqconv_out); + PATTERN_DECL_NODE(eltadd_bias); + PATTERN_DECL_NODE(eltadd_out); + PATTERN_DECL_NODE(relu_out); +}; + // FC with bias // op: mul + elementwise_add // named nodes: @@ -540,6 +612,65 @@ struct ElewiseAddActInplaceGrad : public PatternBase { PATTERN_DECL_NODE(d_ele_y); PATTERN_DECL_NODE(ele_y); }; + +// Conv with Elementwise_add as bias +// op: conv + elementwise_add +// named nodes: +// conv_input, conv_weight, +// conv_out, conv, +// eltwise_bias, eltwise_out, +// elementwise_add +struct ConvBias : public PatternBase { + ConvBias(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "conv_bias") {} + PDNode* operator()(PDNode* conv_input); + // declare operator node's name + PATTERN_DECL_NODE(conv); + PATTERN_DECL_NODE(eltwise); + // declare variable node's name + PATTERN_DECL_NODE(conv_weight); + PATTERN_DECL_NODE(conv_out); + PATTERN_DECL_NODE(eltwise_bias); + PATTERN_DECL_NODE(eltwise_out); +}; + +// Convolution op +// Forward pass for convolution. +// conv_input, conv_bias and conv_filter are inputs. +// conv_output is a result of the operator. +// residual_data is data used by skip connection. +// If residual connection fusion is on, the formula is: +// conv_output = conv_op(conv_filter, conv_input, conv_bias) +// + conv_residual_data +// If the fusion is off, conv_residual_data is not added. +struct Conv : public PatternBase { + Conv(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "convolution") {} + + PDNode* operator()(); + + PATTERN_DECL_NODE(conv_op); + PATTERN_DECL_NODE(conv_input); + PATTERN_DECL_NODE(conv_filter); + PATTERN_DECL_NODE(conv_residual_data); + PATTERN_DECL_NODE(conv_output); +}; + +// ElementwiseAdd used in residual connections. +// y_var is used and convolution output. +// The operator is removed, when residual +// connection fusion is on. +struct ElementwiseAdd : public PatternBase { + ElementwiseAdd(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "elementwise_add") {} + + PDNode* operator()(PDNode* x_var); + + PATTERN_DECL_NODE(elementwise_add_op); + PATTERN_DECL_NODE(elementwise_add_x); + PATTERN_DECL_NODE(elementwise_add_y); + PATTERN_DECL_NODE(elementwise_add_out); +}; } // namespace patterns // Link two ir::Nodes from each other. diff --git a/paddle/fluid/framework/ir/mkldnn_placement_pass.cc b/paddle/fluid/framework/ir/mkldnn_placement_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..65be69b7f5b5e363d5d0753c45f9ff9e3f329fbe --- /dev/null +++ b/paddle/fluid/framework/ir/mkldnn_placement_pass.cc @@ -0,0 +1,37 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/ir/mkldnn_placement_pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +std::unique_ptr MKLDNNPlacementPass::ApplyImpl( + std::unique_ptr graph) const { + VLOG(3) << "Aplies MKL-DNN placement strategy."; + for (const Node* n : graph->Nodes()) { + if (n->IsOp() && n->Op()->HasAttr("use_mkldnn")) { + n->Op()->SetAttr("use_mkldnn", true); + } + } + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(mkldnn_placement_pass, + paddle::framework::ir::MKLDNNPlacementPass); diff --git a/paddle/fluid/framework/ir/mkldnn_placement_pass.h b/paddle/fluid/framework/ir/mkldnn_placement_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..3d4dc9e2b6ecccddea4d63e45710c80d55ef2772 --- /dev/null +++ b/paddle/fluid/framework/ir/mkldnn_placement_pass.h @@ -0,0 +1,31 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +class MKLDNNPlacementPass : public Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..0a1f65d274708dd208d7783c6273160c4c61738a --- /dev/null +++ b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc @@ -0,0 +1,101 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h" +#include +#include "paddle/fluid/framework/lod_tensor.h" + +namespace paddle { +namespace framework { +namespace ir { + +int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope) { + GraphPatternDetector gpd; + auto* pattern = gpd.mutable_pattern(); + + PDNode* x = pattern->NewNode(patterns::PDNodeName(name_scope, "X")) + ->assert_is_op_input("sequence_conv") + ->assert_var_not_persistable(); + patterns::SeqConvEltAddRelu fuse_pattern(pattern, name_scope); + fuse_pattern(x); + + // Create New OpDesc + auto fuse_creator = [&](Node* seqconv, Node* input, Node* seqconv_weight, + Node* eltadd_bias, Node* relu_out) { + OpDesc op_desc; + op_desc.SetType("fusion_seqconv_eltadd_relu"); + op_desc.SetInput("X", {input->Name()}); + op_desc.SetInput("Filter", {seqconv_weight->Name()}); + op_desc.SetInput("Bias", {eltadd_bias->Name()}); + op_desc.SetAttr("contextLength", seqconv->Op()->GetAttr("contextLength")); + op_desc.SetAttr("contextStart", seqconv->Op()->GetAttr("contextStart")); + op_desc.SetAttr("contextStride", seqconv->Op()->GetAttr("contextStride")); + PADDLE_ENFORCE(graph->Has(kParamScopeAttr)); + auto* scope = graph->Get(kParamScopeAttr); + const std::string ColMat = patterns::UniqueKey("SeqConvColMat"); + op_desc.SetOutput("ColMat", {ColMat}); + op_desc.SetOutput("Out", {relu_out->Name()}); + scope->Var(ColMat)->GetMutable(); + + auto* op = graph->CreateOpNode(&op_desc); + IR_NODE_LINK_TO(input, op); + IR_NODE_LINK_TO(seqconv_weight, op); + IR_NODE_LINK_TO(eltadd_bias, op); + IR_NODE_LINK_TO(op, relu_out); + return op; + }; + + int fusion_count{0}; + + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(4) << "handle SeqConv EltAdd Relu fuse"; + GET_IR_NODE_FROM_SUBGRAPH(seqconv, seqconv, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(seqconv_weight, seqconv_weight, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(seqconv_out, seqconv_out, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(eltadd, eltadd, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(eltadd_bias, eltadd_bias, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(eltadd_out, eltadd_out, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(relu, relu, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, fuse_pattern); + + fuse_creator(seqconv, subgraph.at(x), seqconv_weight, eltadd_bias, + relu_out); + std::unordered_set marked_nodes( + {seqconv, seqconv_out, eltadd, eltadd_out, relu}); + GraphSafeRemoveNodes(graph, marked_nodes); + ++fusion_count; + }; + + gpd(graph, handler); + + return fusion_count; +} + +std::unique_ptr SeqConvEltAddReluFusePass::ApplyImpl( + std::unique_ptr graph) const { + FusePassBase::Init(name_scope_, graph.get()); + + int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope()); + AddStatis(fusion_count); + + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(seqconv_eltadd_relu_fuse_pass, + paddle::framework::ir::SeqConvEltAddReluFusePass); diff --git a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..dac9de71930c1768bdf416520caae6468449cd3d --- /dev/null +++ b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h @@ -0,0 +1,38 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +class SeqConvEltAddReluFusePass : public FusePassBase { + public: + virtual ~SeqConvEltAddReluFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + + const std::string name_scope_{"seqconv_eltadd_relu_fuse"}; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/naive_executor.cc b/paddle/fluid/framework/naive_executor.cc index ba10687d65cfbbac89cfc76879c8b202ebd03229..2840d503f1454271afb309efdd435225ab077dc0 100644 --- a/paddle/fluid/framework/naive_executor.cc +++ b/paddle/fluid/framework/naive_executor.cc @@ -37,7 +37,7 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) { } else if (var_type == proto::VarType::FETCH_LIST) { var->GetMutable(); } else if (var_type == proto::VarType::STEP_SCOPES) { - var->GetMutable>(); + var->GetMutable>(); } else if (var_type == proto::VarType::LOD_RANK_TABLE) { var->GetMutable(); } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) { diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index b29ac44699463312a1fdcea55e003daa75997302..121e00b1a3d761f570138092a3e76ae2b722d28a 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -85,10 +85,6 @@ class CompileTimeInferShapeContext : public InferShapeContext { VLOG(3) << "input " << in << " is not LodTensor"; return; } - PADDLE_ENFORCE_EQ(in_var->GetType(), proto::VarType::LOD_TENSOR, - "The %d-th output of Output(%s) must be LoDTensor.", j, - out); - out_var->SetLoDLevel(in_var->GetLoDLevel()); } diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index b4205aba83e774fb9c08193124adb93935c00157..440e0509be727ec2b84abc76fca44edda11f8a0a 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -100,16 +100,6 @@ class OpDesc { std::vector InputNames() const { return MapKeys(inputs_); } std::vector OutputNames() const { return MapKeys(outputs_); } - void SetInputMap(const VariableNameMap &input) { - this->inputs_ = input; - this->need_update_ = true; - } - - void SetOutputMap(const VariableNameMap &output) { - this->outputs_ = output; - this->need_update_ = true; - } - const VariableNameMap &Inputs() const { return inputs_; } const VariableNameMap &Outputs() const { return outputs_; } diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index 9f930065324f13f5aa79c214e820fb6fc2f3a166..14fcde2fe3b1c3acfc0994e9cd37a784c57826d7 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -149,9 +149,17 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { platform::SetDeviceId(dev_id); #endif } - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - platform::RecordEvent record_event(Type(), pool.Get(place)); - RunImpl(scope, place); + + // The profile has a process-wide mutex, results in serious performance issue + // in concurrency scenerio. Here use an `if` to fix this issue. + // Please not remove the `if`, ask @Superjomn if there are any concern. + if (platform::IsProfileEnabled()) { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + platform::RecordEvent record_event(Type(), pool.Get(place)); + RunImpl(scope, place); + } else { + RunImpl(scope, place); + } VLOG(3) << place << " " << DebugStringEx(&scope); } diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index f06bad6c78c05804e583f859906b88fb7b500372..093108cb54779eb0cf35dd83e63eb0b1abb66dcd 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -299,6 +299,12 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes( } ParallelExecutor::~ParallelExecutor() { + const auto dev_ctxs = + platform::DeviceContextPool::Instance().GetAllDeviceContexts(); + for (auto &dev_ctx : dev_ctxs) { + dev_ctx->Wait(); + } + if (member_->own_local_scope_) { for (size_t i = 1; i < member_->local_scopes_.size(); ++i) { Scope *local_scope = member_->local_scopes_[i]; @@ -307,6 +313,10 @@ ParallelExecutor::~ParallelExecutor() { } } } + + // member_ must be destructed before gcs_ since the destructor of + // ReferenceCountOpHandle use raw pointers of gcs_ inside. + member_.reset(); } } // namespace framework diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index fd386a5987f11ff64964e95eb7e9b83572dc790c..ef09b98b2aa91a9d729b94d15dbb676dde4092b6 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -75,7 +75,7 @@ class ParallelExecutor { private: void BCastParamsToDevices(const std::unordered_set &vars) const; - ParallelExecutorPrivate *member_; + std::unique_ptr member_; #ifdef PADDLE_WITH_CUDA // ref_cnts_ is only initialized when ParallelExecutor constructs, and then diff --git a/paddle/fluid/framework/program_desc.cc b/paddle/fluid/framework/program_desc.cc index 589905828f7793c614c0fe12259e9ba5ab11ceac..4b9667113bc7918c1323f0213213a6ffdb7eed8e 100644 --- a/paddle/fluid/framework/program_desc.cc +++ b/paddle/fluid/framework/program_desc.cc @@ -126,7 +126,7 @@ const std::vector ProgramDesc::GetFeedTargetNames() { std::vector feed_target_names; for (auto *op : global_block.AllOps()) { if (op->Type() == kFeedOpType) { - int col = boost::get(op->GetAttr("col")); + size_t col = boost::get(op->GetAttr("col")); if (col >= feed_target_names.size()) { feed_target_names.resize(col + 1); } @@ -143,7 +143,7 @@ const std::vector ProgramDesc::GetFetchTargetNames() { std::vector fetch_target_names; for (auto *op : global_block.AllOps()) { if (op->Type() == kFetchOpType) { - int col = boost::get(op->GetAttr("col")); + size_t col = boost::get(op->GetAttr("col")); if (col >= fetch_target_names.size()) { fetch_target_names.resize(col + 1); } diff --git a/paddle/fluid/framework/reader_test.cc b/paddle/fluid/framework/reader_test.cc index f0d07cb7c1367576084b9494e7758103bb45d1e5..50aca4b5a4ba7a93a1584a03cc16fe5d712a32b5 100644 --- a/paddle/fluid/framework/reader_test.cc +++ b/paddle/fluid/framework/reader_test.cc @@ -39,7 +39,7 @@ TEST(READER, decorate_chain) { { auto endpoints = root->GetEndPoints(); ASSERT_EQ(endpoints.size(), 2U); - ASSERT_NE(endpoints.count(end_point1.get()), 0); + ASSERT_NE(endpoints.count(end_point1.get()), 0UL); ASSERT_NE(endpoints.count(end_point2.get()), 0); } diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc index 1a727a2c8c759d010606d5b605823b7252b35c69..a4abd1b1283f08fb8431fbeea0cea17c8439fdd7 100644 --- a/paddle/fluid/framework/scope.cc +++ b/paddle/fluid/framework/scope.cc @@ -49,18 +49,18 @@ int64_t GetEagerDeletionThreshold() { Scope::~Scope() { DropKids(); } Scope& Scope::NewScope() const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); kids_.push_back(new Scope(this)); return *kids_.back(); } Variable* Scope::Var(const std::string& name) { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); return VarInternal(name); } Variable* Scope::Var(std::string* name) { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); auto new_name = string::Sprintf("%p.%d", this, vars_.size()); if (name != nullptr) { *name = new_name; @@ -69,29 +69,34 @@ Variable* Scope::Var(std::string* name) { } Variable* Scope::FindVar(const std::string& name) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); return FindVarInternal(name); } +Variable* Scope::FindLocalVar(const std::string& name) const { + std::lock_guard lock(mutex_); + return FindVarLocally(name); +} + const Scope* Scope::FindScope(const Variable* var) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); return FindScopeInternal(var); } void Scope::DropKids() { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); for (Scope* s : kids_) delete s; kids_.clear(); } bool Scope::HasKid(const Scope* scope) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); auto it = std::find(this->kids_.begin(), this->kids_.end(), scope); return it != this->kids_.end(); } std::vector Scope::LocalVarNames() const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); std::vector known_vars; known_vars.reserve(this->vars_.size()); for (auto& p : vars_) { @@ -101,7 +106,7 @@ std::vector Scope::LocalVarNames() const { } void Scope::DeleteScope(Scope* scope) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); auto it = std::find(this->kids_.begin(), this->kids_.end(), scope); PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope); this->kids_.erase(it); @@ -114,7 +119,7 @@ void Scope::DeleteScope(Scope* scope) const { } void Scope::EraseVars(const std::vector& var_names) { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); std::set var_set(var_names.begin(), var_names.end()); for (auto it = vars_.begin(); it != vars_.end();) { if (var_set.find(it->first) != var_set.end()) { @@ -127,12 +132,12 @@ void Scope::EraseVars(const std::vector& var_names) { void Scope::Rename(const std::string& origin_name, const std::string& new_name) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); RenameInternal(origin_name, new_name); } std::string Scope::Rename(const std::string& origin_name) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); auto new_name = string::Sprintf("%p.%d", this, vars_.size()); RenameInternal(origin_name, new_name); return new_name; diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h index e42fff1d79d92fb7ed61768a614d8cd98f6775a0..14f9f36812d690fc4a7440f2e7e6a85e9993a535 100644 --- a/paddle/fluid/framework/scope.h +++ b/paddle/fluid/framework/scope.h @@ -63,6 +63,11 @@ class Scope { /// Caller doesn't own the returned Variable. Variable* FindVar(const std::string& name) const; + /// Find a variable in the current scope. + /// Return nullptr if cannot find. + /// Caller doesn't own the returned Variable. + Variable* FindLocalVar(const std::string& name) const; + const Scope* parent() const { return parent_; } /// Find the scope or an ancestor scope that contains the given variable. diff --git a/paddle/fluid/framework/selected_rows_test.cc b/paddle/fluid/framework/selected_rows_test.cc index 928e1ad8b9168e61ddc5782066a4aa29a4296a94..9c427a4ae4c9660b107ca891a60db306cb09301f 100644 --- a/paddle/fluid/framework/selected_rows_test.cc +++ b/paddle/fluid/framework/selected_rows_test.cc @@ -91,7 +91,7 @@ TEST(SelectedRows, SparseTable) { ASSERT_TRUE(table.HasKey(10)); ASSERT_TRUE(table.HasKey(8)); ASSERT_TRUE(table.HasKey(6)); - ASSERT_EQ(table.rows().size(), 3); + ASSERT_EQ(table.rows().size(), 3UL); framework::Tensor ids; ids.Resize(framework::make_ddim({4})); diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index 9fe92831e3aee745f5899ca247259f46e37fbfbb..057482b76dbe76acdb72f9e9e7c28e9892b9d588 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -37,6 +37,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto size = src.numel() * SizeOfType(src.type()); if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) { + if (src_ptr == dst_ptr) { + VLOG(3) << "Skip copy the same data async from " << src_place << " to " + << dst_place; + return; + } memory::Copy(boost::get(dst_place), dst_ptr, boost::get(src_place), src_ptr, size); } @@ -72,6 +77,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto stream = reinterpret_cast(ctx).stream(); if (platform::is_same_place(src_place, dst_place)) { + if (src_ptr == dst_ptr) { + VLOG(3) << "Skip copy the same data async from " << src_place << " to " + << dst_place; + return; + } memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream); } else { @@ -115,6 +125,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, auto dst_ptr = dst->mutable_data(dst_place, src.type()); auto size = src.numel() * SizeOfType(src.type()); if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) { + if (src_ptr == dst_ptr) { + VLOG(3) << "Skip copy the same data from " << src_place << " to " + << dst_place; + return; + } memory::Copy(boost::get(dst_place), dst_ptr, boost::get(src_place), src_ptr, size); } @@ -131,6 +146,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr); } else if (platform::is_gpu_place(src_place) && platform::is_gpu_place(dst_place)) { + if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) { + VLOG(3) << "Skip copy the same data from " << src_place << " to " + << dst_place; + return; + } auto src_gpu_place = boost::get(src_place); auto dst_gpu_place = boost::get(dst_place); memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr); diff --git a/paddle/fluid/framework/tensor_util_test.cc b/paddle/fluid/framework/tensor_util_test.cc index f8aeede3e21a4e32fee8e49a1d5fef1703deffbe..17c55378178325b40e394f4b422c22c1c10bd130 100644 --- a/paddle/fluid/framework/tensor_util_test.cc +++ b/paddle/fluid/framework/tensor_util_test.cc @@ -41,6 +41,11 @@ TEST(TensorCopy, Tensor) { EXPECT_EQ(src_ptr[i], dst_ptr[i]); } + TensorCopy(dst_tensor, *cpu_place, &dst_tensor); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], dst_ptr[i]); + } + EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout()); Tensor slice_tensor = src_tensor.Slice(1, 2); @@ -82,6 +87,15 @@ TEST(TensorCopy, Tensor) { EXPECT_EQ(src_ptr[i], dst_ptr[i]); } + // Copy the same tensor + TensorCopy(gpu_tensor, *gpu_place, gpu_ctx, &gpu_tensor); + gpu_ctx.Wait(); + const int* dst_ptr_tmp = dst_tensor.data(); + EXPECT_NE(src_ptr, dst_ptr_tmp); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], dst_ptr_tmp[i]); + } + Tensor slice_tensor = src_tensor.Slice(1, 2); // CPU Slice Tensor to GPU Tensor diff --git a/paddle/fluid/framework/var_desc.h b/paddle/fluid/framework/var_desc.h index e33849ef502fb10b913e7e28cbd0abdb8b8ff9bb..9d3fb811191c207c75845ef8f8486e8beac7525a 100644 --- a/paddle/fluid/framework/var_desc.h +++ b/paddle/fluid/framework/var_desc.h @@ -59,6 +59,7 @@ class VarDesc { public: explicit VarDesc(const std::string &name) { desc_.set_name(name); + // TODO(paddle-dev): Why default to lodtensor. desc_.mutable_type()->set_type(proto::VarType::LOD_TENSOR); } diff --git a/paddle/fluid/framework/variable.h b/paddle/fluid/framework/variable.h index 067e0c2b8389f88639fd9b95bd680702517efee1..873e1b20a584df3ba90cf5c1a62a3879bf98ce5c 100644 --- a/paddle/fluid/framework/variable.h +++ b/paddle/fluid/framework/variable.h @@ -38,8 +38,12 @@ class Variable { template T* GetMutable() { - if (!IsType()) { + if (!holder_) { holder_.reset(new PlaceholderImpl(new T())); + } else { + PADDLE_ENFORCE(IsType(), + "Variable must be type %s, the holding type is %s", + typeid(T).name(), holder_->Type().name()); } return static_cast(holder_->Ptr()); } diff --git a/paddle/fluid/framework/variable_test.cc b/paddle/fluid/framework/variable_test.cc index c5c1d215f4a6affae0a3bdafacec40a2aee2ca19..003dcfd3dfe5ecfd563a686bb72b061aff602f73 100644 --- a/paddle/fluid/framework/variable_test.cc +++ b/paddle/fluid/framework/variable_test.cc @@ -33,9 +33,10 @@ TEST(Variable, GetMutable) { const Tensor& tt = v->Get(); EXPECT_EQ(1234, tt.content_); - std::string* s = v->GetMutable(); - *s = "hello"; - - const std::string& ss = v->Get(); - EXPECT_EQ("hello", ss); + try { + v->GetMutable(); + } catch (std::exception& e) { + return; + } + EXPECT_TRUE(false); } diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index ec1bc7825dd21628f5c37ea44a154abe7b7e8c73..9794a193bcfaae19552b1f6fbdf2dab2898033d5 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -19,9 +19,19 @@ cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api) add_subdirectory(api) +set(STATIC_INFERENCE_APIS paddle_fluid_api paddle_inference_api analysis_predictor) +set(SHARED_INFERENCE_SRCS + io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc + ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc + ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc) +if (WITH_GPU AND TENSORRT_FOUND) + set(STATIC_INFERENCE_APIS ${STATIC_INFERENCE_APIS} paddle_inference_tensorrt_subgraph_engine) + set(SHARED_INFERENCE_SRCS ${SHARED_INFERENCE_SRCS} ${CMAKE_CURRENT_SOURCE_DIR}/api/api_tensorrt_subgraph_engine.cc) +endif() + # Create static library -cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api - analysis_predictor zero_copy_tensor) +cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor) + if(NOT APPLE) # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac. set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym") @@ -29,10 +39,7 @@ if(NOT APPLE) endif() # Create shared library -cc_library(paddle_fluid_shared SHARED - SRCS io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc - ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc - ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc +cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS} DEPS ${fluid_modules} paddle_fluid_api) set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) diff --git a/paddle/fluid/inference/analysis/analyzer.cc b/paddle/fluid/inference/analysis/analyzer.cc index 8a8aeb5e09a0d9a6746f6d6d61c547363e0e2d30..2e79d495d5ff00000000029ac0f6eb486aaea94a 100644 --- a/paddle/fluid/inference/analysis/analyzer.cc +++ b/paddle/fluid/inference/analysis/analyzer.cc @@ -70,7 +70,7 @@ class DfgPassManagerImpl final : public DfgPassManager { auto trt_teller = [&](const Node* node) { std::unordered_set teller_set( {"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid", - "depthwise_conv2d", "batch_norm", "concat", "tanh", + "depthwise_conv2d", "batch_norm", "concat", "tanh", "pad", "elementwise_add", "dropout"}); if (!node->IsFunction()) return false; @@ -101,7 +101,13 @@ Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); } void Analyzer::Run(Argument* argument) { std::vector passes; - for (auto& pass : all_ir_passes_) { +#ifdef PADDLE_WITH_MKLDNN + if (use_mkldnn_) { + VLOG(3) << "Adding MKL-DNN placement pass"; + passes.push_back("mkldnn_placement_pass"); + } +#endif + for (auto& pass : ir_passes_) { if (!disabled_ir_passes_.count(pass)) { passes.push_back(pass); passes.push_back("graph_viz_pass"); // add graphviz for debug. @@ -117,11 +123,26 @@ void Analyzer::Run(Argument* argument) { } } +Analyzer& Analyzer::IncludeAllIrPasses() { + ir_passes_ = all_ir_passes_; + return *this; +} + Analyzer& Analyzer::DisableIrPasses(const std::vector& passes) { disabled_ir_passes_.insert(passes.begin(), passes.end()); return *this; } +Analyzer& Analyzer::IncludeIrPasses(const std::vector& passes) { + ir_passes_ = passes; + return *this; +} + +Analyzer& Analyzer::SetUseMkldnn(bool use_mkldnn) { + use_mkldnn_ = use_mkldnn; + return *this; +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/analyzer.h b/paddle/fluid/inference/analysis/analyzer.h index 0aa9367bf5692e53e2a1f1247523cf9a4f0b3a1d..c51a4fdb2f6b27e54637481c23bf6f1f6ec97718 100644 --- a/paddle/fluid/inference/analysis/analyzer.h +++ b/paddle/fluid/inference/analysis/analyzer.h @@ -54,6 +54,9 @@ class Analyzer : public OrderedRegistry { void Run(Argument* argument); Analyzer& DisableIrPasses(const std::vector& passes); + Analyzer& IncludeIrPasses(const std::vector& passes); + Analyzer& IncludeAllIrPasses(); + Analyzer& SetUseMkldnn(bool use_mkldnn); DISABLE_COPY_AND_ASSIGN(Analyzer); @@ -64,21 +67,29 @@ class Analyzer : public OrderedRegistry { // larger fusion. const std::vector all_ir_passes_{{ // Manual update the passes here. - "infer_clean_graph_pass", // - "attention_lstm_fuse_pass", // - "embedding_fc_lstm_fuse_pass", // - "fc_lstm_fuse_pass", // - "mul_lstm_fuse_pass", // - "fc_gru_fuse_pass", // - "mul_gru_fuse_pass", // - "seq_concat_fc_fuse_pass", // - "fc_fuse_pass", // + "infer_clean_graph_pass", // + "attention_lstm_fuse_pass", // + "seqconv_eltadd_relu_fuse_pass", // + "embedding_fc_lstm_fuse_pass", // + "fc_lstm_fuse_pass", // + "mul_lstm_fuse_pass", // + "fc_gru_fuse_pass", // + "mul_gru_fuse_pass", // + "seq_concat_fc_fuse_pass", // + "fc_fuse_pass", // + "conv_bn_fuse_pass", // + "conv_eltwiseadd_bn_fuse_pass", // #ifdef PADDLE_WITH_MKLDNN - "conv_relu_mkldnn_fuse_pass", // + "conv_bias_mkldnn_fuse_pass", // + "conv_relu_mkldnn_fuse_pass", // + "conv_elementwise_add_mkldnn_fuse_pass", // #endif }}; std::unordered_set disabled_ir_passes_; + // Ir passes to run + std::vector ir_passes_; + bool use_mkldnn_; }; } // namespace analysis diff --git a/paddle/fluid/inference/analysis/analyzer_tester.cc b/paddle/fluid/inference/analysis/analyzer_tester.cc index f90910ac0d0a897ef01d4ca2bd0bca575baf4c40..5430e5c1ef1c70d27295ebc1a9bd427cd95f006a 100644 --- a/paddle/fluid/inference/analysis/analyzer_tester.cc +++ b/paddle/fluid/inference/analysis/analyzer_tester.cc @@ -51,9 +51,7 @@ void TestWord2vecPrediction(const std::string& model_path) { config.model_dir = model_path; config.use_gpu = false; config.device = 0; - auto predictor = - ::paddle::CreatePaddlePredictor( - config); + auto predictor = ::paddle::CreatePaddlePredictor(config); // One single batch diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc index 3bc6af5241c41bd805699121d614d431d46d863f..eec665767164dc6e79738890947c54d7f7217037 100644 --- a/paddle/fluid/inference/api/analysis_predictor.cc +++ b/paddle/fluid/inference/api/analysis_predictor.cc @@ -25,9 +25,11 @@ #include "paddle/fluid/inference/api/paddle_inference_api.h" #include "paddle/fluid/inference/api/paddle_inference_pass.h" #include "paddle/fluid/inference/utils/singleton.h" +#include "paddle/fluid/platform/cpu_helper.h" #include "paddle/fluid/platform/profiler.h" DECLARE_bool(profile); +DECLARE_int32(paddle_num_threads); namespace paddle { @@ -47,6 +49,9 @@ bool AnalysisPredictor::Init( } #endif + // no matter with or without MKLDNN + paddle::platform::SetNumThreads(FLAGS_paddle_num_threads); + if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); LOG(WARNING) << "ir optimize only supports CPU currently, enable_ir_optim " @@ -72,10 +77,6 @@ bool AnalysisPredictor::Init( inference_program_ = program; } - if (config_._use_mkldnn) { - executor_->EnableMKLDNN(*inference_program_); - } - executor_->Prepare(scope_.get(), *inference_program_, 0, config_.use_feed_fetch_ops); @@ -220,10 +221,24 @@ void AnalysisPredictor::OptimizeInferenceProgram() { argument_.origin_program_desc.reset( new ProgramDesc(*inference_program_->Proto())); - PADDLE_ENFORCE( - config_.ir_mode == contrib::AnalysisConfig::IrPassMode::kExclude, - "Only kExclude is supported yet."); - Analyzer().DisableIrPasses(config_.ir_passes).Run(&argument_); + + switch (config_.ir_mode) { + case contrib::AnalysisConfig::IrPassMode::kExclude: + Analyzer() + .IncludeAllIrPasses() + .SetUseMkldnn(config_._use_mkldnn) + .DisableIrPasses(config_.ir_passes) + .Run(&argument_); + break; + case contrib::AnalysisConfig::IrPassMode::kInclude: + Analyzer() + .SetUseMkldnn(config_._use_mkldnn) + .IncludeIrPasses(config_.ir_passes) + .Run(&argument_); + break; + default: + LOG(ERROR) << "Only kExclude and kInclude modes are supoorted yet."; + } CHECK(argument_.transformed_program_desc); VLOG(5) << "to prepare executor"; @@ -335,6 +350,19 @@ bool AnalysisPredictor::LoadProgramDesc() { } return true; } + +AnalysisPredictor::~AnalysisPredictor() { +#if !defined(_WIN32) + if (FLAGS_profile) { + platform::DisableProfiler(platform::EventSortingKey::kTotal, + "./profile.log"); + } +#endif + if (sub_scope_) { + scope_->DeleteScope(sub_scope_); + } +} + std::unique_ptr AnalysisPredictor::Clone() { auto *x = new AnalysisPredictor(config_); x->Init(scope_, inference_program_); diff --git a/paddle/fluid/inference/api/analysis_predictor.h b/paddle/fluid/inference/api/analysis_predictor.h index 0d01d7ac2b29ea6364b07af9bb3bdeb5ced6bd00..5a9f4d36959d4ee7ca16dec769d9d1283b8787cb 100644 --- a/paddle/fluid/inference/api/analysis_predictor.h +++ b/paddle/fluid/inference/api/analysis_predictor.h @@ -72,6 +72,7 @@ class AnalysisPredictor : public PaddlePredictor { template void GetFetchOne(const framework::LoDTensor &fetchs, PaddleTensor *output_data); + ~AnalysisPredictor(); private: contrib::AnalysisConfig config_; diff --git a/paddle/fluid/inference/api/analysis_predictor_tester.cc b/paddle/fluid/inference/api/analysis_predictor_tester.cc index 1d25f55b3188a684fe38df1417d114348cfa2e8a..13c25da1b52742e6114b294847c21ce735b9fc21 100644 --- a/paddle/fluid/inference/api/analysis_predictor_tester.cc +++ b/paddle/fluid/inference/api/analysis_predictor_tester.cc @@ -27,9 +27,7 @@ TEST(AnalysisPredictor, ZeroCopy) { config.model_dir = FLAGS_dirname + "/word2vec.inference.model"; config.use_feed_fetch_ops = false; - auto predictor = - CreatePaddlePredictor( - config); + auto predictor = CreatePaddlePredictor(config); auto w0 = predictor->GetInputTensor("firstw"); auto w1 = predictor->GetInputTensor("secondw"); diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc index 6682e0a81b20c82aa668a249d37986386d769c83..7cda9c5d8a8366bd097491f37f5352a10e4fb16c 100644 --- a/paddle/fluid/inference/api/api_impl.cc +++ b/paddle/fluid/inference/api/api_impl.cc @@ -23,9 +23,11 @@ limitations under the License. */ #include "paddle/fluid/framework/feed_fetch_method.h" #include "paddle/fluid/inference/api/api_impl.h" #include "paddle/fluid/inference/api/helper.h" +#include "paddle/fluid/platform/cpu_helper.h" #include "paddle/fluid/platform/profiler.h" DEFINE_bool(profile, false, "Turn on profiler for fluid"); +DECLARE_int32(paddle_num_threads); namespace paddle { namespace { @@ -72,6 +74,9 @@ bool NativePaddlePredictor::Init( } #endif + // no matter with or without MKLDNN + paddle::platform::SetNumThreads(FLAGS_paddle_num_threads); + if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); } else { diff --git a/paddle/fluid/inference/api/api_impl_tester.cc b/paddle/fluid/inference/api/api_impl_tester.cc index bed7c871311e476b4ed113d982c690383ad732de..b7b8ee6ea08fe907f3f052ae1118f782ac853ca7 100644 --- a/paddle/fluid/inference/api/api_impl_tester.cc +++ b/paddle/fluid/inference/api/api_impl_tester.cc @@ -205,7 +205,7 @@ void MainThreadsWord2Vec(bool use_gpu) { float* ref_data = refs[tid].data(); EXPECT_EQ(refs[tid].numel(), static_cast(len / sizeof(float))); for (int i = 0; i < refs[tid].numel(); ++i) { - EXPECT_NEAR(ref_data[i], data[i], ACC_DIFF); + EXPECT_NEAR(ref_data[i], data[i], 2e-3); } }); } diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc index 5ee6a5a93168f58770067f76ca7f6bb6f67b2965..7ac468ee4d33f49bba20a07c976055a083743cbc 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc @@ -185,3 +185,4 @@ USE_TRT_CONVERTER(softmax); USE_TRT_CONVERTER(batch_norm); USE_TRT_CONVERTER(concat); USE_TRT_CONVERTER(dropout); +USE_TRT_CONVERTER(pad); diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc index fc6310e90b0257bc84742fb617a00f5778bb1866..702158ea3bcab854eece3ccd40724d92efcbae67 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc @@ -41,11 +41,8 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) { config1.device = 0; config1.max_batch_size = 10; - auto predictor0 = - CreatePaddlePredictor(config0); - auto predictor1 = - CreatePaddlePredictor(config1); + auto predictor0 = CreatePaddlePredictor(config0); + auto predictor1 = CreatePaddlePredictor(config1); for (int batch_id = 0; batch_id < 1; batch_id++) { //# 2. Prepare input. diff --git a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt index d4e6bb3e4a4ceb361ccd35121d0ecf84a764243e..03f0f726eb61c2619c7719a865383090f86b5b7f 100644 --- a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt +++ b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt @@ -3,6 +3,7 @@ project(cpp_inference_demo CXX C) option(WITH_MKL "Compile demo with MKL/OpenBlas support, default use MKL." ON) option(WITH_GPU "Compile demo with GPU/CPU, default use CPU." OFF) option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static." ON) +option(USE_TENSORRT "Compile demo with TensorRT." OFF) macro(safe_set_static_flag) foreach(flag_var @@ -60,6 +61,13 @@ endif(NOT WIN32) include_directories("${PADDLE_LIB}/third_party/boost") include_directories("${PADDLE_LIB}/third_party/eigen3") +if (NOT WIN32) + if (USE_TENSORRT AND WITH_GPU) + include_directories("${TENSORRT_INCLUDE_DIR}") + link_directories("${TENSORRT_LIB_DIR}") + endif() +endif(NOT WIN32) + if (NOT WIN32) link_directories("${PADDLE_LIB}/third_party/install/snappy/lib") link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib") @@ -69,7 +77,7 @@ endif(NOT WIN32) link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib") link_directories("${PADDLE_LIB}/third_party/install/glog/lib") link_directories("${PADDLE_LIB}/third_party/install/gflags/lib") -link_directories("${PADDLE_LIB}/paddle/fluid/inference") +link_directories("${PADDLE_LIB}/paddle/lib") add_executable(${DEMO_NAME} ${DEMO_NAME}.cc) @@ -89,10 +97,10 @@ endif() # Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a if(WITH_STATIC_LIB) set(DEPS - ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) + ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) else() set(DEPS - ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX}) + ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX}) endif() if (NOT WIN32) @@ -112,6 +120,10 @@ endif(NOT WIN32) if(WITH_GPU) if(NOT WIN32) + if (USE_TENSORRT) + set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer${CMAKE_STATIC_LIBRARY_SUFFIX}) + set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer_plugin${CMAKE_STATIC_LIBRARY_SUFFIX}) + endif() set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX}) else() set(DEPS ${DEPS} ${CUDA_LIB}/cudart${CMAKE_STATIC_LIBRARY_SUFFIX} ) diff --git a/paddle/fluid/inference/api/demo_ci/run.sh b/paddle/fluid/inference/api/demo_ci/run.sh index 44335a872f2e00b34e29a9e7601cb390a460362c..67994aad70a40c0e0c8a311914d4ea40b96eaf1e 100755 --- a/paddle/fluid/inference/api/demo_ci/run.sh +++ b/paddle/fluid/inference/api/demo_ci/run.sh @@ -3,11 +3,15 @@ PADDLE_ROOT=$1 TURN_ON_MKL=$2 # use MKL or Openblas TEST_GPU_CPU=$3 # test both GPU/CPU mode or only CPU mode DATA_DIR=$4 # dataset +TENSORRT_INCLUDE_DIR=$5 # TensorRT header file dir, defalut to /usr/local/TensorRT/include +TENSORRT_LIB_DIR=$6 # TensorRT lib file dir, default to /usr/local/TensorRT/lib +inference_install_dir=${PADDLE_ROOT}/build/fluid_inference_install_dir + cd `dirname $0` current_dir=`pwd` if [ $2 == ON ]; then # You can export yourself if move the install path - MKL_LIB=${PADDLE_ROOT}/build/fluid_install_dir/third_party/install/mklml/lib + MKL_LIB=${inference_install_dir}/third_party/install/mklml/lib export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${MKL_LIB} fi if [ $3 == ON ]; then @@ -16,6 +20,11 @@ else use_gpu_list='false' fi +USE_TENSORRT=OFF +if [ [-d"$TENSORRT_INCLUDE_DIR"] -a [-d"$TENSORRT_LIB_DIR"] ]; then + USE_TENSORRT=ON +fi + PREFIX=inference-vis-demos%2F URL_ROOT=http://paddlemodels.cdn.bcebos.com/${PREFIX} @@ -47,7 +56,7 @@ cd build for WITH_STATIC_LIB in ON OFF; do # -----simple_on_word2vec----- rm -rf * - cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \ + cmake .. -DPADDLE_LIB=${inference_install_dir} \ -DWITH_MKL=$TURN_ON_MKL \ -DDEMO_NAME=simple_on_word2vec \ -DWITH_GPU=$TEST_GPU_CPU \ @@ -67,7 +76,7 @@ for WITH_STATIC_LIB in ON OFF; do fi # ---------vis_demo--------- rm -rf * - cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \ + cmake .. -DPADDLE_LIB=${inference_install_dir} \ -DWITH_MKL=$TURN_ON_MKL \ -DDEMO_NAME=vis_demo \ -DWITH_GPU=$TEST_GPU_CPU \ @@ -86,5 +95,23 @@ for WITH_STATIC_LIB in ON OFF; do fi done done + + # --------tensorrt mobilenet------ + if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then + rm -rf * + cmake .. -DPADDLE_LIB=${inference_install_dir} \ + -DWITH_MKL=$TURN_ON_MKL \ + -DDEMO_NAME=trt_mobilenet_demo \ + -DWITH_GPU=$TEST_GPU_CPU \ + -DWITH_STATIC_LIB=$WITH_STATIC_LIB \ + -DUSE_TENSORRT=$USE_TENSORRT \ + -DTENSORRT_INCLUDE_DIR=$TENSORRT_INCLUDE_DIR \ + -DTENSORRT_LIB_DIR=$TENSORRT_LIB_DIR + make -j + ./trt_mobilenet_demo \ + --modeldir=$DATA_DIR/mobilenet/model \ + --data=$DATA_DIR/mobilenet/data.txt \ + --refer=$DATA_DIR/mobilenet/result.txt + fi done set +x diff --git a/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc b/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc index 8058d7e881025b1d806efe187d4523adadff367d..5446fd4d4256c10442a53ea09a447cf308cbd681 100644 --- a/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc +++ b/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc @@ -23,7 +23,7 @@ limitations under the License. */ #include #include //NOLINT -#include "paddle/fluid/inference/paddle_inference_api.h" +#include "paddle/include/paddle_inference_api.h" DEFINE_string(dirname, "", "Directory of the inference model."); DEFINE_bool(use_gpu, false, "Whether use gpu."); @@ -42,8 +42,7 @@ void Main(bool use_gpu) { config.use_gpu = use_gpu; config.fraction_of_gpu_memory = 0.15; config.device = 0; - auto predictor = - CreatePaddlePredictor(config); + auto predictor = CreatePaddlePredictor(config); for (int batch_id = 0; batch_id < 3; batch_id++) { //# 2. Prepare input. @@ -85,8 +84,7 @@ void MainThreads(int num_threads, bool use_gpu) { config.use_gpu = use_gpu; config.fraction_of_gpu_memory = 0.15; config.device = 0; - auto main_predictor = - CreatePaddlePredictor(config); + auto main_predictor = CreatePaddlePredictor(config); std::vector threads; for (int tid = 0; tid < num_threads; ++tid) { diff --git a/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc b/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc new file mode 100644 index 0000000000000000000000000000000000000000..4a8404f21c6ec6a1647e964ac3538b4b49151009 --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc @@ -0,0 +1,82 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file contains demo of mobilenet for tensorrt. + */ + +#include +#include // use glog instead of CHECK to avoid importing other paddle header files. +#include "utils.h" // NOLINT + +DECLARE_double(fraction_of_gpu_memory_to_use); +DEFINE_string(modeldir, "", "Directory of the inference model."); +DEFINE_string(refer, "", "path to reference result for comparison."); +DEFINE_string( + data, "", + "path of data; each line is a record, format is " + "'\t predictor; + paddle::contrib::MixedRTConfig config; + config.param_file = FLAGS_modeldir + "/__params__"; + config.prog_file = FLAGS_modeldir + "/__model__"; + config.use_gpu = true; + config.device = 0; + config.max_batch_size = 1; + config.fraction_of_gpu_memory = 0.1; // set by yourself + predictor = CreatePaddlePredictor(config); + + VLOG(3) << "begin to process data"; + // Just a single batch of data. + std::string line; + std::ifstream file(FLAGS_data); + std::getline(file, line); + auto record = ProcessALine(line); + file.close(); + + // Inference. + PaddleTensor input; + input.shape = record.shape; + input.data = + PaddleBuf(record.data.data(), record.data.size() * sizeof(float)); + input.dtype = PaddleDType::FLOAT32; + + VLOG(3) << "run executor"; + std::vector output; + predictor->Run({input}, &output, 1); + + VLOG(3) << "output.size " << output.size(); + auto& tensor = output.front(); + VLOG(3) << "output: " << SummaryTensor(tensor); + + // compare with reference result + CheckOutput(FLAGS_refer, tensor); +} + +} // namespace demo +} // namespace paddle + +int main(int argc, char** argv) { + google::ParseCommandLineFlags(&argc, &argv, true); + paddle::demo::Main(); + return 0; +} diff --git a/paddle/fluid/inference/api/demo_ci/utils.h b/paddle/fluid/inference/api/demo_ci/utils.h index cb8990671162dff47228736e69617229528cc093..d70c6aea791219a40c3164b51499f9d5e562be71 100644 --- a/paddle/fluid/inference/api/demo_ci/utils.h +++ b/paddle/fluid/inference/api/demo_ci/utils.h @@ -14,13 +14,20 @@ #pragma once #include +#include +#include #include #include -#include "paddle/fluid/inference/paddle_inference_api.h" +#include "paddle/include/paddle_inference_api.h" namespace paddle { namespace demo { +struct Record { + std::vector data; + std::vector shape; +}; + static void split(const std::string& str, char sep, std::vector* pieces) { pieces->clear(); @@ -39,6 +46,58 @@ static void split(const std::string& str, char sep, } } +Record ProcessALine(const std::string& line) { + VLOG(3) << "process a line"; + std::vector columns; + split(line, '\t', &columns); + CHECK_EQ(columns.size(), 2UL) + << "data format error, should be \t"; + + Record record; + std::vector data_strs; + split(columns[0], ' ', &data_strs); + for (auto& d : data_strs) { + record.data.push_back(std::stof(d)); + } + + std::vector shape_strs; + split(columns[1], ' ', &shape_strs); + for (auto& s : shape_strs) { + record.shape.push_back(std::stoi(s)); + } + VLOG(3) << "data size " << record.data.size(); + VLOG(3) << "data shape size " << record.shape.size(); + return record; +} + +void CheckOutput(const std::string& referfile, const PaddleTensor& output) { + std::string line; + std::ifstream file(referfile); + std::getline(file, line); + auto refer = ProcessALine(line); + file.close(); + + size_t numel = output.data.length() / PaddleDtypeSize(output.dtype); + VLOG(3) << "predictor output numel " << numel; + VLOG(3) << "reference output numel " << refer.data.size(); + CHECK_EQ(numel, refer.data.size()); + switch (output.dtype) { + case PaddleDType::INT64: { + for (size_t i = 0; i < numel; ++i) { + CHECK_EQ(static_cast(output.data.data())[i], refer.data[i]); + } + break; + } + case PaddleDType::FLOAT32: + for (size_t i = 0; i < numel; ++i) { + CHECK_LT( + fabs(static_cast(output.data.data())[i] - refer.data[i]), + 1e-5); + } + break; + } +} + /* * Get a summary of a PaddleTensor content. */ diff --git a/paddle/fluid/inference/api/demo_ci/vis_demo.cc b/paddle/fluid/inference/api/demo_ci/vis_demo.cc index fb59cea457027854a099574c867299450690e61c..8d546e3e9c740c10bcf2984e073c956e3612625c 100644 --- a/paddle/fluid/inference/api/demo_ci/vis_demo.cc +++ b/paddle/fluid/inference/api/demo_ci/vis_demo.cc @@ -18,11 +18,7 @@ limitations under the License. */ #include #include // use glog instead of CHECK to avoid importing other paddle header files. -#include -#include - -// #include "paddle/fluid/platform/enforce.h" -#include "paddle/fluid/inference/demo_ci/utils.h" +#include "utils.h" // NOLINT #ifdef PADDLE_WITH_CUDA DECLARE_double(fraction_of_gpu_memory_to_use); @@ -38,70 +34,13 @@ DEFINE_bool(use_gpu, false, "Whether use gpu."); namespace paddle { namespace demo { -struct Record { - std::vector data; - std::vector shape; -}; - -void split(const std::string& str, char sep, std::vector* pieces); - -Record ProcessALine(const std::string& line) { - VLOG(3) << "process a line"; - std::vector columns; - split(line, '\t', &columns); - CHECK_EQ(columns.size(), 2UL) - << "data format error, should be \t"; - - Record record; - std::vector data_strs; - split(columns[0], ' ', &data_strs); - for (auto& d : data_strs) { - record.data.push_back(std::stof(d)); - } - - std::vector shape_strs; - split(columns[1], ' ', &shape_strs); - for (auto& s : shape_strs) { - record.shape.push_back(std::stoi(s)); - } - VLOG(3) << "data size " << record.data.size(); - VLOG(3) << "data shape size " << record.shape.size(); - return record; -} - -void CheckOutput(const std::string& referfile, const PaddleTensor& output) { - std::string line; - std::ifstream file(referfile); - std::getline(file, line); - auto refer = ProcessALine(line); - file.close(); - - size_t numel = output.data.length() / PaddleDtypeSize(output.dtype); - VLOG(3) << "predictor output numel " << numel; - VLOG(3) << "reference output numel " << refer.data.size(); - CHECK_EQ(numel, refer.data.size()); - switch (output.dtype) { - case PaddleDType::INT64: { - for (size_t i = 0; i < numel; ++i) { - CHECK_EQ(static_cast(output.data.data())[i], refer.data[i]); - } - break; - } - case PaddleDType::FLOAT32: - for (size_t i = 0; i < numel; ++i) { - CHECK_LT( - fabs(static_cast(output.data.data())[i] - refer.data[i]), - 1e-5); - } - break; - } -} - +using contrib::AnalysisConfig; /* - * Use the native fluid engine to inference the demo. + * Use the native and analysis fluid engine to inference the demo. */ void Main(bool use_gpu) { - NativeConfig config; + std::unique_ptr predictor, analysis_predictor; + AnalysisConfig config; config.param_file = FLAGS_modeldir + "/__params__"; config.prog_file = FLAGS_modeldir + "/__model__"; config.use_gpu = use_gpu; @@ -111,8 +50,8 @@ void Main(bool use_gpu) { } VLOG(3) << "init predictor"; - auto predictor = - CreatePaddlePredictor(config); + predictor = CreatePaddlePredictor(config); + analysis_predictor = CreatePaddlePredictor(config); VLOG(3) << "begin to process data"; // Just a single batch of data. @@ -130,8 +69,8 @@ void Main(bool use_gpu) { input.dtype = PaddleDType::FLOAT32; VLOG(3) << "run executor"; - std::vector output; - predictor->Run({input}, &output); + std::vector output, analysis_output; + predictor->Run({input}, &output, 1); VLOG(3) << "output.size " << output.size(); auto& tensor = output.front(); @@ -139,6 +78,10 @@ void Main(bool use_gpu) { // compare with reference result CheckOutput(FLAGS_refer, tensor); + + // the analysis_output has some diff with native_output, + // TODO(luotao): add CheckOutput for analysis_output later. + analysis_predictor->Run({input}, &analysis_output, 1); } } // namespace demo @@ -146,9 +89,10 @@ void Main(bool use_gpu) { int main(int argc, char** argv) { google::ParseCommandLineFlags(&argc, &argv, true); - paddle::demo::Main(false /* use_gpu*/); if (FLAGS_use_gpu) { paddle::demo::Main(true /*use_gpu*/); + } else { + paddle::demo::Main(false /*use_gpu*/); } return 0; } diff --git a/paddle/fluid/inference/api/paddle_inference_api.h b/paddle/fluid/inference/api/paddle_inference_api.h index d2876dc27c8826c2f27be21fa0b9fef92d03067a..07ee6e72d1053d2271b8f8d69ce38003f5e038a0 100644 --- a/paddle/fluid/inference/api/paddle_inference_api.h +++ b/paddle/fluid/inference/api/paddle_inference_api.h @@ -259,10 +259,17 @@ struct AnalysisConfig : public NativeConfig { kExclude // Specify the disabled passes in `ir_passes`. }; + void SetIncludeMode() { + ir_mode = IrPassMode::kInclude; + // this pass has to be run at the beginning of all fuse passes + ir_passes = {"infer_clean_graph_pass"}; + } + // Determine whether to perform graph optimization. bool enable_ir_optim = true; // Manually determine the IR passes to run. IrPassMode ir_mode{IrPassMode::kExclude}; + // passes to be excluded/included std::vector ir_passes{"embedding_fc_lstm_fuse_pass"}; // NOT stable yet. diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index fac1babf6ec6131f84d3e3b9fc6efedd9f9f6cfc..0a35e10f6936313928ab21a6f17c40335e8fc882 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,7 +1,7 @@ # Add TRT tests nv_library(tensorrt_converter SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc -batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc +batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc pad_op.cc DEPS tensorrt_engine operator scope framework_proto op_registry) nv_test(test_op_converter SRCS test_op_converter.cc DEPS @@ -26,6 +26,8 @@ nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL) nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL) - nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine dropout_op SERIAL) + +nv_test(test_trt_pad_op SRCS test_pad_op.cc pad_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine pad_op SERIAL) diff --git a/paddle/fluid/inference/tensorrt/convert/pad_op.cc b/paddle/fluid/inference/tensorrt/convert/pad_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..218030a591fcc7e533ef37062265449d4b6044bc --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/pad_op.cc @@ -0,0 +1,68 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * PadOp. + */ +class PadOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(4) << "convert a fluid transpose op to tensorrt tranpose layer"; + + framework::OpDesc op_desc(op, nullptr); + // Declare inputs + auto* input = engine_->GetITensor(op_desc.Input("X")[0]); + + const std::vector paddings = + boost::get>(op_desc.GetAttr("paddings")); + const float pad_value = boost::get(op_desc.GetAttr("pad_value")); + + nvinfer1::Dims input_shape = input->getDimensions(); + int nbDims = input_shape.nbDims; + int pad_size = static_cast(paddings.size()); + PADDLE_ENFORCE_GE(nbDims, 2); + PADDLE_ENFORCE_EQ((nbDims + 1) * 2, pad_size); + PADDLE_ENFORCE(pad_value == 0.0, "The pad layer of TRT only support zero."); + + nvinfer1::DimsHW pre_pad(paddings[pad_size - 4], paddings[pad_size - 2]); + nvinfer1::DimsHW post_pad(paddings[pad_size - 3], paddings[pad_size - 1]); + + auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Padding, + *const_cast(input), + pre_pad, post_pad); + + PADDLE_ENFORCE(layer != nullptr); + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + layer->setName(("scale (Output: " + output_name + ")").c_str()); + layer->getOutput(0)->setName(output_name.c_str()); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +REGISTER_TRT_OP_CONVERTER(pad, PadOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/test_pad_op.cc b/paddle/fluid/inference/tensorrt/convert/test_pad_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ba35d7ddbb2f4e6062713bd82be277e7ad0cb341 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_pad_op.cc @@ -0,0 +1,52 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(PadConverter, main) { + framework::Scope scope; + std::unordered_set parameters; + TRTConvertValidation validator(10, parameters, scope, 1000); + validator.DeclInputVar("pad-X", nvinfer1::Dims3(3, 2, 2)); + validator.DeclOutputVar("pad-Out", nvinfer1::Dims3(3, 3, 5)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("pad"); + desc.SetInput("X", {"pad-X"}); + desc.SetOutput("Out", {"pad-Out"}); + + std::vector paddings = {0, 0, 0, 0, 0, 1, 1, 2}; + float pad_value = 0.0; + desc.SetAttr("paddings", paddings); + desc.SetAttr("pad_value", pad_value); + + LOG(INFO) << "set OP"; + validator.SetOp(*desc.Proto()); + LOG(INFO) << "execute"; + + validator.Execute(2); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(pad); diff --git a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc index 290fb007d8ba94a2d121947fe67c6474586ac0e0..67668298440e9af279e792f786a8123b71172a66 100644 --- a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc @@ -52,9 +52,10 @@ void SetInput(std::vector> *inputs) { } // Easy for profiling independently. -TEST(Analyzer_resnet50, profile) { +void profile(bool use_mkldnn = false) { AnalysisConfig cfg; SetConfig(&cfg); + cfg._use_mkldnn = use_mkldnn; std::vector outputs; std::vector> input_slots_all; @@ -69,6 +70,11 @@ TEST(Analyzer_resnet50, profile) { } } +TEST(Analyzer_resnet50, profile) { profile(); } +#ifndef PADDLE_WITH_MKLDNN +TEST(Analyzer_resnet50, profile_mkldnn) { profile(true /* use_mkldnn */); } +#endif + // Check the fuse status TEST(Analyzer_resnet50, fuse_statis) { AnalysisConfig cfg; @@ -82,15 +88,21 @@ TEST(Analyzer_resnet50, fuse_statis) { } // Compare result of NativeConfig and AnalysisConfig -TEST(Analyzer_resnet50, compare) { +void compare(bool use_mkldnn = false) { AnalysisConfig cfg; SetConfig(&cfg); + cfg._use_mkldnn = use_mkldnn; std::vector> input_slots_all; SetInput(&input_slots_all); CompareNativeAndAnalysis(cfg, input_slots_all); } +TEST(Analyzer_resnet50, compare) { compare(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_resnet50, compare_mkldnn) { compare(true /* use_mkldnn */); } +#endif + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc index c76d72ccd99649913aefcb2aa57fe6061db8ca6d..6399476680c0af83a6d26aea952c58543bdce9ae 100644 --- a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc @@ -308,18 +308,13 @@ TEST(Analyzer_rnn1, ZeroCopy) { PaddlePlace place; int output_size{0}; - auto predictor = - CreatePaddlePredictor( - config); + auto predictor = CreatePaddlePredictor(config); config.use_feed_fetch_ops = true; - auto native_predictor = - CreatePaddlePredictor(config); + auto native_predictor = CreatePaddlePredictor(config); config.use_feed_fetch_ops = true; // the analysis predictor needs feed/fetch. - auto analysis_predictor = - CreatePaddlePredictor( - config); + auto analysis_predictor = CreatePaddlePredictor(config); #define NEW_TENSOR(name__) \ auto name__##_tensor = predictor->GetInputTensor(#name__); diff --git a/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc b/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc index ba04d030b94c0924311dcff5c6a34270a764f877..e0eb919bd896d73a557001982a436fc93f087a74 100644 --- a/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc @@ -18,12 +18,12 @@ namespace paddle { namespace inference { using namespace framework; // NOLINT +static std::vector result_data; struct DataRecord { std::vector>> link_step_data_all; std::vector lod; std::vector> rnn_link_data; - std::vector result_data; size_t num_samples; // total number of samples size_t batch_iter{0}; size_t batch_size{1}; @@ -57,6 +57,7 @@ struct DataRecord { std::ifstream file(path); std::string line; int num_lines = 0; + result_data.clear(); while (std::getline(file, line)) { num_lines++; std::vector data; @@ -135,13 +136,12 @@ TEST(Analyzer_rnn2, profile) { if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { // the first inference result - DataRecord data(FLAGS_infer_data, FLAGS_batch_size); PADDLE_ENFORCE_GT(outputs.size(), 0); size_t size = GetSize(outputs[0]); PADDLE_ENFORCE_GT(size, 0); float *result = static_cast(outputs[0].data.data()); for (size_t i = 0; i < size; i++) { - EXPECT_NEAR(result[i], data.result_data[i], 1e-3); + EXPECT_NEAR(result[i], result_data[i], 1e-3); } } } diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc index cb4671c4379b5f6f144bfd5330866aa38163f4d4..f590ef27967e47ffcb3a97e80dd147efdd1906e6 100644 --- a/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc @@ -183,7 +183,13 @@ TEST(Analyzer_seq_conv1, fuse_statis) { SetConfig(&cfg); int num_ops; auto predictor = CreatePaddlePredictor(cfg); - GetFuseStatis(predictor.get(), &num_ops); + + auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops); + ASSERT_TRUE(fuse_statis.count("fc_fuse")); + ASSERT_TRUE(fuse_statis.count("seqconv_eltadd_relu_fuse")); + EXPECT_EQ(fuse_statis.at("fc_fuse"), 2); + EXPECT_EQ(fuse_statis.at("seqconv_eltadd_relu_fuse"), 6); + EXPECT_EQ(num_ops, 32); } // Compare result of NativeConfig and AnalysisConfig diff --git a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc index 305b8bfe158150d5dfd8bdaee2c0a89afe264de4..8933296490793a7693124eba23f8cf0801881e14 100644 --- a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc @@ -59,9 +59,6 @@ void SetConfig(AnalysisConfig *cfg) { cfg->specify_input_name = true; // TODO(TJ): fix fusion gru cfg->ir_passes.push_back("fc_gru_fuse_pass"); -#ifdef PADDLE_WITH_MKLDNN - cfg->_use_mkldnn = true; -#endif } void SetInput(std::vector> *inputs) { @@ -84,9 +81,10 @@ void SetInput(std::vector> *inputs) { // Easy for profiling independently. // ocr, mobilenet and se_resnext50 -TEST(Analyzer_vis, profile) { +void profile(bool use_mkldnn = false) { AnalysisConfig cfg; SetConfig(&cfg); + cfg._use_mkldnn = use_mkldnn; std::vector outputs; std::vector> input_slots_all; @@ -108,6 +106,12 @@ TEST(Analyzer_vis, profile) { } } +TEST(Analyzer_vis, profile) { profile(); } + +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_vis, profile_mkldnn) { profile(true /* use_mkldnn */); } +#endif + // Check the fuse status TEST(Analyzer_vis, fuse_statis) { AnalysisConfig cfg; @@ -118,15 +122,21 @@ TEST(Analyzer_vis, fuse_statis) { } // Compare result of NativeConfig and AnalysisConfig -TEST(Analyzer_vis, compare) { +void compare(bool use_mkldnn = false) { AnalysisConfig cfg; SetConfig(&cfg); + cfg._use_mkldnn = use_mkldnn; std::vector> input_slots_all; SetInput(&input_slots_all); CompareNativeAndAnalysis(cfg, input_slots_all); } +TEST(Analyzer_vis, compare) { compare(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_vis, compare_mkldnn) { compare(true /* use_mkldnn */); } +#endif + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h index 8603d09cbd09e4cee254faf52c2ccbe8d661994d..b1ee1080030b23e1ef7adefe3a0880f38e9099f5 100644 --- a/paddle/fluid/inference/tests/api/tester_helper.h +++ b/paddle/fluid/inference/tests/api/tester_helper.h @@ -77,11 +77,9 @@ void CompareResult(const std::vector &outputs, std::unique_ptr CreateTestPredictor( const AnalysisConfig &config, bool use_analysis = true) { if (use_analysis) { - return CreatePaddlePredictor(config); + return CreatePaddlePredictor(config); } else { - return CreatePaddlePredictor( - config); + return CreatePaddlePredictor(config); } } @@ -165,7 +163,8 @@ void TestPrediction(const AnalysisConfig &config, const std::vector> &inputs, std::vector *outputs, int num_threads, bool use_analysis = FLAGS_use_analysis) { - LOG(INFO) << "use_analysis: " << use_analysis; + LOG(INFO) << "use_analysis: " << use_analysis + << ", use_mkldnn: " << config._use_mkldnn; if (num_threads == 1) { TestOneThreadPrediction(config, inputs, outputs, use_analysis); } else { @@ -177,6 +176,7 @@ void TestPrediction(const AnalysisConfig &config, void CompareNativeAndAnalysis( const AnalysisConfig &config, const std::vector> &inputs) { + LOG(INFO) << "use_mkldnn: " << config._use_mkldnn; std::vector native_outputs, analysis_outputs; TestOneThreadPrediction(config, inputs, &native_outputs, false); TestOneThreadPrediction(config, inputs, &analysis_outputs, true); diff --git a/paddle/fluid/inference/tests/api/trt_models_tester.cc b/paddle/fluid/inference/tests/api/trt_models_tester.cc index bf320a0cbc2fff5f973c48768281e26d0fde232b..91111f2af56065bbf57ba3a41bddd55ecced1060 100644 --- a/paddle/fluid/inference/tests/api/trt_models_tester.cc +++ b/paddle/fluid/inference/tests/api/trt_models_tester.cc @@ -51,11 +51,8 @@ void CompareTensorRTWithFluid(int batch_size, std::string model_dirname) { config1.model_dir = model_dirname; config1.max_batch_size = batch_size; - auto predictor0 = - CreatePaddlePredictor(config0); - auto predictor1 = - CreatePaddlePredictor(config1); + auto predictor0 = CreatePaddlePredictor(config0); + auto predictor1 = CreatePaddlePredictor(config1); // Prepare inputs int height = 224; int width = 224; diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index 5130d4c2efdb6ec7f2730e71d0998f2f2decd417..c74fcc6c844e8b61dd8d201ee712ce3bc8c3602d 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -86,7 +86,7 @@ function(op_library TARGET) # remove windows unsupported op, because windows has no nccl, no warpctc such ops. foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op" "crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op" - "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op") + "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op") if ("${TARGET}" STREQUAL "${windows_unsupport_op}") return() endif() @@ -230,7 +230,7 @@ if(WITH_DISTRIBUTE) op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) endforeach() - + #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op # listen_and_serv_op sum_op executor SERIAL) @@ -268,6 +268,7 @@ if (WITH_GPU AND TENSORRT_FOUND) else() set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op) endif() +op_library(clip_by_norm_op DEPS selected_rows_functor selected_rows) op_library(sum_op DEPS selected_rows_functor) op_library(sgd_op DEPS selected_rows_functor) op_library(print_op DEPS lod_tensor) @@ -299,11 +300,12 @@ op_library(flatten_op DEPS reshape_op) op_library(sequence_pad_op DEPS sequence_padding) op_library(unstack_op DEPS stack_op) op_library(fake_quantize_op DEPS memory) -op_library(fusion_lstm_op DEPS cpu_lstm_compute) +op_library(fusion_lstm_op DEPS jit_kernel) if (WITH_GPU) op_library(conv_op DEPS vol2col depthwise_conv im2col) op_library(layer_norm_op DEPS cub) op_library(reduce_mean_op DEPS cub) + op_library(affine_channel_op DEPS cub) else() op_library(conv_op DEPS vol2col im2col) endif() diff --git a/paddle/fluid/operators/adadelta_op.cc b/paddle/fluid/operators/adadelta_op.cc index d1970515f58969948b1d2db5847e4344112f77f9..89a7a49e0fa8427826f5d91274912a68f2316b61 100644 --- a/paddle/fluid/operators/adadelta_op.cc +++ b/paddle/fluid/operators/adadelta_op.cc @@ -18,6 +18,7 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; + class AdadeltaOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -31,6 +32,16 @@ class AdadeltaOp : public framework::OperatorWithKernel { "Input(AvgSquaredGrad) of AdadeltaOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("AvgSquaredUpdate"), "Input(AvgSquaredUpdate) of AdadeltaOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of AdadeltaOp should not be null."); @@ -56,6 +67,7 @@ class AdadeltaOp : public framework::OperatorWithKernel { ctx->SetOutputDim("AvgSquaredGradOut", param_dim); ctx->SetOutputDim("AvgSquaredUpdateOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext &ctx) const override { auto input_data_type = diff --git a/paddle/fluid/operators/adadelta_op.h b/paddle/fluid/operators/adadelta_op.h index 822458daf663d99bbb38d99205f51163a0df4c4d..6c616aa03d9809e9b7725a700c7edd5ff5d6dc42 100644 --- a/paddle/fluid/operators/adadelta_op.h +++ b/paddle/fluid/operators/adadelta_op.h @@ -23,6 +23,17 @@ template class AdadeltaOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto avg_squared_grad_out_tensor = ctx.Output("AvgSquaredGradOut"); diff --git a/paddle/fluid/operators/adagrad_op.h b/paddle/fluid/operators/adagrad_op.h index df520fcc898ff5514927dbdd845ecaecdcf3c147..0a16ce00f71586ef55007c3753e024be29d0ed56 100644 --- a/paddle/fluid/operators/adagrad_op.h +++ b/paddle/fluid/operators/adagrad_op.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once + #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" @@ -21,25 +22,31 @@ namespace operators { template struct SparseAdagradFunctor { - void operator()(const DeviceContext& context, - const framework::SelectedRows& grad, - const framework::Tensor& learning_rate, T epsilon, - framework::Tensor* moment, framework::Tensor* param); + void operator()(const DeviceContext &context, + const framework::SelectedRows &grad, + const framework::Tensor &learning_rate, T epsilon, + framework::Tensor *moment, framework::Tensor *param); }; template class AdagradOpKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* param_out_tensor = ctx.Output("ParamOut"); - auto* moment_out_tensor = ctx.Output("MomentOut"); + void Compute(const framework::ExecutionContext &ctx) const override { + const auto *param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + + auto *param_out_tensor = ctx.Output("ParamOut"); + auto *moment_out_tensor = ctx.Output("MomentOut"); param_out_tensor->mutable_data(ctx.GetPlace()); moment_out_tensor->mutable_data(ctx.GetPlace()); T epsilon = static_cast(ctx.Attr("epsilon")); - auto* grad_var = ctx.InputVar("Grad"); + auto *grad_var = ctx.InputVar("Grad"); if (grad_var->IsType()) { auto param = framework::EigenVector::Flatten( *ctx.Input("Param")); @@ -47,16 +54,16 @@ class AdagradOpKernel : public framework::OpKernel { *ctx.Input("Grad")); auto moment = framework::EigenVector::Flatten( *ctx.Input("Moment")); - auto* learning_rate = ctx.Input("LearningRate"); + auto *learning_rate = ctx.Input("LearningRate"); auto param_out = framework::EigenVector::Flatten(*param_out_tensor); auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); - auto* place = ctx.template device_context().eigen_device(); + auto *place = ctx.template device_context().eigen_device(); moment_out.device(*place) = moment + grad * grad; Eigen::DSizes m_dsize(moment_out_tensor->numel()); if (platform::is_cpu_place(ctx.GetPlace())) { - auto* lr = learning_rate->data(); + auto *lr = learning_rate->data(); param_out.device(*place) = param - lr[0] * grad / (moment_out.sqrt() + epsilon); } else { @@ -66,10 +73,10 @@ class AdagradOpKernel : public framework::OpKernel { lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon); } } else if (grad_var->IsType()) { - auto* param_tensor = ctx.Input("Param"); + auto *param_tensor = ctx.Input("Param"); PADDLE_ENFORCE_EQ(param_tensor, param_out_tensor); - auto* moment_tensor = ctx.Input("Moment"); + auto *moment_tensor = ctx.Input("Moment"); PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor); SparseAdagradFunctor functor; diff --git a/paddle/fluid/operators/adam_op.h b/paddle/fluid/operators/adam_op.h index 4cb1f3a80e95bdda79e6451dc3cc87e899b11779..3455d1ee54e8e6e498d0b0e6932ec099af9c0b30 100644 --- a/paddle/fluid/operators/adam_op.h +++ b/paddle/fluid/operators/adam_op.h @@ -18,6 +18,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/operators/math/algorithm.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/for_range.h" @@ -199,23 +200,9 @@ struct SparseAdamFunctor { row_numel_(row_numel), row_count_(row_count) {} - inline HOSTDEVICE int64_t BinarySearchInRows(int64_t row) const { - int64_t beg = 0, end = row_count_ - 1; - while (beg <= end) { - auto mid = ((beg + end) >> 1); - if (rows_[mid] == row) - return mid; - else if (rows_[mid] < row) - beg = mid + 1; - else - end = mid - 1; - } - return -1; - } - inline HOSTDEVICE void operator()(size_t i) const { - int64_t row = i / row_numel_; - auto row_idx = BinarySearchInRows(row); + auto row_idx = + math::BinarySearch(rows_, row_count_, i / row_numel_); T g = row_idx >= 0 ? grad_[row_idx * row_numel_ + i % row_numel_] : 0; // The following code is the same as dense @@ -244,6 +231,12 @@ template class AdamOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + using paddle::framework::LoDTensor; using paddle::operators::detail::Ref; diff --git a/paddle/fluid/operators/adamax_op.cc b/paddle/fluid/operators/adamax_op.cc index 32062574bcf71ff96e451eaa6865b6bbfc3b1c80..d4aa4d338a2379adf985ba7f89b528bc402eda06 100644 --- a/paddle/fluid/operators/adamax_op.cc +++ b/paddle/fluid/operators/adamax_op.cc @@ -35,6 +35,16 @@ class AdamaxOp : public framework::OperatorWithKernel { "Input(LearningRate) of AdamaxOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Beta1Pow"), "Input(Beta1Pow) of AdamaxOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of AdamaxOp should not be null."); diff --git a/paddle/fluid/operators/adamax_op.h b/paddle/fluid/operators/adamax_op.h index de644676fd9c3fabdbf01d2fd9c69858c2627ed3..7137fbd9651b4523f6d1609a0595b30758aa40df 100644 --- a/paddle/fluid/operators/adamax_op.h +++ b/paddle/fluid/operators/adamax_op.h @@ -23,6 +23,17 @@ template class AdamaxOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto moment_out_tensor = ctx.Output("MomentOut"); auto inf_norm_out_tensor = ctx.Output("InfNormOut"); diff --git a/paddle/fluid/operators/affine_channel_op.cc b/paddle/fluid/operators/affine_channel_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..8944a749674c3ba6c83526e4d66f449075716f43 --- /dev/null +++ b/paddle/fluid/operators/affine_channel_op.cc @@ -0,0 +1,255 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +Indicesou may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/data_layout.h" +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class AffineChannelOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor) Feature map input can be a 4D tensor with order NCHW " + "or NHWC. It also can be a 2D tensor and C is the second " + "dimension."); + AddInput("Scale", + "(Tensor) 1D input of shape (C), the c-th element " + "is the scale factor of the affine transformation " + "for the c-th channel of the input."); + AddInput("Bias", + "(Tensor) 1D input of shape (C), the c-th element " + "is the bias of the affine transformation for the " + "c-th channel of the input."); + AddAttr( + "data_layout", + "(string, default NCHW) Only used in " + "An optional string from: \"NHWC\", \"NCHW\". " + "Defaults to \"NHWC\". Specify the data format of the output data, " + "the input will be transformed automatically. ") + .SetDefault("AnyLayout"); + AddOutput("Out", "(Tensor) A tensor of the same shape and order with X."); + AddComment(R"DOC( + +Applies a separate affine transformation to each channel of the input. Useful +for replacing spatial batch norm with its equivalent fixed transformation. +The input also can be 2D tensor and applies a affine transformation in second +dimension. + +$$Out = Scale*X + Bias$$ + +)DOC"); + } +}; + +class AffineChannelOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of AffineChannelOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Scale"), + "Input(Scale) of AffineChannelOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Bias"), + "Input(Bias) of AffineChannelOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of AffineChannelOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", "Out"); + } +}; + +class AffineChannelOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null."); + if (ctx->HasOutput(framework::GradVarName("X"))) { + PADDLE_ENFORCE(ctx->HasInput("Scale"), + "Input(Scale) should not be null."); + ctx->SetOutputDim(framework::GradVarName("X"), + ctx->GetInputDim(framework::GradVarName("Out"))); + } + if (ctx->HasOutput(framework::GradVarName("Scale"))) { + // Scale@GRAD and Bias@GRAD must exist at the same time. + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), + "Output(Scale@GRAD) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + ctx->SetOutputDim(framework::GradVarName("Scale"), + ctx->GetInputDim("Scale")); + ctx->SetOutputDim(framework::GradVarName("Bias"), + ctx->GetInputDim("Scale")); + } + } +}; + +template +using EigenArrayMap = + Eigen::Map>; +template +using ConstEigenArrayMap = + Eigen::Map>; +template +using EigenVectorArrayMap = Eigen::Map>; +template +using ConstEigenVectorArrayMap = + Eigen::Map>; + +template +class AffineChannelKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* scale = ctx.Input("Scale"); + auto* bias = ctx.Input("Bias"); + + auto* y = ctx.Output("Out"); + y->mutable_data(ctx.GetPlace()); + + const framework::DataLayout layout = + framework::StringToDataLayout(ctx.Attr("data_layout")); + + auto dims = x->dims(); + int N = dims[0]; + int C = layout == framework::DataLayout::kNCHW ? dims[1] + : dims[dims.size() - 1]; + int HxW = x->numel() / N / C; + + auto* scale_d = scale->data(); + auto* bias_d = bias->data(); + ConstEigenVectorArrayMap a_e(scale_d, C); + ConstEigenVectorArrayMap b_e(bias_d, C); + + auto* x_d = x->data(); + auto* y_d = y->data(); + if (layout == framework::DataLayout::kNCHW) { + int stride = C * HxW; + for (int i = 0; i < N; i++) { + ConstEigenArrayMap x_e(x_d, HxW, C); + EigenArrayMap y_e(y_d, HxW, C); + y_e = (x_e.rowwise() * a_e.transpose()).rowwise() + b_e.transpose(); + x_d += stride; + y_d += stride; + } + } else { + int num = N * HxW; + ConstEigenArrayMap x_e(x_d, C, num); + EigenArrayMap y_e(y_d, C, num); + y_e = (x_e.colwise() * a_e).colwise() + b_e; + } + } +}; + +template +class AffineChannelGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* scale = ctx.Input("Scale"); + auto* dy = ctx.Input(framework::GradVarName("Out")); + + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dscale = + ctx.Output(framework::GradVarName("Scale")); + auto* dbias = ctx.Output(framework::GradVarName("Bias")); + + const framework::DataLayout layout = + framework::StringToDataLayout(ctx.Attr("data_layout")); + + auto dims = x->dims(); + int N = dims[0]; + int C = layout == framework::DataLayout::kNCHW ? dims[1] + : dims[dims.size() - 1]; + int HxW = x->numel() / N / C; + + auto* x_d = x->data(); + auto* dy_d = dy->data(); + auto* scale_d = scale->data(); + ConstEigenVectorArrayMap scale_e(scale_d, C); + + T* dx_d = dx ? dx->mutable_data(ctx.GetPlace()) : nullptr; + T* dscale_d = dscale ? dscale->mutable_data(ctx.GetPlace()) : nullptr; + T* dbias_d = dbias ? dbias->mutable_data(ctx.GetPlace()) : nullptr; + EigenVectorArrayMap dscale_e(dscale_d, C); + EigenVectorArrayMap dbias_e(dbias_d, C); + + if (layout == framework::DataLayout::kNCHW) { + // compute dx + int stride = C * HxW; + if (dx) { + for (int i = 0; i < N; i++) { + ConstEigenArrayMap dy_e(dy_d, HxW, C); + EigenArrayMap dx_e(dx_d, HxW, C); + dx_e = dy_e.rowwise() * scale_e.transpose(); + dy_d += stride; + dx_d += stride; + } + } + // compute dscale and dbias + if (dscale && dbias) { + dy_d = dy->data(); + for (int i = 0; i < N; i++) { + ConstEigenArrayMap x_e(x_d, HxW, C); + ConstEigenArrayMap dy_e(dy_d, HxW, C); + if (i == 0) { + dscale_e = (x_e * dy_e).colwise().sum(); + } else { + dscale_e += (x_e * dy_e).colwise().sum(); + } + if (i == 0) { + dbias_e = dy_e.colwise().sum(); + } else { + dbias_e += dy_e.colwise().sum(); + } + x_d += stride; + dy_d += stride; + } + } + } else { + int num = N * HxW; + ConstEigenArrayMap dy_e(dy_d, C, num); + // compute dx + if (dx) { + EigenArrayMap dx_e(dx_d, C, num); + dx_e = dy_e.colwise() * scale_e; + } + // compute dscale and dbias + if (dscale && dbias) { + ConstEigenArrayMap x_e(x_d, C, num); + dscale_e = (x_e * dy_e).rowwise().sum(); + dbias_e = dy_e.rowwise().sum(); + } + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +using CPU = paddle::platform::CPUDeviceContext; + +REGISTER_OPERATOR(affine_channel, ops::AffineChannelOp, + ops::AffineChannelOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(affine_channel_grad, ops::AffineChannelOpGrad); + +REGISTER_OP_CPU_KERNEL(affine_channel, ops::AffineChannelKernel, + ops::AffineChannelKernel); +REGISTER_OP_CPU_KERNEL(affine_channel_grad, + ops::AffineChannelGradKernel, + ops::AffineChannelGradKernel); diff --git a/paddle/fluid/operators/affine_channel_op.cu b/paddle/fluid/operators/affine_channel_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..2bebdb345ab324eb0a2dafd54c74833dd21bdb6d --- /dev/null +++ b/paddle/fluid/operators/affine_channel_op.cu @@ -0,0 +1,187 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +Indicesou may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "cub/cub.cuh" +#include "paddle/fluid/framework/data_layout.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/cuda_primitives.h" + +namespace paddle { +namespace operators { + +template +__global__ void KeAffineChannelCUDA(const T* x, const T* scale, const T* bias, + const int C, const int HxW, const int num, + T* y) { + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (int i = gid; i < num; i += stride) { + const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C; + if (HasBias) { + y[i] = scale[c] * x[i] + bias[c]; + } else { + y[i] = scale[c] * x[i]; + } + } +} + +template +class AffineChannelCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* scale = ctx.Input("Scale"); + auto* bias = ctx.Input("Bias"); + + auto* y = ctx.Output("Out"); + y->mutable_data(ctx.GetPlace()); + + const framework::DataLayout layout = + framework::StringToDataLayout(ctx.Attr("data_layout")); + auto& dev_ctx = ctx.template device_context(); + + auto dims = x->dims(); + const int num = x->numel(); + int N = dims[0]; + int C = layout == framework::DataLayout::kNCHW ? dims[1] + : dims[dims.size() - 1]; + int HxW = num / N / C; + + const T* x_d = x->data(); + const T* scale_d = scale->data(); + const T* bias_d = bias->data(); + T* y_d = y->data(); + + int block = 1024; + int grid = (num + block - 1) / block; + if (layout == framework::DataLayout::kNCHW) { + KeAffineChannelCUDA<<>>( + x_d, scale_d, bias_d, C, HxW, num, y_d); + } else { + KeAffineChannelCUDA<<>>( + x_d, scale_d, bias_d, C, HxW, num, y_d); + } + } +}; + +template +__global__ void AffineChannelScaleBiasGradientCUDAKernel( + const T* dy, const T* x, const int N, const int C, const int HxW, T* dscale, + T* dbias) { + const int outer_size = C; + const int inner_size = N * HxW; + typedef cub::BlockReduce BlockReduce; + __shared__ typename BlockReduce::TempStorage ds_storage; + __shared__ typename BlockReduce::TempStorage db_storage; + + for (int i = blockIdx.x; i < outer_size; i += gridDim.x) { + T ds_sum = 0; + T db_sum = 0; + for (int j = threadIdx.x; j < inner_size; j += blockDim.x) { + const int index = layout == framework::DataLayout::kNCHW + ? (j / HxW * C + i) * HxW + j % HxW + : j * outer_size + i; + ds_sum += dy[index] * x[index]; + db_sum += dy[index]; + } + ds_sum = BlockReduce(ds_storage).Reduce(ds_sum, cub::Sum()); + db_sum = BlockReduce(db_storage).Reduce(db_sum, cub::Sum()); + if (threadIdx.x == 0) { + dscale[i] = ds_sum; + dbias[i] = db_sum; + } + __syncthreads(); + } +} + +template +class AffineChannelGradCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* scale = ctx.Input("Scale"); + auto* bias = ctx.Input("Bias"); + auto* dy = ctx.Input(framework::GradVarName("Out")); + + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dscale = + ctx.Output(framework::GradVarName("Scale")); + auto* dbias = ctx.Output(framework::GradVarName("Bias")); + + const framework::DataLayout layout = + framework::StringToDataLayout(ctx.Attr("data_layout")); + auto& dev_ctx = ctx.template device_context(); + + auto dims = x->dims(); + const int num = x->numel(); + int N = dims[0]; + int C = layout == framework::DataLayout::kNCHW ? dims[1] + : dims[dims.size() - 1]; + int HxW = num / N / C; + + const T* x_d = x->data(); + const T* dy_d = dy->data(); + const T* s_d = scale->data(); + + T* dx_d = dx ? dx->mutable_data(ctx.GetPlace()) : nullptr; + T* ds_d = dscale ? dscale->mutable_data(ctx.GetPlace()) : nullptr; + T* db_d = dbias ? dbias->mutable_data(ctx.GetPlace()) : nullptr; + + const int block = 1024; + int max_threads = dev_ctx.GetMaxPhysicalThreadCount(); + const int max_blocks = std::max(max_threads / block, 1); + int grid1 = (num + block - 1) / block; + int grid2 = std::min(C, max_blocks); + if (layout == framework::DataLayout::kNCHW) { + if (dx) { + KeAffineChannelCUDA<<>>( + dy_d, s_d, nullptr, C, HxW, num, dx_d); + } + if (dscale && dbias) { + AffineChannelScaleBiasGradientCUDAKernel< + T, block, framework::DataLayout::kNCHW><<>>( + dy_d, x_d, N, C, HxW, ds_d, db_d); + } + } else { + if (dx) { + KeAffineChannelCUDA<<>>( + dy_d, s_d, nullptr, C, HxW, num, dx_d); + } + if (dscale && dbias) { + AffineChannelScaleBiasGradientCUDAKernel< + T, block, framework::DataLayout::kNHWC><<>>( + dy_d, x_d, N, C, HxW, ds_d, db_d); + } + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +using CUDA = paddle::platform::CUDADeviceContext; + +REGISTER_OP_CUDA_KERNEL(affine_channel, + ops::AffineChannelCUDAKernel, + ops::AffineChannelCUDAKernel); +REGISTER_OP_CUDA_KERNEL(affine_channel_grad, + ops::AffineChannelGradCUDAKernel, + ops::AffineChannelGradCUDAKernel); diff --git a/paddle/fluid/operators/clip_by_norm_op.h b/paddle/fluid/operators/clip_by_norm_op.h index 5af0eb0b2ada66d5ae7d521d80e213f9e61f826f..855c4d70677395992e2bf685c910cbea2d37b20b 100644 --- a/paddle/fluid/operators/clip_by_norm_op.h +++ b/paddle/fluid/operators/clip_by_norm_op.h @@ -16,12 +16,15 @@ limitations under the License. */ #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/transform.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; +using SelectedRows = framework::SelectedRows; template using EigenVector = framework::EigenVector; @@ -31,9 +34,40 @@ class ClipByNormKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto max_norm = context.Attr("max_norm"); - auto* input = context.Input("X"); - auto* output = context.Output("Out"); - output->mutable_data(context.GetPlace()); + auto in_var = context.InputVar("X"); + + Tensor* output = nullptr; + const Tensor* input = nullptr; + if (in_var->IsType()) { + input = context.Input("X"); + + output = context.Output("Out"); + output->mutable_data(context.GetPlace()); + } else if (in_var->IsType()) { + auto* x = context.Input("X"); + + // merge ids in selected rows first + math::scatter::MergeAdd merge_func; + SelectedRows* merged_input = + const_cast(context.scope()) + .Var() + ->GetMutable(); + merge_func(context.template device_context(), *x, + merged_input); + input = &(merged_input->value()); + + SelectedRows* output_selected_rows = context.Output("Out"); + output_selected_rows->set_rows(merged_input->rows()); + output_selected_rows->set_height(merged_input->height()); + output = output_selected_rows->mutable_value(); + output->Resize(merged_input->value().dims()); + output->mutable_data(context.GetPlace()); + } else { + PADDLE_THROW("Unexpected branch, input variable type is %s", + in_var->Type().name()); + } + + PADDLE_ENFORCE_NOT_NULL(input); auto x = EigenVector::Flatten(*input); auto out = EigenVector::Flatten(*output); diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc index 68faa1b2b64b65bb0dbec85b9dd9edd7cc9e6301..71a5db06fce07571944a0003c2eab629e1710b41 100644 --- a/paddle/fluid/operators/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/conv_mkldnn_op.cc @@ -300,10 +300,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { std::vector paddings = ctx.Attr>("paddings"); std::vector dilations = ctx.Attr>("dilations"); bool fuse_relu = ctx.Attr("fuse_relu"); - bool fuse_eltwise = ctx.Attr("fuse_eltwise"); + bool fuse_residual_conn = ctx.Attr("fuse_residual_connection"); int groups = ctx.Attr("groups"); - // TODO: add support for dilation // NOLINT + // TODO(tpatejko): add support for dilation PADDLE_ENFORCE( dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1, "dilation in convolution is not implemented yet"); @@ -369,11 +369,11 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { bias_tz, platform::MKLDNNGetDataType(), memory::format::x); conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md, strides, paddings, mkldnn_engine, - fuse_relu, fuse_eltwise); + fuse_relu, fuse_residual_conn); } else { conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings, - mkldnn_engine, fuse_relu, fuse_eltwise); + mkldnn_engine, fuse_relu, fuse_residual_conn); } // Save conv_pd/src_memory/weights_memory for backward pass dev_ctx.SetBlob(key_conv_pd, conv_pd); @@ -386,9 +386,26 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto user_weights_memory_p = handler.AcquireWeightsMemory( user_weights_md, to_void_cast(filter_data)); - T* output_data = output->mutable_data( - ctx.GetPlace(), paddle::memory::Allocator::kDefault, - handler.GetDstMemorySize()); + T* output_data = nullptr; + + if (fuse_residual_conn) { + auto residual_param = ctx.Input("ResidualData"); + auto residual_param_data = residual_param->data(); + + PADDLE_ENFORCE( + residual_param_data != nullptr, + "Provide data if you want MKLDNN conv+elementwise_add fusion"); + PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(), + "Output and elementwise parameter need to have the " + "same dimension sizes"); + + output->ShareDataWith(*residual_param); + output_data = output->mutable_data(ctx.GetPlace()); + } else { + output_data = output->mutable_data(ctx.GetPlace(), + paddle::memory::Allocator::kDefault, + handler.GetDstMemorySize()); + } // create reorder primitive if the input format is not the preferred one auto src_memory_p = handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline); @@ -425,14 +442,15 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { private: mkldnn::primitive_attr CreatePostOps(bool fuse_relu, - bool fuse_eltwise) const { + bool fuse_residual_conn) const { mkldnn::primitive_attr conv_attr; mkldnn::post_ops post_operations; // Fusion with Elementwise layer relies on adding a sum post-operation with - // the scale parameter. It is assumed that when fuse_eltwise is true, the - // Output tensor contains the data coming from residual connection. The - // result of this post_op is: Output = scale * Output + Conv_Out. - if (fuse_eltwise) { + // the scale parameter. It is assumed that when fuse_residual_connection is + // true, the output tensor contains the data coming from residual + // connection. The result of this post_op is: + // Output = scale * Output + Conv_Out. + if (fuse_residual_conn) { post_operations.append_sum(1.0f); } // Fusion with ReLU layer is executed through the PostOps feature. Create a @@ -453,7 +471,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { const memory::desc& dst, const std::vector& strides, const std::vector& paddings, const mkldnn::engine& engine, const bool fuse_relu, - const bool fuse_eltwise) const { + const bool fuse_residual_conn) const { memory::dims stride_dims = {strides[0], strides[1]}; memory::dims padding_dims = {paddings[0], paddings[1]}; @@ -462,7 +480,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { dst, stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero); - mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise); + mkldnn::primitive_attr conv_attr = + CreatePostOps(fuse_relu, fuse_residual_conn); auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc( conv_desc, conv_attr, engine); @@ -477,7 +496,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { const std::vector& strides, const std::vector& paddings, const mkldnn::engine& engine, const bool fuse_relu, - const bool fuse_eltwise) const { + const bool fuse_residual_conn) const { memory::dims stride_dims = {strides[0], strides[1]}; memory::dims padding_dims = {paddings[0], paddings[1]}; @@ -486,7 +505,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { bias, dst, stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero); - mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise); + mkldnn::primitive_attr conv_attr = + CreatePostOps(fuse_relu, fuse_residual_conn); auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc( conv_desc, conv_attr, engine); diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 8f84bf71a7f77606bed6672f0830e3fc80165a42..8f2561fcc389922f05093055cba4b43dbd4e4536 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -132,6 +132,11 @@ void Conv2DOpMaker::Make() { "(Tensor) The output tensor of convolution operator. " "The format of output tensor is also NCHW.") .Reuse("Input"); + AddInput("ResidualData", + "(Tensor) Tensor with residual data " + "to which convolution output will be added." + "Used with fuse_residual_connection fusion.") + .AsDispensable(); AddAttr>("strides", "(vector default:{1, 1}), the " "strides(h_stride, w_stride) of " @@ -164,10 +169,10 @@ void Conv2DOpMaker::Make() { .SetDefault(false); AddAttr("fuse_relu", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); - AddAttr("fuse_eltwise", + AddAttr("fuse_residual_connection", "(bool, default false) Only used in mkldnn kernel. Used " - "whenever convolution output is connected via skip connection " - "to a previous layer.") + "whenever convolution output is as an input to residual " + "connection.") .SetDefault(false); AddAttr( "data_format", diff --git a/paddle/fluid/operators/decayed_adagrad_op.cc b/paddle/fluid/operators/decayed_adagrad_op.cc index c0f2b49a04d9e88502c4b63bca493cd2b7ad1c5c..d73ae9e2721b388212cb6efa354eb4b480df9cad 100644 --- a/paddle/fluid/operators/decayed_adagrad_op.cc +++ b/paddle/fluid/operators/decayed_adagrad_op.cc @@ -32,6 +32,16 @@ class DecayedAdagradOp : public framework::OperatorWithKernel { PADDLE_ENFORCE( ctx->HasInput("LearningRate"), "Input(LearningRate) of DecayedAdagradOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of DecayedAdagradOp should not be null."); diff --git a/paddle/fluid/operators/decayed_adagrad_op.h b/paddle/fluid/operators/decayed_adagrad_op.h index a46af078e0c6b4bf1faca0570b6a97b026864f13..5df43d33ef9f720fd20d57c53ff37cc85440b24e 100644 --- a/paddle/fluid/operators/decayed_adagrad_op.h +++ b/paddle/fluid/operators/decayed_adagrad_op.h @@ -23,6 +23,17 @@ template class DecayedAdagradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto moment_out_tensor = ctx.Output("MomentOut"); diff --git a/paddle/fluid/operators/detection/CMakeLists.txt b/paddle/fluid/operators/detection/CMakeLists.txt index aa8ed502fc94bd0970dfe5dbf00ef090e799ad30..d5eec148f9b4f76866ec9fca98a596b9bc2860ef 100644 --- a/paddle/fluid/operators/detection/CMakeLists.txt +++ b/paddle/fluid/operators/detection/CMakeLists.txt @@ -20,7 +20,7 @@ detection_library(box_coder_op SRCS box_coder_op.cc box_coder_op.cu) detection_library(iou_similarity_op SRCS iou_similarity_op.cc iou_similarity_op.cu) detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc) -detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc) +detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc poly_util.cc gpc.cc) detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu) detection_library(anchor_generator_op SRCS anchor_generator_op.cc anchor_generator_op.cu) diff --git a/paddle/fluid/operators/detection/generate_proposals_op.cc b/paddle/fluid/operators/detection/generate_proposals_op.cc index e9f966b5778af617c4438fb453366c9cd7c163ef..a69d9c9a529f26b3981ca8d1ba226fda71b8820a 100644 --- a/paddle/fluid/operators/detection/generate_proposals_op.cc +++ b/paddle/fluid/operators/detection/generate_proposals_op.cc @@ -453,33 +453,45 @@ class GenerateProposalsKernel : public framework::OpKernel { class GenerateProposalsOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Scores", "The scores of anchors should be foreground."); - AddInput("BboxDeltas", "bbox_deltas."); - AddInput("ImInfo", "Information for image reshape."); - AddInput("Anchors", "All anchors."); - AddInput("Variances", " variances"); - - AddOutput("RpnRois", "Anchors."); - AddOutput("RpnRoiProbs", "Anchors."); - AddAttr("pre_nms_topN", "pre_nms_topN"); - AddAttr("post_nms_topN", "post_nms_topN"); - AddAttr("nms_thresh", "nms_thres"); - AddAttr("min_size", "min size"); + AddInput("Scores", + "(Tensor) The scores from conv is in shape (N, A, H, W), " + "N is batch size, A is number of anchors, " + "H and W are height and width of the feature map"); + AddInput("BboxDeltas", + "(Tensor) Bounding box deltas from conv is in " + "shape (N, 4*A, H, W)."); + AddInput("ImInfo", + "(Tensor) Information for image reshape is in shape (N, 3), " + "in format (height, width, scale)"); + AddInput("Anchors", + "(Tensor) Bounding box anchors from anchor_generator_op " + "is in shape (A, H, W, 4)."); + AddInput("Variances", + "(Tensor) Bounding box variances with same shape as `Anchors`."); + + AddOutput("RpnRois", + "(LoDTensor), Output proposals with shape (rois_num, 4)."); + AddOutput("RpnRoiProbs", + "(LoDTensor) Scores of proposals with shape (rois_num, 1)."); + AddAttr("pre_nms_topN", + "Number of top scoring RPN proposals to keep before " + "applying NMS."); + AddAttr("post_nms_topN", + "Number of top scoring RPN proposals to keep after " + "applying NMS"); + AddAttr("nms_thresh", "NMS threshold used on RPN proposals."); + AddAttr("min_size", + "Proposal height and width both need to be greater " + "than this min_size."); AddAttr("eta", "The parameter for adaptive NMS."); AddComment(R"DOC( -Generate Proposals OP +This operator Generate bounding box proposals for Faster RCNN. +The propoasls are generated for a list of images based on image +score 'Scores', bounding box regression result 'BboxDeltas' as +well as predefined bounding box shapes 'anchors'. Greedy +non-maximum suppression is applied to generate the final bounding +boxes. -This operator proposes rois according to each box with their probability to be a foreground object and -the box can be calculated by anchors. Bbox_details and scores are the output of RPN. Final proposals -could be used to train detection net. - -Scores is the probability for each box to be an object. In format of (N, A, H, W) where N is batch size, A is number -of anchors, H and W are height and width of the feature map. -BboxDeltas is the differece between predicted box location and anchor location. In format of (N, 4*A, H, W) - -For generating proposals, this operator transposes and resizes scores and bbox_deltas in size of (H*W*A, 1) and (H*W*A, 4) and - calculate box locations as proposals candidates. Then clip boxes to image and remove predicted boxes with small area. -Finally, apply nms to get final proposals as output. )DOC"); } }; diff --git a/paddle/fluid/operators/detection/generate_proposals_op.cu b/paddle/fluid/operators/detection/generate_proposals_op.cu index 0d3817c3e7c43bf7772f16ebc45841ec5ef9a62a..a0b99377109aef4776fadd68101d011a9191b1cc 100644 --- a/paddle/fluid/operators/detection/generate_proposals_op.cu +++ b/paddle/fluid/operators/detection/generate_proposals_op.cu @@ -417,7 +417,7 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel { T *rpn_rois_data = rpn_rois->data(); T *rpn_roi_probs_data = rpn_roi_probs->data(); - auto &place = boost::get(dev_ctx.GetPlace()); + auto place = boost::get(dev_ctx.GetPlace()); int64_t num_proposals = 0; std::vector offset(1, 0); diff --git a/paddle/fluid/operators/detection/gpc.cc b/paddle/fluid/operators/detection/gpc.cc new file mode 100644 index 0000000000000000000000000000000000000000..7c0823c0487d39eece5be08322e7d182b931ba3c --- /dev/null +++ b/paddle/fluid/operators/detection/gpc.cc @@ -0,0 +1,2201 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/** + * @file src/gpc.cpp + * @author huhan02(com@baidu.com) + * @date 2015/12/18 14:17:30 + * @brief + * + * @modified by sunyipeng + * @email sunyipeng@baidu.com + * @date 2018/6/12 + **/ + +#include "paddle/fluid/operators/detection/gpc.h" + +namespace gpc { + +typedef struct lmt_shape { /* Local minima table */ + double y; /* Y coordinate at local minimum */ + edge_node *first_bound; /* Pointer to bound list */ + struct lmt_shape *next; /* Pointer to next local minimum */ +} lmt_node; + +typedef struct sbt_t_shape { /* Scanbeam tree */ + double y; /* Scanbeam node y value */ + struct sbt_t_shape *less; /* Pointer to nodes with lower y */ + struct sbt_t_shape *more; /* Pointer to nodes with higher y */ +} sb_tree; + +typedef struct it_shape { /* Intersection table */ + edge_node *ie[2]; /* Intersecting edge (bundle) pair */ + gpc_vertex point; /* Point of intersection */ + struct it_shape *next; /* The next intersection table node */ +} it_node; + +typedef struct st_shape { /* Sorted edge table */ + edge_node *edge; /* Pointer to AET edge */ + double xb; /* Scanbeam bottom x coordinate */ + double xt; /* Scanbeam top x coordinate */ + double dx; /* Change in x for a unit y increase */ + struct st_shape *prev; /* Previous edge in sorted list */ +} st_node; + +typedef struct bbox_shape { /* Contour axis-aligned bounding box */ + double xmin; /* Minimum x coordinate */ + double ymin; /* Minimum y coordinate */ + double xmax; /* Maximum x coordinate */ + double ymax; /* Maximum y coordinate */ +} bbox; + +/* +=========================================================================== + Global Data +=========================================================================== +*/ + +/* Horizontal edge state transitions within scanbeam boundary */ +const h_state next_h_state[3][6] = { + /* ABOVE BELOW CROSS */ + /* L R L R L R */ + /* NH */ + {BH, TH, TH, BH, NH, NH}, + /* BH */ + {NH, NH, NH, NH, TH, TH}, + /* TH */ + {NH, NH, NH, NH, BH, BH}}; + +/* +=========================================================================== + Private Functions +=========================================================================== +*/ + +static void reset_it(it_node **it) { + it_node *itn; + + while (*it) { + itn = (*it)->next; + gpc_free(*it); + *it = itn; + } +} + +static void reset_lmt(lmt_node **lmt) { + lmt_node *lmtn; + + while (*lmt) { + lmtn = (*lmt)->next; + gpc_free(*lmt); + *lmt = lmtn; + } +} + +static void insert_bound(edge_node **b, edge_node *e) { + edge_node *existing_bound = NULL; + + if (!*b) { + /* Link node e to the tail of the list */ + *b = e; + } else { + /* Do primary sort on the x field */ + if (e[0].bot.x < (*b)[0].bot.x) { + /* Insert a new node mid-list */ + existing_bound = *b; + *b = e; + (*b)->next_bound = existing_bound; + } else { + if (e[0].bot.x == (*b)[0].bot.x) { + /* Do secondary sort on the dx field */ + if (e[0].dx < (*b)[0].dx) { + /* Insert a new node mid-list */ + existing_bound = *b; + *b = e; + (*b)->next_bound = existing_bound; + } else { + /* Head further down the list */ + insert_bound(&((*b)->next_bound), e); + } + } else { + /* Head further down the list */ + insert_bound(&((*b)->next_bound), e); + } + } + } +} + +static edge_node **bound_list(lmt_node **lmt, double y) { + lmt_node *existing_node; + + if (!*lmt) { + /* Add node onto the tail end of the LMT */ + gpc_malloc(*lmt, sizeof(lmt_node), + const_cast("LMT insertion")); + (*lmt)->y = y; + (*lmt)->first_bound = NULL; + (*lmt)->next = NULL; + return &((*lmt)->first_bound); + } else if (y < (*lmt)->y) { + /* Insert a new LMT node before the current node */ + existing_node = *lmt; + gpc_malloc(*lmt, sizeof(lmt_node), + const_cast("LMT insertion")); + (*lmt)->y = y; + (*lmt)->first_bound = NULL; + (*lmt)->next = existing_node; + return &((*lmt)->first_bound); + } else { + if (y > (*lmt)->y) { + /* Head further up the LMT */ + return bound_list(&((*lmt)->next), y); + } else { + /* Use this existing LMT node */ + return &((*lmt)->first_bound); + } + } +} + +static void add_to_sbtree(int *entries, sb_tree **sbtree, double y) { + if (!*sbtree) { + /* Add a new tree node here */ + gpc_malloc(*sbtree, sizeof(sb_tree), + const_cast("scanbeam tree insertion")); + (*sbtree)->y = y; + (*sbtree)->less = NULL; + (*sbtree)->more = NULL; + (*entries)++; + } else { + if ((*sbtree)->y > y) { + /* Head into the 'less' sub-tree */ + add_to_sbtree(entries, &((*sbtree)->less), y); + } else { + if ((*sbtree)->y < y) { + /* Head into the 'more' sub-tree */ + add_to_sbtree(entries, &((*sbtree)->more), y); + } + } + } +} + +static void build_sbt(int *entries, double *sbt, sb_tree *sbtree) { + if (sbtree->less) { + build_sbt(entries, sbt, sbtree->less); + } + sbt[*entries] = sbtree->y; + (*entries)++; + if (sbtree->more) { + build_sbt(entries, sbt, sbtree->more); + } +} + +static void free_sbtree(sb_tree **sbtree) { + if (*sbtree) { + free_sbtree(&((*sbtree)->less)); + free_sbtree(&((*sbtree)->more)); + gpc_free(*sbtree); + } +} + +static int count_optimal_vertices(gpc_vertex_list c) { + int result = 0; + int i = 0; + + /* Ignore non-contributing contours */ + if (c.num_vertices > 0) { + for (i = 0; i < c.num_vertices; i++) { + /* Ignore superfluous vertices embedded in horizontal edges */ + if (gpc_optimal(c.vertex, i, c.num_vertices)) { + result++; + } + } + } + return result; +} + +static edge_node *build_lmt(lmt_node **lmt, sb_tree **sbtree, int *sbt_entries, + gpc_polygon *p, int type, gpc_op op) { + int c = 0; + int i = 0; + int min = 0; + int max = 0; + int num_edges = 0; + int v = 0; + int num_vertices = 0; + int total_vertices = 0; + int e_index = 0; + edge_node *e = NULL; + edge_node *edge_table = NULL; + + for (c = 0; c < p->num_contours; c++) { + total_vertices += count_optimal_vertices(p->contour[c]); + } + + /* Create the entire input polygon edge table in one go */ + gpc_malloc(edge_table, total_vertices * sizeof(edge_node), + const_cast("edge table creation")); + + for (c = 0; c < p->num_contours; c++) { + if (p->contour[c].num_vertices < 0) { + /* Ignore the non-contributing contour and repair the vertex count */ + p->contour[c].num_vertices = -p->contour[c].num_vertices; + } else { + /* Perform contour optimisation */ + num_vertices = 0; + for (i = 0; i < p->contour[c].num_vertices; i++) { + if (gpc_optimal(p->contour[c].vertex, i, p->contour[c].num_vertices)) { + edge_table[num_vertices].vertex.x = p->contour[c].vertex[i].x; + edge_table[num_vertices].vertex.y = p->contour[c].vertex[i].y; + + /* Record vertex in the scanbeam table */ + add_to_sbtree(sbt_entries, sbtree, edge_table[num_vertices].vertex.y); + + num_vertices++; + } + } + + /* Do the contour forward pass */ + for (min = 0; min < num_vertices; min++) { + /* If a forward local minimum... */ + if (gpc_fwd_min(edge_table, min, num_vertices)) { + /* Search for the next local maximum... */ + num_edges = 1; + max = gpc_next_index(min, num_vertices); + while (gpc_not_fmax(edge_table, max, num_vertices)) { + num_edges++; + max = gpc_next_index(max, num_vertices); + } + + /* Build the next edge list */ + e = &edge_table[e_index]; + e_index += num_edges; + v = min; + e[0].bstate[BELOW] = UNBUNDLED; + e[0].bundle[BELOW][CLIP] = 0; + e[0].bundle[BELOW][SUBJ] = 0; + for (i = 0; i < num_edges; i++) { + e[i].xb = edge_table[v].vertex.x; + e[i].bot.x = edge_table[v].vertex.x; + e[i].bot.y = edge_table[v].vertex.y; + + v = gpc_next_index(v, num_vertices); + + e[i].top.x = edge_table[v].vertex.x; + e[i].top.y = edge_table[v].vertex.y; + e[i].dx = (edge_table[v].vertex.x - e[i].bot.x) / + (e[i].top.y - e[i].bot.y); + e[i].type = type; + e[i].outp[ABOVE] = NULL; + e[i].outp[BELOW] = NULL; + e[i].next = NULL; + e[i].prev = NULL; + e[i].succ = + ((num_edges > 1) && (i < (num_edges - 1))) ? &(e[i + 1]) : NULL; + e[i].pred = ((num_edges > 1) && (i > 0)) ? &(e[i - 1]) : NULL; + e[i].next_bound = NULL; + e[i].bside[CLIP] = (op == GPC_DIFF) ? RIGHT : LEFT; + e[i].bside[SUBJ] = LEFT; + } + insert_bound(bound_list(lmt, edge_table[min].vertex.y), e); + } + } + + /* Do the contour reverse pass */ + for (min = 0; min < num_vertices; min++) { + /* If a reverse local minimum... */ + if (gpc_rev_min(edge_table, min, num_vertices)) { + /* Search for the previous local maximum... */ + num_edges = 1; + max = gpc_prev_index(min, num_vertices); + while (gpc_not_rmax(edge_table, max, num_vertices)) { + num_edges++; + max = gpc_prev_index(max, num_vertices); + } + + /* Build the previous edge list */ + e = &edge_table[e_index]; + e_index += num_edges; + v = min; + e[0].bstate[BELOW] = UNBUNDLED; + e[0].bundle[BELOW][CLIP] = 0; + e[0].bundle[BELOW][SUBJ] = 0; + for (i = 0; i < num_edges; i++) { + e[i].xb = edge_table[v].vertex.x; + e[i].bot.x = edge_table[v].vertex.x; + e[i].bot.y = edge_table[v].vertex.y; + + v = gpc_prev_index(v, num_vertices); + + e[i].top.x = edge_table[v].vertex.x; + e[i].top.y = edge_table[v].vertex.y; + e[i].dx = (edge_table[v].vertex.x - e[i].bot.x) / + (e[i].top.y - e[i].bot.y); + e[i].type = type; + e[i].outp[ABOVE] = NULL; + e[i].outp[BELOW] = NULL; + e[i].next = NULL; + e[i].prev = NULL; + e[i].succ = + ((num_edges > 1) && (i < (num_edges - 1))) ? &(e[i + 1]) : NULL; + e[i].pred = ((num_edges > 1) && (i > 0)) ? &(e[i - 1]) : NULL; + e[i].next_bound = NULL; + e[i].bside[CLIP] = (op == GPC_DIFF) ? RIGHT : LEFT; + e[i].bside[SUBJ] = LEFT; + } + insert_bound(bound_list(lmt, edge_table[min].vertex.y), e); + } + } + } + } + return edge_table; +} // NOLINT + +static void add_edge_to_aet(edge_node **aet, edge_node *edge, edge_node *prev) { + if (!*aet) { + /* Append edge onto the tail end of the AET */ + *aet = edge; + edge->prev = prev; + edge->next = NULL; + } else { + /* Do primary sort on the xb field */ + if (edge->xb < (*aet)->xb) { + /* Insert edge here (before the AET edge) */ + edge->prev = prev; + edge->next = *aet; + (*aet)->prev = edge; + *aet = edge; + } else { + if (edge->xb == (*aet)->xb) { + /* Do secondary sort on the dx field */ + if (edge->dx < (*aet)->dx) { + /* Insert edge here (before the AET edge) */ + edge->prev = prev; + edge->next = *aet; + (*aet)->prev = edge; + *aet = edge; + } else { + /* Head further into the AET */ + add_edge_to_aet(&((*aet)->next), edge, *aet); + } + } else { + /* Head further into the AET */ + add_edge_to_aet(&((*aet)->next), edge, *aet); + } + } + } +} + +static void add_intersection(it_node **it, edge_node *edge0, edge_node *edge1, + double x, double y) { + it_node *existing_node; + + if (!*it) { + /* Append a new node to the tail of the list */ + gpc_malloc(*it, sizeof(it_node), + const_cast("IT insertion")); + (*it)->ie[0] = edge0; + (*it)->ie[1] = edge1; + (*it)->point.x = x; + (*it)->point.y = y; + (*it)->next = NULL; + } else { + if ((*it)->point.y > y) { + /* Insert a new node mid-list */ + existing_node = *it; + gpc_malloc(*it, sizeof(it_node), + const_cast("IT insertion")); + (*it)->ie[0] = edge0; + (*it)->ie[1] = edge1; + (*it)->point.x = x; + (*it)->point.y = y; + (*it)->next = existing_node; + } else { + /* Head further down the list */ + add_intersection(&((*it)->next), edge0, edge1, x, y); + } + } +} + +static void add_st_edge(st_node **st, it_node **it, edge_node *edge, + double dy) { + st_node *existing_node; + double den = 0.0; + double r = 0.0; + double x = 0.0; + double y = 0.0; + + if (!*st) { + /* Append edge onto the tail end of the ST */ + gpc_malloc(*st, sizeof(st_node), + const_cast("ST insertion")); + (*st)->edge = edge; + (*st)->xb = edge->xb; + (*st)->xt = edge->xt; + (*st)->dx = edge->dx; + (*st)->prev = NULL; + } else { + den = ((*st)->xt - (*st)->xb) - (edge->xt - edge->xb); + + /* If new edge and ST edge don't cross */ + if ((edge->xt >= (*st)->xt) || (edge->dx == (*st)->dx) || + (fabs(den) <= DBL_EPSILON)) { + /* No intersection - insert edge here (before the ST edge) */ + existing_node = *st; + gpc_malloc(*st, sizeof(st_node), + const_cast("ST insertion")); + (*st)->edge = edge; + (*st)->xb = edge->xb; + (*st)->xt = edge->xt; + (*st)->dx = edge->dx; + (*st)->prev = existing_node; + } else { + /* Compute intersection between new edge and ST edge */ + r = (edge->xb - (*st)->xb) / den; + x = (*st)->xb + r * ((*st)->xt - (*st)->xb); + y = r * dy; + + /* Insert the edge pointers and the intersection point in the IT */ + add_intersection(it, (*st)->edge, edge, x, y); + + /* Head further into the ST */ + add_st_edge(&((*st)->prev), it, edge, dy); + } + } +} + +static void build_intersection_table(it_node **it, edge_node *aet, double dy) { + st_node *st; + st_node *stp; + edge_node *edge = NULL; + + /* Build intersection table for the current scanbeam */ + reset_it(it); + st = NULL; + + /* Process each AET edge */ + for (edge = aet; edge; edge = edge->next) { + if ((edge->bstate[ABOVE] == BUNDLE_HEAD) || edge->bundle[ABOVE][CLIP] || + edge->bundle[ABOVE][SUBJ]) { + add_st_edge(&st, it, edge, dy); + } + } + + /* Free the sorted edge table */ + while (st) { + stp = st->prev; + gpc_free(st); + st = stp; + } +} + +static int count_contours(polygon_node *polygon) { + int nc = 0; + int nv = 0; + vertex_node *v = NULL; + vertex_node *nextv = NULL; + + for (nc = 0; polygon; polygon = polygon->next) { + if (polygon->active) { + /* Count the vertices in the current contour */ + nv = 0; + for (v = polygon->proxy->v[LEFT]; v; v = v->next) { + nv++; + } + + /* Record valid vertex counts in the active field */ + if (nv > 2) { + polygon->active = nv; + nc++; + } else { + /* Invalid contour: just free the heap */ + for (v = polygon->proxy->v[LEFT]; v; v = nextv) { + nextv = v->next; + gpc_free(v); + } + polygon->active = 0; + } + } + } + return nc; +} + +static void add_left(polygon_node *p, double x, double y) { + vertex_node *nv = NULL; + + /* Create a new vertex node and set its fields */ + gpc_malloc(nv, sizeof(vertex_node), + const_cast("vertex node creation")); + nv->x = x; + nv->y = y; + + /* Add vertex nv to the left end of the polygon's vertex list */ + nv->next = p->proxy->v[LEFT]; + + /* Update proxy->[LEFT] to point to nv */ + p->proxy->v[LEFT] = nv; +} + +static void merge_left(polygon_node *p, polygon_node *q, polygon_node *list) { + polygon_node *target = NULL; + + /* Label contour as a hole */ + q->proxy->hole = 1; + + if (p->proxy != q->proxy) { + /* Assign p's vertex list to the left end of q's list */ + p->proxy->v[RIGHT]->next = q->proxy->v[LEFT]; + q->proxy->v[LEFT] = p->proxy->v[LEFT]; + + /* Redirect any p->proxy references to q->proxy */ + + for (target = p->proxy; list; list = list->next) { + if (list->proxy == target) { + list->active = 0; + list->proxy = q->proxy; + } + } + } +} + +static void add_right(polygon_node *p, double x, double y) { + vertex_node *nv = NULL; + + /* Create a new vertex node and set its fields */ + gpc_malloc(nv, sizeof(vertex_node), + const_cast("vertex node creation")); + nv->x = x; + nv->y = y; + nv->next = NULL; + + /* Add vertex nv to the right end of the polygon's vertex list */ + p->proxy->v[RIGHT]->next = nv; + + /* Update proxy->v[RIGHT] to point to nv */ + p->proxy->v[RIGHT] = nv; +} + +static void merge_right(polygon_node *p, polygon_node *q, polygon_node *list) { + polygon_node *target = NULL; + + /* Label contour as external */ + q->proxy->hole = 0; + + if (p->proxy != q->proxy) { + /* Assign p's vertex list to the right end of q's list */ + q->proxy->v[RIGHT]->next = p->proxy->v[LEFT]; + q->proxy->v[RIGHT] = p->proxy->v[RIGHT]; + + /* Redirect any p->proxy references to q->proxy */ + for (target = p->proxy; list; list = list->next) { + if (list->proxy == target) { + list->active = 0; + list->proxy = q->proxy; + } + } + } +} + +static void add_local_min(polygon_node **p, edge_node *edge, double x, + double y) { + polygon_node *existing_min = NULL; + vertex_node *nv = NULL; + + existing_min = *p; + + gpc_malloc(*p, sizeof(polygon_node), + const_cast("polygon node creation")); + + /* Create a new vertex node and set its fields */ + gpc_malloc(nv, sizeof(vertex_node), + const_cast("vertex node creation")); + nv->x = x; + nv->y = y; + nv->next = NULL; + + /* Initialise proxy to point to p itself */ + (*p)->proxy = (*p); + (*p)->active = 1; + (*p)->next = existing_min; + + /* Make v[LEFT] and v[RIGHT] point to new vertex nv */ + (*p)->v[LEFT] = nv; + (*p)->v[RIGHT] = nv; + + /* Assign polygon p to the edge */ + edge->outp[ABOVE] = *p; +} + +static int count_tristrips(polygon_node *tn) { + int total = 0; + + for (total = 0; tn; tn = tn->next) { + if (tn->active > 2) { + total++; + } + } + return total; +} + +void add_vertex(vertex_node **t, double x, double y) { + if (!(*t)) { + gpc_malloc(*t, sizeof(vertex_node), + const_cast("tristrip vertex creation")); + (*t)->x = x; + (*t)->y = y; + (*t)->next = NULL; + } else { + /* Head further down the list */ + add_vertex(&((*t)->next), x, y); + } +} + +void gpc_vertex_create(edge_node *e, int p, int s, double x, double y) { + add_vertex(&(e->outp[p]->v[s]), x, y); + e->outp[p]->active++; +} + +static void new_tristrip(polygon_node **tn, edge_node *edge, double x, + double y) { + if (!(*tn)) { + gpc_malloc(*tn, sizeof(polygon_node), + const_cast("tristrip node creation")); + (*tn)->next = NULL; + (*tn)->v[LEFT] = NULL; + (*tn)->v[RIGHT] = NULL; + (*tn)->active = 1; + add_vertex(&((*tn)->v[LEFT]), x, y); + edge->outp[ABOVE] = *tn; + } else { + /* Head further down the list */ + new_tristrip(&((*tn)->next), edge, x, y); + } +} + +static bbox *create_contour_bboxes(gpc_polygon *p) { + bbox *box; + int c = 0; + int v = 0; + + gpc_malloc(box, p->num_contours * sizeof(bbox), + const_cast("Bounding box creation")); + + /* Construct contour bounding boxes */ + for (c = 0; c < p->num_contours; c++) { + /* Initialise bounding box extent */ + box[c].xmin = DBL_MAX; + box[c].ymin = DBL_MAX; + box[c].xmax = -DBL_MAX; + box[c].ymax = -DBL_MAX; + + for (v = 0; v < p->contour[c].num_vertices; v++) { + /* Adjust bounding box */ + if (p->contour[c].vertex[v].x < box[c].xmin) { + box[c].xmin = p->contour[c].vertex[v].x; + } + if (p->contour[c].vertex[v].y < box[c].ymin) { + box[c].ymin = p->contour[c].vertex[v].y; + } + if (p->contour[c].vertex[v].x > box[c].xmax) { + box[c].xmax = p->contour[c].vertex[v].x; + } + if (p->contour[c].vertex[v].y > box[c].ymax) { + box[c].ymax = p->contour[c].vertex[v].y; + } + } + } + return box; +} + +static void minimax_test(gpc_polygon *subj, gpc_polygon *clip, gpc_op op) { + bbox *s_bbox; + bbox *c_bbox; + int s = 0; + int c = 0; + int *o_table = NULL; + int overlap = 0; + + s_bbox = create_contour_bboxes(subj); + c_bbox = create_contour_bboxes(clip); + + gpc_malloc(o_table, + subj->num_contours * clip->num_contours * sizeof(int), + const_cast("overlap table creation")); + + /* Check all subject contour bounding boxes against clip boxes */ + for (s = 0; s < subj->num_contours; s++) { + for (c = 0; c < clip->num_contours; c++) { + o_table[c * subj->num_contours + s] = + (!((s_bbox[s].xmax < c_bbox[c].xmin) || + (s_bbox[s].xmin > c_bbox[c].xmax))) && + (!((s_bbox[s].ymax < c_bbox[c].ymin) || + (s_bbox[s].ymin > c_bbox[c].ymax))); + } + } + + /* For each clip contour, search for any subject contour overlaps */ + for (c = 0; c < clip->num_contours; c++) { + overlap = 0; + for (s = 0; (!overlap) && (s < subj->num_contours); s++) { + overlap = o_table[c * subj->num_contours + s]; + } + + if (!overlap) { + /* Flag non contributing status by negating vertex count */ + clip->contour[c].num_vertices = -clip->contour[c].num_vertices; + } + } + + if (op == GPC_INT) { + /* For each subject contour, search for any clip contour overlaps */ + for (s = 0; s < subj->num_contours; s++) { + overlap = 0; + for (c = 0; (!overlap) && (c < clip->num_contours); c++) { + overlap = o_table[c * subj->num_contours + s]; + } + + if (!overlap) { + /* Flag non contributing status by negating vertex count */ + subj->contour[s].num_vertices = -subj->contour[s].num_vertices; + } + } + } + + gpc_free(s_bbox); + gpc_free(c_bbox); + gpc_free(o_table); +} + +/* +=========================================================================== + Public Functions +=========================================================================== +*/ + +void gpc_free_polygon(gpc_polygon *p) { + int c = 0; + + for (c = 0; c < p->num_contours; c++) { + gpc_free(p->contour[c].vertex); + } + gpc_free(p->hole); + gpc_free(p->contour); + p->num_contours = 0; +} + +/* +void gpc_read_polygon(FILE *fp, int read_hole_flags, gpc_polygon *p) { + int c = 0; + int v = 0; + + fscanf(fp, "%d", &(p->num_contours)); + gpc_malloc(p->hole, p->num_contours * sizeof(int), + (char *)"hole flag array creation"); + gpc_malloc(p->contour, + p->num_contours * sizeof(gpc_vertex_list), + (char *)"contour creation"); + for (c = 0; c < p->num_contours; c++) { + fscanf(fp, "%d", &(p->contour[c].num_vertices)); + + if (read_hole_flags) { + fscanf(fp, "%d", &(p->hole[c])); + } else { + p->hole[c] = 0; // Assume all contours to be external + } + + gpc_malloc(p->contour[c].vertex, + p->contour[c].num_vertices * sizeof(gpc_vertex), + (char *)"vertex creation"); + for (v = 0; v < p->contour[c].num_vertices; v++) { + fscanf(fp, "%lf %lf", &(p->contour[c].vertex[v].x), + &(p->contour[c].vertex[v].y)); + } + } +} + +void gpc_write_polygon(FILE *fp, int write_hole_flags, gpc_polygon *p) { + int c = 0; + int v = 0; + + fprintf(fp, "%d\n", p->num_contours); + for (c = 0; c < p->num_contours; c++) { + fprintf(fp, "%d\n", p->contour[c].num_vertices); + + if (write_hole_flags) { + fprintf(fp, "%d\n", p->hole[c]); + } + + for (v = 0; v < p->contour[c].num_vertices; v++) { + fprintf(fp, "% .*lf % .*lf\n", DBL_DIG, p->contour[c].vertex[v].x, + DBL_DIG, p->contour[c].vertex[v].y); + } + } +} +*/ + +void gpc_add_contour(gpc_polygon *p, gpc_vertex_list *new_contour, int hole) { + int *extended_hole = NULL; + int c = 0; + int v = 0; + gpc_vertex_list *extended_contour = NULL; + + /* Create an extended hole array */ + gpc_malloc(extended_hole, (p->num_contours + 1) * sizeof(int), + const_cast("contour hole addition")); + + /* Create an extended contour array */ + gpc_malloc(extended_contour, + (p->num_contours + 1) * sizeof(gpc_vertex_list), + const_cast("contour addition")); + + /* Copy the old contour and hole data into the extended arrays */ + for (c = 0; c < p->num_contours; c++) { + extended_hole[c] = p->hole[c]; + extended_contour[c] = p->contour[c]; + } + + /* Copy the new contour and hole onto the end of the extended arrays */ + c = p->num_contours; + extended_hole[c] = hole; + extended_contour[c].num_vertices = new_contour->num_vertices; + gpc_malloc(extended_contour[c].vertex, + new_contour->num_vertices * sizeof(gpc_vertex), + const_cast("contour addition")); + for (v = 0; v < new_contour->num_vertices; v++) { + extended_contour[c].vertex[v] = new_contour->vertex[v]; + } + + /* Dispose of the old contour */ + gpc_free(p->contour); + gpc_free(p->hole); + + /* Update the polygon information */ + p->num_contours++; + p->hole = extended_hole; + p->contour = extended_contour; +} + +// gpc_polygon_clip +void gpc_polygon_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip, + gpc_polygon *result) { + sb_tree *sbtree = NULL; + it_node *it = NULL; + it_node *intersect = NULL; + edge_node *edge = NULL; + edge_node *prev_edge = NULL; + edge_node *next_edge = NULL; + edge_node *succ_edge = NULL; + edge_node *e0 = NULL; + edge_node *e1 = NULL; + edge_node *aet = NULL; + edge_node *c_heap = NULL; + edge_node *s_heap = NULL; + lmt_node *lmt = NULL; + lmt_node *local_min = NULL; + polygon_node *out_poly = NULL; + polygon_node *p = NULL; + polygon_node *q = NULL; + polygon_node *poly = NULL; + polygon_node *npoly = NULL; + polygon_node *cf = NULL; + vertex_node *vtx = NULL; + vertex_node *nv = NULL; + h_state horiz[2]; + int in[2]; + int exists[2]; + int parity[2] = {LEFT, LEFT}; + int c = 0; + int v = 0; + int contributing = 0; + int search = 0; + int scanbeam = 0; + int sbt_entries = 0; + int vclass = 0; + int bl = 0; + int br = 0; + int tl = 0; + int tr = 0; + double *sbt = NULL; + double xb = 0.0; + double px = 0.0; + double yb = 0.0; + double yt = 0.0; + double dy = 0.0; + double ix = 0.0; + double iy = 0.0; + + /* Test for trivial NULL result cases */ + if (((subj->num_contours == 0) && (clip->num_contours == 0)) || + ((subj->num_contours == 0) && ((op == GPC_INT) || (op == GPC_DIFF))) || + ((clip->num_contours == 0) && (op == GPC_INT))) { + result->num_contours = 0; + result->hole = NULL; + result->contour = NULL; + return; + } + /* Identify potentialy contributing contours */ + if (((op == GPC_INT) || (op == GPC_DIFF)) && (subj->num_contours > 0) && + (clip->num_contours > 0)) { + minimax_test(subj, clip, op); + } + /* Build LMT */ + if (subj->num_contours > 0) { + s_heap = build_lmt(&lmt, &sbtree, &sbt_entries, subj, SUBJ, op); + } + if (clip->num_contours > 0) { + c_heap = build_lmt(&lmt, &sbtree, &sbt_entries, clip, CLIP, op); + } + /* Return a NULL result if no contours contribute */ + if (lmt == NULL) { + result->num_contours = 0; + result->hole = NULL; + result->contour = NULL; + reset_lmt(&lmt); + gpc_free(s_heap); + gpc_free(c_heap); + return; + } + + /* Build scanbeam table from scanbeam tree */ + gpc_malloc(sbt, sbt_entries * sizeof(double), + const_cast("sbt creation")); + build_sbt(&scanbeam, sbt, sbtree); + scanbeam = 0; + free_sbtree(&sbtree); + /* Allow pointer re-use without causing memory leak */ + if (subj == result) { + gpc_free_polygon(subj); + } + if (clip == result) { + gpc_free_polygon(clip); + } + /* Invert clip polygon for difference operation */ + if (op == GPC_DIFF) { + parity[CLIP] = RIGHT; + } + local_min = lmt; + + // Process each scanbeam + while (scanbeam < sbt_entries) { + /* Set yb and yt to the bottom and top of the scanbeam */ + yb = sbt[scanbeam++]; + if (scanbeam < sbt_entries) { + yt = sbt[scanbeam]; + dy = yt - yb; + } + /* === SCANBEAM BOUNDARY PROCESSING ================================ */ + /* If LMT node corresponding to yb exists */ + if (local_min) { + if (local_min->y == yb) { + /* Add edges starting at this local minimum to the AET */ + for (edge = local_min->first_bound; edge; edge = edge->next_bound) { + add_edge_to_aet(&aet, edge, NULL); + } + local_min = local_min->next; + } + } + /* Set dummy previous x value */ + px = -DBL_MAX; + /* Create bundles within AET */ + e0 = aet; + e1 = aet; + /* Set up bundle fields of first edge */ + aet->bundle[ABOVE][aet->type] = (aet->top.y != yb); + aet->bundle[ABOVE][!aet->type] = 0; + aet->bstate[ABOVE] = UNBUNDLED; + + for (next_edge = aet->next; next_edge; next_edge = next_edge->next) { + /* Set up bundle fields of next edge */ + next_edge->bundle[ABOVE][next_edge->type] = (next_edge->top.y != yb); + next_edge->bundle[ABOVE][!next_edge->type] = 0; + next_edge->bstate[ABOVE] = UNBUNDLED; + /* Bundle edges above the scanbeam boundary if they coincide */ + if (next_edge->bundle[ABOVE][next_edge->type]) { + if (gpc_eq(e0->xb, next_edge->xb) && gpc_eq(e0->dx, next_edge->dx) && + (e0->top.y != yb)) { + next_edge->bundle[ABOVE][next_edge->type] ^= + e0->bundle[ABOVE][next_edge->type]; + next_edge->bundle[ABOVE][!next_edge->type] = + e0->bundle[ABOVE][!next_edge->type]; + next_edge->bstate[ABOVE] = BUNDLE_HEAD; + e0->bundle[ABOVE][CLIP] = 0; + e0->bundle[ABOVE][SUBJ] = 0; + e0->bstate[ABOVE] = BUNDLE_TAIL; + } + e0 = next_edge; + } + } + horiz[CLIP] = NH; + horiz[SUBJ] = NH; + + // Process each edge at this scanbeam boundary + for (edge = aet; edge; edge = edge->next) { + exists[CLIP] = + edge->bundle[ABOVE][CLIP] + (edge->bundle[BELOW][CLIP] << 1); + exists[SUBJ] = + edge->bundle[ABOVE][SUBJ] + (edge->bundle[BELOW][SUBJ] << 1); + if (exists[CLIP] || exists[SUBJ]) { + /* Set bundle side */ + edge->bside[CLIP] = parity[CLIP]; + edge->bside[SUBJ] = parity[SUBJ]; + /* Determine contributing status and quadrant occupancies */ + switch (op) { + case GPC_DIFF: + case GPC_INT: + contributing = (exists[CLIP] && (parity[SUBJ] || horiz[SUBJ])) || + (exists[SUBJ] && (parity[CLIP] || horiz[CLIP])) || + (exists[CLIP] && exists[SUBJ] && + (parity[CLIP] == parity[SUBJ])); + br = (parity[CLIP]) && (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) && + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) && + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) && + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + case GPC_XOR: + contributing = exists[CLIP] || exists[SUBJ]; + br = (parity[CLIP]) ^ (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) ^ + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) ^ + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) ^ + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + case GPC_UNION: + contributing = (exists[CLIP] && (!parity[SUBJ] || horiz[SUBJ])) || + (exists[SUBJ] && (!parity[CLIP] || horiz[CLIP])) || + (exists[CLIP] && exists[SUBJ] && + (parity[CLIP] == parity[SUBJ])); + br = (parity[CLIP]) || (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) || + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) || + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) || + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + } + // Update parity + parity[CLIP] ^= edge->bundle[ABOVE][CLIP]; + parity[SUBJ] ^= edge->bundle[ABOVE][SUBJ]; + /* Update horizontal state */ + if (exists[CLIP]) { + horiz[CLIP] = next_h_state[horiz[CLIP]] + [((exists[CLIP] - 1) << 1) + parity[CLIP]]; + } + if (exists[SUBJ]) { + horiz[SUBJ] = next_h_state[horiz[SUBJ]] + [((exists[SUBJ] - 1) << 1) + parity[SUBJ]]; + } + vclass = tr + (tl << 1) + (br << 2) + (bl << 3); + if (contributing) { + xb = edge->xb; + switch (vclass) { + case EMN: + case IMN: + add_local_min(&out_poly, edge, xb, yb); + px = xb; + cf = edge->outp[ABOVE]; + break; + case ERI: + if (xb != px) { + add_right(cf, xb, yb); + px = xb; + } + edge->outp[ABOVE] = cf; + cf = NULL; + break; + case ELI: + add_left(edge->outp[BELOW], xb, yb); + px = xb; + cf = edge->outp[BELOW]; + break; + case EMX: + if (xb != px) { + add_left(cf, xb, yb); + px = xb; + } + merge_right(cf, edge->outp[BELOW], out_poly); + cf = NULL; + break; + case ILI: + if (xb != px) { + add_left(cf, xb, yb); + px = xb; + } + edge->outp[ABOVE] = cf; + cf = NULL; + break; + case IRI: + add_right(edge->outp[BELOW], xb, yb); + px = xb; + cf = edge->outp[BELOW]; + edge->outp[BELOW] = NULL; + break; + case IMX: + if (xb != px) { + add_right(cf, xb, yb); + px = xb; + } + merge_left(cf, edge->outp[BELOW], out_poly); + cf = NULL; + edge->outp[BELOW] = NULL; + break; + case IMM: + if (xb != px) { + add_right(cf, xb, yb); + px = xb; + } + merge_left(cf, edge->outp[BELOW], out_poly); + edge->outp[BELOW] = NULL; + add_local_min(&out_poly, edge, xb, yb); + cf = edge->outp[ABOVE]; + break; + case EMM: + if (xb != px) { + add_left(cf, xb, yb); + px = xb; + } + merge_right(cf, edge->outp[BELOW], out_poly); + edge->outp[BELOW] = NULL; + add_local_min(&out_poly, edge, xb, yb); + cf = edge->outp[ABOVE]; + break; + case LED: + if (edge->bot.y == yb) { + add_left(edge->outp[BELOW], xb, yb); + } + edge->outp[ABOVE] = edge->outp[BELOW]; + px = xb; + break; + case RED: + if (edge->bot.y == yb) { + add_right(edge->outp[BELOW], xb, yb); + } + edge->outp[ABOVE] = edge->outp[BELOW]; + px = xb; + break; + default: + break; + } /* End of switch */ + } /* End of contributing conditional */ + } /* End of edge exists conditional */ + } // End of AET loop + + /* Delete terminating edges from the AET, otherwise compute xt */ + for (edge = aet; edge; edge = edge->next) { + if (edge->top.y == yb) { + prev_edge = edge->prev; + next_edge = edge->next; + if (prev_edge) { + prev_edge->next = next_edge; + } else { + aet = next_edge; + } + if (next_edge) { + next_edge->prev = prev_edge; + } + /* Copy bundle head state to the adjacent tail edge if required */ + if ((edge->bstate[BELOW] == BUNDLE_HEAD) && prev_edge) { + if (prev_edge->bstate[BELOW] == BUNDLE_TAIL) { + prev_edge->outp[BELOW] = edge->outp[BELOW]; + prev_edge->bstate[BELOW] = UNBUNDLED; + if (prev_edge->prev) { + if (prev_edge->prev->bstate[BELOW] == BUNDLE_TAIL) { + prev_edge->bstate[BELOW] = BUNDLE_HEAD; + } + } + } + } + } else { + if (edge->top.y == yt) { + edge->xt = edge->top.x; + } else { + edge->xt = edge->bot.x + edge->dx * (yt - edge->bot.y); + } + } + } + + if (scanbeam < sbt_entries) { + /* === SCANBEAM INTERIOR PROCESSING ============================== */ + build_intersection_table(&it, aet, dy); + /* Process each node in the intersection table */ + for (intersect = it; intersect; intersect = intersect->next) { + e0 = intersect->ie[0]; + e1 = intersect->ie[1]; + /* Only generate output for contributing intersections */ + if ((e0->bundle[ABOVE][CLIP] || e0->bundle[ABOVE][SUBJ]) && + (e1->bundle[ABOVE][CLIP] || e1->bundle[ABOVE][SUBJ])) { + p = e0->outp[ABOVE]; + q = e1->outp[ABOVE]; + ix = intersect->point.x; + iy = intersect->point.y + yb; + + in[CLIP] = (e0->bundle[ABOVE][CLIP] && !e0->bside[CLIP]) || + (e1->bundle[ABOVE][CLIP] && e1->bside[CLIP]) || + (!e0->bundle[ABOVE][CLIP] && !e1->bundle[ABOVE][CLIP] && + e0->bside[CLIP] && e1->bside[CLIP]); + in[SUBJ] = (e0->bundle[ABOVE][SUBJ] && !e0->bside[SUBJ]) || + (e1->bundle[ABOVE][SUBJ] && e1->bside[SUBJ]) || + (!e0->bundle[ABOVE][SUBJ] && !e1->bundle[ABOVE][SUBJ] && + e0->bside[SUBJ] && e1->bside[SUBJ]); + + // Determine quadrant occupancies + switch (op) { + case GPC_DIFF: + case GPC_INT: + tr = (in[CLIP]) && (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + case GPC_XOR: + tr = (in[CLIP]) ^ (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + case GPC_UNION: + tr = (in[CLIP]) || (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + } + vclass = tr + (tl << 1) + (br << 2) + (bl << 3); + switch (vclass) { + case EMN: + add_local_min(&out_poly, e0, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + break; + case ERI: + if (p) { + add_right(p, ix, iy); + e1->outp[ABOVE] = p; + e0->outp[ABOVE] = NULL; + } + break; + case ELI: + if (q) { + add_left(q, ix, iy); + e0->outp[ABOVE] = q; + e1->outp[ABOVE] = NULL; + } + break; + case EMX: + if (p && q) { + add_left(p, ix, iy); + merge_right(p, q, out_poly); + e0->outp[ABOVE] = NULL; + e1->outp[ABOVE] = NULL; + } + break; + case IMN: + add_local_min(&out_poly, e0, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + break; + case ILI: + if (p) { + add_left(p, ix, iy); + e1->outp[ABOVE] = p; + e0->outp[ABOVE] = NULL; + } + break; + case IRI: + if (q) { + add_right(q, ix, iy); + e0->outp[ABOVE] = q; + e1->outp[ABOVE] = NULL; + } + break; + case IMX: + if (p && q) { + add_right(p, ix, iy); + merge_left(p, q, out_poly); + e0->outp[ABOVE] = NULL; + e1->outp[ABOVE] = NULL; + } + break; + case IMM: + if (p && q) { + add_right(p, ix, iy); + merge_left(p, q, out_poly); + add_local_min(&out_poly, e0, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + } + break; + case EMM: + if (p && q) { + add_left(p, ix, iy); + merge_right(p, q, out_poly); + add_local_min(&out_poly, e0, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + } + break; + default: + break; + } // End of switch + } /* End of contributing intersection conditional */ + + /* Swap bundle sides in response to edge crossing */ + if (e0->bundle[ABOVE][CLIP]) { + e1->bside[CLIP] = !e1->bside[CLIP]; + } + if (e1->bundle[ABOVE][CLIP]) { + e0->bside[CLIP] = !e0->bside[CLIP]; + } + if (e0->bundle[ABOVE][SUBJ]) { + e1->bside[SUBJ] = !e1->bside[SUBJ]; + } + if (e1->bundle[ABOVE][SUBJ]) { + e0->bside[SUBJ] = !e0->bside[SUBJ]; + } + + /* Swap e0 and e1 bundles in the AET */ + prev_edge = e0->prev; + next_edge = e1->next; + if (next_edge) { + next_edge->prev = e0; + } + if (e0->bstate[ABOVE] == BUNDLE_HEAD) { + search = 1; + while (search) { + prev_edge = prev_edge->prev; + if (prev_edge) { + if (prev_edge->bstate[ABOVE] != BUNDLE_TAIL) { + search = 0; + } + } else { + search = 0; + } + } + } + if (!prev_edge) { + aet->prev = e1; + e1->next = aet; + aet = e0->next; + } else { + prev_edge->next->prev = e1; + e1->next = prev_edge->next; + prev_edge->next = e0->next; + } + e0->next->prev = prev_edge; + e1->next->prev = e1; + e0->next = next_edge; + } /* End of IT loop*/ + + // Prepare for next scanbeam + for (edge = aet; edge; edge = next_edge) { + next_edge = edge->next; + succ_edge = edge->succ; + if ((edge->top.y == yt) && succ_edge) { + /* Replace AET edge by its successor */ + succ_edge->outp[BELOW] = edge->outp[ABOVE]; + succ_edge->bstate[BELOW] = edge->bstate[ABOVE]; + succ_edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP]; + succ_edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ]; + prev_edge = edge->prev; + if (prev_edge) { + prev_edge->next = succ_edge; + } else { + aet = succ_edge; + } + if (next_edge) { + next_edge->prev = succ_edge; + } + succ_edge->prev = prev_edge; + succ_edge->next = next_edge; + } else { + /* Update this edge */ + edge->outp[BELOW] = edge->outp[ABOVE]; + edge->bstate[BELOW] = edge->bstate[ABOVE]; + edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP]; + edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ]; + edge->xb = edge->xt; + } + edge->outp[ABOVE] = NULL; + } + } + } /* === END OF SCANBEAM PROCESSING ================================== */ + // Generate result polygon from out_poly + result->contour = NULL; + result->hole = NULL; + result->num_contours = count_contours(out_poly); + if (result->num_contours > 0) { + gpc_malloc(result->hole, result->num_contours * sizeof(int), + const_cast("hole flag table creation")); + gpc_malloc(result->contour, + result->num_contours * sizeof(gpc_vertex_list), + const_cast("contour creation")); + + c = 0; + for (poly = out_poly; poly; poly = npoly) { + npoly = poly->next; + if (poly->active) { + result->hole[c] = poly->proxy->hole; + result->contour[c].num_vertices = poly->active; + gpc_malloc( + result->contour[c].vertex, + result->contour[c].num_vertices * sizeof(gpc_vertex), + const_cast("vertex creation")); + + v = result->contour[c].num_vertices - 1; + for (vtx = poly->proxy->v[LEFT]; vtx; vtx = nv) { + nv = vtx->next; + result->contour[c].vertex[v].x = vtx->x; + result->contour[c].vertex[v].y = vtx->y; + gpc_free(vtx); + v--; + } + c++; + } + gpc_free(poly); + } + } else { + for (poly = out_poly; poly; poly = npoly) { + npoly = poly->next; + gpc_free(poly); + } + } + + // Tidy up + reset_it(&it); + reset_lmt(&lmt); + gpc_free(c_heap); + gpc_free(s_heap); + gpc_free(sbt); +} // NOLINT + +void gpc_free_tristrip(gpc_tristrip *t) { + int s = 0; + for (s = 0; s < t->num_strips; s++) { + gpc_free(t->strip[s].vertex); + } + gpc_free(t->strip); + t->num_strips = 0; +} + +void gpc_polygon_to_tristrip(gpc_polygon *s, gpc_tristrip *t) { + gpc_polygon c; + c.num_contours = 0; + c.hole = NULL; + c.contour = NULL; + gpc_tristrip_clip(GPC_DIFF, s, &c, t); +} + +// gpc_tristrip_clip +void gpc_tristrip_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip, + gpc_tristrip *result) { + sb_tree *sbtree = NULL; + it_node *it = NULL; + it_node *intersect = NULL; + edge_node *edge = NULL; + edge_node *prev_edge = NULL; + edge_node *next_edge = NULL; + edge_node *succ_edge = NULL; + edge_node *e0 = NULL; + edge_node *e1 = NULL; + edge_node *aet = NULL; + edge_node *c_heap = NULL; + edge_node *s_heap = NULL; + edge_node *cf = NULL; + lmt_node *lmt = NULL; + lmt_node *local_min = NULL; + polygon_node *tlist = NULL; + polygon_node *tn = NULL; + polygon_node *tnn = NULL; + polygon_node *p = NULL; + polygon_node *q = NULL; + vertex_node *lt = NULL; + vertex_node *ltn = NULL; + vertex_node *rt = NULL; + vertex_node *rtn = NULL; + h_state horiz[2]; + vertex_type cft = NUL; + int in[2]; + int exists[2]; + int parity[2] = {LEFT, LEFT}; + int s = 0; + int v = 0; + int contributing = 0; + int search = 0; + int scanbeam = 0; + int sbt_entries = 0; + int vclass = 0; + int bl = 0; + int br = 0; + int tl = 0; + int tr = 0; + double *sbt = NULL; + double xb = 0.0; + double px = 0.0; + double nx = 0.0; + double yb = 0.0; + double yt = 0.0; + double dy = 0.0; + double ix = 0.0; + double iy = 0.0; + + /* Test for trivial NULL result cases */ + if (((subj->num_contours == 0) && (clip->num_contours == 0)) || + ((subj->num_contours == 0) && ((op == GPC_INT) || (op == GPC_DIFF))) || + ((clip->num_contours == 0) && (op == GPC_INT))) { + result->num_strips = 0; + result->strip = NULL; + return; + } + + /* Identify potentialy contributing contours */ + if (((op == GPC_INT) || (op == GPC_DIFF)) && (subj->num_contours > 0) && + (clip->num_contours > 0)) { + minimax_test(subj, clip, op); + } + /* Build LMT */ + if (subj->num_contours > 0) { + s_heap = build_lmt(&lmt, &sbtree, &sbt_entries, subj, SUBJ, op); + } + if (clip->num_contours > 0) { + c_heap = build_lmt(&lmt, &sbtree, &sbt_entries, clip, CLIP, op); + } + /* Return a NULL result if no contours contribute */ + if (lmt == NULL) { + result->num_strips = 0; + result->strip = NULL; + reset_lmt(&lmt); + gpc_free(s_heap); + gpc_free(c_heap); + return; + } + + /* Build scanbeam table from scanbeam tree */ + gpc_malloc(sbt, sbt_entries * sizeof(double), + const_cast("sbt creation")); + build_sbt(&scanbeam, sbt, sbtree); + scanbeam = 0; + free_sbtree(&sbtree); + + /* Invert clip polygon for difference operation */ + if (op == GPC_DIFF) { + parity[CLIP] = RIGHT; + } + local_min = lmt; + + // Process each scanbeam + while (scanbeam < sbt_entries) { + /* Set yb and yt to the bottom and top of the scanbeam */ + yb = sbt[scanbeam++]; + if (scanbeam < sbt_entries) { + yt = sbt[scanbeam]; + dy = yt - yb; + } + + /* === SCANBEAM BOUNDARY PROCESSING ================================ */ + /* If LMT node corresponding to yb exists */ + if (local_min) { + if (local_min->y == yb) { + /* Add edges starting at this local minimum to the AET */ + for (edge = local_min->first_bound; edge; edge = edge->next_bound) { + add_edge_to_aet(&aet, edge, NULL); + } + local_min = local_min->next; + } + } + /* Set dummy previous x value */ + /* Create bundles within AET */ + px = -DBL_MAX; + e0 = aet; + e1 = aet; + + /* Set up bundle fields of first edge */ + aet->bundle[ABOVE][aet->type] = (aet->top.y != yb); + aet->bundle[ABOVE][!aet->type] = 0; + aet->bstate[ABOVE] = UNBUNDLED; + + for (next_edge = aet->next; next_edge; next_edge = next_edge->next) { + /* Set up bundle fields of next edge */ + next_edge->bundle[ABOVE][next_edge->type] = (next_edge->top.y != yb); + next_edge->bundle[ABOVE][!next_edge->type] = 0; + next_edge->bstate[ABOVE] = UNBUNDLED; + + /* Bundle edges above the scanbeam boundary if they coincide */ + if (next_edge->bundle[ABOVE][next_edge->type]) { + if (gpc_eq(e0->xb, next_edge->xb) && gpc_eq(e0->dx, next_edge->dx) && + (e0->top.y != yb)) { + next_edge->bundle[ABOVE][next_edge->type] ^= + e0->bundle[ABOVE][next_edge->type]; + next_edge->bundle[ABOVE][!next_edge->type] = + e0->bundle[ABOVE][!next_edge->type]; + next_edge->bstate[ABOVE] = BUNDLE_HEAD; + e0->bundle[ABOVE][CLIP] = 0; + e0->bundle[ABOVE][SUBJ] = 0; + e0->bstate[ABOVE] = BUNDLE_TAIL; + } + e0 = next_edge; + } + } + horiz[CLIP] = NH; + horiz[SUBJ] = NH; + + /* Process each edge at this scanbeam boundary */ + for (edge = aet; edge; edge = edge->next) { + exists[CLIP] = + edge->bundle[ABOVE][CLIP] + (edge->bundle[BELOW][CLIP] << 1); + exists[SUBJ] = + edge->bundle[ABOVE][SUBJ] + (edge->bundle[BELOW][SUBJ] << 1); + + if (exists[CLIP] || exists[SUBJ]) { + /* Set bundle side */ + edge->bside[CLIP] = parity[CLIP]; + edge->bside[SUBJ] = parity[SUBJ]; + + /* Determine contributing status and quadrant occupancies */ + switch (op) { + case GPC_DIFF: + case GPC_INT: + contributing = (exists[CLIP] && (parity[SUBJ] || horiz[SUBJ])) || + (exists[SUBJ] && (parity[CLIP] || horiz[CLIP])) || + (exists[CLIP] && exists[SUBJ] && + (parity[CLIP] == parity[SUBJ])); + br = (parity[CLIP]) && (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) && + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) && + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) && + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + case GPC_XOR: + contributing = exists[CLIP] || exists[SUBJ]; + br = (parity[CLIP]) ^ (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) ^ + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) ^ + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) ^ + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + case GPC_UNION: + contributing = (exists[CLIP] && (!parity[SUBJ] || horiz[SUBJ])) || + (exists[SUBJ] && (!parity[CLIP] || horiz[CLIP])) || + (exists[CLIP] && exists[SUBJ] && + (parity[CLIP] == parity[SUBJ])); + br = (parity[CLIP]) || (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) || + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) || + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) || + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + } + + // Update parity + parity[CLIP] ^= edge->bundle[ABOVE][CLIP]; + parity[SUBJ] ^= edge->bundle[ABOVE][SUBJ]; + + /* Update horizontal state */ + if (exists[CLIP]) { + horiz[CLIP] = next_h_state[horiz[CLIP]] + [((exists[CLIP] - 1) << 1) + parity[CLIP]]; + } + if (exists[SUBJ]) { + horiz[SUBJ] = next_h_state[horiz[SUBJ]] + [((exists[SUBJ] - 1) << 1) + parity[SUBJ]]; + } + vclass = tr + (tl << 1) + (br << 2) + (bl << 3); + + if (contributing) { + xb = edge->xb; + switch (vclass) { + case EMN: + new_tristrip(&tlist, edge, xb, yb); + cf = edge; + break; + case ERI: + edge->outp[ABOVE] = cf->outp[ABOVE]; + if (xb != cf->xb) { + gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb); + } + cf = NULL; + break; + case ELI: + gpc_vertex_create(edge, BELOW, LEFT, xb, yb); + edge->outp[ABOVE] = NULL; + cf = edge; + break; + case EMX: + if (xb != cf->xb) { + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + } + edge->outp[ABOVE] = NULL; + cf = NULL; + break; + case IMN: + if (cft == LED) { + if (cf->bot.y != yb) { + gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb); + } + new_tristrip(&tlist, cf, cf->xb, yb); + } + edge->outp[ABOVE] = cf->outp[ABOVE]; + gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb); + break; + case ILI: + new_tristrip(&tlist, edge, xb, yb); + cf = edge; + cft = ILI; + break; + case IRI: + if (cft == LED) { + if (cf->bot.y != yb) { + gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb); + } + new_tristrip(&tlist, cf, cf->xb, yb); + } + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + edge->outp[ABOVE] = NULL; + break; + case IMX: + gpc_vertex_create(edge, BELOW, LEFT, xb, yb); + edge->outp[ABOVE] = NULL; + cft = IMX; + break; + case IMM: + gpc_vertex_create(edge, BELOW, LEFT, xb, yb); + edge->outp[ABOVE] = cf->outp[ABOVE]; + if (xb != cf->xb) { + gpc_vertex_create(cf, ABOVE, RIGHT, xb, yb); + } + cf = edge; + break; + case EMM: + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + edge->outp[ABOVE] = NULL; + new_tristrip(&tlist, edge, xb, yb); + cf = edge; + break; + case LED: + if (edge->bot.y == yb) { + gpc_vertex_create(edge, BELOW, LEFT, xb, yb); + } + edge->outp[ABOVE] = edge->outp[BELOW]; + cf = edge; + cft = LED; + break; + case RED: + edge->outp[ABOVE] = cf->outp[ABOVE]; + if (cft == LED) { + if (cf->bot.y == yb) { + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + } else { + if (edge->bot.y == yb) { + gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb); + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + } + } + } else { + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb); + } + cf = NULL; + break; + default: + break; + } /* End of switch */ + } /* End of contributing conditional */ + } /* End of edge exists conditional */ + } // End of AET loop + + /* Delete terminating edges from the AET, otherwise compute xt */ + for (edge = aet; edge; edge = edge->next) { + if (edge->top.y == yb) { + prev_edge = edge->prev; + next_edge = edge->next; + if (prev_edge) { + prev_edge->next = next_edge; + } else { + aet = next_edge; + } + if (next_edge) { + next_edge->prev = prev_edge; + } + + /* Copy bundle head state to the adjacent tail edge if required */ + if ((edge->bstate[BELOW] == BUNDLE_HEAD) && prev_edge) { + if (prev_edge->bstate[BELOW] == BUNDLE_TAIL) { + prev_edge->outp[BELOW] = edge->outp[BELOW]; + prev_edge->bstate[BELOW] = UNBUNDLED; + if (prev_edge->prev) { + if (prev_edge->prev->bstate[BELOW] == BUNDLE_TAIL) { + prev_edge->bstate[BELOW] = BUNDLE_HEAD; + } + } + } + } + } else { + if (edge->top.y == yt) { + edge->xt = edge->top.x; + } else { + edge->xt = edge->bot.x + edge->dx * (yt - edge->bot.y); + } + } + } + + if (scanbeam < sbt_entries) { + /* === SCANBEAM INTERIOR PROCESSING ============================== */ + build_intersection_table(&it, aet, dy); + /* Process each node in the intersection table */ + for (intersect = it; intersect; intersect = intersect->next) { + e0 = intersect->ie[0]; + e1 = intersect->ie[1]; + + /* Only generate output for contributing intersections */ + if ((e0->bundle[ABOVE][CLIP] || e0->bundle[ABOVE][SUBJ]) && + (e1->bundle[ABOVE][CLIP] || e1->bundle[ABOVE][SUBJ])) { + p = e0->outp[ABOVE]; + q = e1->outp[ABOVE]; + ix = intersect->point.x; + iy = intersect->point.y + yb; + + in[CLIP] = (e0->bundle[ABOVE][CLIP] && !e0->bside[CLIP]) || + (e1->bundle[ABOVE][CLIP] && e1->bside[CLIP]) || + (!e0->bundle[ABOVE][CLIP] && !e1->bundle[ABOVE][CLIP] && + e0->bside[CLIP] && e1->bside[CLIP]); + in[SUBJ] = (e0->bundle[ABOVE][SUBJ] && !e0->bside[SUBJ]) || + (e1->bundle[ABOVE][SUBJ] && e1->bside[SUBJ]) || + (!e0->bundle[ABOVE][SUBJ] && !e1->bundle[ABOVE][SUBJ] && + e0->bside[SUBJ] && e1->bside[SUBJ]); + + switch (op) { // Determine quadrant occupancies + case GPC_DIFF: + case GPC_INT: + tr = (in[CLIP]) && (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + case GPC_XOR: + tr = (in[CLIP]) ^ (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + case GPC_UNION: + tr = (in[CLIP]) || (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + } + + vclass = tr + (tl << 1) + (br << 2) + (bl << 3); + switch (vclass) { + case EMN: + new_tristrip(&tlist, e1, ix, iy); + e0->outp[ABOVE] = e1->outp[ABOVE]; + break; + case ERI: + if (p) { + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + e0->outp[ABOVE] = NULL; + } + break; + case ELI: + if (q) { + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(e1, ABOVE, LEFT, ix, iy); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + e0->outp[ABOVE] = e1->outp[ABOVE]; + e1->outp[ABOVE] = NULL; + } + break; + case EMX: + if (p && q) { + gpc_vertex_create(e0, ABOVE, LEFT, ix, iy); + e0->outp[ABOVE] = NULL; + e1->outp[ABOVE] = NULL; + } + break; + case IMN: + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + new_tristrip(&tlist, prev_edge, px, iy); + e1->outp[ABOVE] = prev_edge->outp[ABOVE]; + gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy); + new_tristrip(&tlist, e0, ix, iy); + next_edge->outp[ABOVE] = e0->outp[ABOVE]; + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + break; + case ILI: + if (p) { + gpc_vertex_create(e0, ABOVE, LEFT, ix, iy); + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + e0->outp[ABOVE] = NULL; + } + break; + case IRI: + if (q) { + gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy); + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + e0->outp[ABOVE] = e1->outp[ABOVE]; + e1->outp[ABOVE] = NULL; + } + break; + case IMX: + if (p && q) { + gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy); + gpc_vertex_create(e1, ABOVE, LEFT, ix, iy); + e0->outp[ABOVE] = NULL; + e1->outp[ABOVE] = NULL; + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + new_tristrip(&tlist, prev_edge, px, iy); + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + next_edge->outp[ABOVE] = prev_edge->outp[ABOVE]; + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + } + break; + case IMM: + if (p && q) { + gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy); + gpc_vertex_create(e1, ABOVE, LEFT, ix, iy); + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + new_tristrip(&tlist, prev_edge, px, iy); + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + e1->outp[ABOVE] = prev_edge->outp[ABOVE]; + gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy); + new_tristrip(&tlist, e0, ix, iy); + next_edge->outp[ABOVE] = e0->outp[ABOVE]; + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + } + break; + case EMM: + if (p && q) { + gpc_vertex_create(e0, ABOVE, LEFT, ix, iy); + new_tristrip(&tlist, e1, ix, iy); + e0->outp[ABOVE] = e1->outp[ABOVE]; + } + break; + default: + break; + } /* End of switch */ + } /* End of contributing intersection conditional */ + + // Swap bundle sides in response to edge crossing + if (e0->bundle[ABOVE][CLIP]) { + e1->bside[CLIP] = !e1->bside[CLIP]; + } + if (e1->bundle[ABOVE][CLIP]) { + e0->bside[CLIP] = !e0->bside[CLIP]; + } + if (e0->bundle[ABOVE][SUBJ]) { + e1->bside[SUBJ] = !e1->bside[SUBJ]; + } + if (e1->bundle[ABOVE][SUBJ]) { + e0->bside[SUBJ] = !e0->bside[SUBJ]; + } + + /* Swap e0 and e1 bundles in the AET */ + prev_edge = e0->prev; + next_edge = e1->next; + if (e1->next) { + e1->next->prev = e0; + } + + if (e0->bstate[ABOVE] == BUNDLE_HEAD) { + search = 1; + while (search) { + prev_edge = prev_edge->prev; + if (prev_edge) { + if (prev_edge->bundle[ABOVE][CLIP] || + prev_edge->bundle[ABOVE][SUBJ] || + (prev_edge->bstate[ABOVE] == BUNDLE_HEAD)) { + search = 0; + } + } else { + search = 0; + } + } + } + if (!prev_edge) { + e1->next = aet; + aet = e0->next; + } else { + e1->next = prev_edge->next; + prev_edge->next = e0->next; + } + e0->next->prev = prev_edge; + e1->next->prev = e1; + e0->next = next_edge; + } /* End of IT loop*/ + + /* Prepare for next scanbeam */ + for (edge = aet; edge; edge = next_edge) { + next_edge = edge->next; + succ_edge = edge->succ; + + if ((edge->top.y == yt) && succ_edge) { + /* Replace AET edge by its successor */ + succ_edge->outp[BELOW] = edge->outp[ABOVE]; + succ_edge->bstate[BELOW] = edge->bstate[ABOVE]; + succ_edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP]; + succ_edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ]; + prev_edge = edge->prev; + if (prev_edge) { + prev_edge->next = succ_edge; + } else { + aet = succ_edge; + } + if (next_edge) { + next_edge->prev = succ_edge; + } + succ_edge->prev = prev_edge; + succ_edge->next = next_edge; + } else { + /* Update this edge */ + edge->outp[BELOW] = edge->outp[ABOVE]; + edge->bstate[BELOW] = edge->bstate[ABOVE]; + edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP]; + edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ]; + edge->xb = edge->xt; + } + edge->outp[ABOVE] = NULL; + } + } + } /* === END OF SCANBEAM PROCESSING ================================== */ + + // Generate result tristrip from tlist + result->strip = NULL; + result->num_strips = count_tristrips(tlist); + if (result->num_strips > 0) { + gpc_malloc(result->strip, + result->num_strips * sizeof(gpc_vertex_list), + const_cast("tristrip list creation")); + + s = 0; + for (tn = tlist; tn; tn = tnn) { + tnn = tn->next; + if (tn->active > 2) { + /* Valid tristrip: copy the vertices and free the heap */ + result->strip[s].num_vertices = tn->active; + gpc_malloc(result->strip[s].vertex, + tn->active * sizeof(gpc_vertex), + const_cast("tristrip creation")); + v = 0; + if (0) { + lt = tn->v[RIGHT]; + rt = tn->v[LEFT]; + } else { + lt = tn->v[LEFT]; + rt = tn->v[RIGHT]; + } + while (lt || rt) { + if (lt) { + ltn = lt->next; + result->strip[s].vertex[v].x = lt->x; + result->strip[s].vertex[v].y = lt->y; + v++; + gpc_free(lt); + lt = ltn; + } + if (rt) { + rtn = rt->next; + result->strip[s].vertex[v].x = rt->x; + result->strip[s].vertex[v].y = rt->y; + v++; + gpc_free(rt); + rt = rtn; + } + } + s++; + } else { + /* Invalid tristrip: just free the heap */ + for (lt = tn->v[LEFT]; lt; lt = ltn) { + ltn = lt->next; + gpc_free(lt); + } + for (rt = tn->v[RIGHT]; rt; rt = rtn) { + rtn = rt->next; + gpc_free(rt); + } + } + gpc_free(tn); + } + } + // Tidy up + reset_it(&it); + reset_lmt(&lmt); + gpc_free(c_heap); + gpc_free(s_heap); + gpc_free(sbt); +} // NOLINT + +} // namespace gpc + +/* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */ diff --git a/paddle/fluid/operators/detection/gpc.h b/paddle/fluid/operators/detection/gpc.h new file mode 100644 index 0000000000000000000000000000000000000000..ee86262ef2c486e4eaeeeaf56c2392d2a1c5851b --- /dev/null +++ b/paddle/fluid/operators/detection/gpc.h @@ -0,0 +1,246 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/*************************************************************************** + * + * Copyright (c) 2015 Baidu.com, Inc. All Rights Reserved + * + **************************************************************************/ + +/** + * @file include/gpc.h + * @author huhan02(com@baidu.com) + * @date 2015/12/18 13:52:10 + * @brief + * + * @modified by sunyipeng + * @email sunyipeng@baidu.com + * @date 2018/6/12 + **/ + +#ifndef PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ // GPC_H_ +#define PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ // GPC_H_ + +#include +#include +#include +#include + +namespace gpc { + +typedef enum { // Set operation type + GPC_DIFF, // Difference + GPC_INT, // Intersection + GPC_XOR, // Exclusive or + GPC_UNION // Union +} gpc_op; + +typedef struct { // Polygon vertex structure + double x; // Vertex x component + double y; // vertex y component +} gpc_vertex; + +typedef struct { // Vertex list structure + int num_vertices; // Number of vertices in list + gpc_vertex *vertex; // Vertex array pointer +} gpc_vertex_list; + +typedef struct { // Polygon set structure + int num_contours; // Number of contours in polygon + int *hole; // Hole external contour flags + gpc_vertex_list *contour; // Contour array pointer +} gpc_polygon; + +typedef struct { // Tristrip set structure + int num_strips; // Number of tristrips + gpc_vertex_list *strip; // Tristrip array pointer +} gpc_tristrip; + +typedef enum { LEFT, RIGHT } gpc_left_right; + +typedef enum { ABOVE, BELOW } gpc_above_below; + +typedef enum { CLIP, SUBJ } gpc_clip_subj; + +typedef enum { /* Edge intersection classes */ + NUL, /* Empty non-intersection */ + EMX, /* External maximum */ + ELI, /* External left intermediate */ + TED, /* Top edge */ + ERI, /* External right intermediate */ + RED, /* Right edge */ + IMM, /* Internal maximum and minimum */ + IMN, /* Internal minimum */ + EMN, /* External minimum */ + EMM, /* External maximum and minimum */ + LED, /* Left edge */ + ILI, /* Internal left intermediate */ + BED, /* Bottom edge */ + IRI, /* Internal right intermediate */ + IMX, /* Internal maximum */ + FUL /* Full non-intersection */ +} vertex_type; + +typedef enum { /* Horizontal edge states */ + NH, /* No horizontal edge */ + BH, /* Bottom horizontal edge */ + TH /* Top horizontal edge */ +} h_state; + +typedef enum { /* Edge bundle state */ + UNBUNDLED, /* Isolated edge not within a bundle */ + BUNDLE_HEAD, /* Bundle head node */ + BUNDLE_TAIL /* Passive bundle tail node */ +} bundle_state; + +typedef struct v_shape { /* Internal vertex list datatype */ + double x; /* X coordinate component */ + double y; /* Y coordinate component */ + struct v_shape *next; /* Pointer to next vertex in list */ +} vertex_node; + +typedef struct p_shape { /* Internal contour / tristrip type */ + int active; /* Active flag / vertex count */ + int hole; /* Hole / external contour flag */ + vertex_node *v[2]; /* Left and right vertex list ptrs */ + struct p_shape *next; /* Pointer to next polygon contour */ + struct p_shape *proxy; /* Pointer to actual structure used */ +} polygon_node; + +typedef struct edge_shape { + gpc_vertex vertex; /* Piggy-backed contour vertex data */ + gpc_vertex bot; /* Edge lower (x, y) coordinate */ + gpc_vertex top; /* Edge upper (x, y) coordinate */ + double xb; /* Scanbeam bottom x coordinate */ + double xt; /* Scanbeam top x coordinate */ + double dx; /* Change in x for a unit y increase */ + int type; /* Clip / subject edge flag */ + int bundle[2][2]; /* Bundle edge flags */ + int bside[2]; /* Bundle left / right indicators */ + bundle_state bstate[2]; /* Edge bundle state */ + polygon_node *outp[2]; /* Output polygon / tristrip pointer */ + struct edge_shape *prev; /* Previous edge in the AET */ + struct edge_shape *next; /* Next edge in the AET */ + struct edge_shape *pred; /* Edge connected at the lower end */ + struct edge_shape *succ; /* Edge connected at the upper end */ + struct edge_shape *next_bound; /* Pointer to next bound in LMT */ +} edge_node; + +inline bool gpc_eq(float a, float b) { return (fabs(a - b) <= 1e-6); } + +inline bool gpc_prev_index(float a, float b) { return (fabs(a - b) <= 1e-6); } + +inline int gpc_prev_index(int i, int n) { return ((i - 1 + n) % n); } + +inline int gpc_next_index(int i, int n) { return ((i + 1) % n); } + +inline int gpc_optimal(gpc_vertex *v, int i, int n) { + return (v[(i + 1) % n].y != v[i].y || v[(i - 1 + n) % n].y != v[i].y); +} + +inline int gpc_fwd_min(edge_node *v, int i, int n) { + return (v[(i + 1) % n].vertex.y > v[i].vertex.y && + v[(i - 1 + n) % n].vertex.y >= v[i].vertex.y); +} + +inline int gpc_not_fmax(edge_node *v, int i, int n) { + return (v[(i + 1) % n].vertex.y > v[i].vertex.y); +} + +inline int gpc_rev_min(edge_node *v, int i, int n) { + return (v[(i + 1) % n].vertex.y >= v[i].vertex.y && + v[(i - 1 + n) % n].vertex.y > v[i].vertex.y); +} + +inline int gpc_not_rmax(edge_node *v, int i, int n) { + return (v[(i - 1 + n) % n].vertex.y > v[i].vertex.y); +} + +// inline void gpc_p_edge(edge_node *d, edge_node *e, int p, double i, double j) +// { +inline void gpc_p_edge(edge_node *d, edge_node *e, int p) { + d = e; + do { + d = d->prev; + } while (!d->outp[p]); + // i = d->bot.x + d->dx * (j - d->bot.y); +} + +// inline void gpc_n_edge(edge_node *d, edge_node *e, int p, double i, double j) +// { +inline void gpc_n_edge(edge_node *d, edge_node *e, int p) { + d = e; + do { + d = d->next; + } while (!d->outp[p]); + // i = d->bot.x + d->dx * (j - d->bot.y); +} + +template +void gpc_malloc(T *&p, int b, char *s) { + if (b > 0) { + p = (T *)malloc(b); + + if (!p) { + fprintf(stderr, "gpc malloc failure: %s\n", s); + exit(0); + } + } else { + p = NULL; + } +} +template +void gpc_free(T *&p) { + if (p) { + free(p); + p = NULL; + } +} + +/* +=========================================================================== + Public Function Prototypes +=========================================================================== +*/ + +void add_vertex(vertex_node **t, double x, double y); + +void gpc_vertex_create(edge_node *e, int p, int s, double x, double y); + +/* +void gpc_read_polygon(FILE *infile_ptr, int read_hole_flags, + gpc_polygon *polygon); + +void gpc_write_polygon(FILE *outfile_ptr, int write_hole_flags, + gpc_polygon *polygon); +*/ +void gpc_add_contour(gpc_polygon *polygon, gpc_vertex_list *contour, int hole); + +void gpc_polygon_clip(gpc_op set_operation, gpc_polygon *subject_polygon, + gpc_polygon *clip_polygon, gpc_polygon *result_polygon); + +void gpc_tristrip_clip(gpc_op set_operation, gpc_polygon *subject_polygon, + gpc_polygon *clip_polygon, + gpc_tristrip *result_tristrip); + +void gpc_polygon_to_tristrip(gpc_polygon *polygon, gpc_tristrip *tristrip); + +void gpc_free_polygon(gpc_polygon *polygon); + +void gpc_free_tristrip(gpc_tristrip *tristrip); + +} // namespace gpc + +#endif // PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ +/* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */ diff --git a/paddle/fluid/operators/detection/multiclass_nms_op.cc b/paddle/fluid/operators/detection/multiclass_nms_op.cc index 60b93efdce810f8552374449fe5a6fc79b1a92c1..9e78b28a6011bb7bd299ca3438eb407f600d7000 100644 --- a/paddle/fluid/operators/detection/multiclass_nms_op.cc +++ b/paddle/fluid/operators/detection/multiclass_nms_op.cc @@ -9,10 +9,11 @@ http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and + limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detection/poly_util.h" namespace paddle { namespace operators { @@ -20,9 +21,6 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -constexpr int64_t kOutputDim = 6; -constexpr int64_t kBBoxSize = 4; - class MultiClassNMSOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -42,10 +40,15 @@ class MultiClassNMSOp : public framework::OperatorWithKernel { "The rank of Input(BBoxes) must be 3."); PADDLE_ENFORCE_EQ(score_dims.size(), 3, "The rank of Input(Scores) must be 3."); - PADDLE_ENFORCE_EQ(box_dims[2], 4, - "The 2nd dimension of Input(BBoxes) must be 4, " - "represents the layout of coordinate " - "[xmin, ymin, xmax, ymax]"); + PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 || box_dims[2] == 16 || + box_dims[2] == 24 || box_dims[2] == 32, + "The 2nd dimension of Input(BBoxes) must be 4 or 8, " + "represents the layout of coordinate " + "[xmin, ymin, xmax, ymax] or " + "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or " + "8 points: [xi, yi] i= 1,2,...,8 or " + "12 points: [xi, yi] i= 1,2,...,12 or " + "16 points: [xi, yi] i= 1,2,...,16"); PADDLE_ENFORCE_EQ(box_dims[1], score_dims[2], "The 1st dimensiong of Input(BBoxes) must be equal to " "3rd dimension of Input(Scores), which represents the " @@ -53,7 +56,7 @@ class MultiClassNMSOp : public framework::OperatorWithKernel { // Here the box_dims[0] is not the real dimension of output. // It will be rewritten in the computing kernel. - ctx->SetOutputDim("Out", {box_dims[1], 6}); + ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2}); } protected: @@ -128,6 +131,21 @@ static inline T JaccardOverlap(const T* box1, const T* box2, } } +template +T PolyIoU(const T* box1, const T* box2, const size_t box_size, + const bool normalized) { + T bbox1_area = PolyArea(box1, box_size, normalized); + T bbox2_area = PolyArea(box2, box_size, normalized); + T inter_area = PolyOverlapArea(box1, box2, box_size, normalized); + if (bbox1_area == 0 || bbox2_area == 0 || inter_area == 0) { + // If coordinate values are is invalid + // if area size <= 0, return 0. + return T(0.); + } else { + return inter_area / (bbox1_area + bbox2_area - inter_area); + } +} + template class MultiClassNMSKernel : public framework::OpKernel { public: @@ -137,6 +155,8 @@ class MultiClassNMSKernel : public framework::OpKernel { // The total boxes for each instance. int64_t num_boxes = bbox.dims()[0]; // 4: [xmin ymin xmax ymax] + // 8: [x1 y1 x2 y2 x3 y3 x4 y4] + // 16, 24, or 32: [x1 y1 x2 y2 ... xn yn], n = 8, 12 or 16 int64_t box_size = bbox.dims()[1]; std::vector scores_data(num_boxes); @@ -154,8 +174,19 @@ class MultiClassNMSKernel : public framework::OpKernel { for (size_t k = 0; k < selected_indices->size(); ++k) { if (keep) { const int kept_idx = (*selected_indices)[k]; - T overlap = JaccardOverlap(bbox_data + idx * box_size, + T overlap = T(0.); + // 4: [xmin ymin xmax ymax] + if (box_size == 4) { + overlap = JaccardOverlap(bbox_data + idx * box_size, bbox_data + kept_idx * box_size, true); + } + // 8: [x1 y1 x2 y2 x3 y3 x4 y4] or 16, 24, 32 + if (box_size == 8 || box_size == 16 || box_size == 24 || + box_size == 32) { + overlap = + PolyIoU(bbox_data + idx * box_size, + bbox_data + kept_idx * box_size, box_size, true); + } keep = overlap <= adaptive_threshold; } else { break; @@ -228,7 +259,9 @@ class MultiClassNMSKernel : public framework::OpKernel { void MultiClassOutput(const Tensor& scores, const Tensor& bboxes, const std::map>& selected_indices, Tensor* outs) const { - int predict_dim = scores.dims()[1]; + int64_t predict_dim = scores.dims()[1]; + int64_t box_size = bboxes.dims()[1]; + int64_t out_dim = bboxes.dims()[1] + 2; auto* scores_data = scores.data(); auto* bboxes_data = bboxes.data(); auto* odata = outs->data(); @@ -240,11 +273,11 @@ class MultiClassNMSKernel : public framework::OpKernel { const std::vector& indices = it.second; for (size_t j = 0; j < indices.size(); ++j) { int idx = indices[j]; - const T* bdata = bboxes_data + idx * kBBoxSize; - odata[count * kOutputDim] = label; // label - odata[count * kOutputDim + 1] = sdata[idx]; // score - // xmin, ymin, xmax, ymax - std::memcpy(odata + count * kOutputDim + 2, bdata, 4 * sizeof(T)); + const T* bdata = bboxes_data + idx * box_size; + odata[count * out_dim] = label; // label + odata[count * out_dim + 1] = sdata[idx]; // score + // xmin, ymin, xmax, ymax or multi-points coordinates + std::memcpy(odata + count * out_dim + 2, bdata, box_size * sizeof(T)); count++; } } @@ -261,6 +294,7 @@ class MultiClassNMSKernel : public framework::OpKernel { int64_t class_num = score_dims[1]; int64_t predict_dim = score_dims[2]; int64_t box_dim = boxes->dims()[2]; + int64_t out_dim = boxes->dims()[2] + 2; std::vector>> all_indices; std::vector batch_starts = {0}; @@ -283,7 +317,7 @@ class MultiClassNMSKernel : public framework::OpKernel { T* od = outs->mutable_data({1}, ctx.GetPlace()); od[0] = -1; } else { - outs->mutable_data({num_kept, kOutputDim}, ctx.GetPlace()); + outs->mutable_data({num_kept, out_dim}, ctx.GetPlace()); for (int64_t i = 0; i < batch_size; ++i) { Tensor ins_score = scores->Slice(i, i + 1); ins_score.Resize({class_num, predict_dim}); @@ -311,10 +345,11 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("BBoxes", - "(Tensor) A 3-D Tensor with shape [N, M, 4] represents the " + "(Tensor) A 3-D Tensor with shape " + "[N, M, 4 or 8 16 24 32] represents the " "predicted locations of M bounding bboxes, N is the batch size. " "Each bounding box has four coordinate values and the layout is " - "[xmin, ymin, xmax, ymax]."); + "[xmin, ymin, xmax, ymax], when box size equals to 4."); AddInput("Scores", "(Tensor) A 3-D Tensor with shape [N, C, M] represents the " "predicted confidence predictions. N is the batch size, C is the " @@ -351,8 +386,12 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Out", "(LoDTensor) A 2-D LoDTensor with shape [No, 6] represents the " "detections. Each row has 6 values: " - "[label, confidence, xmin, ymin, xmax, ymax], No is the total " - "number of detections in this mini-batch. For each instance, " + "[label, confidence, xmin, ymin, xmax, ymax] or " + "(LoDTensor) A 2-D LoDTensor with shape [No, 10] represents the " + "detections. Each row has 10 values: " + "[label, confidence, x1, y1, x2, y2, x3, y3, x4, y4]. No is the " + "total number of detections in this mini-batch." + "For each instance, " "the offsets in first dimension are called LoD, the number of " "offset is N + 1, if LoD[i + 1] - LoD[i] == 0, means there is " "no detected bbox."); diff --git a/paddle/fluid/operators/detection/poly_util.cc b/paddle/fluid/operators/detection/poly_util.cc new file mode 100644 index 0000000000000000000000000000000000000000..1af2c95c6cf526d651b196b54614a21a9cddde8c --- /dev/null +++ b/paddle/fluid/operators/detection/poly_util.cc @@ -0,0 +1,132 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifndef POLY_UTIL_CC_ +#define POLY_UTIL_CC_ + +#include "paddle/fluid/operators/detection/poly_util.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using gpc::gpc_polygon_clip; +using gpc::gpc_free_polygon; + +template +void Array2PointVec(const T*& box, const size_t box_size, + std::vector>& vec) { + size_t pts_num = box_size / 2; + vec.resize(pts_num); + for (size_t i = 0; i < pts_num; i++) { + vec.at(i).x = box[2 * i]; + vec.at(i).y = box[2 * i + 1]; + } +} + +template +void Array2Poly(const T*& box, const size_t box_size, gpc::gpc_polygon& poly) { + size_t pts_num = box_size / 2; + poly.num_contours = 1; + poly.hole = (int*)malloc(sizeof(int)); + poly.hole[0] = 0; + poly.contour = (gpc::gpc_vertex_list*)malloc(sizeof(gpc::gpc_vertex_list)); + poly.contour->num_vertices = pts_num; + poly.contour->vertex = + (gpc::gpc_vertex*)malloc(sizeof(gpc::gpc_vertex) * pts_num); + for (size_t i = 0; i < pts_num; ++i) { + poly.contour->vertex[i].x = box[2 * i]; + poly.contour->vertex[i].y = box[2 * i + 1]; + } +} + +template +void PointVec2Poly(const std::vector>& vec, gpc::gpc_polygon& poly) { + int pts_num = vec.size(); + poly.num_contours = 1; + poly.hole = (int*)malloc(sizeof(int)); + poly.hole[0] = 0; + poly.contour = (gpc::gpc_vertex_list*)malloc(sizeof(gpc::gpc_vertex_list)); + poly.contour->num_vertices = pts_num; + poly.contour->vertex = + (gpc::gpc_vertex*)malloc(sizeof(gpc::gpc_vertex) * pts_num); + for (size_t i = 0; i < pts_num; ++i) { + poly.contour->vertex[i].x = vec[i].x; + poly.contour->vertex[i].y = vec[i].y; + } +} + +template +void Poly2PointVec(const gpc::gpc_vertex_list& contour, + std::vector>& vec) { + int pts_num = contour.num_vertices; + vec.resize(pts_num); + for (int i = 0; i < pts_num; i++) { + vec.at(i).x = contour.vertex[i].x; + vec.at(i).y = contour.vertex[i].y; + } +} + +template +T GetContourArea(std::vector>& vec) { + size_t pts_num = vec.size(); + if (pts_num < 3) return T(0.); + T area = T(0.); + for (size_t i = 0; i < pts_num; ++i) { + area += vec[i].x * vec[(i + 1) % pts_num].y - + vec[i].y * vec[(i + 1) % pts_num].x; + } + return std::fabs(area / 2.0); +} + +template +T PolyArea(const T* box, const size_t box_size, const bool normalized) { + // If coordinate values are is invalid + // if area size <= 0, return 0. + std::vector> vec; + Array2PointVec(box, box_size, vec); + return GetContourArea(vec); +} + +template +T PolyOverlapArea(const T* box1, const T* box2, const size_t box_size, + const bool normalized) { + gpc::gpc_polygon poly1; + gpc::gpc_polygon poly2; + Array2Poly(box1, box_size, poly1); + Array2Poly(box2, box_size, poly2); + gpc::gpc_polygon respoly; + gpc::gpc_op op = gpc::GPC_INT; + gpc::gpc_polygon_clip(op, &poly2, &poly1, &respoly); + + T inter_area = T(0.); + int contour_num = respoly.num_contours; + for (int i = 0; i < contour_num; ++i) { + std::vector> resvec; + Poly2PointVec(respoly.contour[i], resvec); + // inter_area += std::fabs(cv::contourArea(resvec)) + 0.5f * + // (cv::arcLength(resvec, true)); + inter_area += GetContourArea(resvec); + } + + gpc::gpc_free_polygon(&poly1); + gpc::gpc_free_polygon(&poly2); + gpc::gpc_free_polygon(&respoly); + return inter_area; +} + +} // namespace operators +} // namespace paddle + +#endif diff --git a/paddle/fluid/operators/detection/poly_util.h b/paddle/fluid/operators/detection/poly_util.h new file mode 100644 index 0000000000000000000000000000000000000000..f07baf72d9ff07b8fcb45dcfb2a35741fb1aeed0 --- /dev/null +++ b/paddle/fluid/operators/detection/poly_util.h @@ -0,0 +1,73 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifndef POLY_UTIL_H_ +#define POLY_UTIL_H_ + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detection/gpc.h" + +namespace paddle { +namespace operators { + +template +class Point_ { + public: + // default constructor + Point_() {} + Point_(T _x, T _y) {} + Point_(const Point_& pt) {} + + Point_& operator=(const Point_& pt); + // conversion to another data type + // template operator Point_<_T>() const; + // conversion to the old-style C structures + // operator Vec() const; + + // checks whether the point is inside the specified rectangle + // bool inside(const Rect_& r) const; + T x; //!< x coordinate of the point + T y; //!< y coordinate of the point +}; + +template +void Array2PointVec(const T*& box, const size_t box_size, + std::vector>& vec); + +template +void Array2Poly(const T*& box, const size_t box_size, gpc::gpc_polygon& poly); + +template +void PointVec2Poly(const std::vector>& vec, gpc::gpc_polygon& poly); + +template +void Poly2PointVec(const gpc::gpc_vertex_list& contour, + std::vector>& vec); + +template +T GetContourArea(std::vector>& vec); + +template +T PolyArea(const T* box, const size_t box_size, const bool normalized); + +template +T PolyOverlapArea(const T* box1, const T* box2, const size_t box_size, + const bool normalized); +} // namespace operators +} // namespace paddle + +#include "paddle/fluid/operators/detection/poly_util.cc" + +#endif // POLY_UTIL_H_ diff --git a/paddle/fluid/operators/detection/polygon_box_transform_op.cc b/paddle/fluid/operators/detection/polygon_box_transform_op.cc index 568d50d457d838d5f11605710c0d3b987af01d10..4b3bc2edb58fe23393d906094c41b6ad62c71155 100644 --- a/paddle/fluid/operators/detection/polygon_box_transform_op.cc +++ b/paddle/fluid/operators/detection/polygon_box_transform_op.cc @@ -41,9 +41,9 @@ class PolygonBoxTransformCPUKernel : public framework::OpKernel { for (int id_w = 0; id_w < width; ++id_w) { id = id_n * height * width + width * id_h + id_w; if (id_n % 2 == 0) { - out_data[id] = id_w - in_data[id]; + out_data[id] = id_w * 4 - in_data[id]; } else { - out_data[id] = id_h - in_data[id]; + out_data[id] = id_h * 4 - in_data[id]; } } } diff --git a/paddle/fluid/operators/detection/polygon_box_transform_op.cu b/paddle/fluid/operators/detection/polygon_box_transform_op.cu index 6187ac6622c65d2bbc525c3fe2cb397cf74ac612..e1eaf084a3413dd1d13514e2d7b22572d21dd119 100644 --- a/paddle/fluid/operators/detection/polygon_box_transform_op.cu +++ b/paddle/fluid/operators/detection/polygon_box_transform_op.cu @@ -32,9 +32,9 @@ __global__ void PolygonBoxTransformKernel(const int n, const int h, const int w, if (id_n < n && id_h < h && id_w < w) { int id = id_n * h * w + w * id_h + id_w; if (id_n % 2 == 0) { - output[id] = id_w - input[id]; + output[id] = id_w * 4 - input[id]; } else { - output[id] = id_h - input[id]; + output[id] = id_h * 4 - input[id]; } } } diff --git a/paddle/fluid/operators/distributed/CMakeLists.txt b/paddle/fluid/operators/distributed/CMakeLists.txt index 56734b81e8716a0c0c37a11e35c9118ee7b55020..21db93958a4a586c74a1e060f1f04b5af1dcd889 100644 --- a/paddle/fluid/operators/distributed/CMakeLists.txt +++ b/paddle/fluid/operators/distributed/CMakeLists.txt @@ -20,7 +20,7 @@ if(WITH_GRPC) DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL) cc_test(rpc_server_test SRCS rpc_server_test.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_sparse_table_op SERIAL) - cc_test(varhandle_test SRCS varhandle_test.cc) + cc_test(varhandle_test SRCS varhandle_test.cc DEPS profiler) return() endif() diff --git a/paddle/fluid/operators/distributed/grpc_client.cc b/paddle/fluid/operators/distributed/grpc_client.cc index 13682b78f0eccf049daa315f3a26aafd22e42a41..f5d5627815c7320dad5051b0f7d95b8ec6703687 100644 --- a/paddle/fluid/operators/distributed/grpc_client.cc +++ b/paddle/fluid/operators/distributed/grpc_client.cc @@ -12,14 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/distributed/grpc_client.h" - #include - #include #include "glog/logging.h" // For VLOG #include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/operators/distributed/grpc_client.h" #include "paddle/fluid/operators/distributed/grpc_serde.h" #include "paddle/fluid/operators/distributed/request_handler.h" #include "paddle/fluid/platform/profiler.h" @@ -73,10 +71,11 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, const framework::Scope* p_scope = &scope; const auto ch = GetChannel(ep_val); SendProcessor* s = new SendProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "Send", var_name_val, p_ctx, p_scope)); + const std::string method = "SendRPC"; + VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); - framework::AsyncIO([var_name_val, p_scope, p_ctx, s, this] { + framework::AsyncIO([var_name_val, p_scope, p_ctx, s, method, h, this] { auto* var = p_scope->FindVar(var_name_val); ::grpc::ByteBuffer req; @@ -87,10 +86,16 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, // stub context s->response_call_back_ = nullptr; + platform::RecordRPCEvent record_event(method, p_ctx); + auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } }); req_count_++; @@ -122,10 +127,11 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep, const framework::Scope* p_scope = &scope; const auto ch = GetChannel(ep_val); GetProcessor* s = new GetProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "Get", var_name_val, p_ctx, p_scope)); + const std::string method = "GetRPC"; + VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); - framework::AsyncIO([var_name_val, s, this] { + framework::AsyncIO([var_name_val, s, method, p_ctx, h, this] { // prepare input sendrecv::VariableMessage req; req.set_varname(var_name_val); @@ -137,10 +143,16 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep, // stub context s->response_call_back_ = ProcGetResponse; + platform::RecordRPCEvent record_event(method, p_ctx); + auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } }); req_count_++; @@ -161,12 +173,14 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep, const framework::Scope* p_scope = &scope; const auto ch = GetChannel(ep_val); GetProcessor* s = new GetProcessor(ch); - VarHandlePtr h( - new VarHandle(ep, "Prefetch", out_var_name_val, p_ctx, p_scope)); + + const std::string method = "PrefetchRPC"; + + VarHandlePtr h(new VarHandle(ep, method, out_var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, - s, this] { + s, method, h, this] { auto* var = p_scope->FindVar(in_var_name_val); ::grpc::ByteBuffer req; @@ -177,11 +191,17 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep, // stub context s->response_call_back_ = ProcGetResponse; + platform::RecordRPCEvent record_event(method, p_ctx); + auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, static_cast(s)); + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } }); req_count_++; @@ -193,15 +213,24 @@ VarHandlePtr GRPCClient::AsyncSendBatchBarrier(const std::string& ep, const auto ch = GetChannel(ep); BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "BatchBarrier", BATCH_BARRIER_MESSAGE, - nullptr, nullptr)); + const std::string method = "BatchBarrierRPC"; + VarHandlePtr h( + new VarHandle(ep, method, BATCH_BARRIER_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(BATCH_BARRIER_MESSAGE); + + platform::RecordRPCEvent record_event(method, nullptr); + auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + return h; } @@ -209,15 +238,24 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { const auto ch = GetChannel(ep); FetchBarrierProcessor* s = new FetchBarrierProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "FetchBarrier", FETCH_BARRIER_MESSAGE, - nullptr, nullptr)); + const std::string method = "FetchBarrierRPC"; + VarHandlePtr h( + new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(FETCH_BARRIER_MESSAGE); + + platform::RecordRPCEvent record_event(method, nullptr); + auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + return h; } @@ -226,15 +264,23 @@ VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep, const auto ch = GetChannel(ep); BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); - VarHandlePtr h( - new VarHandle(ep, "SendComplete", COMPLETE_MESSAGE, nullptr, nullptr)); + const std::string method = "SendCompleteRPC"; + VarHandlePtr h(new VarHandle(ep, method, COMPLETE_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(COMPLETE_MESSAGE); + + platform::RecordRPCEvent record_event(method, nullptr); + auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + return h; } @@ -244,17 +290,27 @@ VarHandlePtr GRPCClient::AsyncCheckpointNotify(const std::string& ep, const auto ch = GetChannel(ep); CheckpointNotifyProcessor* s = new CheckpointNotifyProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "CheckPointNotify", CHECKPOINT_SAVE_MESSAGE, - nullptr, nullptr)); + + const std::string method = "CheckPointNotifyRPC"; + + VarHandlePtr h( + new VarHandle(ep, method, CHECKPOINT_SAVE_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(CHECKPOINT_SAVE_MESSAGE); req.set_out_varname(dir); + platform::RecordRPCEvent record_event(method, nullptr); + auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + return h; } @@ -273,12 +329,16 @@ void GRPCClient::Proceed() { BaseProcessor* c = static_cast(tag); GPR_ASSERT(ok); PADDLE_ENFORCE(c); + if (c->status_.ok()) { VLOG(3) << c->GetVarHandlePtr()->String() << " process"; c->Process(); } else if (c->status_.error_code() == grpc::StatusCode::DEADLINE_EXCEEDED) { + // FIXME(gongwb): parse error_details? LOG(ERROR) << c->GetVarHandlePtr()->String() - << " meets grpc error:" << c->status_.error_message(); + << " meets grpc error, error_code:" << c->status_.error_code() + << " error_message:" << c->status_.error_message() + << " error_details:" << c->status_.error_details(); { std::lock_guard lk(sync_mutex_); ok_ = false; @@ -286,7 +346,10 @@ void GRPCClient::Proceed() { c->Finish(false); } else { LOG(FATAL) << c->GetVarHandlePtr()->String() - << " meets grpc error:" << c->status_.error_message(); + << " meets grpc error, error_code:" << c->status_.error_code() + << " error_message:" << c->status_.error_message() + << " error_details:" << c->status_.error_details(); + c->Finish(false); } diff --git a/paddle/fluid/operators/distributed/grpc_serde.cc b/paddle/fluid/operators/distributed/grpc_serde.cc index 3f8796713a6b89a308113981614673e07e8d367f..bac098b892658beece85271765eb31eeb3eeda17 100644 --- a/paddle/fluid/operators/distributed/grpc_serde.cc +++ b/paddle/fluid/operators/distributed/grpc_serde.cc @@ -36,6 +36,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, ::grpc::ByteBuffer* msg, const std::string& out_name) { + platform::RecordRPCEvent record_event("serial", &ctx); // Default DestroyCallback does nothing, When using GPU // the CPU buffer need to be freed. DestroyCallback destroy_callback = [](void* backing) {}; @@ -147,6 +148,7 @@ void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, const framework::Scope* scope, framework::Variable** var) { + platform::RecordRPCEvent record_event("deserial", &ctx); operators::distributed::GRPCVariableResponse resp(scope, &ctx); PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!"); *var = resp.GetVar(); diff --git a/paddle/fluid/operators/fill_constant_op.cc b/paddle/fluid/operators/fill_constant_op.cc index 2826b82117db113d4d8c10095e89f610ca895775..e04a68717b351ddb0be5a7e70aa9297e5eb0125f 100644 --- a/paddle/fluid/operators/fill_constant_op.cc +++ b/paddle/fluid/operators/fill_constant_op.cc @@ -70,6 +70,12 @@ class FillConstantOp : public framework::OperatorBase { } }; +class FillConstantOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override {} +}; + class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -102,4 +108,5 @@ Fill up a variable with specified constant value. namespace ops = paddle::operators; REGISTER_OPERATOR(fill_constant, ops::FillConstantOp, ops::FillConstantInferShape, ops::FillConstantOpMaker, - paddle::framework::EmptyGradOpMaker); + paddle::framework::EmptyGradOpMaker, + ops::FillConstantOpVarTypeInference); diff --git a/paddle/fluid/operators/ftrl_op.cc b/paddle/fluid/operators/ftrl_op.cc index 70ba25c213046cc934f46be067080d5fdbb42f9e..b77e12d6508eb07ae137b313ca91eac951afbcbe 100644 --- a/paddle/fluid/operators/ftrl_op.cc +++ b/paddle/fluid/operators/ftrl_op.cc @@ -34,6 +34,16 @@ class FTRLOp : public framework::OperatorWithKernel { "Input(Grad) of FTRL should not be null."); PADDLE_ENFORCE(ctx->HasInput("LearningRate"), "Input(LearningRate) of FTRL should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of FTRL should not be null."); diff --git a/paddle/fluid/operators/ftrl_op.h b/paddle/fluid/operators/ftrl_op.h index 6f821e7e9944214fc5ebdf6bc7db8789b8ada6b9..8f812c9a037bfac8c1e29e32a5ad5b077c8153d1 100644 --- a/paddle/fluid/operators/ftrl_op.h +++ b/paddle/fluid/operators/ftrl_op.h @@ -28,6 +28,17 @@ template class FTRLOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto* param_out = ctx.Output("ParamOut"); auto* sq_accum_out = ctx.Output("SquaredAccumOut"); auto* lin_accum_out = ctx.Output("LinearAccumOut"); diff --git a/paddle/fluid/operators/fused_embedding_fc_lstm_op.cc b/paddle/fluid/operators/fused_embedding_fc_lstm_op.cc index 0b917a403620e2ffb2cbb4ca7856cce9584e1eef..fdc9cb4888b3468b85abfa0c693ed8ac5b0d450b 100644 --- a/paddle/fluid/operators/fused_embedding_fc_lstm_op.cc +++ b/paddle/fluid/operators/fused_embedding_fc_lstm_op.cc @@ -93,11 +93,7 @@ void FusedEmbeddingFCLSTMOp::InferShape( ctx->SetOutputDim("Cell", out_dims); ctx->ShareLoD("Ids", "Hidden"); ctx->ShareLoD("Ids", "Cell"); - int xx_width; - if (ctx->Attrs().Get("use_seq")) { - xx_width = wh_dims[1]; - } else { - xx_width = x_dims[1] > wh_dims[1] ? wh_dims[1] : x_dims[1]; + if (!ctx->Attrs().Get("use_seq")) { PADDLE_ENFORCE(ctx->HasOutput("BatchedInput"), "Assert only one Output(BatchedInput) of LSTM."); PADDLE_ENFORCE(ctx->HasOutput("BatchedHidden"), @@ -112,7 +108,7 @@ void FusedEmbeddingFCLSTMOp::InferShape( ctx->SetOutputDim("BatchedHidden", out_dims); ctx->SetOutputDim("BatchedCell", out_dims); } - ctx->SetOutputDim("XX", {x_dims[0], xx_width}); + ctx->SetOutputDim("XX", {x_dims[0], wh_dims[1]}); ctx->ShareLoD("Ids", "XX"); } @@ -435,8 +431,6 @@ class FusedEmbeddingFCLSTMKernel : public framework::OpKernel { INIT_VEC_FUNC INIT_BASE_INPUT_DATAS - // std::cout << "===> Batch Compute" << std::endl; - auto* reordered_h0 = ctx.Output("ReorderedH0"); auto* reordered_c0 = ctx.Output("ReorderedC0"); auto* batched_input = ctx.Output("BatchedInput"); diff --git a/paddle/fluid/operators/fusion_lstm_op.cc b/paddle/fluid/operators/fusion_lstm_op.cc index ae1f6d8e489039667d861a69acabf2c632ef2061..067e6a3e7cccc1f15ebdd984f3a2441339a989ab 100644 --- a/paddle/fluid/operators/fusion_lstm_op.cc +++ b/paddle/fluid/operators/fusion_lstm_op.cc @@ -15,11 +15,9 @@ limitations under the License. */ #include "paddle/fluid/operators/fusion_lstm_op.h" #include #include "paddle/fluid/operators/math/blas.h" -#include "paddle/fluid/operators/math/cpu_lstm_compute.h" -#include "paddle/fluid/operators/math/cpu_vec.h" #include "paddle/fluid/operators/math/fc_compute.h" +#include "paddle/fluid/operators/math/jit_kernel.h" #include "paddle/fluid/operators/math/sequence2batch.h" -#include "paddle/fluid/platform/cpu_info.h" namespace paddle { namespace operators { @@ -219,121 +217,55 @@ This operator fuse the X into LSTM, more details can refer to LSTM op. template class FuisonLSTMKernel : public framework::OpKernel { public: -#define INIT_VEC_FUNC \ - std::function act_gate, act_cell, act_cand; \ - auto& act_gate_str = ctx.Attr("gate_activation"); \ - auto& act_cell_str = ctx.Attr("cell_activation"); \ - auto& act_cand_str = ctx.Attr("candidate_activation"); \ - if (platform::jit::MayIUse(platform::jit::avx)) { \ - math::VecActivations act_functor; \ - act_gate = act_functor(act_gate_str); \ - act_cell = act_functor(act_cell_str); \ - act_cand = act_functor(act_cand_str); \ - } else { \ - math::VecActivations act_functor; \ - act_gate = act_functor(act_gate_str); \ - act_cell = act_functor(act_cell_str); \ - act_cand = act_functor(act_cand_str); \ - } - -#define INIT_BASE_INPUT_OUTPUT \ - auto* x = ctx.Input("X"); \ - auto* h0 = ctx.Input("H0"); \ - auto* c0 = ctx.Input("C0"); \ - auto* wx = ctx.Input("WeightX"); \ - auto* wh = ctx.Input("WeightH"); \ - auto* bias = ctx.Input("Bias"); \ - auto* xx = ctx.Output("XX"); \ - auto* hidden_out = ctx.Output("Hidden"); \ - auto* cell_out = ctx.Output("Cell"); \ - bool is_reverse = ctx.Attr("is_reverse"); \ - bool use_peepholes = ctx.Attr("use_peepholes"); - -#define INIT_BASE_SIZES \ - auto x_dims = x->dims(); /* T x M*/ \ - auto wh_dims = wh->dims(); /* D x 4D*/ \ - const int M = x_dims[1]; \ - const int D = wh_dims[0]; \ - const int D2 = D * 2; \ - const int D3 = D * 3; \ - const int D4 = wh_dims[1]; - -#define INIT_BASE_INPUT_DATAS \ - const T* x_data = x->data(); \ - const T* wx_data = wx->data(); \ - const T* wh_data = wh->data(); \ - /* diagonal weight*/ \ - const T* wc_data = bias->data() + D4; \ - /* for peephole only*/ \ - T* checked_cell_data = nullptr; \ - auto place = ctx.GetPlace(); \ - if (use_peepholes) { \ - /* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/ \ - auto* checked_cell = ctx.Output("CheckedCell"); \ - checked_cell_data = checked_cell->mutable_data(place); \ - } - -/// Compute LSTM +#define INIT_BASE_DEFINES \ + using DeviceContext = paddle::platform::CPUDeviceContext; \ + auto* x = ctx.Input("X"); \ + auto* h0 = ctx.Input("H0"); \ + auto* c0 = ctx.Input("C0"); \ + auto* wx = ctx.Input("WeightX"); \ + auto* wh = ctx.Input("WeightH"); \ + auto* bias = ctx.Input("Bias"); \ + auto* xx = ctx.Output("XX"); \ + auto* hidden_out = ctx.Output("Hidden"); \ + auto* cell_out = ctx.Output("Cell"); \ + bool is_reverse = ctx.Attr("is_reverse"); \ + bool use_peepholes = ctx.Attr("use_peepholes"); \ + auto x_dims = x->dims(); /* T x M*/ \ + auto wh_dims = wh->dims(); /* D x 4D*/ \ + const int M = x_dims[1]; \ + const int D = wh_dims[0]; \ + const int D4 = wh_dims[1] + +#define INIT_OTHER_DEFINES \ + const T* x_data = x->data(); \ + const T* wx_data = wx->data(); \ + const T* wh_data = wh->data(); \ + /* diagonal weight*/ \ + const T* wp_data = bias->data() + D4; \ + /* for peephole only*/ \ + T* checked_cell_data = nullptr; \ + auto place = ctx.GetPlace(); \ + if (use_peepholes) { \ + /* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/ \ + auto* checked_cell = ctx.Output("CheckedCell"); \ + checked_cell_data = checked_cell->mutable_data(place); \ + } \ + const auto& ker = \ + math::jitkernel::KernelPool::Instance() \ + .template Get, const std::string&, \ + const std::string&, const std::string&>( \ + ctx.Attr("gate_activation"), \ + ctx.Attr("candidate_activation"), \ + ctx.Attr("cell_activation"), D, use_peepholes) + +// Wh GEMM #define GEMM_WH_ADDON(bs, prev, out) \ blas.GEMM(CblasNoTrans, CblasNoTrans, bs, D4, D, static_cast(1), prev, D, \ wh_data, D4, static_cast(1), out, D4) -#define GET_Ct(ct_1, gates, ct) \ - /* C_t = C_t-1 * fgated + cand_gated * igated*/ \ - act_cand(D, gates, gates); \ - blas.VMUL(D, gates, gates + D, gates + D); \ - blas.VMUL(D, ct_1, gates + D2, gates + D2); \ - blas.VADD(D, gates + D, gates + D2, ct) - -#define GET_Ht(ct, gates, ht) \ - /* H_t = act_cell(C_t) * ogated */ \ - act_cell(D, ct, gates + D2); \ - blas.VMUL(D, gates + D2, gates + D3, ht) - -#define GET_Ct_NOH0C0(gates, ct) \ - /* C_t = igated * cgated*/ \ - act_gate(D, gates + D, gates + D); \ - act_cand(D, gates, gates); \ - blas.VMUL(D, gates, gates + D, ct) - -#define COMPUTE_CtHt_NOH0C0(gates, ct, ht) \ - GET_Ct_NOH0C0(gates, ct); \ - act_gate(D, gates + D3, gates + D3); \ - GET_Ht(ct, gates, ht) - -#define COMPUTE_CtHt_PEEPHOLE_NOH0C0(gates, ct, ht) \ - GET_Ct_NOH0C0(gates, ct); \ - /* get outgated, put W_oc * C_t on igated */ \ - blas.VMUL(D, wc_data + D2, ct, gates + D); \ - blas.VADD(D, gates + D, gates + D3, gates + D3); \ - act_gate(D, gates + D3, gates + D3); \ - GET_Ht(ct, gates, ht) - -#define COMPUTE_CtHt(gates, ct_1, ct, ht) \ - act_gate(D3, gates + D, gates + D); \ - GET_Ct(ct_1, gates, ct); \ - GET_Ht(ct, gates, ht) - -#define COMPUTE_CtHt_PEEPHOLE(gates, ct_1, ct, ht) \ - /* get fgated and igated*/ \ - blas.VMUL(D, wc_data, ct_1, checked_cell_data); \ - blas.VMUL(D, wc_data + D, ct_1, checked_cell_data + D); \ - blas.VADD(D2, checked_cell_data, gates + D, gates + D); \ - act_gate(D2, gates + D, gates + D); \ - GET_Ct(ct_1, gates, ct); \ - /* get ogated*/ \ - blas.VMUL(D, wc_data + D2, ct, gates + D); \ - blas.VADD(D, gates + D, gates + D3, gates + D3); \ - act_gate(D, gates + D3, gates + D3); \ - GET_Ht(ct, gates, ht) - void SeqCompute(const framework::ExecutionContext& ctx) const { - using DeviceContext = paddle::platform::CPUDeviceContext; - INIT_BASE_INPUT_OUTPUT - INIT_BASE_SIZES - INIT_VEC_FUNC - INIT_BASE_INPUT_DATAS - + INIT_BASE_DEFINES; + INIT_OTHER_DEFINES; auto x_lod = x->lod(); const int total_T = x_dims[0]; const int N = x_lod[0].size() - 1; @@ -357,89 +289,47 @@ class FuisonLSTMKernel : public framework::OpKernel { gate_offset = -D; } -#define MOVE_ONE_STEP \ - prev_h_data = h_out_data; \ - prev_c_data = c_out_data; \ - xx_data = xx_data + xx_offset; \ - h_out_data = h_out_data + gate_offset; \ - c_out_data = c_out_data + gate_offset - -#define PROCESS_H0C0_DEFINES \ - int bid = is_reverse ? N - 1 - i : i; \ - int seq_len = x_lod[0][bid + 1] - x_lod[0][bid]; \ - const T* prev_c_data = nullptr; \ - const T* prev_h_data = nullptr; \ - int tstart = 0 - -#define PROCESS_H0C0_PEEPHOLE \ - PROCESS_H0C0_DEFINES; \ - if (h0_data) { \ - prev_h_data = h0_data + bid * D; \ - prev_c_data = c0_data + bid * D; \ - } else { \ - COMPUTE_CtHt_PEEPHOLE_NOH0C0(xx_data, c_out_data, h_out_data); \ - MOVE_ONE_STEP; \ - tstart = 1; \ - } - -#define PROCESS_H0C0 \ - PROCESS_H0C0_DEFINES; \ - if (h0_data) { \ - prev_h_data = h0_data + bid * D; \ - prev_c_data = c0_data + bid * D; \ - } else { \ - COMPUTE_CtHt_NOH0C0(xx_data, c_out_data, h_out_data); \ - MOVE_ONE_STEP; \ - tstart = 1; \ - } - - if (use_peepholes) { - for (int i = 0; i < N; ++i) { - PROCESS_H0C0_PEEPHOLE - for (int step = tstart; step < seq_len; ++step) { - GEMM_WH_ADDON(1, prev_h_data, xx_data); - COMPUTE_CtHt_PEEPHOLE(xx_data, prev_c_data, c_out_data, h_out_data); - MOVE_ONE_STEP; - } - } - } else { - // TODO(TJ): unly workaround, clean me - std::function compute_ctht; - if (platform::jit::MayIUse(platform::jit::avx) && - act_gate_str == "sigmoid" && act_cand_str == "tanh" && - act_cell_str == "tanh" && D == 8) { - compute_ctht = math::lstm_compute_ctht; + for (int i = 0; i < N; ++i) { + int bid = is_reverse ? N - 1 - i : i; + int seq_len = x_lod[0][bid + 1] - x_lod[0][bid]; + const T* prev_c_data = nullptr; + const T* prev_h_data = nullptr; + int tstart = 0; + if (h0_data) { + prev_h_data = h0_data + bid * D; + prev_c_data = c0_data + bid * D; } else { - compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) { - COMPUTE_CtHt(gates, ct_1, ct, ht); - }; + ker->ComputeC1H1(xx_data, c_out_data, h_out_data, wp_data); + tstart = 1; + // move one step + prev_h_data = h_out_data; + prev_c_data = c_out_data; + xx_data = xx_data + xx_offset; + h_out_data = h_out_data + gate_offset; + c_out_data = c_out_data + gate_offset; } - for (int i = 0; i < N; ++i) { - PROCESS_H0C0 - for (int step = tstart; step < seq_len; ++step) { - GEMM_WH_ADDON(1, prev_h_data, xx_data); - compute_ctht(xx_data, prev_c_data, c_out_data, h_out_data); - MOVE_ONE_STEP; - } + for (int step = tstart; step < seq_len; ++step) { + GEMM_WH_ADDON(1, prev_h_data, xx_data); + ker->ComputeCtHt(xx_data, prev_c_data, c_out_data, h_out_data, wp_data, + checked_cell_data); + // move one step + prev_h_data = h_out_data; + prev_c_data = c_out_data; + xx_data = xx_data + xx_offset; + h_out_data = h_out_data + gate_offset; + c_out_data = c_out_data + gate_offset; } } -#undef PROCESS_H0C0_DEFINES -#undef PROCESS_H0C0_PEEPHOLE -#undef PROCESS_H0C0 -#undef MOVE_ONE_STEP } void BatchCompute(const framework::ExecutionContext& ctx) const { - using DeviceContext = platform::CPUDeviceContext; - INIT_BASE_INPUT_OUTPUT - INIT_BASE_SIZES + INIT_BASE_DEFINES; if (x->lod()[0].size() == 2) { xx->Resize({x_dims[0], D4}); SeqCompute(ctx); return; } - INIT_VEC_FUNC - INIT_BASE_INPUT_DATAS + INIT_OTHER_DEFINES; auto* reordered_h0 = ctx.Output("ReorderedH0"); auto* reordered_c0 = ctx.Output("ReorderedC0"); @@ -487,8 +377,8 @@ class FuisonLSTMKernel : public framework::OpKernel { prev_c_data = reordered_c0_data; size_t sz = sizeof(T) * D; for (int i = 0; i < max_bs; ++i) { - std::memcpy(reordered_h0_data, h0_data + seq_order[i] * D, sz); - std::memcpy(reordered_c0_data, c0_data + seq_order[i] * D, sz); + blas.VCOPY(sz, h0_data + seq_order[i] * D, reordered_h0_data); + blas.VCOPY(sz, c0_data + seq_order[i] * D, reordered_c0_data); reordered_h0_data += D; reordered_c0_data += D; } @@ -498,13 +388,7 @@ class FuisonLSTMKernel : public framework::OpKernel { T* cur_h_out_data = batched_h_out_data; T* cur_c_out_data = batched_c_out_data; for (int i = 0; i < max_bs; ++i) { - GET_Ct_NOH0C0(cur_in_data, cur_c_out_data); - if (use_peepholes) { - blas.VMUL(D, wc_data + D2, cur_c_out_data, cur_in_data + D); - blas.VADD(D, cur_in_data + D, cur_in_data + D3, cur_in_data + D3); - } - act_gate(D, cur_in_data + D3, cur_in_data + D3); - GET_Ht(cur_c_out_data, cur_in_data, cur_h_out_data); + ker->ComputeC1H1(cur_in_data, cur_c_out_data, cur_h_out_data, wp_data); cur_in_data += D4; cur_c_out_data += D; cur_h_out_data += D; @@ -513,71 +397,37 @@ class FuisonLSTMKernel : public framework::OpKernel { prev_h_data = batched_h_out_data; prev_c_data = batched_c_out_data; } + + // compute kernel part const auto& batch_starts = batched_lod[0]; const int max_seq_len = batch_starts.size() - 1; const int offset = tstart * max_bs * D; batched_input_data = batched_input_data + offset * 4; batched_h_out_data = batched_h_out_data + offset; batched_c_out_data = batched_c_out_data + offset; - -#define DEFINE_CUR \ - T* cur_in_data = batched_input_data; \ - T* cur_prev_c_data = prev_c_data; \ - T* cur_c_out_data = batched_c_out_data; \ - T* cur_h_out_data = batched_h_out_data - -#define MOVE_ONE_BATCH \ - cur_in_data += D4; \ - cur_prev_c_data += D; \ - cur_c_out_data += D; \ - cur_h_out_data += D - -#define MOVE_ONE_STEP \ - prev_c_data = batched_c_out_data; \ - prev_h_data = batched_h_out_data; \ - batched_c_out_data = cur_c_out_data; \ - batched_h_out_data = cur_h_out_data; \ - batched_input_data = cur_in_data - - if (use_peepholes) { - for (int step = tstart; step < max_seq_len; ++step) { - const int cur_bs = batch_starts[step + 1] - batch_starts[step]; - GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data); - DEFINE_CUR; - for (int i = 0; i < cur_bs; ++i) { - COMPUTE_CtHt_PEEPHOLE(cur_in_data, cur_prev_c_data, cur_c_out_data, - cur_h_out_data); - MOVE_ONE_BATCH; - } - MOVE_ONE_STEP; - } - } else { - // TODO(TJ): unly workaround, clean me - std::function compute_ctht; - if (platform::jit::MayIUse(platform::jit::avx) && - act_gate_str == "sigmoid" && act_cand_str == "tanh" && - act_cell_str == "tanh" && D == 8) { - compute_ctht = math::lstm_compute_ctht; - } else { - compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) { - COMPUTE_CtHt(gates, ct_1, ct, ht); - }; - } - for (int step = tstart; step < max_seq_len; ++step) { - const int cur_bs = batch_starts[step + 1] - batch_starts[step]; - GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data); - DEFINE_CUR; - for (int i = 0; i < cur_bs; ++i) { - compute_ctht(cur_in_data, cur_prev_c_data, cur_c_out_data, - cur_h_out_data); - MOVE_ONE_BATCH; - } - MOVE_ONE_STEP; + for (int step = tstart; step < max_seq_len; ++step) { + const int cur_bs = batch_starts[step + 1] - batch_starts[step]; + GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data); + T* cur_in_data = batched_input_data; + T* cur_prev_c_data = prev_c_data; + T* cur_c_out_data = batched_c_out_data; + T* cur_h_out_data = batched_h_out_data; + for (int i = 0; i < cur_bs; ++i) { + ker->ComputeCtHt(cur_in_data, cur_prev_c_data, cur_c_out_data, + cur_h_out_data, wp_data, checked_cell_data); + // move one batch + cur_in_data += D4; + cur_prev_c_data += D; + cur_c_out_data += D; + cur_h_out_data += D; } + // move one step + prev_c_data = batched_c_out_data; + prev_h_data = batched_h_out_data; + batched_c_out_data = cur_c_out_data; + batched_h_out_data = cur_h_out_data; + batched_input_data = cur_in_data; } -#undef MOVE_ONE_STEP -#undef MOVE_ONE_BATCH -#undef DEFINE_CUR math::Batch2LoDTensorFunctor to_seq; batched_h_out->set_lod(batched_lod); @@ -594,18 +444,9 @@ class FuisonLSTMKernel : public framework::OpKernel { } } -#undef COMPUTE_CtHt_PEEPHOLE -#undef COMPUTE_CtHt -#undef GET_Ct_NOH0C0 -#undef COMPUTE_CtHt_NOH0C0 -#undef COMPUTE_CtHt_PEEPHOLE_NOH0C0 -#undef GET_Ht -#undef GET_Ct #undef GEMM_WH_ADDON -#undef INIT_BASE_INPUT_DATAS -#undef INIT_BASE_SIZES -#undef INIT_BASE_INPUT_OUTPUT -#undef INIT_VEC_FUNC +#undef INIT_OTHER_DEFINES +#undef INIT_BASE_DEFINES }; } // namespace operators diff --git a/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc b/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..b0910dc19edb246d9acfe3bdb15071c64cbdaba7 --- /dev/null +++ b/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc @@ -0,0 +1,229 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h" +#include // for min, max +#include +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/fc_compute.h" + +namespace paddle { +namespace operators { + +void FusionSeqConvEltAddReluOp::InferShape( + framework::InferShapeContext* ctx) const { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of FusionSeqConvEltAddReluOp should not be null."); + PADDLE_ENFORCE( + ctx->HasInput("Filter"), + "Input(Filter) of FusionSeqConvEltAddReluOp should not be null."); + PADDLE_ENFORCE( + ctx->HasInput("Bias"), + "Input(Bias) of FusionSeqConvEltAddReluOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Out"), + "Output(Out) of FusionSeqConvEltAddReluOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("ColMat"), + "Output(ColMat) of FusionSeqConvEltAddReluOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto w_dims = ctx->GetInputDim("Filter"); + int context_length = ctx->Attrs().Get("contextLength"); + PADDLE_ENFORCE( + ctx->Attrs().Get("contextStride") == 1, + "Currently, FusionSeqConvEltAddReluOp only supports contextStride=1."); + PADDLE_ENFORCE(x_dims.size() == 2 && w_dims.size() == 2, + "Input(X, Filter) should be 2-D tensor."); + PADDLE_ENFORCE(x_dims.size() == 2 && w_dims.size() == 2, + "Input(X, Filter) should be 2-D tensor."); + PADDLE_ENFORCE(w_dims[0] == context_length * x_dims[1], + "Filter's height should be context_length * " + "input_hidden_size ."); + PADDLE_ENFORCE_GT(context_length + ctx->Attrs().Get("contextStart"), 0, + "contextStart size should be smaller than contextLength."); + + ctx->SetOutputDim("Out", {x_dims[0], w_dims[1]}); + ctx->SetOutputDim("ColMat", {x_dims[0], w_dims[0]}); + ctx->ShareLoD("X", "Out"); +} + +framework::OpKernelType FusionSeqConvEltAddReluOp::GetExpectedKernelType( + const framework::ExecutionContext& ctx) const { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); +} + +void FusionSeqConvEltAddReluOpMaker::Make() { + AddInput("X", + "(LoDTensor) the input is a LodTensor, which support " + "variable-time length input sequence. The underlying tensor in " + "this LoDTensor is a matrix with shape (T X M), where T is the " + "total time steps in this mini-batch, M is the dim size of x."); + // PaddingData only support false yet, should be ensured at pass. + AddInput("Filter", + "(Tensor) same as the input(Filter) of sequence conv op is an " + "learnable parameter." + "This is a tensor with shape (K, N), where K is the " + "context_length * dim size of x, N is the output feature size."); + AddInput("Bias", + "(Tensor) the learnable weights. shape (1, N), where N is the " + "output feature size"); + AddOutput( + "Out", + "(LoDTensor) the output(Out) is a LodTensor, which support " + "variable-time length output sequence. The underlying tensor in " + "this LoDTensor is a matrix with shape (T, N), where, T is the " + "total time steps in this mini-batch, N is the output feature size."); + AddOutput("ColMat", + "(Tensor) (T, K), where T is where T is the " + "total time steps in this mini-batch, K is height of Filter") + .AsIntermediate(); + AddAttr("contextLength", + "(int) the contextLength of FusionSeqConvEltAddReluOp is the " + "height of the convolution kernel.") + .GreaterThan(0); + AddAttr("contextStart", + "(int, default:0) the contextStart of FusionSeqConvEltAddReluOp " + "represents the beginning of the convolution of the number of " + "rows of sequence, which can be negative. The negative number " + "means to pad contextStart time-steps of zeros or learnable " + "parameters at the beginning of each instance. The positive " + "number means to skip contextStart time-steps of each " + "instance.") + .SetDefault(0); + AddAttr( + "contextStride", + "(int, default:1) the contextStride of FusionSeqConvEltAddReluOp " + "represents the stride length of convolution kernel. " + "Currently, FusionSeqConvEltAddReluOp only supports" + "contextStride=1.") + .SetDefault(1) + .GreaterThan(0); + AddComment(R"DOC( +Fusion Sequence Conv and ElementwiseAdd Operator. +)DOC"); +} + +template +class FusionSeqConvEltAddReluKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + using DeviceContext = paddle::platform::CPUDeviceContext; + auto* x = ctx.Input("X"); + auto* w = ctx.Input("Filter"); + auto* b = ctx.Input("Bias"); + auto* y = ctx.Output("Out"); + auto* col = ctx.Output("ColMat"); + + auto x_lod = x->lod(); + auto x_dims = x->dims(); + auto w_dims = w->dims(); + PADDLE_ENFORCE_EQ(b->numel(), w_dims[1], + "bias size should be equal to output feature size."); + PADDLE_ENFORCE_EQ(x_lod.size(), 1UL, + "Only support one level sequence now."); + + const T* x_data = x->data(); + const T* w_data = w->data(); + const T* b_data = b->data(); + T* y_data = y->mutable_data(ctx.GetPlace()); + T* col_data = col->mutable_data(ctx.GetPlace()); + + int context_start = ctx.Attr("contextStart"); + int context_length = ctx.Attr("contextLength"); + int up_pad = std::max(0, -context_start); + int down_pad = std::max(0, context_start + context_length - 1); + // im2col + int src_mat_w = static_cast(x_dims[1]); + int src_mat_w_sz = src_mat_w * sizeof(T); + int col_mat_w = static_cast(w_dims[0]); + int col_mat_w_sz = col_mat_w * sizeof(T); + for (int i = 0; i < static_cast(x_lod[0].size()) - 1; ++i) { + int st = x_lod[0][i]; + int ed = x_lod[0][i + 1]; + const T* src_data = x_data + st * src_mat_w; + T* dst_data = col_data + st * col_mat_w; + int seq_len = ed - st; + if (seq_len > up_pad + down_pad) { + // zero all up_pad and fill data + std::memset(dst_data, 0, up_pad * col_mat_w_sz); + dst_data = dst_data + up_pad * src_mat_w; + int copy_size = col_mat_w_sz - up_pad * src_mat_w_sz; + for (int j = 0; j < up_pad; ++j) { + // blas.VCOPY? + std::memcpy(dst_data, src_data, copy_size); + dst_data += (col_mat_w - src_mat_w); + copy_size += src_mat_w_sz; + } + // fill data + for (int j = 0; j < seq_len - up_pad - down_pad; ++j) { + std::memcpy(dst_data, src_data, copy_size); + dst_data += col_mat_w; + src_data += src_mat_w; + } + // zero all down_pad and fill data + std::memset(dst_data, 0, down_pad * col_mat_w_sz); + copy_size -= src_mat_w_sz; + for (int j = 0; j < down_pad; ++j) { + std::memcpy(dst_data, src_data, copy_size); + dst_data += col_mat_w; + src_data += src_mat_w; + copy_size -= src_mat_w_sz; + } + } else { + PADDLE_ENFORCE_GE(context_length, up_pad + down_pad + 1); + std::memset(dst_data, 0, seq_len * col_mat_w_sz); + dst_data = dst_data + up_pad * src_mat_w; + int zero_sz = up_pad * src_mat_w_sz; + int cur_src_sz = seq_len * src_mat_w_sz; + for (int j = 0; j < std::min(up_pad, seq_len); ++j) { + int copy_size = std::min(cur_src_sz, col_mat_w_sz - zero_sz); + std::memcpy(dst_data, src_data, copy_size); + dst_data += (col_mat_w - src_mat_w); + zero_sz -= src_mat_w_sz; + } + // from bottom + dst_data = col_data + ed * col_mat_w; + src_data = x_data + st * src_mat_w; + zero_sz = down_pad * src_mat_w_sz; + for (int j = 1; j <= std::min(down_pad, seq_len); ++j) { + int copy_size = std::min(cur_src_sz, col_mat_w_sz - zero_sz); + std::memcpy(dst_data - (zero_sz + copy_size) / sizeof(T), + src_data + std::max(seq_len - j - up_pad, 0) * src_mat_w, + copy_size); + dst_data -= col_mat_w; + zero_sz -= src_mat_w_sz; + } + } + } + auto& dev_ctx = ctx.template device_context(); + auto blas = math::GetBlas(dev_ctx); + math::FCCompute(blas, x_dims[0], w_dims[1], w_dims[0], + col_data, w_data, y_data, b_data, true); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(fusion_seqconv_eltadd_relu, ops::FusionSeqConvEltAddReluOp, + ops::FusionSeqConvEltAddReluOpMaker, + paddle::framework::DefaultGradOpDescMaker); + +REGISTER_OP_CPU_KERNEL(fusion_seqconv_eltadd_relu, + ops::FusionSeqConvEltAddReluKernel, + ops::FusionSeqConvEltAddReluKernel); diff --git a/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h b/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h new file mode 100644 index 0000000000000000000000000000000000000000..028d79dc2a1ee8d789fe4b8724b320442041a71b --- /dev/null +++ b/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h @@ -0,0 +1,42 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using Tensor = framework::Tensor; + +class FusionSeqConvEltAddReluOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override; + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override; +}; + +class FusionSeqConvEltAddReluOpMaker + : public framework::OpProtoAndCheckerMaker { + public: + void Make() override; +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc b/paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc index 0cd3d3887cf5167c779a8b20442fdb458cd7eab4..8d2f055d53a0c5bbef624ff3b01b01724d0b3a21 100644 --- a/paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc +++ b/paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc @@ -136,9 +136,9 @@ class FusionSeqExpandConcatFCOpKernel : public framework::OpKernel { // since infershape can not get lod info PADDLE_ENFORCE_EQ(ref_lod.size(), 1UL, "Only support input lod size is 1."); PADDLE_ENFORCE_EQ(in1_lod.size(), 1UL, "Only support input lod size is 1."); - PADDLE_ENFORCE_EQ(in1_lod[0].size() - 1, N, + PADDLE_ENFORCE_EQ(static_cast(in1_lod[0].size() - 1), N, "Batch size of all inputs should be equal."); - PADDLE_ENFORCE_EQ(in1_lod[0][N], N, + PADDLE_ENFORCE_EQ(static_cast(in1_lod[0][N]), N, "Seq_length of other inputs should be 1."); PADDLE_ENFORCE_EQ(in1_dims[0], N, "input height should be batch size."); for (size_t i = 2; i < ins.size(); ++i) { diff --git a/paddle/fluid/operators/isfinite_op.cc b/paddle/fluid/operators/isfinite_op.cc index 248c7793560db99c0af06421bf74808422016061..7b42efd623b31a703bf51d2d157130b3120b42a4 100644 --- a/paddle/fluid/operators/isfinite_op.cc +++ b/paddle/fluid/operators/isfinite_op.cc @@ -60,7 +60,7 @@ class OverflowOpMaker : public framework::OpProtoAndCheckerMaker { "(Tensor) 1-dim tensor, contains a bool scalar. The output " "tensor of overflow operator."); AddComment(string::Sprintf(R"DOC( -Overflow operator. +Overflow %s operator. $$Out = any(X)$$ @@ -69,6 +69,8 @@ Out = Inf if any X contains Inf, Out = Nan if any X contains Nan, Out = 0 if no Inf/Nan detected. If X contains both Inf/Nan, it will return the first indicator it meeted. + +%s )DOC", GetName(), GetComments())); } diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index dc008d16971bc762b401ddece56f9ec56f7a47d6..26f09c46c2224a4a46d302dff4b2ec594f0be103 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -66,7 +66,7 @@ static void ParallelExecuteBlocks( << "pointer: " << prepared[run_block].get(); executor->RunPreparedContext(prepared[run_block].get(), scope); } catch (const std::exception &e) { - LOG(ERROR) << "run sub program error " << e.what(); + LOG(FATAL) << "run sub program:" << idx << " error " << e.what(); } })); } diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt index 0f7ce471f0f40edb3b34c4fdad9c2dbdc278505e..878c5d59c571cd9e387fc08b061bfd2727db908f 100644 --- a/paddle/fluid/operators/math/CMakeLists.txt +++ b/paddle/fluid/operators/math/CMakeLists.txt @@ -3,8 +3,8 @@ add_subdirectory(detail) endif(NOT WIN32) function(math_library TARGET) - # math_library is a function to create math library. - # The interface is the same as cc_library. + # math_library is a function to create math library. + # The interface is the same as cc_library. # But it handle split GPU/CPU code and link some common library. set(cc_srcs) set(cu_srcs) @@ -45,15 +45,13 @@ math_library(im2col) if (NOT WIN32) # windows do not support avx functions yet. math_library(gru_compute DEPS activation_functions math_function) math_library(lstm_compute DEPS activation_functions) -# TODO(TJ): ugly workaround, clean me -cc_library(cpu_lstm_compute SRCS cpu_lstm_compute.cc DEPS activation_functions cblas cpu_info) endif (NOT WIN32) cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context) math_library(math_function DEPS blas) math_library(maxouting) math_library(pooling) -math_library(selected_rows_functor DEPS selected_rows math_function) +math_library(selected_rows_functor DEPS selected_rows math_function blas) math_library(sequence2batch) math_library(sequence_padding) math_library(sequence_pooling DEPS math_function) @@ -76,3 +74,7 @@ if(WITH_GPU) endif() cc_test(concat_test SRCS concat_test.cc DEPS concat) cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info) +cc_library(jit_kernel + SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_lstm.cc + DEPS cpu_info cblas) +cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel) diff --git a/paddle/fluid/operators/math/algorithm.h b/paddle/fluid/operators/math/algorithm.h new file mode 100644 index 0000000000000000000000000000000000000000..262469beea7449eb5820b86de1ac4f790a833e79 --- /dev/null +++ b/paddle/fluid/operators/math/algorithm.h @@ -0,0 +1,44 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include // for int64_t +#include + +#include "paddle/fluid/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { + +template +HOSTDEVICE inline int64_t BinarySearch(const T *x, int64_t num, const T &val) { + int64_t beg = 0, end = num - 1; + while (beg <= end) { + auto mid = ((beg + end) >> 1); + if (x[mid] == val) + return mid; + else if (x[mid] < val) + beg = mid + 1; + else + end = mid - 1; + } + return -1; +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/cpu_lstm_compute.cc b/paddle/fluid/operators/math/cpu_lstm_compute.cc deleted file mode 100644 index e96d1879331974e0873e13f171414bcfa8c45953..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/math/cpu_lstm_compute.cc +++ /dev/null @@ -1,43 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at -http://www.apache.org/licenses/LICENSE-2.0 -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/fluid/operators/math/cpu_lstm_compute.h" - -namespace paddle { -namespace operators { -namespace math { -#ifdef __AVX__ -template <> -void lstm_compute_ctht(float* gates, const float* ct_1, float* ct, - float* ht) { - namespace act = detail::forward::avx; - // gates: W_ch, W_ih, W_fh, W_oh - __m256 c, i, f, o; - c = _mm256_loadu_ps(gates); - i = _mm256_loadu_ps(gates + 8); - f = _mm256_loadu_ps(gates + 16); - o = _mm256_loadu_ps(gates + 24); - - /* C_t = C_t-1 * fgated + cand_gated * igated*/ - c = _mm256_mul_ps(act::Tanh(c), act::Sigmoid(i)); - i = _mm256_loadu_ps(ct_1); - f = _mm256_mul_ps(i, act::Sigmoid(f)); - f = _mm256_add_ps(c, f); - _mm256_storeu_ps(ct, f); - - /* H_t = act_cell(C_t) * ogated */ - o = _mm256_mul_ps(act::Tanh(f), act::Sigmoid(o)); - _mm256_storeu_ps(ht, o); -} -#endif -} // namespace math -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/math/cpu_lstm_compute.h b/paddle/fluid/operators/math/cpu_lstm_compute.h deleted file mode 100644 index 169a9e4b47f54851ad436428416eca879b78e186..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/math/cpu_lstm_compute.h +++ /dev/null @@ -1,64 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at -http://www.apache.org/licenses/LICENSE-2.0 -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include "paddle/fluid/operators/math/cpu_vec.h" -#include "paddle/fluid/platform/cpu_info.h" -#ifdef __AVX__ -#include -#endif - -namespace paddle { -namespace operators { -namespace math { - -// TODO(TJ): ugly workaround, clean me -template -void lstm_compute_ctht(T* gates, const T* ct_1, T* ct, T* ht) { - // gates: W_ch, W_ih, W_fh, W_oh - vec_sigmoid(24, gates + 8, gates + 8); - vec_tanh(8, gates, gates); - const T *i = gates + 8, *f = gates + 16, *o = gates + 24; - const T min = SIGMOID_THRESHOLD_MIN; - const T max = SIGMOID_THRESHOLD_MAX; - for (int d = 0; d < 8; ++d) { - // C_t = C_t-1 * fgated + cand_gated * igated - ct[d] = ct_1[d] * f[d] + gates[d] * i[d]; - // H_t = act_cell(C_t) * ogated - T tmp = ct[d] * 2; - tmp = static_cast(0) - ((tmp < min) ? min : ((tmp > max) ? max : tmp)); - vec_exp(1, &tmp, &tmp); - tmp = static_cast(2) / (static_cast(1) + tmp) - static_cast(1); - ht[d] = tmp * o[d]; - } -} - -#ifdef __AVX__ -namespace detail { -namespace forward { -namespace avx { -__m256 Sigmoid(const __m256 a); -__m256 Tanh(const __m256 a); - -} // namespace avx -} // namespace forward -} // namespace detail - -template <> -void lstm_compute_ctht(float* gates, const float* ct_1, float* ct, - float* ht); - -#endif - -} // namespace math -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/math/cpu_vec.h b/paddle/fluid/operators/math/cpu_vec.h index 6a059968b79189458349e466079cc7a663a8e5ff..0aed253c80fc28560716cbcfa70f74ef9c84f9b6 100644 --- a/paddle/fluid/operators/math/cpu_vec.h +++ b/paddle/fluid/operators/math/cpu_vec.h @@ -125,10 +125,8 @@ inline void vec_scal(const int n, const float a, } template <> -inline void vec_scal(const int n, - const float a, - const float* x, - float* y) { +inline void vec_scal(const int n, const float a, + const float* x, float* y) { // TODO(TJ): enable me vec_scal(n, a, x, y); } @@ -181,10 +179,10 @@ inline void vec_bias_sub(const int n, const float a, } template <> -inline void vec_bias_sub(const int n, - const float a, - const float* x, - float* y) { +inline void vec_bias_sub(const int n, + const float a, + const float* x, + float* y) { // TODO(TJ): enable me vec_bias_sub(n, a, x, y); } @@ -242,7 +240,7 @@ inline void vec_cross(const int n, const float* x, } template <> -inline void vec_cross( +inline void vec_cross( const int n, const float* x, const float* y, const float* z, float* out) { // TODO(TJ): enable me vec_cross(n, x, y, z, out); @@ -296,10 +294,10 @@ inline void vec_add_bias(const int n, const float a, } template <> -inline void vec_add_bias(const int n, - const float a, - const float* x, - float* y) { +inline void vec_add_bias(const int n, + const float a, + const float* x, + float* y) { // TODO(TJ): enable me vec_add_bias(n, a, x, y); } @@ -390,9 +388,9 @@ inline void vec_sigmoid(const int n, const float* x, } template <> -inline void vec_sigmoid(const int n, - const float* x, - float* y) { +inline void vec_sigmoid(const int n, + const float* x, + float* y) { // TODO(TJ): enable me vec_sigmoid(n, x, y); } @@ -454,9 +452,8 @@ inline void vec_relu(const int n, const float* x, } template <> -inline void vec_relu(const int n, - const float* x, - float* y) { +inline void vec_relu(const int n, const float* x, + float* y) { // TODO(TJ): enable me vec_relu(n, x, y); } diff --git a/paddle/fluid/operators/math/cpu_vec_test.cc b/paddle/fluid/operators/math/cpu_vec_test.cc index 3ce66f49ed8354c49e8af26ca6eb48fef654a40b..cd40f1b2f984126663a5711efac24fdf6d680b32 100644 --- a/paddle/fluid/operators/math/cpu_vec_test.cc +++ b/paddle/fluid/operators/math/cpu_vec_test.cc @@ -110,7 +110,7 @@ TEST(CpuVecTest, sigmoid) { TestAndBench(sz, vec_sigmoid, ref_sigmoid); TestAndBench(sz, vec_sigmoid, ref_sigmoid); TestAndBench(sz, vec_sigmoid, ref_sigmoid); - TestAndBench(sz, vec_sigmoid, + TestAndBench(sz, vec_sigmoid, ref_sigmoid); } TestAndBench(30, vec_sigmoid, ref_sigmoid); @@ -123,8 +123,7 @@ TEST(CpuVecTest, tanh) { TestAndBench(sz, vec_tanh, ref_tanh); TestAndBench(sz, vec_tanh, ref_tanh); TestAndBench(sz, vec_tanh, ref_tanh); - TestAndBench(sz, vec_tanh, - ref_tanh); + TestAndBench(sz, vec_tanh, ref_tanh); } TestAndBench(30, vec_tanh, ref_tanh); } @@ -136,8 +135,7 @@ TEST(CpuVecTest, relu) { TestAndBench(sz, vec_relu, ref_relu); TestAndBench(sz, vec_relu, ref_relu); TestAndBench(sz, vec_relu, ref_relu); - TestAndBench(sz, vec_relu, - ref_relu); + TestAndBench(sz, vec_relu, ref_relu); } TestAndBench(30, vec_relu, ref_relu); } @@ -170,7 +168,7 @@ TEST(CpuVecTest, inplace_sigmoid) { TestInplace(sz, vec_sigmoid, ref_sigmoid); TestInplace(sz, vec_sigmoid, ref_sigmoid); TestInplace(sz, vec_sigmoid, ref_sigmoid); - TestInplace(sz, vec_sigmoid, + TestInplace(sz, vec_sigmoid, ref_sigmoid); } TestInplace(30, vec_sigmoid, ref_sigmoid); @@ -183,8 +181,7 @@ TEST(CpuVecTest, inplace_tanh) { TestInplace(sz, vec_tanh, ref_tanh); TestInplace(sz, vec_tanh, ref_tanh); TestInplace(sz, vec_tanh, ref_tanh); - TestInplace(sz, vec_tanh, - ref_tanh); + TestInplace(sz, vec_tanh, ref_tanh); } TestInplace(30, vec_tanh, ref_tanh); } @@ -196,8 +193,7 @@ TEST(CpuVecTest, inplace_relu) { TestInplace(sz, vec_relu, ref_relu); TestInplace(sz, vec_relu, ref_relu); TestInplace(sz, vec_relu, ref_relu); - TestInplace(sz, vec_relu, - ref_relu); + TestInplace(sz, vec_relu, ref_relu); } TestInplace(30, vec_relu, ref_relu); } diff --git a/paddle/fluid/operators/math/depthwise_conv.cu b/paddle/fluid/operators/math/depthwise_conv.cu index 3be389912307f7aac6dda6d1018943eb8f08696d..66d37c3bf31ffa420cc527cb576dcdc5505a0960 100644 --- a/paddle/fluid/operators/math/depthwise_conv.cu +++ b/paddle/fluid/operators/math/depthwise_conv.cu @@ -46,17 +46,20 @@ __forceinline__ __device__ unsigned warp_id() { return ret; } +#define ARG_DEFINE_KernelDepthwiseConv \ + const T *const input_data, const T *const filter_data, const int batch_size, \ + const int output_channels, const int output_height, \ + const int output_width, const int input_channels, \ + const int input_height, const int input_width, \ + const int filter_multiplier, const int filter_height, \ + const int filter_width, const int stride_height, const int stride_width, \ + const int padding_height, const int padding_width, \ + const int dilate_height, const int dilate_width, T *const output_data + // A Cuda kernel to compute the depthwise convolution forward pass // in NCHW format. template -__device__ __inline__ void KernelDepthwiseConv( - const T* const input_data, const T* const filter_data, const int batch_size, - const int output_channels, const int output_height, const int output_width, - const int input_channels, const int input_height, const int input_width, - const int filter_multiplier, const int filter_height, - const int filter_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, const int dilate_height, - const int dilate_width, T* const output_data) { +__device__ __inline__ void KernelDepthwiseConv(ARG_DEFINE_KernelDepthwiseConv) { for (int w_out = threadIdx.x; w_out < output_width; w_out += blockDim.x) { for (int h_out = threadIdx.y; h_out < output_height; h_out += blockDim.y) { const int batch = blockIdx.y; @@ -97,42 +100,105 @@ __device__ __inline__ void KernelDepthwiseConv( } } -template -__global__ void KernelDepthwiseConvSp( - const T* const input_data, const T* const filter_data, const int batch_size, - const int output_channels, const int output_height, const int output_width, - const int input_channels, const int input_height, const int input_width, - const int filter_multiplier, const int filter_height, - const int filter_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, const int dilate_height, - const int dilate_width, T* const output_data) { - if (c_filter_multiplier == 0) - KernelDepthwiseConv(input_data, filter_data, batch_size, output_channels, - output_height, output_width, input_channels, - input_height, input_width, filter_multiplier, - filter_height, filter_width, stride_height, - stride_width, padding_height, padding_width, - dilate_height, dilate_width, output_data); +template +__device__ __inline__ void KernelDepthwiseConvCFilter( + ARG_DEFINE_KernelDepthwiseConv) { + const int kWeghtSize = c_filter * c_filter; + T r_weight[kWeghtSize]; + const int batch = blockIdx.y; + const int c_out = blockIdx.x; + const T* weight = filter_data + c_out * c_filter * c_filter; + for (int i = 0; i < c_filter * c_filter; i++) r_weight[i] = weight[i]; - else - KernelDepthwiseConv(input_data, filter_data, batch_size, output_channels, - output_height, output_width, input_channels, - input_height, input_width, c_filter_multiplier, - filter_height, filter_height, c_stride, c_stride, - padding_height, padding_width, dilate_height, - dilate_width, output_data); + for (int w_out = threadIdx.x; w_out < output_width; w_out += blockDim.x) { + for (int h_out = threadIdx.y; h_out < output_height; h_out += blockDim.y) { + const int batch = blockIdx.y; + const int c_out = blockIdx.x; + + const int c_in = c_out / filter_multiplier; + T value = 0; + const int h_in_start = -padding_height + h_out * stride_height; + const int w_in_start = -padding_width + w_out * stride_width; + const int h_in_end = h_in_start + c_filter * dilate_height; + const int w_in_end = w_in_start + c_filter * dilate_width; + + const int in_offset = + ((batch * input_channels + c_in) * input_height) * input_width; + + const int h_end = h_in_end < input_height ? h_in_end : input_height; + const int w_end = w_in_end < input_width ? w_in_end : input_width; + const int h_start = h_in_start > 0 ? h_in_start : 0; + const int w_start = w_in_start > 0 ? w_in_start : 0; + + for (int h_in = h_in_start, h_f = 0; h_f < c_filter; + h_in += dilate_height, h_f++) { + for (int w_in = w_in_start, w_f = 0; w_f < c_filter; + w_in += dilate_width, w_f++) { + if (h_in >= 0 && h_in < input_height && w_in >= 0 && + w_in < input_width) { + const int offset = in_offset + h_in * input_width + w_in; + value += r_weight[h_f * c_filter + w_f] * input_data[offset]; + } + } + } + int index = + ((batch * gridDim.x + c_out) * output_height + h_out) * output_width + + w_out; + output_data[index] = value; + } + } +} + +template +__global__ void KernelDepthwiseConvSp(ARG_DEFINE_KernelDepthwiseConv) { + if (c_filter_multiplier == 0) { + if (c_filter == -1) + KernelDepthwiseConv( + input_data, filter_data, batch_size, output_channels, output_height, + output_width, input_channels, input_height, input_width, + filter_multiplier, filter_height, filter_width, stride_height, + stride_width, padding_height, padding_width, dilate_height, + dilate_width, output_data); + else + KernelDepthwiseConvCFilter( + input_data, filter_data, batch_size, output_channels, output_height, + output_width, input_channels, input_height, input_width, + filter_multiplier, filter_height, filter_width, stride_height, + stride_width, padding_height, padding_width, dilate_height, + dilate_width, output_data); + } else { + if (c_filter == -1) + KernelDepthwiseConv(input_data, filter_data, batch_size, + output_channels, output_height, output_width, + input_channels, input_height, input_width, + c_filter_multiplier, filter_height, filter_height, + c_stride, c_stride, padding_height, padding_width, + dilate_height, dilate_width, output_data); + else + KernelDepthwiseConvCFilter( + input_data, filter_data, batch_size, output_channels, output_height, + output_width, input_channels, input_height, input_width, + c_filter_multiplier, filter_height, filter_height, c_stride, c_stride, + padding_height, padding_width, dilate_height, dilate_width, + output_data); + } } // CUDA kernel to compute the depthwise convolution backprop w.r.t input. +#define ARG_DEFINE_KernelDepthwiseConvInputGrad \ + const T *const output_grad_data, const T *const filter_data, \ + const int batch_size, const int output_channels, \ + const int output_height, const int output_width, \ + const int input_channels, const int input_height, const int input_width, \ + const int filter_multiplier, const int filter_height, \ + const int filter_width, const int stride_height, const int stride_width, \ + const int padding_height, const int padding_width, \ + const int dilate_height, const int dilate_width, \ + T *const input_grad_data + template __device__ __inline__ void KernelDepthwiseConvInputGrad( - const T* const output_grad_data, const T* const filter_data, - const int batch_size, const int output_channels, const int output_height, - const int output_width, const int input_channels, const int input_height, - const int input_width, const int filter_multiplier, const int filter_height, - const int filter_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, const int dilate_height, - const int dilate_width, T* const input_grad_data) { + ARG_DEFINE_KernelDepthwiseConvInputGrad) { for (int w_in = threadIdx.x; w_in < input_width; w_in += blockDim.x) { for (int h_in = threadIdx.y; h_in < input_height; h_in += blockDim.y) { const int batch = blockIdx.y; @@ -184,15 +250,67 @@ __device__ __inline__ void KernelDepthwiseConvInputGrad( } } -template +template +__device__ __inline__ void KernelDepthwiseConvInputGradCFilter( + ARG_DEFINE_KernelDepthwiseConvInputGrad) { + const int kWeghtSize = c_filter * c_filter * c_filter_multiplier + 1; + T r_weight[kWeghtSize]; + const int batch = blockIdx.y; + const int c_in = blockIdx.x; + + for (int c_i = 0; c_i < filter_multiplier; c_i++) { + int c_out = c_in * filter_multiplier + c_i; + const T* weight = filter_data + c_out * c_filter * c_filter; + for (int i = 0; i < c_filter * c_filter; i++) + r_weight[i + c_i * c_filter * c_filter] = + weight[c_filter * c_filter - i - 1]; + } + + for (int w_in = threadIdx.x; w_in < input_width; w_in += blockDim.x) { + for (int h_in = threadIdx.y; h_in < input_height; h_in += blockDim.y) { + const int batch = blockIdx.y; + const int c_in = blockIdx.x; + + int h_out_start = h_in - (c_filter - 1) * dilate_height + padding_height; + + int w_out_start = w_in - (c_filter - 1) * dilate_width + padding_width; + + T value = 0; + + for (int c_i = 0; c_i < filter_multiplier; c_i++) { + int c_out = c_in * filter_multiplier + c_i; + for (int h_out = h_out_start, h_f = 0; h_f < c_filter; + h_out += dilate_height, h_f++) { + for (int w_out = w_out_start, w_f = 0; w_f < c_filter; + w_out += dilate_width, w_f++) { + int s_h_out = h_out / stride_height; + int s_w_out = w_out / stride_width; + if (h_out % stride_height == 0 && w_out % stride_width == 0 && + s_h_out >= 0 && s_h_out < output_height && s_w_out >= 0 && + s_w_out < output_width) { + const int output_grad_offset = + ((batch * output_channels + c_out) * output_height + + s_h_out) * + output_width + + s_w_out; + value += + output_grad_data[output_grad_offset] * + r_weight[h_f * c_filter + w_f + c_i * c_filter * c_filter]; + } + } + } + } + int index = + ((batch * gridDim.x + c_in) * input_height + h_in) * input_width + + w_in; + input_grad_data[index] = value; + } + } +} + +template __global__ void KernelDepthwiseConvInputGradSp( - const T* const output_grad_data, const T* const filter_data, - const int batch_size, const int output_channels, const int output_height, - const int output_width, const int input_channels, const int input_height, - const int input_width, const int filter_multiplier, const int filter_height, - const int filter_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, const int dilate_height, - const int dilate_width, T* const input_grad_data) { + ARG_DEFINE_KernelDepthwiseConvInputGrad) { if (c_filter_multiplier == 0) KernelDepthwiseConvInputGrad( output_grad_data, filter_data, batch_size, output_channels, @@ -200,13 +318,20 @@ __global__ void KernelDepthwiseConvInputGradSp( filter_multiplier, filter_height, filter_width, stride_height, stride_width, padding_height, padding_width, dilate_height, dilate_width, input_grad_data); - else + else if (c_filter == -1) KernelDepthwiseConvInputGrad( output_grad_data, filter_data, batch_size, output_channels, output_height, output_width, input_channels, input_height, input_width, c_filter_multiplier, filter_height, filter_width, c_stride, c_stride, padding_height, padding_width, dilate_height, dilate_width, input_grad_data); + else + KernelDepthwiseConvInputGradCFilter( + output_grad_data, filter_data, batch_size, output_channels, + output_height, output_width, input_channels, input_height, input_width, + c_filter_multiplier, filter_height, filter_width, c_stride, c_stride, + padding_height, padding_width, dilate_height, dilate_width, + input_grad_data); } // Cuda kernel to compute the depthwise convolution backprop w.r.t. filter. @@ -325,12 +450,14 @@ class DepthwiseConvFunctor { dim3 threads(std::min(output_width, thread), blocks, 1); dim3 grid(output_channels, batch_size, 1); int filter_multiplier = output_channels / input_channels; -#define check_case(c_filter_multiplier, c_stride) \ +#define check_case(c_filter_multiplier, c_stride, c_filter) \ if (c_filter_multiplier == 0 || \ filter_multiplier == c_filter_multiplier && \ - stride_height == stride_width && stride_height == c_stride) { \ - KernelDepthwiseConvSp<<>>( \ + stride_height == stride_width && stride_height == c_stride && \ + (ksize_height == ksize_width && ksize_height == c_filter || \ + c_filter == -1)) { \ + KernelDepthwiseConvSp<<>>( \ input_data, filter_data, batch_size, output_channels, output_height, \ output_width, input_channels, input_height, input_width, \ filter_multiplier, ksize_height, ksize_width, stride_height, \ @@ -338,11 +465,17 @@ class DepthwiseConvFunctor { dilate_width, output_data); \ return; \ } - check_case(1, 1); - check_case(1, 2); - // NOTE(liangdun): 0,0 for other case - // add other case if needed, e.g. check_case(2^n,1) - check_case(0, 0); + check_case(1, 1, 3); + check_case(1, 1, 5); + check_case(1, 1, -1); + check_case(1, 2, 3); + check_case(1, 2, 5); + check_case(1, 2, -1); + check_case(0, 0, 3); + check_case(0, 0, 5); + check_case(0, 0, -1); +// NOTE(liangdun): 0,0 for other case +// add other case if needed, e.g. check_case(2^n,1) #undef check_case } }; @@ -384,13 +517,15 @@ class DepthwiseConvInputGradFunctor { dim3 grid(input_channels, batch_size, 1); int filter_multiplier = output_channels / input_channels; -#define check_case(c_filter_multiplier, c_stride) \ +#define check_case(c_filter_multiplier, c_stride, c_filter) \ if (c_filter_multiplier == 0 || \ filter_multiplier == c_filter_multiplier && \ - stride_height == stride_width && stride_height == c_stride) { \ + stride_height == stride_width && stride_height == c_stride && \ + (ksize_height == ksize_width && ksize_height == c_filter || \ + c_filter == -1)) { \ KernelDepthwiseConvInputGradSp< \ - T, c_filter_multiplier, \ - c_stride><<>>( \ + T, c_filter_multiplier, c_stride, \ + c_filter><<>>( \ output_grad_data, filter_data, batch_size, output_channels, \ output_height, output_width, input_channels, input_height, \ input_width, filter_multiplier, ksize_height, ksize_width, \ @@ -398,11 +533,21 @@ class DepthwiseConvInputGradFunctor { dilate_height, dilate_width, input_grad_data); \ return; \ } - check_case(1, 1); - check_case(1, 2); - // NOTE(liangdun): 0,0 for other case - // add other case if needed, e.g. check_case(2^n,1) - check_case(0, 0); + check_case(1, 1, 3); + check_case(1, 1, 5); + check_case(1, 1, -1); + check_case(1, 2, 3); + check_case(1, 2, 5); + check_case(1, 2, -1); + check_case(2, 1, 3); + check_case(2, 1, 5); + check_case(2, 1, -1); + check_case(2, 2, 3); + check_case(2, 2, 5); + check_case(2, 2, -1); + check_case(0, 0, -1); +// NOTE(liangdun): 0,0 for other case +// add other case if needed, e.g. check_case(2^n,1) #undef check_case } }; diff --git a/paddle/fluid/operators/math/fc_compute.h b/paddle/fluid/operators/math/fc_compute.h index 1f5a49c0ab5a10b0d7dc1febd258ce76c467cb1c..87220d4019fc9337fb8355172ca4f1372cfd4558 100644 --- a/paddle/fluid/operators/math/fc_compute.h +++ b/paddle/fluid/operators/math/fc_compute.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/jit_kernel.h" DECLARE_int32(paddle_num_threads); @@ -30,20 +31,25 @@ inline void FCCompute(const BlasT& blas, const int M, if (B == NULL) { return; } + if (relu) { + const auto& vaddrelu = jitkernel::KernelPool::Instance() + .template Get>(N); + for (int i = 0; i < M; i++) { + T* dst = Y + i * N; + vaddrelu->Compute(B, dst, dst); + } + } else { + const auto& vadd = jitkernel::KernelPool::Instance() + .template Get>(N); #ifdef PADDLE_WITH_MKLML #pragma omp parallel for if (FLAGS_paddle_num_threads > 1) #endif - for (int i = 0; i < M; i++) { - blas.AXPY(N, static_cast(1), B, Y + i * N); + for (int i = 0; i < M; i++) { + T* dst = Y + i * N; + vadd->Compute(B, dst, dst); + } } - - if (!relu) { - return; - } - - // TODO(TJ): fuse relu - LOG(FATAL) << "Not implemented!"; } } // namespace math diff --git a/paddle/fluid/operators/math/jit_kernel.cc b/paddle/fluid/operators/math/jit_kernel.cc new file mode 100644 index 0000000000000000000000000000000000000000..68b708b345334bc63b5e2e88c308d20ca6378e6b --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel.cc @@ -0,0 +1,41 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/jit_kernel.h" +#include +#include + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +namespace jit = platform::jit; + +KernelPool& KernelPool::Instance() { + static thread_local KernelPool g_jit_kernels; + return g_jit_kernels; +} + +std::shared_ptr KernelPool::Get(const std::string& key) const { + if (kers_.find(key) == kers_.end()) { + return nullptr; + } + return kers_.at(key); +} + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel.h b/paddle/fluid/operators/math/jit_kernel.h new file mode 100644 index 0000000000000000000000000000000000000000..e91e4e8e5adfdfff8163efe7fc1451bc602504e0 --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel.h @@ -0,0 +1,148 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include // for shared_ptr +#include +#include +#include "paddle/fluid/platform/cpu_info.h" +#include "paddle/fluid/platform/macros.h" + +// Note: Only support on CPU yet. +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +#define SIGMOID_THRESHOLD_MIN -40.0 +#define SIGMOID_THRESHOLD_MAX 13.0 +#define EXP_MAX_INPUT 40.0 +#define AVX_FLOAT_BLOCK 8 +#define AVX2_FLOAT_BLOCK 8 +#define AVX512_FLOAT_BLOCK 16 + +typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block; + +class Kernel { + public: + Kernel() = default; + virtual ~Kernel() = default; + int num_{0}; + int end_{0}; + int rest_{0}; + DISABLE_COPY_AND_ASSIGN(Kernel); +}; + +class KernelPool { + public: + static KernelPool &Instance(); + + template + std::shared_ptr Get(ARGS... args); + + std::shared_ptr Get(const std::string &key) const; + + private: + KernelPool() = default; + std::unordered_map> kers_; + + DISABLE_COPY_AND_ASSIGN(KernelPool); +}; + +template +class VMulKernel : public Kernel { + public: + virtual void Compute(const T *x, const T *y, T *z) const = 0; +}; + +template +class VAddKernel : public Kernel { + public: + virtual void Compute(const T *x, const T *y, T *z) const = 0; +}; + +template +class VScalKernel : public Kernel { + public: + virtual void Compute(const T a, const T *x, T *y) const = 0; + virtual void Compute(const T a, T *x) const = 0; +}; + +template +class VAddBiasKernel : public Kernel { + public: + virtual void Compute(const T a, const T *x, T *y) const = 0; +}; + +template +class VAddReluKernel : public Kernel { + public: + virtual void Compute(const T *x, const T *y, T *z) const = 0; +}; + +template +class VActKernel : public Kernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VReluKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VIdentityKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VExpKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VSigmoidKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VTanhKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class LSTMKernel : public Kernel { + public: + virtual void ComputeCtHt(T *gates, const T *ct_1, T *ct, T *ht, + /* below only used in peephole*/ + const T *wp_data = nullptr, + T *checked = nullptr) const = 0; + + // compute c1 and h1 without c0 or h0 + virtual void ComputeC1H1(T *gates, T *ct, T *ht, + /* below only used in peephole*/ + const T *wp_data = nullptr) const = 0; +}; + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_blas.cc b/paddle/fluid/operators/math/jit_kernel_blas.cc new file mode 100644 index 0000000000000000000000000000000000000000..c88b17b012d1b9cd59220f6b37cb2ecb8a1551a4 --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_blas.cc @@ -0,0 +1,479 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/jit_kernel.h" +#include +#include "paddle/fluid/operators/math/jit_kernel_macro.h" +#ifdef PADDLE_WITH_MKLML +#include "paddle/fluid/platform/dynload/mklml.h" +#endif + +#ifdef __AVX__ +#include +#endif + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +namespace jit = platform::jit; + +/* VMUL JitKernel */ +template +class VMulKernelImpl : public VMulKernel { + public: + explicit VMulKernelImpl(int d) : VMulKernel() { this->num_ = d; } + void Compute(const T* x, const T* y, T* z) const override { + for (int i = 0; i < this->num_; ++i) { + z[i] = x[i] * y[i]; + } + } +}; + +#ifdef PADDLE_WITH_MKLML +#define MKL_FLOAT(isa, block) \ + template <> \ + void VMulKernelImpl::Compute( \ + const float* x, const float* y, float* z) const { \ + platform::dynload::vsMul(this->num_, x, y, z); \ + } + +#define MKL_DOUBLE(isa, block) \ + template <> \ + void VMulKernelImpl::Compute( \ + const double* x, const double* y, double* z) const { \ + platform::dynload::vdMul(this->num_, x, y, z); \ + } + +FOR_EACH_ISA(MKL_FLOAT, kGT16); +FOR_EACH_ISA_BLOCK(MKL_DOUBLE); +#endif + +#define INTRI8_FLOAT(isa) \ + template <> \ + void VMulKernelImpl::Compute( \ + const float* x, const float* y, float* z) const { \ + __m256 tmpx, tmpy; \ + tmpx = _mm256_loadu_ps(x); \ + tmpy = _mm256_loadu_ps(y); \ + tmpx = _mm256_mul_ps(tmpx, tmpy); \ + _mm256_storeu_ps(z, tmpx); \ + } + +// avx > for > mkl +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f); +#endif +// TODO(TJ): eq16 test and complete avx512 +#undef INTRI8_FLOAT +#undef MKL_FLOAT +#undef MKL_DOUBLE + +/* VADD JitKernel */ +template +class VAddKernelImpl : public VAddKernel { + public: + explicit VAddKernelImpl(int d) : VAddKernel() { this->num_ = d; } + void Compute(const T* x, const T* y, T* z) const override { + for (int i = 0; i < this->num_; ++i) { + z[i] = x[i] + y[i]; + } + } +}; + +#ifdef PADDLE_WITH_MKLML +#define MKL_FLOAT(isa, block) \ + template <> \ + void VAddKernelImpl::Compute( \ + const float* x, const float* y, float* z) const { \ + platform::dynload::vsAdd(this->num_, x, y, z); \ + } + +#define MKL_DOUBLE(isa, block) \ + template <> \ + void VAddKernelImpl::Compute( \ + const double* x, const double* y, double* z) const { \ + platform::dynload::vdAdd(this->num_, x, y, z); \ + } + +FOR_EACH_ISA(MKL_FLOAT, kGT16); +FOR_EACH_ISA_BLOCK(MKL_DOUBLE); +#endif + +#define INTRI8_FLOAT(isa) \ + template <> \ + void VAddKernelImpl::Compute( \ + const float* x, const float* y, float* z) const { \ + __m256 tmpx, tmpy; \ + tmpx = _mm256_loadu_ps(x); \ + tmpy = _mm256_loadu_ps(y); \ + tmpx = _mm256_add_ps(tmpx, tmpy); \ + _mm256_storeu_ps(z, tmpx); \ + } +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f); +#endif +// TODO(TJ): eq16 test and complete avx512 + +#undef INTRI8_FLOAT +#undef MKL_FLOAT +#undef MKL_DOUBLE + +/* VSCAL JitKernel */ +template +class VScalKernelImpl : public VScalKernel { + public: + explicit VScalKernelImpl(int d) : VScalKernel() { this->num_ = d; } + void Compute(const T a, const T* x, T* y) const override { + for (int i = 0; i < this->num_; ++i) { + y[i] = a * x[i]; + } + } + void Compute(const T a, T* x) const override { + for (int i = 0; i < this->num_; ++i) { + x[i] = a * x[i]; + } + } +}; + +#ifdef PADDLE_WITH_MKLML +#define MKL_FLOAT(isa, block) \ + template <> \ + void VScalKernelImpl::Compute(const float a, float* x) \ + const { \ + platform::dynload::cblas_sscal(this->num_, a, x, 1); \ + } + +#define MKL_DOUBLE(isa, block) \ + template <> \ + void VScalKernelImpl::Compute(const double a, double* x) \ + const { \ + platform::dynload::cblas_dscal(this->num_, a, x, 1); \ + } + +FOR_EACH_ISA(MKL_FLOAT, kGT16); +FOR_EACH_ISA_BLOCK(MKL_DOUBLE); +#endif + +#define INTRI8_FLOAT(isa) \ + template <> \ + void VScalKernelImpl::Compute( \ + const float a, const float* x, float* y) const { \ + __m256 tmp; \ + __m256 scalar = _mm256_set1_ps(a); \ + tmp = _mm256_loadu_ps(x); \ + tmp = _mm256_mul_ps(tmp, scalar); \ + _mm256_storeu_ps(y, tmp); \ + } +#define INTRI8_INPLACE_FLOAT(isa) \ + template <> \ + void VScalKernelImpl::Compute(const float a, float* x) \ + const { \ + __m256 tmp; \ + __m256 scalar = _mm256_set1_ps(a); \ + tmp = _mm256_loadu_ps(x); \ + tmp = _mm256_mul_ps(tmp, scalar); \ + _mm256_storeu_ps(x, tmp); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +INTRI8_INPLACE_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +INTRI8_INPLACE_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f); +INTRI8_INPLACE_FLOAT(jit::avx512f); +#endif +// TODO(TJ): eq16 test and complete avx512 + +#undef INTRI8_FLOAT +#undef INTRI8_INPLACE_FLOAT +#undef MKL_FLOAT +#undef MKL_DOUBLE + +/* VAddBias JitKernel */ +template +class VAddBiasKernelImpl : public VAddBiasKernel { + public: + explicit VAddBiasKernelImpl(int d) : VAddBiasKernel() { this->num_ = d; } + void Compute(const T a, const T* x, T* y) const override { + for (int i = 0; i < this->num_; ++i) { + y[i] = x[i] + a; + } + } +}; + +#define INTRI8_FLOAT(isa) \ + template <> \ + void VAddBiasKernelImpl::Compute( \ + const float a, const float* x, float* y) const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + tmp = _mm256_add_ps(tmp, _mm256_set1_ps(a)); \ + _mm256_storeu_ps(y, tmp); \ + } + +#define INTRI16_FLOAT(isa) \ + template <> \ + void VAddBiasKernelImpl::Compute( \ + const float a, const float* x, float* y) const { \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + tmp0 = _mm256_add_ps(tmp0, _mm256_set1_ps(a)); \ + tmp1 = _mm256_add_ps(tmp1, _mm256_set1_ps(a)); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +INTRI16_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +INTRI16_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f); +INTRI16_FLOAT(jit::avx512f); +#endif +// TODO(TJ): eq16 test and complete avx512 + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT + +/* VRelu JitKernel */ +template +class VReluKernelImpl : public VReluKernel { + public: + explicit VReluKernelImpl(int d) : VReluKernel() { this->num_ = d; } + void Compute(const T* x, T* y) const override { + for (int i = 0; i < this->num_; ++i) { + y[i] = x[i] > 0 ? x[i] : 0; + } + } +}; + +#define INTRI8_FLOAT(isa) \ + template <> \ + void VReluKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + tmp = _mm256_max_ps(tmp, _mm256_setzero_ps()); \ + _mm256_storeu_ps(y, tmp); \ + } + +#define INTRI16_FLOAT(isa) \ + template <> \ + void VReluKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 zeros = _mm256_setzero_ps(); \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + tmp0 = _mm256_max_ps(tmp0, zeros); \ + tmp1 = _mm256_max_ps(tmp1, zeros); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#define INTRI_GT8LT16_FLOAT(isa) \ + template <> \ + VReluKernelImpl::VReluKernelImpl(int d) \ + : VReluKernel() { \ + this->num_ = d; \ + this->end_ = AVX_FLOAT_BLOCK; \ + this->rest_ = d - AVX_FLOAT_BLOCK; \ + } \ + template <> \ + void VReluKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 zeros = _mm256_setzero_ps(); \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + this->rest_); \ + tmp0 = _mm256_max_ps(tmp0, zeros); \ + tmp1 = _mm256_max_ps(tmp1, zeros); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + this->rest_, tmp1); \ + } + +#define INTRI_GT16_FLOAT(isa) \ + template <> \ + VReluKernelImpl::VReluKernelImpl(int d) \ + : VReluKernel() { \ + this->num_ = d; \ + this->end_ = d - d % AVX_FLOAT_BLOCK; \ + this->rest_ = d - AVX_FLOAT_BLOCK; \ + } \ + template <> \ + void VReluKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 zeros = _mm256_setzero_ps(); \ + for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \ + __m256 tmp = _mm256_loadu_ps(x + i); \ + tmp = _mm256_max_ps(tmp, zeros); \ + _mm256_storeu_ps(y + i, tmp); \ + } \ + __m256 tmp = _mm256_loadu_ps(x + this->rest_); \ + tmp = _mm256_max_ps(tmp, zeros); \ + _mm256_storeu_ps(y + this->rest_, tmp); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +INTRI16_FLOAT(jit::avx); +INTRI_GT8LT16_FLOAT(jit::avx); +INTRI_GT16_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +INTRI16_FLOAT(jit::avx2); +INTRI_GT8LT16_FLOAT(jit::avx2); +INTRI_GT16_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +// TODO(TJ): refine avx512 +INTRI8_FLOAT(jit::avx512f); +INTRI16_FLOAT(jit::avx512f); +INTRI_GT8LT16_FLOAT(jit::avx512f); +INTRI_GT16_FLOAT(jit::avx512f); +#endif + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef INTRI_GT8LT16_FLOAT +#undef INTRI_GT16_FLOAT + +/* An empty JitKernel */ +template +class VIdentityKernelImpl : public VIdentityKernel { + public: + explicit VIdentityKernelImpl(int d) : VIdentityKernel() { this->num_ = d; } + void Compute(const T* x, T* y) const override {} +}; + +/* VAddRelu JitKernel */ +template +class VAddReluKernelImpl : public VAddReluKernel { + public: + explicit VAddReluKernelImpl(int d) : VAddReluKernel() { this->num_ = d; } + void Compute(const T* x, const T* y, T* z) const override { + for (int i = 0; i < this->num_; ++i) { + z[i] = x[i] + y[i]; + z[i] = z[i] > 0 ? z[i] : 0; + } + } +}; + +#define INTRI8_FLOAT(isa) \ + template <> \ + void VAddReluKernelImpl::Compute( \ + const float* x, const float* y, float* z) const { \ + __m256 tmpx = _mm256_loadu_ps(x); \ + __m256 tmpy = _mm256_loadu_ps(y); \ + tmpy = _mm256_add_ps(tmpx, tmpy); \ + tmpy = _mm256_max_ps(tmpy, _mm256_setzero_ps()); \ + _mm256_storeu_ps(z, tmpy); \ + } + +#define INTRI16_FLOAT(isa) \ + template <> \ + void VAddReluKernelImpl::Compute( \ + const float* x, const float* y, float* z) const { \ + __m256 zeros = _mm256_setzero_ps(); \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(y); \ + tmp0 = _mm256_add_ps(tmp0, tmp1); \ + tmp0 = _mm256_max_ps(tmp0, zeros); \ + tmp1 = _mm256_loadu_ps(x + 8); \ + __m256 tmp2 = _mm256_loadu_ps(y + 8); \ + tmp1 = _mm256_add_ps(tmp1, tmp2); \ + tmp1 = _mm256_max_ps(tmp1, zeros); \ + _mm256_storeu_ps(z, tmp0); \ + _mm256_storeu_ps(z + 8, tmp1); \ + } + +#define INTRI_COMMON_FLOAT(isa, block) \ + template <> \ + VAddReluKernelImpl::VAddReluKernelImpl(int d) \ + : VAddReluKernel() { \ + this->num_ = d; \ + this->end_ = d - d % AVX_FLOAT_BLOCK; \ + this->rest_ = d - this->end_; \ + } \ + template <> \ + void VAddReluKernelImpl::Compute( \ + const float* x, const float* y, float* z) const { \ + __m256 zeros = _mm256_setzero_ps(); \ + for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \ + __m256 tmpx = _mm256_loadu_ps(x + i); \ + __m256 tmpy = _mm256_loadu_ps(y + i); \ + tmpy = _mm256_add_ps(tmpx, tmpy); \ + tmpy = _mm256_max_ps(tmpy, zeros); \ + _mm256_storeu_ps(z + i, tmpy); \ + } \ + for (int i = this->end_; i < this->num_; ++i) { \ + z[i] = x[i] + y[i]; \ + z[i] = z[i] > 0 ? z[i] : 0; \ + } \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +INTRI16_FLOAT(jit::avx); +INTRI_COMMON_FLOAT(jit::avx, kGT16); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +INTRI16_FLOAT(jit::avx2); +INTRI_COMMON_FLOAT(jit::avx2, kGT16); +#endif +#ifdef __AVX512F__ +// TODO(TJ): refine avx512 +INTRI8_FLOAT(jit::avx512f); +INTRI16_FLOAT(jit::avx512f); +INTRI_COMMON_FLOAT(jit::avx512f, kGT16); +#endif + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef INTRI_COMMON_FLOAT + +REGISTER_JITKERNEL(vmul, VMulKernel); +REGISTER_JITKERNEL(vadd, VAddKernel); +REGISTER_JITKERNEL(vscal, VScalKernel); +REGISTER_JITKERNEL(vaddb, VAddBiasKernel); +REGISTER_JITKERNEL(vrelu, VReluKernel); +REGISTER_JITKERNEL(vaddrelu, VAddReluKernel); +REGISTER_JITKERNEL(videntity, VIdentityKernel); + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_exp.cc b/paddle/fluid/operators/math/jit_kernel_exp.cc new file mode 100644 index 0000000000000000000000000000000000000000..c4247580f491a7ca26259528ca74dd92e35785a9 --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_exp.cc @@ -0,0 +1,541 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/jit_kernel.h" +#include // for exp +#include +#include "paddle/fluid/operators/math/jit_kernel_macro.h" +#ifdef PADDLE_WITH_MKLML +#include "paddle/fluid/platform/dynload/mklml.h" +#endif + +#ifdef __AVX__ +#include +#endif + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace jit = platform::jit; + +/* VExp JitKernel */ +template +class VExpKernelImpl : public VExpKernel { + public: + explicit VExpKernelImpl(int d) : VExpKernel() { this->num_ = d; } + void Compute(const T* x, T* y) const override { + for (int i = 0; i < this->num_; ++i) { + y[i] = std::exp(x[i]); + } + } +}; + +#ifdef PADDLE_WITH_MKLML +#define MKL_FLOAT(isa, block) \ + template <> \ + void VExpKernelImpl::Compute(const float* x, float* y) \ + const { \ + platform::dynload::vsExp(this->num_, x, y); \ + } + +#define MKL_DOUBLE(isa, block) \ + template <> \ + void VExpKernelImpl::Compute(const double* x, double* y) \ + const { \ + platform::dynload::vdExp(this->num_, x, y); \ + } +FOR_EACH_ISA(MKL_FLOAT, kLT8); +FOR_EACH_ISA(MKL_FLOAT, kGT8LT16); +FOR_EACH_ISA(MKL_FLOAT, kGT16); +FOR_EACH_ISA_BLOCK(MKL_DOUBLE); +#endif + +namespace detail { + +#ifdef __AVX__ + +#define ALIGN32 __attribute__((aligned(32))) + +#define _PS256_CONST(Name, Val) \ + static const float _ps256_##Name[8] ALIGN32 = {Val, Val, Val, Val, \ + Val, Val, Val, Val} + +#define _PI256_CONST(Name, Val) \ + static const int _pi256_##Name[8] ALIGN32 = {Val, Val, Val, Val, \ + Val, Val, Val, Val} + +_PI256_CONST(0x7f, 0x7f); +_PS256_CONST(one, 1.f); +_PS256_CONST(0p5, 0.5f); +_PS256_CONST(exp_hi, 88.3762626647949f); +_PS256_CONST(exp_lo, -88.3762626647949f); +_PS256_CONST(cephes_LOG2EF, 1.44269504088896341); +_PS256_CONST(cephes_exp_C1, 0.693359375); +_PS256_CONST(cephes_exp_C2, -2.12194440e-4); +_PS256_CONST(cephes_exp_p0, 1.9875691500E-4); +_PS256_CONST(cephes_exp_p1, 1.3981999507E-3); +_PS256_CONST(cephes_exp_p2, 8.3334519073E-3); +_PS256_CONST(cephes_exp_p3, 4.1665795894E-2); +_PS256_CONST(cephes_exp_p4, 1.6666665459E-1); +_PS256_CONST(cephes_exp_p5, 5.0000001201E-1); + +typedef union imm_xmm_union { + __m256i imm; + __m128i xmm[2]; +} imm_xmm_union; + +#define COPY_IMM_TO_XMM(imm_, xmm0_, xmm1_) \ + { \ + imm_xmm_union u ALIGN32; \ + u.imm = imm_; \ + xmm0_ = u.xmm[0]; \ + xmm1_ = u.xmm[1]; \ + } + +#define COPY_XMM_TO_IMM(xmm0_, xmm1_, imm_) \ + { \ + imm_xmm_union u ALIGN32; \ + u.xmm[0] = xmm0_; \ + u.xmm[1] = xmm1_; \ + imm_ = u.imm; \ + } + +#define AVX2_BITOP_USING_SSE2(fn) \ + static inline __m256i avx2_mm256_##fn(__m256i x, int y) { \ + /* use SSE2 to perform the bitop AVX2 */ \ + __m128i x1, x2; \ + __m256i ret; \ + COPY_IMM_TO_XMM(x, x1, x2); \ + x1 = _mm_##fn(x1, y); \ + x2 = _mm_##fn(x2, y); \ + COPY_XMM_TO_IMM(x1, x2, ret); \ + return ret; \ + } + +#define AVX2_INTOP_USING_SSE2(fn) \ + static inline __m256i avx2_mm256_add_epi32(__m256i x, __m256i y) { \ + /* use SSE2 to perform the AVX2 integer operation */ \ + __m128i x1, x2; \ + __m128i y1, y2; \ + __m256i ret; \ + COPY_IMM_TO_XMM(x, x1, x2); \ + COPY_IMM_TO_XMM(y, y1, y2); \ + x1 = _mm_##fn(x1, y1); \ + x2 = _mm_##fn(x2, y2); \ + COPY_XMM_TO_IMM(x1, x2, ret); \ + return ret; \ + } + +AVX2_BITOP_USING_SSE2(slli_epi32); +AVX2_INTOP_USING_SSE2(add_epi32); + +#define AVXEXP_BASE \ + __m256 tmp = _mm256_setzero_ps(), fx; \ + __m256 one = *reinterpret_cast(_ps256_one); \ + __m256i imm0; \ + x = _mm256_min_ps(x, *reinterpret_cast(_ps256_exp_hi)); \ + x = _mm256_max_ps(x, *reinterpret_cast(_ps256_exp_lo)); \ + /* express exp(x) as exp(g + n*log(2)) */ \ + fx = _mm256_mul_ps(x, \ + *reinterpret_cast(_ps256_cephes_LOG2EF)); \ + fx = _mm256_add_ps(fx, *reinterpret_cast(_ps256_0p5)); \ + tmp = _mm256_floor_ps(fx); \ + /* if greater, substract 1 */ \ + __m256 mask = _mm256_cmp_ps(tmp, fx, _CMP_GT_OS); \ + mask = _mm256_and_ps(mask, one); \ + fx = _mm256_sub_ps(tmp, mask); \ + tmp = _mm256_mul_ps(fx, \ + *reinterpret_cast(_ps256_cephes_exp_C1)); \ + __m256 z = _mm256_mul_ps( \ + fx, *reinterpret_cast(_ps256_cephes_exp_C2)); \ + x = _mm256_sub_ps(x, tmp); \ + x = _mm256_sub_ps(x, z); \ + z = _mm256_mul_ps(x, x); \ + __m256 y = *reinterpret_cast(_ps256_cephes_exp_p0); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p1)); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p2)); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p3)); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p4)); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p5)); \ + y = _mm256_mul_ps(y, z); \ + y = _mm256_add_ps(y, x); \ + y = _mm256_add_ps(y, one); \ + /* build 2^n */ \ + imm0 = _mm256_cvttps_epi32(fx) + +__m256 ExpAVX(__m256 x) { + AVXEXP_BASE; + // two AVX2 instructions using SSE2 + imm0 = avx2_mm256_add_epi32(imm0, + *reinterpret_cast(_pi256_0x7f)); + imm0 = avx2_mm256_slli_epi32(imm0, 23); + __m256 pow2n = _mm256_castsi256_ps(imm0); + y = _mm256_mul_ps(y, pow2n); + return y; +} +#endif + +#ifdef __AVX2__ +__m256 ExpAVX2(__m256 x) { + AVXEXP_BASE; + // two AVX2 instructions + imm0 = _mm256_add_epi32(imm0, *reinterpret_cast(_pi256_0x7f)); + imm0 = _mm256_slli_epi32(imm0, 23); + __m256 pow2n = _mm256_castsi256_ps(imm0); + y = _mm256_mul_ps(y, pow2n); + return y; +} +#endif + +} // namespace detail + +#define INTRI8_FLOAT(isa, expisa) \ + template <> \ + void VExpKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + _mm256_storeu_ps(y, expisa(tmp)); \ + } + +#define INTRI16_FLOAT(isa, expisa) \ + template <> \ + void VExpKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + tmp0 = expisa(tmp0); \ + tmp1 = expisa(tmp1); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx, detail::ExpAVX); +INTRI16_FLOAT(jit::avx, detail::ExpAVX); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx2, detail::ExpAVX2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); +#endif +// TODO(TJ): eq16 test and complete avx512 + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef MKL_FLOAT +#undef MKL_DOUBLE + +REGISTER_JITKERNEL(vexp, VExpKernel); + +/* VSigmoid JitKernel */ +template +class VSigmoidKernelImpl : public VSigmoidKernel { + public: + explicit VSigmoidKernelImpl(int d) : VSigmoidKernel() { + this->num_ = d; + vexp_ = KernelPool::Instance().template Get>(d); + } + void Compute(const T* x, T* y) const override { + const T min = SIGMOID_THRESHOLD_MIN; + const T max = SIGMOID_THRESHOLD_MAX; + for (int i = 0; i < this->num_; ++i) { + y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]); + y[i] = static_cast(0) - y[i]; + } + vexp_->Compute(y, y); + for (int i = 0; i < this->num_; ++i) { + y[i] = static_cast(1) / (static_cast(1) + y[i]); + } + } + + private: + std::shared_ptr> vexp_; +}; + +#define INTRI_SIGMOID(tmp, min, max, expisa) \ + tmp = _mm256_max_ps(tmp, min); \ + tmp = _mm256_min_ps(tmp, max); \ + tmp = _mm256_sub_ps(_mm256_set1_ps(0.0f), tmp); \ + tmp = expisa(tmp); \ + tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp); \ + tmp = _mm256_div_ps(_mm256_set1_ps(1.0f), tmp) + +#define INTRI8_FLOAT(isa, expisa) \ + template <> \ + void VSigmoidKernelImpl::Compute(const float* x, float* y) \ + const { \ + /* TODO(TJ): try to use static const*/ \ + __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \ + __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \ + __m256 tmp = _mm256_loadu_ps(x); \ + INTRI_SIGMOID(tmp, min, max, expisa); \ + _mm256_storeu_ps(y, tmp); \ + } + +#define INTRI16_FLOAT(isa, expisa) \ + template <> \ + void VSigmoidKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \ + __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + INTRI_SIGMOID(tmp0, min, max, expisa); \ + INTRI_SIGMOID(tmp1, min, max, expisa); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#define INTRI_GT8LT16_FLOAT(isa, expisa) \ + template <> \ + VSigmoidKernelImpl::VSigmoidKernelImpl(int d) \ + : VSigmoidKernel() { \ + this->num_ = d; \ + this->end_ = AVX_FLOAT_BLOCK; \ + this->rest_ = d - this->end_; \ + vexp_ = \ + KernelPool::Instance().template Get>(this->rest_); \ + } \ + template <> \ + void VSigmoidKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \ + __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \ + __m256 tmp = _mm256_loadu_ps(x); \ + INTRI_SIGMOID(tmp, min, max, expisa); \ + _mm256_storeu_ps(y, tmp); \ + const float min_ = SIGMOID_THRESHOLD_MIN; \ + const float max_ = SIGMOID_THRESHOLD_MAX; \ + for (int i = this->end_; i < this->num_; ++i) { \ + y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]); \ + y[i] = 0.f - y[i]; \ + } \ + vexp_->Compute(y + this->end_, y + this->end_); \ + for (int i = this->end_; i < this->num_; ++i) { \ + y[i] = 1.f / (1.f + y[i]); \ + } \ + } + +#define INTRI_GT16_FLOAT(isa, expisa) \ + template <> \ + VSigmoidKernelImpl::VSigmoidKernelImpl(int d) \ + : VSigmoidKernel() { \ + this->num_ = d; \ + this->rest_ = d % AVX_FLOAT_BLOCK; \ + this->end_ = d - this->rest_; \ + vexp_ = \ + KernelPool::Instance().template Get>(this->rest_); \ + } \ + template <> \ + void VSigmoidKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \ + __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \ + for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \ + __m256 tmp = _mm256_loadu_ps(x + i); \ + INTRI_SIGMOID(tmp, min, max, expisa); \ + _mm256_storeu_ps(y + i, tmp); \ + } \ + const float min_ = SIGMOID_THRESHOLD_MIN; \ + const float max_ = SIGMOID_THRESHOLD_MAX; \ + for (int i = this->end_; i < this->num_; ++i) { \ + y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]); \ + y[i] = 0.f - y[i]; \ + } \ + vexp_->Compute(y + this->end_, y + this->end_); \ + for (int i = this->end_; i < this->num_; ++i) { \ + y[i] = 1.f / (1.f + y[i]); \ + } \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx, detail::ExpAVX); +INTRI16_FLOAT(jit::avx, detail::ExpAVX); +INTRI_GT8LT16_FLOAT(jit::avx, detail::ExpAVX); +INTRI_GT16_FLOAT(jit::avx, detail::ExpAVX); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx2, detail::ExpAVX2); +// maybe use avx at gt8lt16 and gt16 +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); +// maybe use avx2 at gt8lt16 and gt16 +#endif + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef INTRI_GT8LT16_FLOAT +#undef INTRI_GT16_FLOAT +#undef INTRI_VSIGMOID + +REGISTER_JITKERNEL(vsigmoid, VSigmoidKernel); + +/* VTanh JitKernel */ +template +class VTanhKernelImpl : public VTanhKernel { + public: + explicit VTanhKernelImpl(int d) : VTanhKernel() { + this->num_ = d; + vscal_ = KernelPool::Instance().template Get>(d); + vsigmoid_ = KernelPool::Instance().template Get>(d); + vaddbias_ = KernelPool::Instance().template Get>(d); + } + void Compute(const T* x, T* y) const override { + vscal_->Compute(static_cast(2), x, y); + vsigmoid_->Compute(y, y); + vscal_->Compute(static_cast(2), y); + vaddbias_->Compute(static_cast(-1), y, y); + } + + private: + std::shared_ptr> vscal_; + std::shared_ptr> vsigmoid_; + std::shared_ptr> vaddbias_; +}; + +#define INTRI_VTANH(tmp, expisa) \ + tmp = _mm256_mul_ps(_mm256_set1_ps(-2.0f), tmp); \ + tmp = _mm256_min_ps(tmp, _mm256_set1_ps(EXP_MAX_INPUT)); \ + tmp = expisa(tmp); \ + tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp); \ + tmp = _mm256_div_ps(_mm256_set1_ps(2.0f), tmp); \ + tmp = _mm256_sub_ps(tmp, _mm256_set1_ps(1.0f)) + +#define INTRI8_FLOAT(isa, expisa) \ + template <> \ + void VTanhKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + INTRI_VTANH(tmp, expisa); \ + _mm256_storeu_ps(y, tmp); \ + } + +#define INTRI16_FLOAT(isa, expisa) \ + template <> \ + void VTanhKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + INTRI_VTANH(tmp0, expisa); \ + INTRI_VTANH(tmp1, expisa); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#define INTRI_GT8LT16_FLOAT(isa, expisa) \ + template <> \ + VTanhKernelImpl::VTanhKernelImpl(int d) \ + : VTanhKernel() { \ + this->num_ = d; \ + this->end_ = AVX_FLOAT_BLOCK; \ + this->rest_ = d - this->end_; \ + vscal_ = \ + KernelPool::Instance().template Get>(this->rest_); \ + vsigmoid_ = KernelPool::Instance().template Get>( \ + this->rest_); \ + vaddbias_ = KernelPool::Instance().template Get>( \ + this->rest_); \ + } \ + template <> \ + void VTanhKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + INTRI_VTANH(tmp, expisa); \ + _mm256_storeu_ps(y, tmp); \ + x += AVX_FLOAT_BLOCK; \ + y += AVX_FLOAT_BLOCK; \ + vscal_->Compute(2.f, x, y); \ + vsigmoid_->Compute(y, y); \ + vscal_->Compute(2.f, y); \ + vaddbias_->Compute(-1.f, y, y); \ + } + +#define INTRI_GT16_FLOAT(isa, expisa) \ + template <> \ + VTanhKernelImpl::VTanhKernelImpl(int d) \ + : VTanhKernel() { \ + this->num_ = d; \ + this->rest_ = d % AVX_FLOAT_BLOCK; \ + this->end_ = d - this->rest_; \ + vscal_ = \ + KernelPool::Instance().template Get>(this->rest_); \ + vsigmoid_ = KernelPool::Instance().template Get>( \ + this->rest_); \ + vaddbias_ = KernelPool::Instance().template Get>( \ + this->rest_); \ + } \ + template <> \ + void VTanhKernelImpl::Compute(const float* x, float* y) \ + const { \ + for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \ + __m256 tmp = _mm256_loadu_ps(x + i); \ + INTRI_VTANH(tmp, expisa); \ + _mm256_storeu_ps(y + i, tmp); \ + } \ + x += this->end_; \ + y += this->end_; \ + vscal_->Compute(2.f, x, y); \ + vsigmoid_->Compute(y, y); \ + vscal_->Compute(2.f, y); \ + vaddbias_->Compute(-1.f, y, y); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx, detail::ExpAVX); +INTRI16_FLOAT(jit::avx, detail::ExpAVX); +INTRI_GT8LT16_FLOAT(jit::avx, detail::ExpAVX); +INTRI_GT16_FLOAT(jit::avx, detail::ExpAVX); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx2, detail::ExpAVX2); +// maybe use avx at gt8lt16 and gt16 +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); +// maybe use avx at gt8lt16 and gt16 +#endif + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef INTRI_GT8LT16_FLOAT +#undef INTRI_GT16_FLOAT +#undef INTRI_VTANH + +REGISTER_JITKERNEL(vtanh, VTanhKernel); + +#undef JITKERNEL_NEW_ACT_IMPL + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_lstm.cc b/paddle/fluid/operators/math/jit_kernel_lstm.cc new file mode 100644 index 0000000000000000000000000000000000000000..26bd26e2e171feea569fbd646a9caf03bebbaa46 --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_lstm.cc @@ -0,0 +1,360 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/jit_kernel.h" +#include +#include "paddle/fluid/operators/math/jit_kernel_macro.h" +#include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/platform/macros.h" + +#ifdef __AVX__ +#include +#endif + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace detail { +#ifdef __AVX__ +__m256 ExpAVX(__m256 x); +#endif + +#ifdef __AVX2__ +__m256 ExpAVX2(__m256 x); +#endif + +} // namespace detail + +namespace jit = platform::jit; + +#ifdef __AVX__ +typedef enum { kSigmoid, kRelu, kTanh, kIdentity } act_type; + +class AVXAct { + public: + virtual ~AVXAct() = default; + virtual __m256 Compute(__m256 x) const = 0; +}; + +template +class AVXActImpl : public AVXAct { + public: + __m256 Compute(__m256 x) const override { PADDLE_THROW("Unkown type!"); } +}; + +#define AVX_SIGMOID(isa, expisa) \ + template <> \ + __m256 AVXActImpl::Compute(__m256 x) const { \ + __m256 ones = _mm256_set1_ps(1.0f); \ + x = _mm256_max_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MIN)); \ + x = _mm256_min_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MAX)); \ + x = _mm256_sub_ps(_mm256_set1_ps(0.0f), x); \ + x = expisa(x); \ + x = _mm256_add_ps(ones, x); \ + return _mm256_div_ps(ones, x); \ + } + +#define AVX_TANH(isa, expisa) \ + template <> \ + __m256 AVXActImpl::Compute(__m256 x) const { \ + __m256 ones = _mm256_set1_ps(1.0f); \ + x = _mm256_mul_ps(_mm256_set1_ps(-2.0f), x); \ + x = _mm256_min_ps(x, _mm256_set1_ps(EXP_MAX_INPUT)); \ + x = expisa(x); \ + x = _mm256_add_ps(ones, x); \ + x = _mm256_div_ps(_mm256_set1_ps(2.0f), x); \ + return _mm256_sub_ps(x, ones); \ + } + +#define AVX_RELU(isa) \ + template <> \ + __m256 AVXActImpl::Compute(__m256 x) const { \ + return _mm256_max_ps(x, _mm256_setzero_ps()); \ + } + +#define AVX_IDENTITY(isa) \ + template <> \ + __m256 AVXActImpl::Compute(__m256 x) const { \ + return x; \ + } + +#define FOR_EACH_AVX_ISA(macro_) \ + macro_(jit::avx); \ + macro_(jit::avx2); \ + macro_(jit::avx512f) + +FOR_EACH_AVX_ISA(AVX_RELU); +FOR_EACH_AVX_ISA(AVX_IDENTITY); + +AVX_SIGMOID(jit::avx, detail::ExpAVX); +AVX_TANH(jit::avx, detail::ExpAVX); + +#ifdef __AVX2__ +AVX_SIGMOID(jit::avx2, detail::ExpAVX2); +AVX_SIGMOID(jit::avx512f, detail::ExpAVX2); +AVX_TANH(jit::avx2, detail::ExpAVX2); +AVX_TANH(jit::avx512f, detail::ExpAVX2); +#endif + +#undef FOR_EACH_AVX_ISA +#undef AVX_IDENTITY +#undef AVX_RELU +#undef AVX_TANH +#undef AVX_SIGMOID + +#endif + +template +static std::shared_ptr> GetActKernel( + const std::string& type, int n) { + if (type == "sigmoid") { + return std::dynamic_pointer_cast>( + KernelPool::Instance().template Get>(n)); + } else if (type == "relu") { + return std::dynamic_pointer_cast>( + KernelPool::Instance().template Get>(n)); + } else if (type == "tanh") { + return std::dynamic_pointer_cast>( + KernelPool::Instance().template Get>(n)); + } else if (type == "identity" || type == "") { + return std::dynamic_pointer_cast>( + KernelPool::Instance().template Get>(n)); + } + PADDLE_THROW("Not support type: %s", type); + return nullptr; +} + +/* LSTM JitKernel */ +template +class LSTMKernelImpl : public LSTMKernel { + public: + explicit LSTMKernelImpl(const std::string& act_gate, + const std::string& act_cand, + const std::string& act_cell, int d) + : LSTMKernel() { + d_ = d; + d2_ = d * 2; + d3_ = d * 3; + act_gate_d3_ = GetActKernel(act_gate, d3_); + act_gate_d_ = GetActKernel(act_gate, d); + act_cand_d_ = GetActKernel(act_cand, d); + act_cell_d_ = GetActKernel(act_cell, d); + vmul_d_ = KernelPool::Instance().template Get>(d); + vadd_d_ = KernelPool::Instance().template Get>(d); + } + + void ComputeCtHt(T* gates, const T* ct_1, T* ct, T* ht, const T* wp_data, + T* checked) const override { + // gates: W_ch, W_ih, W_fh, W_oh + act_gate_d3_->Compute(gates + d_, gates + d_); + + /* C_t = C_t-1 * fgated + cand_gated * igated */ + act_cand_d_->Compute(gates, gates); + vmul_d_->Compute(gates, gates + d_, gates + d_); + vmul_d_->Compute(ct_1, gates + d2_, gates + d2_); + vadd_d_->Compute(gates + d_, gates + d2_, ct); + + /* H_t = act_cell(C_t) * ogated */ + act_cell_d_->Compute(ct, gates + d2_); + vmul_d_->Compute(gates + d2_, gates + d3_, ht); + } + void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override { + /* C_t = igated * cgated*/ + act_gate_d_->Compute(gates + d_, gates + d_); + act_cand_d_->Compute(gates, gates); + vmul_d_->Compute(gates, gates + d_, ct); + /* H_t = act_cell(C_t) * ogated */ + act_gate_d_->Compute(gates + d3_, gates + d3_); + act_cell_d_->Compute(ct, gates + d2_); + vmul_d_->Compute(gates + d2_, gates + d3_, ht); + } + + private: + int d_, d2_, d3_; + std::shared_ptr> act_gate_d3_, act_gate_d_, act_cand_d_, + act_cell_d_; + std::shared_ptr> vmul_d_; + std::shared_ptr> vadd_d_; +#ifdef __AVX__ + std::unique_ptr avx_act_gate_, avx_act_cand_, avx_act_cell_; +#endif +}; + +#define INTRI8_FLOAT(isa) \ + template <> \ + LSTMKernelImpl::LSTMKernelImpl( \ + const std::string& act_gate, const std::string& act_cand, \ + const std::string& act_cell, int d) \ + : LSTMKernel() { \ + auto GetAVXAct = [&](const std::string& type) -> std::unique_ptr { \ + if (type == "sigmoid") { \ + return std::unique_ptr(new AVXActImpl()); \ + } else if (type == "relu") { \ + return std::unique_ptr(new AVXActImpl()); \ + } else if (type == "tanh") { \ + return std::unique_ptr(new AVXActImpl()); \ + } else if (type == "identity" || type == "") { \ + return std::unique_ptr(new AVXActImpl()); \ + } \ + PADDLE_THROW("Not support type: %s", type); \ + }; \ + avx_act_gate_ = GetAVXAct(act_gate); \ + avx_act_cand_ = GetAVXAct(act_cand); \ + avx_act_cell_ = GetAVXAct(act_cell); \ + } \ + template <> \ + void LSTMKernelImpl::ComputeCtHt( \ + float* gates, const float* ct_1, float* ct, float* ht, \ + const float* wp_data, float* checked) const { \ + /* gates: W_ch, W_ih, W_fh, W_oh */ \ + __m256 c, i, f, o; \ + c = _mm256_loadu_ps(gates); \ + i = _mm256_loadu_ps(gates + 8); \ + f = _mm256_loadu_ps(gates + 16); \ + o = _mm256_loadu_ps(gates + 24); \ + /* C_t = C_t-1 * fgated + cand_gated * igated*/ \ + c = _mm256_mul_ps(avx_act_cand_->Compute(c), avx_act_gate_->Compute(i)); \ + i = _mm256_loadu_ps(ct_1); \ + f = _mm256_mul_ps(i, avx_act_gate_->Compute(f)); \ + f = _mm256_add_ps(c, f); \ + _mm256_storeu_ps(ct, f); \ + /* H_t = act_cell(C_t) * ogated */ \ + o = _mm256_mul_ps(avx_act_cell_->Compute(f), avx_act_gate_->Compute(o)); \ + _mm256_storeu_ps(ht, o); \ + } \ + template <> \ + void LSTMKernelImpl::ComputeC1H1( \ + float* gates, float* ct, float* ht, const float* wp_data) const { \ + __m256 c, i, o; \ + c = _mm256_loadu_ps(gates); \ + i = _mm256_loadu_ps(gates + 8); \ + o = _mm256_loadu_ps(gates + 24); \ + /* C_t = igated * cgated*/ \ + c = _mm256_mul_ps(avx_act_gate_->Compute(i), avx_act_cand_->Compute(c)); \ + _mm256_storeu_ps(ct, c); \ + /* H_t = act_cell(C_t) * ogated */ \ + o = _mm256_mul_ps(avx_act_cell_->Compute(c), avx_act_gate_->Compute(o)); \ + _mm256_storeu_ps(ht, o); \ + } + +// TODO(TJ): optimize keq16 + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f); +#endif + +/* Peephole JitKernel */ +template +class PeepholeKernelImpl : public LSTMKernel { + public: + explicit PeepholeKernelImpl(const std::string& act_gate, + const std::string& act_cand, + const std::string& act_cell, int d) + : LSTMKernel() { + d_ = d; + d2_ = d * 2; + d3_ = d * 3; + act_gate_d_ = GetActKernel(act_gate, d); + act_cand_d_ = GetActKernel(act_cand, d); + act_cell_d_ = GetActKernel(act_cell, d); + vmul_d_ = KernelPool::Instance().template Get>(d); + vadd_d_ = KernelPool::Instance().template Get>(d); + vadd_d2_ = KernelPool::Instance().template Get>(d2_); + act_gate_d2_ = GetActKernel(act_gate, d2_); + } + + void ComputeCtHt(T* gates, const T* ct_1, T* ct, T* ht, const T* wp_data, + T* checked) const override { + /* get fgated and igated*/ + vmul_d_->Compute(wp_data, ct_1, checked); + vmul_d_->Compute(wp_data + d_, ct_1, checked + d_); + vadd_d2_->Compute(checked, gates + d_, gates + d_); + act_gate_d2_->Compute(gates + d_, gates + d_); + /* C_t = C_t-1 * fgated + cand_gated * igated*/ + act_cand_d_->Compute(gates, gates); + vmul_d_->Compute(gates, gates + d_, gates + d_); + vmul_d_->Compute(ct_1, gates + d2_, gates + d2_); + vadd_d_->Compute(gates + d_, gates + d2_, ct); + /* get ogated*/ + vmul_d_->Compute(wp_data + d2_, ct, gates + d_); + vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_); + act_gate_d_->Compute(gates + d3_, gates + d3_); + /* H_t = act_cell(C_t) * ogated */ + act_cell_d_->Compute(ct, gates + d2_); + vmul_d_->Compute(gates + d2_, gates + d3_, ht); + } + + void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override { + /* C_t = igated * cgated*/ + act_gate_d_->Compute(gates + d_, gates + d_); + act_cand_d_->Compute(gates, gates); + vmul_d_->Compute(gates, gates + d_, ct); + /* get outgated, put W_oc * C_t on igated */ + vmul_d_->Compute(wp_data + d2_, ct, gates + d_); + vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_); + /* H_t = act_cell(C_t) * ogated */ + act_gate_d_->Compute(gates + d3_, gates + d3_); + act_cell_d_->Compute(ct, gates + d2_); + vmul_d_->Compute(gates + d2_, gates + d3_, ht); + } + + private: + int d_, d2_, d3_; + std::shared_ptr> act_gate_d2_, act_gate_d_, act_cand_d_, + act_cell_d_; + std::shared_ptr> vmul_d_; + std::shared_ptr> vadd_d_, vadd_d2_; +}; + +#define JITKERNEL_DECLARE_LSTM(ker_class, ker_dtype) \ + template <> \ + std::shared_ptr> \ + KernelPool::Get, const std::string&, \ + const std::string&, const std::string&, int, bool>( \ + const std::string& act_gate, const std::string& act_cand, \ + const std::string& act_cell, int d, bool use_peephole) + +#define JITKERNEL_KEY_LSTM(ker_key, dtype_key) \ + #ker_key #dtype_key + std::to_string(d) + act_gate + act_cand + act_cell + \ + (use_peephole ? "p" : "n") + +#define JITKERNEL_NEW_LSTM_IMPL(ker, dtype, isa, k) \ + if (use_peephole) { \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>( \ + act_gate, act_cand, act_cell, d)); \ + } else { \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>(act_gate, act_cand, \ + act_cell, d)); \ + } + +REGISTER_JITKERNEL_ARGS(lstm, LSTMKernel, JITKERNEL_DECLARE_LSTM, + JITKERNEL_KEY_LSTM, JITKERNEL_NEW_LSTM_IMPL); + +#undef INTRI8_FLOAT +#undef JITKERNEL_DECLARE_LSTM +#undef JITKERNEL_KEY_LSTM +#undef JITKERNEL_NEW_LSTM_IMPL +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_macro.h b/paddle/fluid/operators/math/jit_kernel_macro.h new file mode 100644 index 0000000000000000000000000000000000000000..d8e55f2673560ff6afa34376b73275b57a8ceea1 --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_macro.h @@ -0,0 +1,111 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "paddle/fluid/platform/cpu_info.h" + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +namespace jit = platform::jit; + +#define SEARCH_BLOCK(macro_, ker, dtype, isa) \ + if (d < AVX_FLOAT_BLOCK) { \ + macro_(ker, dtype, isa, kLT8); \ + } else if (d == AVX_FLOAT_BLOCK) { \ + macro_(ker, dtype, isa, kEQ8); \ + } else if (d > AVX_FLOAT_BLOCK && d < AVX512_FLOAT_BLOCK) { \ + macro_(ker, dtype, isa, kGT8LT16); \ + } else if (d == AVX512_FLOAT_BLOCK) { \ + macro_(ker, dtype, isa, kEQ16); \ + } else { \ + macro_(ker, dtype, isa, kGT16); \ + } + +#define SEARCH_ISA_BLOCK(macro_, ker, dtype) \ + if (jit::MayIUse(jit::avx512f)) { \ + SEARCH_BLOCK(macro_, ker, dtype, jit::avx512f); \ + } else if (jit::MayIUse(jit::avx2)) { \ + SEARCH_BLOCK(macro_, ker, dtype, jit::avx2); \ + } else if (jit::MayIUse(jit::avx)) { \ + SEARCH_BLOCK(macro_, ker, dtype, jit::avx); \ + } else { \ + SEARCH_BLOCK(macro_, ker, dtype, jit::isa_any); \ + } + +#define JITKERNEL_DECLARE(ker_class, ker_dtype) \ + template <> \ + std::shared_ptr> \ + KernelPool::Get, int>(int d) + +#define JITKERNEL_KEY(ker_key, dtype_key) \ + #ker_key #dtype_key + std::to_string(d) + +#define JITKERNEL_NEW_IMPL(ker, dtype, isa, k) \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>(d)) + +#define JITKERNEL_WITH_DTYPE(ker_key, ker_class, ker_dtype, dtype_key, \ + marco_declare, macro_key, macro_impl) \ + marco_declare(ker_class, ker_dtype) { \ + std::string key = macro_key(ker_key, dtype_key); \ + if (kers_.find(key) == kers_.end()) { \ + std::shared_ptr> p; \ + SEARCH_ISA_BLOCK(macro_impl, ker_class, ker_dtype); \ + kers_.insert({key, std::dynamic_pointer_cast(p)}); \ + return p; \ + } \ + return std::dynamic_pointer_cast>( \ + kers_.at(key)); \ + } + +#define REGISTER_JITKERNEL(ker_key, ker_class) \ + JITKERNEL_WITH_DTYPE(ker_key, ker_class, float, f, JITKERNEL_DECLARE, \ + JITKERNEL_KEY, JITKERNEL_NEW_IMPL); \ + JITKERNEL_WITH_DTYPE(ker_key, ker_class, double, d, JITKERNEL_DECLARE, \ + JITKERNEL_KEY, JITKERNEL_NEW_IMPL) + +#define REGISTER_JITKERNEL_ARGS(ker_key, ker_class, marco_declare, macro_key, \ + macro_impl) \ + JITKERNEL_WITH_DTYPE(ker_key, ker_class, float, f, marco_declare, macro_key, \ + macro_impl); \ + JITKERNEL_WITH_DTYPE(ker_key, ker_class, double, d, marco_declare, \ + macro_key, macro_impl) + +#define FOR_EACH_ISA(macro_, block) \ + macro_(jit::avx512f, block); \ + macro_(jit::avx2, block); \ + macro_(jit::avx, block); \ + macro_(jit::isa_any, block) + +#define FOR_EACH_BLOCK(macro_, isa) \ + macro_(isa, kLT8); \ + macro_(isa, kEQ8); \ + macro_(isa, kGT8LT16); \ + macro_(isa, kEQ16); \ + macro_(isa, kGT16) + +#define FOR_EACH_ISA_BLOCK(macro_) \ + FOR_EACH_BLOCK(macro_, jit::avx512f); \ + FOR_EACH_BLOCK(macro_, jit::avx2); \ + FOR_EACH_BLOCK(macro_, jit::avx); \ + FOR_EACH_BLOCK(macro_, jit::isa_any) + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_test.cc b/paddle/fluid/operators/math/jit_kernel_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..c9e6ab740da4e39de3dc41f9df352b88e696c38d --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_test.cc @@ -0,0 +1,807 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/jit_kernel.h" +#include +#include // for exp +#include // for memcpy +#include +#include +#include +#include "gflags/gflags.h" +#include "glog/logging.h" +#include "gtest/gtest.h" + +#ifdef PADDLE_WITH_MKLML +#include "paddle/fluid/platform/dynload/mklml.h" +#endif + +#ifdef __AVX__ +#include +#endif + +constexpr int repeat = 20000; + +inline double GetCurrentUS() { + struct timeval time; + gettimeofday(&time, NULL); + return 1e+6 * time.tv_sec + time.tv_usec; +} + +template +void RandomVec(const int n, T* a, const T lower = static_cast(-20.f), + const T upper = static_cast(20.f)) { + static unsigned int seed = 100; + std::mt19937 rng(seed++); + std::uniform_real_distribution uniform_dist(0, 1); + for (int i = 0; i < n; ++i) { + a[i] = static_cast(uniform_dist(rng) * (upper - lower) + lower); + } +} + +void vrelu_ref(const int n, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = x[i] > 0.f ? x[i] : 0.f; + } +} + +#if defined __AVX__ || defined __AVX2__ +void vrelu_intri8(const int n, const float* x, float* y) { + __m256 tmp = _mm256_loadu_ps(x); + tmp = _mm256_max_ps(tmp, _mm256_setzero_ps()); + _mm256_storeu_ps(y, tmp); +} +#endif + +TEST(JitKernel, vrelu) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -10.f, 1.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vrelu_ref(d, x_data, zref_data); + } + auto trefe = GetCurrentUS(); +#if defined __AVX__ || defined __AVX2__ + if (d == 8) { + auto si0 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vrelu_intri8(d, x_data, zref_data); + } + auto si1 = GetCurrentUS(); + VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat; + } +#endif + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vaddbias_ref(const int n, const float a, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = x[i] + a; + } +} + +TEST(JitKernel, vaddbias) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 64, 100, 128, 256}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -2.f, 2.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float a = 2.f; + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vaddbias_ref(d, a, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(a, x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vexp_ref(const int n, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = std::exp(x[i]); + } +} + +#ifdef PADDLE_WITH_MKLML +void vexp_mkl(const int n, const float* x, float* y) { + paddle::platform::dynload::vsExp(n, x, y); +} +#endif + +TEST(JitKernel, vexp) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 128, 256}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -2.f, 2.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vexp_ref(d, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + +#ifdef PADDLE_WITH_MKLML + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vexp_mkl(d, x_data, zref_data); + } + auto tmkle = GetCurrentUS(); +#endif + + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat +#ifdef PADDLE_WITH_MKLML + << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, " +#else + << " us, " +#endif + << "tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +inline float _sigmoid(float x) { + const float min = SIGMOID_THRESHOLD_MIN; + const float max = SIGMOID_THRESHOLD_MAX; + float tmp = (x < min) ? min : ((x > max) ? max : x); + return 1.f / (1.f + std::exp(-tmp)); +} + +void vsigmoid_ref(const int n, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = _sigmoid(x[i]); + } +} + +void vsigmoid_better( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VExpKernel>& vexp, + const int n, const float* x, float* y) { + const float min = SIGMOID_THRESHOLD_MIN; + const float max = SIGMOID_THRESHOLD_MAX; + for (int i = 0; i < n; ++i) { + y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]); + y[i] = 0.f - y[i]; + } + vexp->Compute(y, y); + for (int i = 0; i < n; ++i) { + y[i] = 1.f / (1.f + y[i]); + } +} + +TEST(JitKernel, vsigmoid) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -2.f, 2.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const auto& vexp = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vsigmoid_better(vexp, d, x_data, zref_data); + } + auto tmkle = GetCurrentUS(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vsigmoid_ref(d, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat + << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +inline float _tanh(float x) { return 2.f * _sigmoid(2.f * x) - 1.f; } + +void vtanh_ref(const int n, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = _tanh(x[i]); + } +} + +void vtanh_better( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VScalKernel>& vscal, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VSigmoidKernel>& + vsigmoid, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VAddBiasKernel>& + vaddbias, + const int n, const float* x, float* y) { + vscal->Compute(2.f, x, y); + vsigmoid->Compute(y, y); + vscal->Compute(2.f, y); + vaddbias->Compute(-1.f, y, y); +} + +TEST(JitKernel, vtanh) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -2.f, 2.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const auto& vscal = + jit::KernelPool::Instance().template Get>(d); + const auto& vsigmoid = + jit::KernelPool::Instance().template Get>(d); + const auto& vaddbias = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vtanh_better(vscal, vsigmoid, vaddbias, d, x_data, zref_data); + } + auto tmkle = GetCurrentUS(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vtanh_ref(d, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat + << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void lstm_ctht_ref( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VSigmoidKernel>& + vsigmoid_3d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VTanhKernel>& vtanh_d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VExpKernel>& vexp_1, + const int d, float* gates, const float* ct_1, float* ct, float* ht) { + vsigmoid_3d->Compute(gates + d, gates + d); + vtanh_d->Compute(gates, gates); + const float *i = gates + d, *f = gates + d * 2, *o = gates + d * 3; + const float min = SIGMOID_THRESHOLD_MIN; + const float max = SIGMOID_THRESHOLD_MAX; + for (int k = 0; k < d; ++k) { + // C_t = C_t-1 * fgated + cand_gated * igated + ct[k] = ct_1[k] * f[k] + gates[k] * i[k]; + // H_t = act_cell(C_t) * ogated + float tmp = ct[k] * 2; + tmp = 0.f - ((tmp < min) ? min : ((tmp > max) ? max : tmp)); + vexp_1->Compute(&tmp, &tmp); + tmp = 2.f / (1.f + tmp) - 1.f; + ht[k] = tmp * o[k]; + } +} + +void lstm_ctht_better( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VSigmoidKernel>& + vsigmoid_3d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VTanhKernel>& vtanh_d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VMulKernel>& vmul_d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VAddKernel>& vadd_d, + const int d, float* gates, const float* ct_1, float* ct, float* ht) { + int d2 = d * 2; + vsigmoid_3d->Compute(gates + d, gates + d); + vtanh_d->Compute(gates, gates); + vmul_d->Compute(gates, gates + d, gates + d); + vmul_d->Compute(ct_1, gates + d2, gates + d2); + vadd_d->Compute(gates + d, gates + d2, ct); + /* H_t = act_cell(C_t) * ogated */ + vtanh_d->Compute(ct, gates + d2); + vmul_d->Compute(gates + d2, gates + d * 3, ht); +} + +TEST(JitKernel, lstm) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 32, 64, 100}) { + int d4 = d * 4; + int d3 = d * 3; + std::vector x(d4), xref(d4); + std::vector ct_1(d), ct_tgt(d), ht_tgt(d); + std::vector ct_ref(d), ht_ref(d); + RandomVec(d4, x.data(), -2.f, 2.f); + RandomVec(d, ct_1.data(), -2.f, 2.f); + memcpy(xref.data(), x.data(), sizeof(float) * d4); + std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh"; + const auto& ker = + jit::KernelPool::Instance() + .template Get, const std::string&, + const std::string&, const std::string&>( + act_gate, act_cand, act_cell, d, false); + // below kernels are used to compute refer + const auto& vsigmoid_3d = + jit::KernelPool::Instance().template Get>( + d3); + const auto& vtanh_d = + jit::KernelPool::Instance().template Get>(d); + const auto& vexp_1 = + jit::KernelPool::Instance().template Get>(1); + const auto& vmul_d = + jit::KernelPool::Instance().template Get>(d); + const auto& vadd_d = + jit::KernelPool::Instance().template Get>(d); + + float* x_data = x.data(); + float* xref_data = xref.data(); + const float* ct_1_data = ct_1.data(); + float* ct_tgt_data = ct_tgt.data(); + float* ht_tgt_data = ht_tgt.data(); + float* ct_ref_data = ct_ref.data(); + float* ht_ref_data = ht_ref.data(); + // compute once to check correctness + lstm_ctht_ref(vsigmoid_3d, vtanh_d, vexp_1, d, xref_data, ct_1_data, + ct_ref_data, ht_ref_data); + ker->ComputeCtHt(x_data, ct_1_data, ct_tgt_data, ht_tgt_data); + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ct_tgt_data[i], ct_ref_data[i], 1e-3); + EXPECT_NEAR(ht_tgt_data[i], ht_ref_data[i], 1e-3); + } + + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + lstm_ctht_better(vsigmoid_3d, vtanh_d, vmul_d, vadd_d, d, xref_data, + ct_1_data, ct_ref_data, ht_ref_data); + } + auto tmkle = GetCurrentUS(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + lstm_ctht_ref(vsigmoid_3d, vtanh_d, vexp_1, d, xref_data, ct_1_data, + ct_ref_data, ht_ref_data); + } + auto trefe = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->ComputeCtHt(x_data, ct_1_data, ct_tgt_data, ht_tgt_data); + } + auto ttgte = GetCurrentUS(); + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat + << " us, better(jit) takes: " << (tmkle - tmkls) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + } +} + +void vscal_ref(const int n, const float a, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = a * x[i]; + } +} +void vscal_inp_ref(const int n, const float a, float* x) { + for (int i = 0; i < n; ++i) { + x[i] = a * x[i]; + } +} +#if defined __AVX__ || defined __AVX2__ +void vscal_intri8(const int n, const float a, const float* x, float* y) { + __m256 tmp; + __m256 scalar = _mm256_set1_ps(a); + tmp = _mm256_loadu_ps(x); + tmp = _mm256_mul_ps(tmp, scalar); + _mm256_storeu_ps(y, tmp); +} +void vscal_inp_intri8(const int n, const float a, float* x) { + __m256 tmp; + __m256 scalar = _mm256_set1_ps(a); + tmp = _mm256_loadu_ps(x); + tmp = _mm256_mul_ps(tmp, scalar); + _mm256_storeu_ps(x, tmp); +} +#endif + +#ifdef PADDLE_WITH_MKLML +void vscal_inp_mkl(const int n, const float a, float* x) { + paddle::platform::dynload::cblas_sscal(n, a, x, 1); +} +#endif + +TEST(JitKernel, vscal) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d), y(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data()); + std::memcpy(y.data(), x.data(), sizeof(float) * d); + float a = 2.f; + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* y_data = y.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_ref(d, a, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto trefs1 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_inp_ref(d, a, y_data); + } + auto trefe1 = GetCurrentUS(); + +#ifdef PADDLE_WITH_MKLML + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_inp_mkl(d, a, y_data); + } + auto tmkle = GetCurrentUS(); +#endif + +#if defined __AVX__ || defined __AVX2__ + if (d == 8) { + auto si0 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_intri8(d, a, x_data, zref_data); + } + auto si1 = GetCurrentUS(); + auto si2 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_inp_intri8(d, a, y_data); + } + auto si3 = GetCurrentUS(); + VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat + << " us, inplace: " << (si3 - si2) / repeat; + } +#endif + + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(a, x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + auto ttgts1 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(a, y_data); + } + auto ttgte1 = GetCurrentUS(); + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat + << " us, inplace takes: " << (trefe1 - trefs1) / repeat +#ifdef PADDLE_WITH_MKLML + << " us, mkl inplace takes: " << (tmkle - tmkls) / repeat << " us, " +#else + << " us, " +#endif + << "tgt takes: " << (ttgte - ttgts) / repeat + << "us, tgt inplace takes: " << (ttgte1 - ttgts1) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vmul_ref(const int n, const float* x, const float* y, float* z) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] * y[i]; + } +} + +#if defined __AVX__ || defined __AVX2__ +void vmul_intri8(const int n, const float* x, const float* y, float* z) { + __m256 tmpx, tmpy; + tmpx = _mm256_loadu_ps(x); + tmpy = _mm256_loadu_ps(y); + tmpx = _mm256_mul_ps(tmpx, tmpy); + _mm256_storeu_ps(z, tmpx); +} +#endif + +#ifdef PADDLE_WITH_MKLML +void vmul_mkl(const int n, const float* x, const float* y, float* z) { + paddle::platform::dynload::vsMul(n, x, y, z); +} +#endif + +TEST(JitKernel, vmul) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d), y(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data()); + RandomVec(d, y.data()); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + const float* y_data = y.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vmul_ref(d, x_data, y_data, zref_data); + } + auto trefe = GetCurrentUS(); + +#ifdef PADDLE_WITH_MKLML + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vmul_mkl(d, x_data, y_data, zref_data); + } + auto tmkle = GetCurrentUS(); +#endif + +#if defined __AVX__ || defined __AVX2__ + if (d == 8) { + auto si0 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vmul_intri8(d, x_data, y_data, zref_data); + } + auto si1 = GetCurrentUS(); + VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat; + } +#endif + + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, y_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat +#ifdef PADDLE_WITH_MKLML + << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, " +#else + << " us, " +#endif + << "tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vadd_ref(const int n, const float* x, const float* y, float* z) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] + y[i]; + } +} + +#if defined __AVX__ || defined __AVX2__ +void vadd_intri8(const int n, const float* x, const float* y, float* z) { + __m256 tmpx, tmpy; + tmpx = _mm256_loadu_ps(x); + tmpy = _mm256_loadu_ps(y); + tmpx = _mm256_add_ps(tmpx, tmpy); + _mm256_storeu_ps(z, tmpx); +} +#endif + +#ifdef PADDLE_WITH_MKLML +void vadd_mkl(const int n, const float* x, const float* y, float* z) { + paddle::platform::dynload::vsAdd(n, x, y, z); +} +#endif + +TEST(JitKernel, vadd) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d), y(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data()); + RandomVec(d, y.data()); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + const float* y_data = y.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vadd_ref(d, x_data, y_data, zref_data); + } + auto trefe = GetCurrentUS(); + +#ifdef PADDLE_WITH_MKLML + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vadd_mkl(d, x_data, y_data, zref_data); + } + auto tmkle = GetCurrentUS(); +#endif + +#if defined __AVX__ || defined __AVX2__ + if (d == 8) { + auto si0 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vadd_intri8(d, x_data, y_data, zref_data); + } + auto si1 = GetCurrentUS(); + VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat; + } +#endif + + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, y_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat +#ifdef PADDLE_WITH_MKLML + << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, " +#else + << " us, " +#endif + << "tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vaddrelu_ref(const int n, const float* x, const float* y, float* z) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] + y[i]; + z[i] = z[i] > 0 ? z[i] : 0; + } +} +void vaddrelu_better( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VAddKernel>& vadd, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VReluKernel>& vrelu, + const float* x, const float* y, float* z) { + vadd->Compute(x, y, z); + vrelu->Compute(z, z); +} + +TEST(JitKernel, vaddrelu) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d), y(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data()); + RandomVec(d, y.data()); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const auto& vadd = + jit::KernelPool::Instance().template Get>(d); + const auto& vrelu = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + const float* y_data = y.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vadd_ref(d, x_data, y_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vaddrelu_better(vadd, vrelu, x_data, y_data, zref_data); + } + auto tmkle = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, y_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat + << " us, better takes: " << (tmkle - tmkls) / repeat << " us, " + << "tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +TEST(JitKernel, pool) { + namespace jit = paddle::operators::math::jitkernel; + const int frame_size = 4; + std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh"; + const auto& plstm1 = + jit::KernelPool::Instance() + .template Get, const std::string&, + const std::string&, const std::string&>( + act_gate, act_cand, act_cell, frame_size, false); + const auto& plstm2 = + jit::KernelPool::Instance() + .template Get, const std::string&, + const std::string&, const std::string&>( + act_gate, act_cand, act_cell, frame_size, false); + const auto& peephole = + jit::KernelPool::Instance() + .template Get, const std::string&, + const std::string&, const std::string&>( + act_gate, act_cand, act_cell, frame_size, true); + EXPECT_TRUE(plstm1 != peephole); + + const auto& pvmul_f = + jit::KernelPool::Instance().template Get>(4); + EXPECT_TRUE(std::dynamic_pointer_cast(plstm2) != + std::dynamic_pointer_cast(pvmul_f)); + + const auto& pvmul_d = + jit::KernelPool::Instance().template Get>(4); + EXPECT_TRUE(std::dynamic_pointer_cast(pvmul_f) != + std::dynamic_pointer_cast(pvmul_d)); + + const auto& pvmul_from_key = jit::KernelPool::Instance().Get("vmulf4"); + EXPECT_EQ(pvmul_f, pvmul_from_key); + const auto& pvmul_from_key2 = jit::KernelPool::Instance().Get("vmulf5"); + EXPECT_TRUE(pvmul_from_key2 == nullptr); +} diff --git a/paddle/fluid/operators/math/selected_rows_functor.cc b/paddle/fluid/operators/math/selected_rows_functor.cc index 8e8baf49b2330e95ff1a868b0b0a03bc10d84484..08f57dd45ad76946cbcafb98a3414003ed9d67a9 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cc +++ b/paddle/fluid/operators/math/selected_rows_functor.cc @@ -12,10 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include #include #include -#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" namespace paddle { @@ -150,6 +151,45 @@ template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; +template +struct SelectedRowsSumTo { + void operator()(const platform::CPUDeviceContext& context, + const std::vector& input1, + const std::vector& input2_offsets, + framework::SelectedRows* input2) { + // Ensure all selected rows have the same height + size_t size = 0u; + for (auto iter = input1.begin(); iter != input1.end(); ++iter) { + auto& in_rows = (*iter)->rows(); + size += in_rows.end() - in_rows.begin(); + auto in1_height = (*iter)->height(); + PADDLE_ENFORCE_EQ(in1_height, input2->height()); + } + // concat rows + std::vector in2_rows; + in2_rows.reserve(in2_rows.size() + size); + for (auto iter = input1.begin(); iter != input1.end(); ++iter) { + const framework::Vector& in_rows = (*iter)->rows(); + in2_rows.insert(in2_rows.end(), in_rows.begin(), in_rows.end()); + } + input2->set_rows(in2_rows); + + auto* in2_value = input2->mutable_value(); + auto* in2_data = in2_value->data(); + auto blas = math::GetBlas(context); + size_t offset = 0u; + for (size_t i = 0u; i != input1.size(); ++i) { + auto& in_value = input1[i]->value(); + const auto* in_data = in_value.data(); + offset += input2_offsets[i]; + blas.VCOPY(in_value.numel(), in_data, in2_data + offset); + } + } +}; + +template struct SelectedRowsSumTo; +template struct SelectedRowsSumTo; + template struct SelectedRowsAddToTensor { void operator()(const platform::CPUDeviceContext& context, @@ -207,35 +247,45 @@ struct MergeAdd { const framework::SelectedRows& input, framework::SelectedRows* output) { framework::SelectedRows& out = *output; - auto input_rows = input.rows(); - std::set row_set(input_rows.begin(), input_rows.end()); - std::vector merge_rows(row_set.begin(), row_set.end()); + std::vector input_rows(input.rows()); - auto input_width = input.value().dims()[1]; - out.set_rows(merge_rows); + std::map> merge_row_map; + for (size_t i = 0; i < input_rows.size(); ++i) { + merge_row_map[input_rows[i]].push_back(i); + } + + std::vector merge_rows(merge_row_map.size()); + size_t idx = 0; + int64_t input_width = input.value().dims()[1]; out.set_height(input.height()); - out.mutable_value()->mutable_data( + + T* out_data = out.mutable_value()->mutable_data( framework::make_ddim( {static_cast(merge_rows.size()), input_width}), context.GetPlace()); - - math::SetConstant constant_functor; - constant_functor(context, out.mutable_value(), 0.0); - - auto* out_data = out.mutable_value()->data(); - auto* input_data = input.value().data(); - - for (size_t i = 0; i < input_rows.size(); i++) { - size_t out_i = FindPos(merge_rows, input_rows[i]); - for (int64_t j = 0; j < input_width; j++) { - out_data[out_i * input_width + j] += input_data[i * input_width + j]; + const T* in_data = input.value().data(); + + for (auto& row_pair : merge_row_map) { + auto* out_ptr = out_data + idx * input_width; + auto& rows = row_pair.second; + merge_rows[idx] = row_pair.first; + ++idx; + // rows.size() is always larger than 0 + std::memcpy(out_ptr, in_data + rows[0] * input_width, + sizeof(T) * input_width); + + for (size_t i = 1; i < rows.size(); ++i) { + auto* in_ptr = in_data + rows[i] * input_width; + for (int64_t j = 0; j < input_width; ++j) { + out_ptr[j] += in_ptr[j]; + } } } + + out.set_rows(merge_rows); } }; -template struct MergeAdd; -template struct MergeAdd; template struct MergeAdd; template struct MergeAdd; diff --git a/paddle/fluid/operators/math/selected_rows_functor.h b/paddle/fluid/operators/math/selected_rows_functor.h index aa419f74fcd2a53cdd734ec270bc154b78c9f2ff..900be86f91c6658a5265189a6745316c6471209e 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.h +++ b/paddle/fluid/operators/math/selected_rows_functor.h @@ -12,8 +12,14 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ #pragma once + +#include +#include + #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/platform/device_context.h" #define INLINE_FOR2(sizei, sizej) \ @@ -49,6 +55,15 @@ struct SelectedRowsAddTo { const int64_t input2_offset, framework::SelectedRows* input2); }; +// input2 = [all input in input1] + input2 +template +struct SelectedRowsSumTo { + void operator()(const DeviceContext& context, + const std::vector& input1, + const std::vector& input2_offsets, + framework::SelectedRows* input2); +}; + // input2 = input1 + input2 template struct SelectedRowsAddToTensor { @@ -70,6 +85,104 @@ struct MergeAdd { framework::SelectedRows* output); }; +template <> +struct MergeAdd { + framework::SelectedRows operator()(const platform::CPUDeviceContext& context, + const framework::SelectedRows& input) { + framework::SelectedRows out; + (*this)(context, input, &out); + return out; + } + + void operator()(const platform::CPUDeviceContext& context, + const framework::SelectedRows& input, + framework::SelectedRows* output) { + framework::SelectedRows& out = *output; + std::vector input_rows(input.rows()); + + std::map> merge_row_map; + for (size_t i = 0; i < input_rows.size(); ++i) { + merge_row_map[input_rows[i]].push_back(i); + } + + std::vector merge_rows(merge_row_map.size()); + size_t idx = 0; + int64_t input_width = input.value().dims()[1]; + out.set_height(input.height()); + + auto* out_data = out.mutable_value()->mutable_data( + framework::make_ddim( + {static_cast(merge_rows.size()), input_width}), + context.GetPlace()); + auto* in_data = input.value().data(); + + auto blas = GetBlas(context); + for (auto& row_pair : merge_row_map) { + auto* out_ptr = out_data + idx * input_width; + auto& rows = row_pair.second; + merge_rows[idx] = row_pair.first; + ++idx; + // rows.size() is always larger than 0 + blas.VCOPY(input_width, in_data + rows[0] * input_width, out_ptr); + + for (size_t i = 1; i < rows.size(); ++i) { + blas.AXPY(input_width, 1., in_data + rows[i] * input_width, out_ptr); + } + } + + out.set_rows(merge_rows); + } +}; + +template <> +struct MergeAdd { + framework::SelectedRows operator()(const platform::CPUDeviceContext& context, + const framework::SelectedRows& input) { + framework::SelectedRows out; + (*this)(context, input, &out); + return out; + } + + void operator()(const platform::CPUDeviceContext& context, + const framework::SelectedRows& input, + framework::SelectedRows* output) { + framework::SelectedRows& out = *output; + std::vector input_rows(input.rows()); + + std::map> merge_row_map; + for (size_t i = 0; i < input_rows.size(); ++i) { + merge_row_map[input_rows[i]].push_back(i); + } + + std::vector merge_rows(merge_row_map.size()); + size_t idx = 0; + int64_t input_width = input.value().dims()[1]; + out.set_height(input.height()); + + auto* out_data = out.mutable_value()->mutable_data( + framework::make_ddim( + {static_cast(merge_rows.size()), input_width}), + context.GetPlace()); + auto* in_data = input.value().data(); + + auto blas = GetBlas(context); + for (auto& row_pair : merge_row_map) { + auto* out_ptr = out_data + idx * input_width; + auto& rows = row_pair.second; + merge_rows[idx] = row_pair.first; + ++idx; + // rows.size() is always larger than 0 + blas.VCOPY(input_width, in_data + rows[0] * input_width, out_ptr); + + for (size_t i = 1; i < rows.size(); ++i) { + blas.AXPY(input_width, 1., in_data + rows[i] * input_width, out_ptr); + } + } + + out.set_rows(merge_rows); + } +}; + template struct Add { framework::SelectedRows operator()(const DeviceContext& context, diff --git a/paddle/fluid/operators/math/selected_rows_functor_test.cc b/paddle/fluid/operators/math/selected_rows_functor_test.cc index 70bed820ee58885861fa8c5535c931f258625572..835589356042b44c9fa5988aed726434fd66910a 100644 --- a/paddle/fluid/operators/math/selected_rows_functor_test.cc +++ b/paddle/fluid/operators/math/selected_rows_functor_test.cc @@ -219,3 +219,174 @@ TEST(selected_rows_functor, cpu_add_to) { // row9: 2.0 + 3.0 EXPECT_EQ(tensor1_data[9 * row_numel + 6], 5.0); } + +TEST(selected_rows_functor, cpu_merge_add_float) { + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + functor; + int64_t height = 10; + int64_t row_numel = 10; + + std::vector rows{0, 4, 4, 7}; + std::unique_ptr selected_rows{ + new paddle::framework::SelectedRows(rows, height)}; + auto* in_value = selected_rows->mutable_value(); + in_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows.size()), row_numel}), + cpu_place); + functor(ctx, in_value, 1.0); + + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + + paddle::operators::math::scatter::MergeAdd + merge_add_functor; + merge_add_functor(ctx, *selected_rows, output.get()); + + auto out_height = output->height(); + EXPECT_EQ(out_height, height); + + auto& out_rows = output->rows(); + EXPECT_EQ(out_rows[0], 0); + EXPECT_EQ(out_rows[1], 4); + EXPECT_EQ(out_rows[2], 7); + + auto* out_data = output->value().data(); + + EXPECT_EQ(out_data[0 * row_numel], 1.0); + EXPECT_EQ(out_data[1 * row_numel], 2.0); + EXPECT_EQ(out_data[2 * row_numel], 1.0); +} + +TEST(selected_rows_functor, cpu_merge_add_int) { + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + functor; + int64_t height = 10; + int64_t row_numel = 10; + + std::vector rows{0, 4, 4, 7}; + std::unique_ptr selected_rows{ + new paddle::framework::SelectedRows(rows, height)}; + auto* in_value = selected_rows->mutable_value(); + in_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows.size()), row_numel}), + cpu_place); + functor(ctx, in_value, 1); + + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + + paddle::operators::math::scatter::MergeAdd + merge_add_functor; + merge_add_functor(ctx, *selected_rows, output.get()); + + auto out_height = output->height(); + EXPECT_EQ(out_height, height); + + auto& out_rows = output->rows(); + EXPECT_EQ(out_rows[0], 0); + EXPECT_EQ(out_rows[1], 4); + EXPECT_EQ(out_rows[2], 7); + + auto* out_data = output->value().data(); + + EXPECT_EQ(out_data[0 * row_numel], 1); + EXPECT_EQ(out_data[1 * row_numel], 2); + EXPECT_EQ(out_data[2 * row_numel], 1); +} +TEST(selected_rows_functor, cpu_sum_to) { + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + functor; + int64_t height = 10; + int64_t row_numel = 10; + std::vector rows1{0, 4, 7}; + std::unique_ptr selected_rows1{ + new paddle::framework::SelectedRows(rows1, height)}; + auto* in1_value = selected_rows1->mutable_value(); + in1_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows1.size()), row_numel}), + cpu_place); + functor(ctx, in1_value, 1.0); + std::vector rows2{0, 5, 7, 9}; + std::unique_ptr selected_rows2{ + new paddle::framework::SelectedRows(rows2, height)}; + auto* in2_value = selected_rows2->mutable_value(); + in2_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows2.size()), row_numel}), + cpu_place); + functor(ctx, in2_value, 2.0); + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + output->set_height(height); + auto* out_value = output->mutable_value(); + // simplely concat two SelectedRows + out_value->mutable_data(paddle::framework::make_ddim({7, 10}), + cpu_place); + paddle::operators::math::SelectedRowsSumTo + sum_to_functor; + sum_to_functor(ctx, std::vector( + {selected_rows1.get(), selected_rows2.get()}), + std::vector({0, in1_value->numel()}), output.get()); + auto out_height = output->height(); + EXPECT_EQ(out_height, height); + auto& out_rows = output->rows(); + // input1 rows + EXPECT_EQ(out_rows[0], 0); + EXPECT_EQ(out_rows[1], 4); + EXPECT_EQ(out_rows[2], 7); + // input2 rows + EXPECT_EQ(out_rows[3], 0); + EXPECT_EQ(out_rows[4], 5); + EXPECT_EQ(out_rows[5], 7); + EXPECT_EQ(out_rows[6], 9); + auto* out_data = output->value().data(); + // input1 value + EXPECT_EQ(out_data[0 * row_numel + 0], 1.0); + EXPECT_EQ(out_data[0 * row_numel + 8], 1.0); + EXPECT_EQ(out_data[1 * row_numel + 1], 1.0); + EXPECT_EQ(out_data[2 * row_numel + 6], 1.0); + // input2 value + EXPECT_EQ(out_data[3 * row_numel + 3], 2.0); + EXPECT_EQ(out_data[3 * row_numel + 8], 2.0); + EXPECT_EQ(out_data[4 * row_numel + 4], 2.0); + EXPECT_EQ(out_data[5 * row_numel + 7], 2.0); + EXPECT_EQ(out_data[6 * row_numel + 9], 2.0); + std::unique_ptr tensor1{ + new paddle::framework::Tensor()}; + tensor1->mutable_data( + paddle::framework::make_ddim({height, row_numel}), cpu_place); + functor(ctx, tensor1.get(), 3.0); + paddle::operators::math::SelectedRowsAddToTensor< + paddle::platform::CPUDeviceContext, float> + add_to_tensor_functor; + add_to_tensor_functor(ctx, *output, tensor1.get()); + auto* tensor1_data = tensor1->data(); + // row0: 1.0 + 2.0 + 3.0 + EXPECT_EQ(tensor1_data[0 * row_numel + 0], 6.0); + // row1: 3.0 + EXPECT_EQ(tensor1_data[1 * row_numel + 1], 3.0); + // row4 : 1.0 + 3.0 + EXPECT_EQ(tensor1_data[4 * row_numel + 6], 4.0); + // row5: 2.0 + 3.0 + EXPECT_EQ(tensor1_data[5 * row_numel + 7], 5.0); + // row6: 3.0 + EXPECT_EQ(tensor1_data[6 * row_numel + 1], 3.0); + // row7: 1.0 + 2.0 + 3.0 + EXPECT_EQ(tensor1_data[7 * row_numel + 3], 6.0); + // row9: 2.0 + 3.0 + EXPECT_EQ(tensor1_data[9 * row_numel + 6], 5.0); +} diff --git a/paddle/fluid/operators/math/sequence_pooling.cc b/paddle/fluid/operators/math/sequence_pooling.cc index 69318a6598c8c69eceab7216df6382537153d34f..235b5405fb7d016f4bd8c738f75b303522183116 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cc +++ b/paddle/fluid/operators/math/sequence_pooling.cc @@ -12,9 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/math/sequence_pooling.h" #include + +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/sequence_pooling.h" namespace paddle { namespace operators { @@ -180,6 +182,7 @@ class SequencePoolFunctor { } auto lod = input.lod()[0]; auto& place = *context.eigen_device(); + auto blas = math::GetBlas(context); for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { Tensor in_t = input.Slice(static_cast(lod[i]), static_cast(lod[i + 1])); @@ -191,7 +194,14 @@ class SequencePoolFunctor { if (pooltype == "AVERAGE") { out_e.device(place) = in_e.mean(Eigen::array({{0}})); } else if (pooltype == "SUM") { - out_e.device(place) = in_e.sum(Eigen::array({{0}})); + if (h > 0) { + const T* in_data = in_t.data(); + T* out_data = out_t.mutable_data(context.GetPlace()); + blas.VCOPY(w, in_data, out_data); + for (int64_t r = 1; r != h; ++r) { + blas.AXPY(w, 1., in_data + r * w, out_data); + } + } } else if (pooltype == "SQRT") { out_e.device(place) = in_e.sum(Eigen::array({{0}})) / std::sqrt(static_cast(h)); @@ -223,6 +233,7 @@ class SequencePoolGradFunctor { } auto lod = in_grad->lod()[0]; auto& place = *context.eigen_device(); + auto blas = math::GetBlas(context); for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { auto in_g_t = in_grad->Slice(static_cast(lod[i]), static_cast(lod[i + 1])); @@ -237,7 +248,11 @@ class SequencePoolGradFunctor { if (pooltype == "AVERAGE") { in_g_e.device(place) = (out_g_e / static_cast(h)).broadcast(bcast); } else if (pooltype == "SUM") { - in_g_e.device(place) = (out_g_e).broadcast(bcast); + const T* out_g_data = out_g_t.data(); + T* in_g_data = in_g_t.mutable_data(context.GetPlace()); + for (int r = 0; r != h; ++r) { + blas.VCOPY(w, out_g_data, in_g_data + r * w); + } } else if (pooltype == "SQRT") { in_g_e.device(place) = (out_g_e / std::sqrt(static_cast(h))).broadcast(bcast); diff --git a/paddle/fluid/operators/momentum_op.cc b/paddle/fluid/operators/momentum_op.cc index 5f43c5810812260c4384349bdb709716c9a182f5..12b916fcebd425bd4a03d920f947829098a924a1 100644 --- a/paddle/fluid/operators/momentum_op.cc +++ b/paddle/fluid/operators/momentum_op.cc @@ -24,7 +24,7 @@ class MomentumOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(framework::InferShapeContext *ctx) const override { + void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("Param"), "Input(param) of Momentum should not be null."); PADDLE_ENFORCE(ctx->HasInput("Grad"), @@ -33,6 +33,11 @@ class MomentumOp : public framework::OperatorWithKernel { "Input(velocity) of Momentum should not be null."); PADDLE_ENFORCE(ctx->HasInput("LearningRate"), "Input(LearningRate) of Momentum should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of Momentum should not be null."); @@ -40,12 +45,15 @@ class MomentumOp : public framework::OperatorWithKernel { "Output(VelocityOut) of Momentum should not be null."); auto param_dim = ctx->GetInputDim("Param"); - PADDLE_ENFORCE_EQ( - param_dim, ctx->GetInputDim("Grad"), - "Param and Grad input of MomentumOp should have the same dimension."); - PADDLE_ENFORCE_EQ( - param_dim, ctx->GetInputDim("Velocity"), - "Param and Velocity of MomentumOp should have the same dimension."); + if (ctx->GetInputsVarType("Grad")[0] == + framework::proto::VarType::LOD_TENSOR) { + PADDLE_ENFORCE_EQ( + param_dim, ctx->GetInputDim("Grad"), + "Param and Grad input of MomentumOp should have the same dimension."); + PADDLE_ENFORCE_EQ( + param_dim, ctx->GetInputDim("Velocity"), + "Param and Velocity of MomentumOp should have the same dimension."); + } PADDLE_ENFORCE_EQ(framework::product(ctx->GetInputDim("LearningRate")), 1, "Learning_rate should be a scalar"); @@ -53,13 +61,34 @@ class MomentumOp : public framework::OperatorWithKernel { ctx->SetOutputDim("VelocityOut", param_dim); } framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext &ctx) const override { - auto input_data_type = - framework::ToDataType(ctx.Input("Param")->type()); + const framework::ExecutionContext& ctx) const override { + auto input_data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param")); return framework::OpKernelType(input_data_type, ctx.GetPlace()); } }; +class MomentumOpInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto input_var = op_desc.Input("Param")[0]; + for (auto& out_var : op_desc.Output("ParamOut")) { + if (block->FindRecursiveOrCreateVar(input_var).GetType() == + framework::proto::VarType::SELECTED_ROWS) { + block->FindRecursiveOrCreateVar(out_var).SetType( + framework::proto::VarType::SELECTED_ROWS); + } else if (block->FindRecursiveOrCreateVar(input_var).GetType() == + framework::proto::VarType::LOD_TENSOR) { + block->FindRecursiveOrCreateVar(out_var).SetType( + framework::proto::VarType::LOD_TENSOR); + } else { + PADDLE_THROW( + "Only support LodTensor and SelectedRows, Unexpected Input Type."); + } + } + } +}; + class MomentumOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -110,6 +139,9 @@ $$ } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(momentum, ops::MomentumOp, ops::MomentumOpMaker); -REGISTER_OP_CPU_KERNEL(momentum, ops::MomentumOpKernel, - ops::MomentumOpKernel); +REGISTER_OPERATOR(momentum, ops::MomentumOp, ops::MomentumOpMaker, + paddle::framework::EmptyGradOpMaker, + ops::MomentumOpInferVarType); +REGISTER_OP_CPU_KERNEL( + momentum, ops::MomentumOpKernel, + ops::MomentumOpKernel); diff --git a/paddle/fluid/operators/momentum_op.cu b/paddle/fluid/operators/momentum_op.cu index a3932db1f3a50305d585cd3d5e86fa1b527df78b..b68fec34d43f0dee834f1045f192d5c6089d9356 100644 --- a/paddle/fluid/operators/momentum_op.cu +++ b/paddle/fluid/operators/momentum_op.cu @@ -15,65 +15,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/momentum_op.h" -namespace paddle { -namespace operators { - -template -__global__ void MomentumKernel(const T* p, const T* g, const T* v, - const T* learning_rate, const T mu, - const int64_t num, bool use_nesterov, T* p_out, - T* v_out) { - T lr = learning_rate[0]; - if (use_nesterov) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; - i += blockDim.x * gridDim.x) { - T g_val = g[i]; - T v_new = v[i] * mu + g_val; - v_out[i] = v_new; - p_out[i] = p[i] - (g_val + v_new * mu) * lr; - } - } else { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; - i += blockDim.x * gridDim.x) { - T v_new = v[i] * mu + g[i]; - v_out[i] = v_new; - p_out[i] = p[i] - lr * v_new; - } - } -} - -template -class MomentumOpCUDAKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto param_out = ctx.Output("ParamOut"); - auto velocity_out = ctx.Output("VelocityOut"); - auto param = ctx.Input("Param"); - auto velocity = ctx.Input("Velocity"); - auto grad = ctx.Input("Grad"); - auto learning_rate = ctx.Input("LearningRate"); - - T* p_out = param_out->mutable_data(ctx.GetPlace()); - T* v_out = velocity_out->mutable_data(ctx.GetPlace()); - - T mu = static_cast(ctx.Attr("mu")); - bool use_nesterov = ctx.Attr("use_nesterov"); - - auto* p = param->data(); - auto* v = velocity->data(); - auto* g = grad->data(); - auto* lr = learning_rate->data(); - - int block = 512; - int grid = (param->numel() + block - 1) / block; - MomentumKernel<<>>( - p, g, v, lr, mu, param->numel(), use_nesterov, p_out, v_out); - } -}; - -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(momentum, ops::MomentumOpCUDAKernel, - ops::MomentumOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL( + momentum, ops::MomentumOpKernel, + ops::MomentumOpKernel); diff --git a/paddle/fluid/operators/momentum_op.h b/paddle/fluid/operators/momentum_op.h index 264726040fb566a52b8c0cdee0a1524197d2a675..6b4d00f56ca06c402c07ecf770a390e88ae3edf1 100644 --- a/paddle/fluid/operators/momentum_op.h +++ b/paddle/fluid/operators/momentum_op.h @@ -13,29 +13,48 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/algorithm.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { -template -class MomentumOpKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto param_out = ctx.Output("ParamOut"); - auto velocity_out = ctx.Output("VelocityOut"); - auto param = ctx.Input("Param"); - auto velocity = ctx.Input("Velocity"); - auto grad = ctx.Input("Grad"); - auto learning_rate = ctx.Input("LearningRate"); +using framework::Tensor; +using framework::SelectedRows; +struct NoNesterov; +struct UseNesterov; - param_out->mutable_data(ctx.GetPlace()); - velocity_out->mutable_data(ctx.GetPlace()); +template +class CPUDenseMomentumFunctor { + private: + const Tensor* param; + const Tensor* grad; + const Tensor* velocity; + const Tensor* learning_rate; + const T mu; + const T use_nesterov; + Tensor* param_out; + Tensor* velocity_out; - T mu = static_cast(ctx.Attr("mu")); - bool use_nesterov = ctx.Attr("use_nesterov"); + public: + CPUDenseMomentumFunctor(const Tensor* param, const Tensor* grad, + const Tensor* velocity, const Tensor* learning_rate, + const T mu, const bool use_nesterov, + Tensor* param_out, Tensor* velocity_out) + : param(param), + grad(grad), + velocity(velocity), + learning_rate(learning_rate), + mu(mu), + use_nesterov(use_nesterov), + param_out(param_out), + velocity_out(velocity_out) {} + inline void operator()() { auto p_out = framework::EigenVector::Flatten(*param_out); auto v_out = framework::EigenVector::Flatten(*velocity_out); @@ -53,5 +72,283 @@ class MomentumOpKernel : public framework::OpKernel { } }; +template +class DenseMomentumFunctor; + +// NOTE(dzh) for performance. +// avoid if/else in inside kernel, implement GPU UseNesterov/NoNesterov as two +// functor. +template +class DenseMomentumFunctor { + private: + const T* p_; + const T* g_; + const T* v_; + const T* lr_; + const T mu_; + const int64_t num_; + T* p_out_; + T* v_out_; + + public: + DenseMomentumFunctor(const T* p, const T* g, const T* v, + const T* learning_rate, const T mu, const int64_t num, + T* p_out, T* v_out) + : p_(p), + g_(g), + v_(v), + lr_(learning_rate), + mu_(mu), + num_(num), + p_out_(p_out), + v_out_(v_out) {} + inline HOSTDEVICE void operator()(size_t i) const { + // put memory access in register + const T p = p_[i]; + const T g = g_[i]; + const T lr = lr_[0]; + const T v = v_[i]; + T v_out = v * mu_ + g; + T p_out = p - (g + v_out * mu_) * lr; + // write reigster to memory + v_out_[i] = v_out; + p_out_[i] = p_out; + } +}; + +template +class DenseMomentumFunctor { + private: + const T* p_; + const T* g_; + const T* v_; + const T* lr_; + const T mu_; + const int64_t num_; + T* p_out_; + T* v_out_; + + public: + DenseMomentumFunctor(const T* p, const T* g, const T* v, + const T* learning_rate, const T mu, const int64_t num, + T* p_out, T* v_out) + : p_(p), + g_(g), + v_(v), + lr_(learning_rate), + mu_(mu), + num_(num), + p_out_(p_out), + v_out_(v_out) {} + inline HOSTDEVICE void operator()(size_t i) const { + // put memory access in register + const T p = p_[i]; + const T g = g_[i]; + const T lr = lr_[0]; + const T v = v_[i]; + T v_out = v * mu_ + g; + T p_out = p - lr * v_out; + // write reigster to memory + v_out_[i] = v_out; + p_out_[i] = p_out; + } +}; + +template +class SparseMomentumFunctor; + +template +class SparseMomentumFunctor { + private: + const T* p_; + const T* g_; + const T* v_; + const T* lr_; + const T mu_; + const int64_t* rows_; + const int64_t row_numel_; + const int64_t row_height_; + T* p_out_; + T* v_out_; + + public: + SparseMomentumFunctor(const T* p, const T* g, const T* v, const T* lr, + const T mu, const int64_t* rows, int64_t row_numel, + int64_t row_height, T* p_out, T* v_out) + : p_(p), + g_(g), + v_(v), + lr_(lr), + mu_(mu), + rows_(rows), + row_numel_(row_numel), + row_height_(row_height), + p_out_(p_out), + v_out_(v_out) {} + + inline HOSTDEVICE void operator()(size_t i) { + auto row_idx = + math::BinarySearch(rows_, row_height_, i / row_numel_); + T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0; + // put memory access in register + const T p = p_[i]; + const T lr = lr_[0]; + const T v = v_[i]; + T v_out = v * mu_ + g; + T p_out = p - (g + v_out * mu_) * lr; + // write reigster to memory + v_out_[i] = v_out; + p_out_[i] = p_out; + } +}; + +template +class SparseMomentumFunctor { + private: + const T* p_; + const T* g_; + const T* v_; + const T* lr_; + const T mu_; + const int64_t* rows_; + const int64_t row_numel_; + const int64_t row_height_; + T* p_out_; + T* v_out_; + + public: + SparseMomentumFunctor(const T* p, const T* g, const T* v, const T* lr, + const T mu, const int64_t* rows, int64_t row_numel, + int64_t row_height, T* p_out, T* v_out) + : p_(p), + g_(g), + v_(v), + lr_(lr), + mu_(mu), + rows_(rows), + row_numel_(row_numel), + row_height_(row_height), + p_out_(p_out), + v_out_(v_out) {} + + inline HOSTDEVICE void operator()(size_t i) { + auto row_idx = + math::BinarySearch(rows_, row_height_, i / row_numel_); + T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0; + // put memory access in register + const T p = p_[i]; + const T lr = lr_[0]; + const T v = v_[i]; + T v_out = v * mu_ + g; + T p_out = p - v_out * lr; + // write reigster to memory + v_out_[i] = v_out; + p_out_[i] = p_out; + } +}; + +template +class MomentumOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + T mu = static_cast(ctx.Attr("mu")); + bool use_nesterov = ctx.Attr("use_nesterov"); + + auto learning_rate = ctx.Input("LearningRate"); + auto param = ctx.Input("Param"); + auto param_out = ctx.Output("ParamOut"); + auto* velocity = ctx.Input("Velocity"); + auto velocity_out = ctx.Output("VelocityOut"); + param_out->mutable_data(ctx.GetPlace()); + velocity_out->mutable_data(ctx.GetPlace()); + + auto* grad_var = ctx.InputVar("Grad"); + if (grad_var->IsType()) { + auto grad = ctx.Input("Grad"); + if (platform::is_cpu_place(ctx.GetPlace())) { + CPUDenseMomentumFunctor functor(param, grad, velocity, learning_rate, + mu, use_nesterov, param_out, + velocity_out); + functor(); + } else if (platform::is_gpu_place(ctx.GetPlace())) { + platform::ForRange for_range( + static_cast(ctx.device_context()), + param->numel()); + if (use_nesterov) { + DenseMomentumFunctor functor( + param->data(), grad->data(), velocity->data(), + learning_rate->data(), mu, param->numel(), + param_out->mutable_data(ctx.GetPlace()), + velocity_out->mutable_data(ctx.GetPlace())); + for_range(functor); + + } else { + DenseMomentumFunctor functor( + param->data(), grad->data(), velocity->data(), + learning_rate->data(), mu, param->numel(), + param_out->mutable_data(ctx.GetPlace()), + velocity_out->mutable_data(ctx.GetPlace())); + for_range(functor); + } + } + + } else if (grad_var->IsType()) { + // sparse update embedding with selectedrows + auto grad = ctx.Input("Grad"); + + // sparse update maybe empty. + if (grad->rows().size() == 0) { + VLOG(3) << "Grad SelectedRows contains no data!"; + return; + } + auto* merged_grad = const_cast(ctx.scope()) + .Var() + ->GetMutable(); + math::scatter::MergeAdd merge_func; + merge_func(ctx.template device_context(), *grad, + merged_grad); + + const int64_t* rows = nullptr; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + rows = merged_grad->rows().CUDAData(ctx.GetPlace()); + } else { +#endif + rows = merged_grad->rows().data(); +#ifdef PADDLE_WITH_CUDA + } +#endif + int64_t row_numel = + merged_grad->value().numel() / merged_grad->rows().size(); + platform::ForRange for_range( + static_cast(ctx.device_context()), + param->numel()); + if (use_nesterov) { + SparseMomentumFunctor functor( + param->data(), merged_grad->value().data(), + velocity->data(), learning_rate->data(), mu, rows, row_numel, + static_cast(merged_grad->rows().size()), + param_out->mutable_data(ctx.GetPlace()), + velocity_out->mutable_data(ctx.GetPlace())); + for_range(functor); + + } else { + SparseMomentumFunctor functor( + param->data(), merged_grad->value().data(), + velocity->data(), learning_rate->data(), mu, rows, row_numel, + static_cast(merged_grad->rows().size()), + param_out->mutable_data(ctx.GetPlace()), + velocity_out->mutable_data(ctx.GetPlace())); + for_range(functor); + } + } else { + PADDLE_THROW( + string::Sprintf("MomentumOp only supports LoDTensor or SelectedRows " + "gradient, but the received Variable Type is %s", + grad_var->Type().name())); + } + } +}; + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/parallel_do_op.cc b/paddle/fluid/operators/parallel_do_op.cc index 97c36a83fc5eff421725d05f66fca05f5169d1bb..ab25628d45699dbcfc1fc5792958bae9e42e72a3 100644 --- a/paddle/fluid/operators/parallel_do_op.cc +++ b/paddle/fluid/operators/parallel_do_op.cc @@ -397,6 +397,24 @@ class ParallelDoGradOpShapeInference : public framework::InferShapeBase { } }; +class ParallelDoGradOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + framework::BlockDesc *sub_block = + boost::get(op_desc.GetAttr(kParallelBlock)); + for (auto &out_vars : op_desc.Outputs()) { + for (auto &out_var : out_vars.second) { + auto &var = block->FindRecursiveOrCreateVar(out_var); + auto sub_var = sub_block->FindRecursiveOrCreateVar(out_var); + if (sub_var.GetType() != var.GetType()) { + var.SetType(sub_var.GetType()); + } + } + } + } +}; + } // namespace operators } // namespace paddle @@ -404,4 +422,5 @@ REGISTER_OPERATOR(parallel_do, paddle::operators::ParallelDoOp, paddle::operators::ParallelDoOpProtoMaker, paddle::operators::ParallelDoGradOpDescMaker); REGISTER_OPERATOR(parallel_do_grad, paddle::operators::ParallelDoGradOp, - paddle::operators::ParallelDoGradOpShapeInference); + paddle::operators::ParallelDoGradOpShapeInference, + paddle::operators::ParallelDoGradOpVarTypeInference); diff --git a/paddle/fluid/operators/reader/blocking_queue.h b/paddle/fluid/operators/reader/blocking_queue.h index 28cc91a5ed5d74994e5b960a0a4dd3c6a5e6cdcc..51b980acb5a08d431d96a3a92479dec09119c27e 100644 --- a/paddle/fluid/operators/reader/blocking_queue.h +++ b/paddle/fluid/operators/reader/blocking_queue.h @@ -31,8 +31,8 @@ class BlockingQueue { // is a workaround and a simplified version of framework::Channel as it // doesn't support GPU and it implements on buffered blocking queue. public: - explicit BlockingQueue(size_t capacity) - : capacity_(capacity), closed_(false) { + explicit BlockingQueue(size_t capacity, bool speed_test_mode = false) + : capacity_(capacity), speed_test_mode_(speed_test_mode), closed_(false) { PADDLE_ENFORCE_GT( capacity_, 0, "The capacity of a reader::BlockingQueue must be greater than 0."); @@ -72,7 +72,9 @@ class BlockingQueue { if (!queue_.empty()) { PADDLE_ENFORCE_NOT_NULL(elem); *elem = queue_.front(); - queue_.pop_front(); + if (LIKELY(!speed_test_mode_)) { + queue_.pop_front(); + } send_cv_.notify_one(); return true; } else { @@ -114,6 +116,7 @@ class BlockingQueue { private: size_t capacity_; + bool speed_test_mode_; bool closed_; std::deque queue_; diff --git a/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h b/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h index 4f7cfc24ec035349f3c85e84d876ad9b5b5493a6..3f041ff7e4e32b407729a22aab25d3aab199fee0 100644 --- a/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h +++ b/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h @@ -33,8 +33,9 @@ class LoDTensorBlockingQueue { private: LoDTensorBlockingQueue(size_t capacity, - const std::vector& dims) - : queue_(capacity), dims_(dims) {} + const std::vector& dims, + bool speed_test_mode = false) + : queue_(capacity, speed_test_mode), dims_(dims) {} public: bool Push(const std::vector& lod_tensor_vec) { @@ -69,11 +70,12 @@ class LoDTensorBlockingQueue { class LoDTensorBlockingQueueHolder { public: - void InitOnce(size_t capacity, const std::vector& dims) { + void InitOnce(size_t capacity, const std::vector& dims, + bool speed_test_mode = false) { PADDLE_ENFORCE( queue_ == nullptr, "LoDTensorBlockingQueueHolder::InitOnce() can only be called once"); - queue_.reset(new LoDTensorBlockingQueue(capacity, dims)); + queue_.reset(new LoDTensorBlockingQueue(capacity, dims, speed_test_mode)); } inline const std::shared_ptr& GetQueue() const { diff --git a/paddle/fluid/operators/reader/reader_blocking_queue_test.cc b/paddle/fluid/operators/reader/reader_blocking_queue_test.cc index 7d1b381d56c8cdc1e79e594b18c1a1ed59ab5284..8cd505806056f1af33712e2c92b7661d87485708 100644 --- a/paddle/fluid/operators/reader/reader_blocking_queue_test.cc +++ b/paddle/fluid/operators/reader/reader_blocking_queue_test.cc @@ -217,3 +217,27 @@ TEST(BlockingQueue, MyClassTest) { q.Receive(&b); EXPECT_EQ(a.val_, b.val_); } + +TEST(BlockingQueue, speed_test_mode) { + size_t queue_size = 10; + BlockingQueue q1(queue_size, false); + for (size_t i = 0; i < queue_size; ++i) { + q1.Send(i); + } + size_t b; + for (size_t i = 0; i < queue_size; ++i) { + q1.Receive(&b); + EXPECT_EQ(b, i); + } + EXPECT_EQ(q1.Size(), 0UL); + + BlockingQueue q2(queue_size, true); + for (size_t i = 0; i < queue_size; ++i) { + q2.Send(i); + } + for (size_t i = 0; i < queue_size; ++i) { + q2.Receive(&b); + EXPECT_EQ(b, 0); + } + EXPECT_EQ(q2.Size(), queue_size); +} diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc index d72f85f2c44db2fa887732cfc05e1376a6a79e4a..500d86fec33830fc2cfb0412f1f2c7780d08eb02 100644 --- a/paddle/fluid/operators/reshape_op.cc +++ b/paddle/fluid/operators/reshape_op.cc @@ -164,7 +164,7 @@ dimension value will be copied from Input(X) at runtime. Note that the index of [2, 3, 4], Attr(shape) = [2, 3, 2, 0] is an invalid input. 3. Input(Shape) has a higher priority than Attr(shape) if it is provided, while -Attr(shape) still should be set correctly to gurantee shape inference in +Attr(shape) still should be set correctly to gurantee shape inference in compile-time. )DOC"); @@ -259,7 +259,6 @@ class Reshape2Op : public ReshapeOp { : ReshapeOp(type, inputs, outputs, attrs) {} void InferShape(framework::InferShapeContext *ctx) const override { - ReshapeOp::InferShape(ctx); PADDLE_ENFORCE(ctx->HasOutput("XShape"), "Output(XShape) of ReshapeOp should not be null."); const auto &x_dims = ctx->GetInputDim("X"); @@ -270,6 +269,8 @@ class Reshape2Op : public ReshapeOp { } ctx->SetOutputDim("XShape", framework::make_ddim(xshape_dims)); ctx->ShareLoD("X", /*->*/ "XShape"); + + ReshapeOp::InferShape(ctx); } }; diff --git a/paddle/fluid/operators/rmsprop_op.cc b/paddle/fluid/operators/rmsprop_op.cc index 2f773f222e50a440801b06a4fd997bf237b34772..f06f87e61d3a4d1fc8b864b9dd84e697fb12a006 100644 --- a/paddle/fluid/operators/rmsprop_op.cc +++ b/paddle/fluid/operators/rmsprop_op.cc @@ -32,6 +32,11 @@ class RmspropOp : public framework::OperatorWithKernel { "Input(Grad) of RmspropOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Moment"), "Input(Moment) of RmspropOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(param_out) of RmspropOp should not be null."); diff --git a/paddle/fluid/operators/rmsprop_op.h b/paddle/fluid/operators/rmsprop_op.h index 25ed32c5ebb2ff5be962ac1e3e38c970623d705c..797cd45fdcdbd5c3567d1676f37e148304ee6e2d 100644 --- a/paddle/fluid/operators/rmsprop_op.h +++ b/paddle/fluid/operators/rmsprop_op.h @@ -13,66 +13,254 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/algorithm.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { -using Tensor = framework::Tensor; template using EigenVector = framework::EigenVector; +template +struct DenseRmspropGradFunctor { + inline explicit DenseRmspropGradFunctor(const T *grad) : grad_(grad) {} + + HOSTDEVICE inline T operator()(int64_t idx) const { return grad_[idx]; } + + const T *grad_; +}; + +template +struct SparseRmspropGradFunctor { + inline SparseRmspropGradFunctor(const T *grad, const int64_t *rows, + int64_t row_numel, int64_t row_count) + : grad_(grad), + rows_(rows), + row_numel_(row_numel), + row_count_(row_count) {} + + HOSTDEVICE inline T operator()(int64_t idx) const { + auto row_idx = math::BinarySearch(rows_, row_count_, idx / row_numel_); + return row_idx >= 0 ? grad_[row_idx * row_numel_ + idx % row_numel_] : 0; + } + + const T *grad_; + const int64_t *rows_; + int64_t row_numel_; + int64_t row_count_; +}; + +template +struct UncenteredRmspropFunctor { + UncenteredRmspropFunctor(T *param, T *ms, T *mom, const T *lr, T rho, + T epsilon, T momentum, + const GradFunctor &grad_functor) + : param_(param), + ms_(ms), + mom_(mom), + lr_(lr), + rho_(rho), + epsilon_(epsilon), + momentum_(momentum), + grad_functor_(grad_functor) {} + + HOSTDEVICE inline void operator()(int64_t idx) const { + T g = grad_functor_(idx); + T ms_out = rho_ * ms_[idx] + (1 - rho_) * g * g; + T mom_out = momentum_ * mom_[idx] + lr_[0] * g / sqrt(ms_out + epsilon_); + param_[idx] -= mom_out; + ms_[idx] = ms_out; + mom_[idx] = mom_out; + } + + T *param_; + T *ms_; + T *mom_; + const T *lr_; + T rho_; + T epsilon_; + T momentum_; + GradFunctor grad_functor_; +}; + +template +struct CenteredRmspropFunctor { + CenteredRmspropFunctor(T *param, T *ms, T *mom, T *mean_grad, const T *lr, + T rho, T epsilon, T momentum, + const GradFunctor &grad_functor) + : param_(param), + ms_(ms), + mom_(mom), + mean_grad_(mean_grad), + lr_(lr), + rho_(rho), + epsilon_(epsilon), + momentum_(momentum), + grad_functor_(grad_functor) {} + + HOSTDEVICE inline void operator()(int64_t idx) const { + T g = grad_functor_(idx); + T ms_out = rho_ * ms_[idx] + (1 - rho_) * g * g; + T mg_out = rho_ * mean_grad_[idx] + (1 - rho_) * g; + T mom_out = momentum_ * mom_[idx] + + lr_[0] * g / sqrt(ms_out - mg_out * mg_out + epsilon_); + param_[idx] -= mom_out; + ms_[idx] = ms_out; + mom_[idx] = mom_out; + mean_grad_[idx] = mg_out; + } + + T *param_; + T *ms_; + T *mom_; + T *mean_grad_; + const T *lr_; + T rho_; + T epsilon_; + T momentum_; + GradFunctor grad_functor_; +}; + template class RmspropOpKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* param_out = ctx.Output("ParamOut"); - auto* moment_out = ctx.Output("MomentOut"); - auto* mean_square_out = ctx.Output("MeanSquareOut"); + void Compute(const framework::ExecutionContext &ctx) const override { + using LoDTensor = framework::LoDTensor; + auto *grad_var = ctx.InputVar("Grad"); + auto *param_out = ctx.Output("ParamOut"); + auto *moment_out = ctx.Output("MomentOut"); + auto *mean_square_out = ctx.Output("MeanSquareOut"); - auto grad = ctx.Input("Grad"); + auto epsilon = static_cast(ctx.Attr("epsilon")); + auto rho = static_cast(ctx.Attr("decay")); + auto momentum = static_cast(ctx.Attr("momentum")); + bool centered = ctx.Attr("centered"); - param_out->mutable_data(ctx.GetPlace()); - moment_out->mutable_data(ctx.GetPlace()); - mean_square_out->mutable_data(ctx.GetPlace()); + auto &p_tensor = *ctx.Input("Param"); + auto &ms_tensor = *ctx.Input("MeanSquare"); + auto &lr_tensor = *ctx.Input("LearningRate"); + auto &mom_tensor = *ctx.Input("Moment"); - float epsilon = ctx.Attr("epsilon"); - float rho = ctx.Attr("decay"); - float momentum = ctx.Attr("momentum"); - bool centered = ctx.Attr("centered"); + PADDLE_ENFORCE_EQ(&p_tensor, param_out, + "Param and ParamOut must be the same Tensor"); + PADDLE_ENFORCE_EQ(&mom_tensor, moment_out, + "Moment and MomentOut must be the same Tensor"); + PADDLE_ENFORCE_EQ(&ms_tensor, mean_square_out, + "MeanSquare and MeanSquareOut must be the same Tensor"); + + auto &dev_ctx = ctx.template device_context(); + size_t limit = static_cast(ms_tensor.numel()); + + if (grad_var->IsType()) { + auto &grad_tensor = grad_var->Get(); + + if (std::is_same::value) { + auto &place = + *ctx.template device_context().eigen_device(); + auto lr_value = lr_tensor.data()[0]; + + auto p = EigenVector::Flatten(p_tensor); + auto ms = EigenVector::Flatten(ms_tensor); + auto g = EigenVector::Flatten(grad_tensor); + auto mom = EigenVector::Flatten(mom_tensor); + + auto p_out = EigenVector::Flatten(*param_out); + auto mom_out = EigenVector::Flatten(*moment_out); + auto ms_out = EigenVector::Flatten(*mean_square_out); + + ms_out.device(place) = rho * ms + (1 - rho) * g * g; + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto mg = EigenVector::Flatten(mg_tensor); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + auto mg_out = EigenVector::Flatten(*mean_grad_out); + + mg_out.device(place) = rho * mg + (1 - rho) * g; + mom_out.device(place) = + momentum * mom + + lr_value * g / (ms_out - mg_out.square() + epsilon).sqrt(); + } else { + mom_out.device(place) = + momentum * mom + lr_value * g / (ms_out + epsilon).sqrt(); + } + p_out.device(place) = p - mom_out; + } else { + DenseRmspropGradFunctor grad_func(grad_tensor.data()); + platform::ForRange for_range(dev_ctx, limit); + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + for_range(CenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), + mean_grad_out->mutable_data(ctx.GetPlace()), + lr_tensor.data(), rho, epsilon, momentum, grad_func)); + } else { + for_range(UncenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } + } + } else if (grad_var->IsType()) { + auto &grad = grad_var->Get(); + auto *merged_grad = const_cast(ctx.scope()) + .Var() + ->GetMutable(); + + math::scatter::MergeAdd merge_func; + merge_func(dev_ctx, grad, merged_grad); + + platform::ForRange for_range(dev_ctx, limit); + const int64_t *rows; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + rows = merged_grad->rows().CUDAData(ctx.GetPlace()); + } else { +#endif + rows = merged_grad->rows().data(); +#ifdef PADDLE_WITH_CUDA + } +#endif + auto &merged_tensor = merged_grad->value(); + int64_t row_count = merged_grad->rows().size(); + int64_t row_numel = merged_tensor.numel() / row_count; + SparseRmspropGradFunctor grad_func(merged_tensor.data(), rows, + row_numel, row_count); - auto p = EigenVector::Flatten(*ctx.Input("Param")); - auto ms = EigenVector::Flatten(*ctx.Input("MeanSquare")); - auto lr = EigenVector::Flatten(*ctx.Input("LearningRate")); - auto g = EigenVector::Flatten(*grad); - auto mom = EigenVector::Flatten(*ctx.Input("Moment")); - - auto p_out = EigenVector::Flatten(*param_out); - auto mom_out = EigenVector::Flatten(*moment_out); - auto ms_out = EigenVector::Flatten(*mean_square_out); - auto& place = *ctx.template device_context().eigen_device(); - - Eigen::DSizes grad_dsize(static_cast(grad->numel())); - - ms_out.device(place) = rho * ms + (1 - rho) * g * g; - if (centered) { - auto mg = EigenVector::Flatten(*ctx.Input("MeanGrad")); - auto* mean_grad_out = ctx.Output("MeanGradOut"); - mean_grad_out->mutable_data(ctx.GetPlace()); - auto mg_out = EigenVector::Flatten(*mean_grad_out); - - mg_out.device(place) = rho * mg + (1 - rho) * g; - mom_out.device(place) = momentum * mom + - lr.broadcast(grad_dsize) * g / - (ms_out - mg_out.square() + epsilon).sqrt(); + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + for_range(CenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), + mean_grad_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } else { + for_range(UncenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } } else { - mom_out.device(place) = - momentum * mom + - lr.broadcast(grad_dsize) * g / (ms_out + epsilon).sqrt(); + PADDLE_THROW("RMSProp only supports LoDTensor or SelectedRows gradient"); } - p_out.device(place) = p - mom_out; } }; diff --git a/paddle/fluid/operators/roi_align_op.cc b/paddle/fluid/operators/roi_align_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c57a34c3a745e8fc03ca57dce478ecf60058a9a9 --- /dev/null +++ b/paddle/fluid/operators/roi_align_op.cc @@ -0,0 +1,166 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/roi_align_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +class ROIAlignOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ROIAlignOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("ROIs"), + "Input(ROIs) of ROIAlignOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ROIAlignOp should not be null."); + auto input_dims = ctx->GetInputDim("X"); + auto rois_dims = ctx->GetInputDim("ROIs"); + + PADDLE_ENFORCE(input_dims.size() == 4, + "The format of input tensor is NCHW."); + PADDLE_ENFORCE(rois_dims.size() == 2, + "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]."); + PADDLE_ENFORCE(rois_dims[1] == 4, + "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]."); + int pooled_height = ctx->Attrs().Get("pooled_height"); + int pooled_width = ctx->Attrs().Get("pooled_width"); + float spatial_scale = ctx->Attrs().Get("spatial_scale"); + + PADDLE_ENFORCE_GT(pooled_height, 0, + "The pooled output height must greater than 0"); + PADDLE_ENFORCE_GT(pooled_width, 0, + "The pooled output width must greater than 0"); + PADDLE_ENFORCE_GT(spatial_scale, 0.0f, + "The spatial scale must greater than 0"); + + auto out_dims = input_dims; + out_dims[0] = rois_dims[0]; + out_dims[1] = input_dims[1]; + out_dims[2] = pooled_height; + out_dims[3] = pooled_width; + + ctx->SetOutputDim("Out", out_dims); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + +class ROIAlignGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "The GRAD@Out of ROIAlignGradOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName("X")), + "The GRAD@X of ROIAlignGradOp should not be null."); + ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X")); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + +class ROIAlignOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor), " + "The input of ROIAlignOp. " + "The format of input tensor is NCHW. Where N is batch size, " + "C is the number of input channels, " + "H is the height of the feature, and " + "W is the width of the feature."); + AddInput("ROIs", + "(LoDTensor), " + "ROIs (Regions of Interest) to pool over. " + "should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]. " + "(x1, y1) is the top left coordinates, and " + "(x2, y2) is the bottom right coordinates."); + AddOutput("Out", + "(Tensor), " + "The output of ROIAlignOp is a 4-D tensor with shape " + "(num_rois, channels, pooled_h, pooled_w)."); + AddAttr("spatial_scale", + "(float, default 1.0), " + "Multiplicative spatial scale factor " + "to translate ROI coords from their input scale " + "to the scale used when pooling.") + .SetDefault(1.0); + AddAttr("pooled_height", + "(int, default 1), " + "The pooled output height.") + .SetDefault(1); + AddAttr("pooled_width", + "(int, default 1), " + "The pooled output width.") + .SetDefault(1); + AddAttr("sampling_ratio", + "(int,default -1)," + "number of sampling points in the interpolation grid" + "If <=0, then grid points are adaptive to roi_width " + "and pooled_w, likewise for height") + .SetDefault(-1); + AddComment(R"DOC( +**RoIAlign Operator** + +Region of interest align (also known as RoI align) is to perform +bilinear interpolation on inputs of nonuniform sizes to obtain +fixed-size feature maps (e.g. 7*7) + +Dividing each region proposal into equal-sized sections with +the pooled_width and pooled_height. Location remains the origin +result. + +In each ROI bin, the value of the four regularly sampled locations +are computed directly through bilinear interpolation. The output is +the mean of four locations. +Thus avoid the misaligned problem. + )DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(roi_align, ops::ROIAlignOp, ops::ROIAlignOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(roi_align_grad, ops::ROIAlignGradOp); +REGISTER_OP_CPU_KERNEL( + roi_align, + ops::CPUROIAlignOpKernel, + ops::CPUROIAlignOpKernel); +REGISTER_OP_CPU_KERNEL( + roi_align_grad, + ops::CPUROIAlignGradOpKernel, + ops::CPUROIAlignGradOpKernel); diff --git a/paddle/fluid/operators/roi_align_op.cu b/paddle/fluid/operators/roi_align_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..bcec6f3563df7f4e1e48554cc891d596f9e56024 --- /dev/null +++ b/paddle/fluid/operators/roi_align_op.cu @@ -0,0 +1,353 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/roi_align_op.h" +#include "paddle/fluid/platform/cuda_primitives.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +static constexpr int kNumCUDAThreads = 512; +static constexpr int kNumMaxinumNumBlocks = 4096; + +static inline int NumBlocks(const int N) { + return std::min((N + kNumCUDAThreads - 1) / kNumCUDAThreads, + kNumMaxinumNumBlocks); +} + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__device__ T BilinearInterpolate(const T* input_data, const int height, + const int width, T y, T x) { + if (y < -1.0 || y > height || x < -1.0 || x > width) { + return 0; + } + y = y <= 0 ? 0 : y; + x = x <= 0 ? 0 : x; + int y_low = static_cast(y); + int x_low = static_cast(x); + int y_high; + int x_high; + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = static_cast(y_low); + } else { + y_high = y_low + 1; + } + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = static_cast(x_low); + } else { + x_high = x_low + 1; + } + T ly = y - y_low, lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + + T v1 = input_data[y_low * width + x_low]; + T v2 = input_data[y_low * width + x_high]; + T v3 = input_data[y_high * width + x_low]; + T v4 = input_data[y_high * width + x_high]; + T w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + + T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + return val; +} + +template +__device__ void BilinearInterpolateGradient(const int height, const int width, + T y, T x, T* w1, T* w2, T* w3, + T* w4, int* x_low, int* x_high, + int* y_low, int* y_high) { + if (y < -1.0 || y > height || x < -1.0 || x > width) { + return; + } + + y = y <= 0 ? 0 : y; + x = x <= 0 ? 0 : x; + *y_low = static_cast(y); + *x_low = static_cast(x); + if (*y_low >= height - 1) { + *y_high = *y_low = height - 1; + y = static_cast(*y_low); + } else { + *y_high = *y_low + 1; + } + if (*x_low >= width - 1) { + *x_high = *x_low = width - 1; + x = static_cast(*x_low); + } else { + *x_high = *x_low + 1; + } + T ly = y - *y_low, lx = x - *x_low; + T hy = 1. - ly, hx = 1. - lx; + *w1 = hy * hx, *w2 = hy * lx, *w3 = ly * hx, *w4 = ly * lx; + + return; +} + +template +__global__ void GPUROIAlignForward( + const int nthreads, const T* input_data, const T* input_rois, + const float spatial_scale, const int channels, const int height, + const int width, const int pooled_height, const int pooled_width, + const int sampling_ratio, int* roi_batch_id_data, T* output_data) { + CUDA_1D_KERNEL_LOOP(i, nthreads) { + int pw = i % pooled_width; + int ph = (i / pooled_width) % pooled_height; + int c = (i / pooled_width / pooled_height) % channels; + int n = i / pooled_width / pooled_height / channels; + + const T* offset_input_rois = input_rois + n * kROISize; + int roi_batch_ind = roi_batch_id_data[n]; + + T roi_xmin = offset_input_rois[0] * spatial_scale; + T roi_ymin = offset_input_rois[1] * spatial_scale; + T roi_xmax = offset_input_rois[2] * spatial_scale; + T roi_ymax = offset_input_rois[3] * spatial_scale; + + T roi_width = max(roi_xmax - roi_xmin, static_cast(1.)); + T roi_height = max(roi_ymax - roi_ymin, static_cast(1.)); + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + + const T* offset_input_data = + input_data + (roi_batch_ind * channels + c) * height * width; + + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); + int roi_bin_grid_w = + (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); + const T count = roi_bin_grid_h * roi_bin_grid_w; + T output_val = 0; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + const T y = roi_ymin + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T x = roi_xmin + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + T val = BilinearInterpolate(offset_input_data, height, width, y, x); + output_val += val; + } + } + output_val /= count; + output_data[i] = output_val; + } +} + +template +__global__ void GPUROIAlignBackward(const int nthreads, const T* input_rois, + const T* out_grad, const int num_rois, + const float spatial_scale, + const int channels, const int height, + const int width, const int pooled_height, + const int pooled_width, + const int sampling_ratio, + int* roi_batch_id_data, T* input_grad) { + CUDA_1D_KERNEL_LOOP(i, nthreads) { + int pw = i % pooled_width; + int ph = (i / pooled_width) % pooled_height; + int c = (i / pooled_width / pooled_height) % channels; + int n = i / pooled_width / pooled_height / channels; + const T* offset_input_rois = input_rois + n * kROISize; + int roi_batch_ind = roi_batch_id_data[n]; + + T roi_xmin = offset_input_rois[0] * spatial_scale; + T roi_ymin = offset_input_rois[1] * spatial_scale; + T roi_xmax = offset_input_rois[2] * spatial_scale; + T roi_ymax = offset_input_rois[3] * spatial_scale; + + T roi_width = max(roi_xmax - roi_xmin, static_cast(1.)); + T roi_height = max(roi_ymax - roi_ymin, static_cast(1.)); + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + + T* offset_input_grad = + input_grad + (roi_batch_ind * channels + c) * height * width; + + const T* offset_out_grad = + out_grad + (n * channels + c) * pooled_height * pooled_width; + const T out_grad_this_bin = offset_out_grad[ph * pooled_width + pw]; + + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); + int roi_bin_grid_w = + (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); + + const T count = roi_bin_grid_h * roi_bin_grid_w; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + const T y = roi_ymin + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T x = roi_xmin + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + T w1 = 0, w2 = 0, w3 = 0, w4 = 0; + int x_low = -1, x_high = -1, y_low = -1, y_high = -1; + BilinearInterpolateGradient(height, width, y, x, &w1, &w2, &w3, &w4, + &x_low, &x_high, &y_low, &y_high); + T diff1 = out_grad_this_bin * w1 / count; + T diff2 = out_grad_this_bin * w2 / count; + T diff3 = out_grad_this_bin * w3 / count; + T diff4 = out_grad_this_bin * w4 / count; + if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) { + platform::CudaAtomicAdd(offset_input_grad + y_low * width + x_low, + diff1); + platform::CudaAtomicAdd(offset_input_grad + y_low * width + x_high, + diff2); + platform::CudaAtomicAdd(offset_input_grad + y_high * width + x_low, + diff3); + platform::CudaAtomicAdd(offset_input_grad + y_high * width + x_high, + diff4); + } + } + } + } +} + +template +class GPUROIAlignOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* rois = ctx.Input("ROIs"); + auto* out = ctx.Output("Out"); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + auto sampling_ratio = ctx.Attr("sampling_ratio"); + + auto in_dims = in->dims(); + int batch_size = in_dims[0]; + int channels = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + + int rois_num = rois->dims()[0]; + + if (rois_num == 0) return; + + int output_size = out->numel(); + int blocks = NumBlocks(output_size); + int threads = kNumCUDAThreads; + + Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(platform::CPUPlace()); + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + PADDLE_ENFORCE_EQ( + rois_batch_size, batch_size, + "The rois_batch_size and imgs batch_size must be the same."); + int rois_num_with_lod = rois_lod[rois_batch_size]; + PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod, + "The rois_num from input and lod must be the same."); + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + Tensor roi_batch_id_list_gpu; + framework::TensorCopySync(roi_batch_id_list, ctx.GetPlace(), + &roi_batch_id_list_gpu); + GPUROIAlignForward< + T><<>>( + output_size, in->data(), rois->data(), spatial_scale, channels, + height, width, pooled_height, pooled_width, sampling_ratio, + roi_batch_id_list_gpu.data(), + out->mutable_data(ctx.GetPlace())); + } +}; + +template +class GPUROIAlignGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* rois = ctx.Input("ROIs"); + + auto* out_grad = ctx.Input(framework::GradVarName("Out")); + auto* in_grad = ctx.Output(framework::GradVarName("X")); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + auto sampling_ratio = ctx.Attr("sampling_ratio"); + + int rois_num = rois->dims()[0]; + int channels = in->dims()[1]; + int height = in->dims()[2]; + int width = in->dims()[3]; + + if (!in_grad) { + return; + } + Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(platform::CPUPlace()); + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + Tensor roi_batch_id_list_gpu; + framework::TensorCopySync(roi_batch_id_list, ctx.GetPlace(), + &roi_batch_id_list_gpu); + + in_grad->mutable_data(ctx.GetPlace()); + math::SetConstant set_zero; + set_zero(ctx.cuda_device_context(), in_grad, static_cast(0)); + + int output_grad_size = out_grad->numel(); + int blocks = NumBlocks(output_grad_size); + int threads = kNumCUDAThreads; + + if (output_grad_size > 0) { + GPUROIAlignBackward< + T><<>>( + output_grad_size, rois->data(), out_grad->data(), rois_num, + spatial_scale, channels, height, width, pooled_height, pooled_width, + sampling_ratio, roi_batch_id_list_gpu.data(), + in_grad->mutable_data(ctx.GetPlace())); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + roi_align, + ops::GPUROIAlignOpKernel, + ops::GPUROIAlignOpKernel); +REGISTER_OP_CUDA_KERNEL( + roi_align_grad, + ops::GPUROIAlignGradOpKernel, + ops::GPUROIAlignGradOpKernel); diff --git a/paddle/fluid/operators/roi_align_op.h b/paddle/fluid/operators/roi_align_op.h new file mode 100644 index 0000000000000000000000000000000000000000..a18aee1b86283cbb48f0b804ccfc476d7cd78f3b --- /dev/null +++ b/paddle/fluid/operators/roi_align_op.h @@ -0,0 +1,332 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +static constexpr int kROISize = 4; + +template +void PreCalcForBilinearInterpolate( + const platform::DeviceContext& ctx, const int height, const int width, + const int pooled_height, const int pooled_width, const int iy_upper, + const int ix_upper, T roi_ymin, T roi_xmin, T bin_size_h, T bin_size_w, + int roi_bin_grid_h, int roi_bin_grid_w, Tensor* pre_pos, Tensor* pre_w) { + int pre_calc_index = 0; + int* pre_pos_data = pre_pos->mutable_data(ctx.GetPlace()); + T* pre_w_data = pre_w->mutable_data(ctx.GetPlace()); + for (int ph = 0; ph < pooled_height; ph++) { + for (int pw = 0; pw < pooled_width; pw++) { + for (int iy = 0; iy < iy_upper; iy++) { + // calculate y of sample points + T y = roi_ymin + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); + // calculate x of samle points + for (int ix = 0; ix < ix_upper; ix++) { + T x = roi_xmin + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + // deal with elements out of map + if (y < -1.0 || y > height || x < -1.0 || x > width) { + for (int i = 0; i < kROISize; ++i) { + pre_pos_data[i + pre_calc_index * kROISize] = 0; + pre_w_data[i + pre_calc_index * kROISize] = 0; + } + pre_calc_index += 1; + continue; + } + y = y <= 0 ? 0 : y; + x = x <= 0 ? 0 : x; + + int y_low = static_cast(y); + int x_low = static_cast(x); + int y_high; + int x_high; + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = static_cast(y_low); + } else { + y_high = y_low + 1; + } + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = static_cast(x_low); + } else { + x_high = x_low + 1; + } + T ly = y - y_low, lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + pre_pos_data[pre_calc_index * kROISize] = y_low * width + x_low; + pre_pos_data[pre_calc_index * kROISize + 1] = y_low * width + x_high; + pre_pos_data[pre_calc_index * kROISize + 2] = y_high * width + x_low; + pre_pos_data[pre_calc_index * kROISize + 3] = y_high * width + x_high; + pre_w_data[pre_calc_index * kROISize] = hy * hx; + pre_w_data[pre_calc_index * kROISize + 1] = hy * lx; + pre_w_data[pre_calc_index * kROISize + 2] = ly * hx; + pre_w_data[pre_calc_index * kROISize + 3] = ly * lx; + pre_calc_index += 1; + } + } + } + } +} + +template +void bilinear_interpolate_gradient(const int height, const int width, T y, T x, + const T out_grad_this_bin, const T count, + T* batch_grad_data) { + int x_low, y_low, x_high, y_high; + T w1, w2, w3, w4; + if (y < -1.0 || y > height || x < -1.0 || x > width) { + w1 = w2 = w3 = w4 = 0; + x_low = x_high = y_low = y_high = -1; + return; + } + y = y <= 0 ? 0 : y; + x = x <= 0 ? 0 : x; + y_low = static_cast(y); + x_low = static_cast(x); + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = static_cast(y_low); + } else { + y_high = y_low + 1; + } + + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = static_cast(x_low); + } else { + x_high = x_low + 1; + } + + T ly = y - y_low, lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + T diff1 = out_grad_this_bin * w1 / count; + T diff2 = out_grad_this_bin * w2 / count; + T diff3 = out_grad_this_bin * w3 / count; + T diff4 = out_grad_this_bin * w4 / count; + if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) { + *(batch_grad_data + y_low * width + x_low) += diff1; + *(batch_grad_data + y_low * width + x_high) += diff2; + *(batch_grad_data + y_high * width + x_low) += diff3; + *(batch_grad_data + y_high * width + x_high) += diff4; + } +} + +template +class CPUROIAlignOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* rois = ctx.Input("ROIs"); + auto* out = ctx.Output("Out"); + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + auto sampling_ratio = ctx.Attr("sampling_ratio"); + + auto& dev_ctx = ctx.template device_context(); + + auto in_dims = in->dims(); + int batch_size = in_dims[0]; + int channels = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + int rois_num = rois->dims()[0]; + + auto in_stride = framework::stride(in_dims); + auto roi_stride = framework::stride(rois->dims()); + auto out_stride = framework::stride(out->dims()); + + const T* input_data = in->data(); + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(ctx.GetPlace()); + + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + PADDLE_ENFORCE_EQ( + rois_batch_size, batch_size, + "The rois_batch_size and imgs batch_size must be the same."); + int rois_num_with_lod = rois_lod[rois_batch_size]; + PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod, + "The rois_num from input and lod must be the same."); + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + T* output_data = out->mutable_data(ctx.GetPlace()); + const T* rois_data = rois->data(); + for (int n = 0; n < rois_num; ++n) { + int roi_batch_id = roi_batch_id_data[n]; + T roi_xmin = rois_data[0] * spatial_scale; + T roi_ymin = rois_data[1] * spatial_scale; + T roi_xmax = rois_data[2] * spatial_scale; + T roi_ymax = rois_data[3] * spatial_scale; + + T roi_width = std::max(roi_xmax - roi_xmin, static_cast(1.)); + T roi_height = std::max(roi_ymax - roi_ymin, static_cast(1.)); + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + const T* batch_data = input_data + roi_batch_id * in_stride[0]; + + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); + int roi_bin_grid_w = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_width / pooled_width); + const T count = roi_bin_grid_h * roi_bin_grid_w; + Tensor pre_pos; + Tensor pre_w; + int pre_size = count * out_stride[1]; + pre_pos.Resize({pre_size, kROISize}); + pre_w.Resize({pre_size, kROISize}); + + PreCalcForBilinearInterpolate( + dev_ctx, height, width, pooled_height, pooled_width, roi_bin_grid_h, + roi_bin_grid_w, roi_ymin, roi_xmin, bin_size_h, bin_size_w, + roi_bin_grid_h, roi_bin_grid_w, &pre_pos, &pre_w); + const int* pre_pos_data = pre_pos.data(); + const T* pre_w_data = pre_w.data(); + for (int c = 0; c < channels; c++) { + int pre_calc_index = 0; + for (int ph = 0; ph < pooled_height; ph++) { + for (int pw = 0; pw < pooled_width; pw++) { + const int pool_index = ph * pooled_width + pw; + T output_val = 0; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + for (int i = 0; i < kROISize; i++) { + int pos = pre_pos_data[pre_calc_index * kROISize + i]; + T w = pre_w_data[pre_calc_index * kROISize + i]; + output_val += w * batch_data[pos]; + } + pre_calc_index += 1; + } + } + output_val /= count; + output_data[pool_index] = output_val; + } + } + batch_data += in_stride[1]; + output_data += out_stride[1]; + } + rois_data += roi_stride[0]; + } + } +}; + +template +class CPUROIAlignGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* rois = ctx.Input("ROIs"); + auto* out_grad = + ctx.Input(framework::GradVarName("Out")); + auto* in_grad = ctx.Output(framework::GradVarName("X")); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + auto sampling_ratio = ctx.Attr("sampling_ratio"); + auto in_dims = in->dims(); + if (!in_grad) { + return; + } + int channels = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + int rois_num = rois->dims()[0]; + Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(ctx.GetPlace()); + + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + + const T* rois_data = rois->data(); + const T* out_grad_data = out_grad->data(); + T* in_grad_data = in_grad->mutable_data(ctx.GetPlace()); + + auto in_stride = framework::stride(in->dims()); + auto roi_stride = framework::stride(rois->dims()); + auto out_stride = framework::stride(out_grad->dims()); + + for (int n = 0; n < rois_num; ++n) { + int roi_batch_idx = roi_batch_id_data[n]; + T roi_xmin = rois_data[0] * spatial_scale; + T roi_ymin = rois_data[1] * spatial_scale; + T roi_xmax = rois_data[2] * spatial_scale; + T roi_ymax = rois_data[3] * spatial_scale; + T roi_width = std::max(roi_xmax - roi_xmin, static_cast(1.)); + T roi_height = std::max(roi_ymax - roi_ymin, static_cast(1.)); + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + for (int c = 0; c < channels; ++c) { + T* batch_grad_data = + in_grad_data + roi_batch_idx * in_stride[0] + c * in_stride[1]; + const T* batch_out_grad_data = + out_grad_data + n * out_stride[0] + c * out_stride[1]; + for (int ph = 0; ph < pooled_height; ++ph) { + for (int pw = 0; pw < pooled_width; ++pw) { + int pool_index = ph * pooled_width + pw; + T out_grad_this_bin = batch_out_grad_data[pool_index]; + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); + int roi_bin_grid_w = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_width / pooled_width); + T count = roi_bin_grid_h * roi_bin_grid_w; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + const T y = roi_ymin + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T x = roi_xmin + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + bilinear_interpolate_gradient(height, width, y, x, + out_grad_this_bin, count, + batch_grad_data); + } + } + } + } + } + rois_data += roi_stride[0]; + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/roi_pool_op.cc b/paddle/fluid/operators/roi_pool_op.cc index d6d209d5de041500a9b4893d70800a58e8ee1e1d..8e29761ec208764e263e357a0b3c9456c932d093 100644 --- a/paddle/fluid/operators/roi_pool_op.cc +++ b/paddle/fluid/operators/roi_pool_op.cc @@ -174,4 +174,4 @@ REGISTER_OP_CPU_KERNEL( REGISTER_OP_CPU_KERNEL( roi_pool_grad, ops::CPUROIPoolGradOpKernel, - ops::CPUROIPoolOpKernel); + ops::CPUROIPoolGradOpKernel); diff --git a/paddle/fluid/operators/roi_pool_op.cu b/paddle/fluid/operators/roi_pool_op.cu index 46e20285db6d7acd39dead3994409645adddf494..75c3dd6bc498e35c6249f79a1c24cfe17316670e 100644 --- a/paddle/fluid/operators/roi_pool_op.cu +++ b/paddle/fluid/operators/roi_pool_op.cu @@ -249,4 +249,4 @@ REGISTER_OP_CUDA_KERNEL( REGISTER_OP_CUDA_KERNEL( roi_pool_grad, ops::GPUROIPoolGradOpKernel, - ops::GPUROIPoolOpKernel); + ops::GPUROIPoolGradOpKernel); diff --git a/paddle/fluid/operators/sequence_concat_op.cc b/paddle/fluid/operators/sequence_concat_op.cc index 397a3182953e3f1afaeadeff6d53a4f22fb95d26..3234b60861da3d0c6a8434eb11fd0488a95e171f 100644 --- a/paddle/fluid/operators/sequence_concat_op.cc +++ b/paddle/fluid/operators/sequence_concat_op.cc @@ -90,11 +90,13 @@ REGISTER_OPERATOR(sequence_concat, paddle::framework::OperatorWithKernel, paddle::framework::DefaultGradOpDescMaker); template using Kernel = op::SeqConcatKernel; -REGISTER_OP_CPU_KERNEL(sequence_concat, Kernel, Kernel); +REGISTER_OP_CPU_KERNEL(sequence_concat, Kernel, Kernel, + Kernel); + REGISTER_OPERATOR(sequence_concat_grad, paddle::framework::OperatorWithKernel, op::SeqConcatGradShapeInferer); template using GradKernel = op::SeqConcatGradKernel; REGISTER_OP_CPU_KERNEL(sequence_concat_grad, GradKernel, - GradKernel); + GradKernel, GradKernel); diff --git a/paddle/fluid/operators/sequence_unpad_op.cc b/paddle/fluid/operators/sequence_unpad_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..e633e378a226ece8adea2e150cc6c1e9aa874331 --- /dev/null +++ b/paddle/fluid/operators/sequence_unpad_op.cc @@ -0,0 +1,153 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/sequence_unpad_op.h" + +namespace paddle { +namespace operators { + +class SequenceUnpadOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceUnpadOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Length"), + "Input(Length) of SequenceUnpadOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SequenceUnpadOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_GE(x_dims.size(), 2, + "The rank of Input(X) can't be less than 2."); + + auto len_dims = ctx->GetInputDim("Length"); + PADDLE_ENFORCE(len_dims.size() == 2 && len_dims[1] == 1, + "The shape of Input(Length) should be [batch_size, 1]."); + PADDLE_ENFORCE( + len_dims[0] == x_dims[0], + "Input(X) and Input(Length) should have the same first dimension."); + + int64_t out_dim_0 = -1; + if (ctx->IsRuntime()) { + out_dim_0 = x_dims[0] * x_dims[1]; + } + + std::vector out_dims_vec{out_dim_0}; + if (x_dims.size() == 2) { + out_dims_vec.push_back(1); + } else { + for (int i = 2; i < x_dims.size(); ++i) { + out_dims_vec.push_back(x_dims[i]); + } + } + ctx->SetOutputDim("Out", framework::make_ddim(out_dims_vec)); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("X")); + return framework::OpKernelType(data_type, ctx.device_context()); + } +}; + +class SequenceUnpadOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(LoDTensor, default LoDTensor) Input tensor which " + "contains the padded sequences with equal length."); + AddInput("Length", + "(LoDTensor) The input tensor which specifies the actual ength of " + "sequences after unpadding."); + AddOutput( + "Out", + "(LoDTensor) The output tensor which contains unpadded sequences."); + AddComment(R"DOC( + Sequence Unpad Operator + + This operator removes the padding data in the input sequences and convert + them into sequences with actual length as output, identitied by lod + information. + + Example: + + Given input tensor Input(X): + X.data = [[ 1.0, 2.0, 3.0, 4.0, 5.0], + [ 6.0, 7.0, 8.0, 9.0, 10.0], + [11.0, 12.0, 13.0, 14.0, 15.0]], +` + in which there are 3 sequences padded to length 5, and the acutal length + specified by Input(Length): + + Length.data = [[2], [3], [4]], + + after unpadding, Output(Out) will be: + + Out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]] + Out.lod = [[0, 2, 5, 9]] + + )DOC"); + } +}; + +class SequenceUnpadGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceUnpadGradOp should not be null."); + PADDLE_ENFORCE( + ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) of SequenceUnpadGradOp should not be null."); + + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ framework::GradVarName("X")); + } + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("X")); + return framework::OpKernelType(data_type, ctx.device_context()); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(sequence_unpad, ops::SequenceUnpadOp, + ops::SequenceUnpadOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(sequence_unpad_grad, ops::SequenceUnpadGradOp); +REGISTER_OP_CPU_KERNEL( + sequence_unpad, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel); +REGISTER_OP_CPU_KERNEL( + sequence_unpad_grad, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel); diff --git a/paddle/fluid/operators/sequence_unpad_op.cu b/paddle/fluid/operators/sequence_unpad_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..75248372237ec2cb23122f6b16e64f6ce750ebf9 --- /dev/null +++ b/paddle/fluid/operators/sequence_unpad_op.cu @@ -0,0 +1,30 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/sequence_unpad_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + sequence_unpad, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel); +REGISTER_OP_CUDA_KERNEL( + sequence_unpad_grad, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel); diff --git a/paddle/fluid/operators/sequence_unpad_op.h b/paddle/fluid/operators/sequence_unpad_op.h new file mode 100644 index 0000000000000000000000000000000000000000..07df3dca831d7e646050ae57402c1a493c2e50e9 --- /dev/null +++ b/paddle/fluid/operators/sequence_unpad_op.h @@ -0,0 +1,104 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/memory/memcpy.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/sequence_padding.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using LoD = framework::LoD; + +template +class SequenceUnpadOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x_t = ctx.Input("X"); + auto* len_t = ctx.Input("Length"); + auto* out_t = ctx.Output("Out"); + out_t->mutable_data(ctx.GetPlace()); + + const int64_t* seq_len_ptr = nullptr; + if (platform::is_gpu_place(ctx.GetPlace())) { + LoDTensor seq_len_cpu; + seq_len_cpu.Resize(len_t->dims()); + seq_len_ptr = seq_len_cpu.mutable_data(platform::CPUPlace()); + framework::TensorCopy(*len_t, platform::CPUPlace(), + ctx.template device_context(), + &seq_len_cpu); + } else { + seq_len_ptr = len_t->data(); + } + + size_t batch_size = x_t->dims()[0]; + std::vector out_lod0(batch_size + 1, 0); + for (size_t i = 0; i < batch_size; ++i) { + out_lod0[i + 1] = out_lod0[i] + seq_len_ptr[i]; + } + + framework::LoD out_lod; + out_lod.push_back(out_lod0); + out_t->set_lod(out_lod); + + std::vector out_dims_vec{static_cast(out_lod0.back())}; + if (x_t->dims().size() == 2) { + out_dims_vec.push_back(1); + } else { + for (int i = 2; i < x_t->dims().size(); ++i) { + out_dims_vec.push_back(x_t->dims()[i]); + } + } + out_t->Resize(framework::make_ddim(out_dims_vec)); + + int64_t padded_length = x_t->dims()[1]; + math::UnpaddingLoDTensorFunctor()( + ctx.template device_context(), *x_t, out_t, + padded_length, 0, false, math::kBatchLengthWidth); + } +}; + +template +class SequenceUnpadGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* d_x = ctx.Output(framework::GradVarName("X")); + if (d_x) { + const auto* d_out = ctx.Input(framework::GradVarName("Out")); + const auto* x_t = ctx.Input("X"); + d_x->mutable_data(ctx.GetPlace()); + + int padded_length = x_t->dims()[1]; + + LoDTensor zero_pads; + zero_pads.Resize({1, 1}); + zero_pads.mutable_data(ctx.GetPlace()); + math::SetConstant set_zero; + auto& dev_ctx = ctx.template device_context(); + set_zero(dev_ctx, &zero_pads, static_cast(0)); + + math::PaddingLoDTensorFunctor()( + ctx.template device_context(), *d_out, d_x, zero_pads, + padded_length, 0, false, math::kBatchLengthWidth); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/sgd_op.cc b/paddle/fluid/operators/sgd_op.cc index fef230e42d07a5ed73b7a7a6ab682694675bb9d2..411a126bc8e2b3a8d25f436489c13970568ccae4 100644 --- a/paddle/fluid/operators/sgd_op.cc +++ b/paddle/fluid/operators/sgd_op.cc @@ -21,7 +21,7 @@ class SGDOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("Param"), "Input(Param) of SGDOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Grad"), @@ -42,7 +42,7 @@ class SGDOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { + const framework::ExecutionContext &ctx) const override { auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param")); return framework::OpKernelType(data_type, ctx.device_context()); } @@ -50,17 +50,20 @@ class SGDOp : public framework::OperatorWithKernel { class SGDOpInferVarType : public framework::VarTypeInference { public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto input_var = op_desc.Input("Param")[0]; - for (auto& out_var : op_desc.Output("ParamOut")) { - if (block->FindRecursiveOrCreateVar(input_var).GetType() == - framework::proto::VarType::SELECTED_ROWS) { - block->FindRecursiveOrCreateVar(out_var).SetType( - framework::proto::VarType::SELECTED_ROWS); - } else { - block->FindRecursiveOrCreateVar(out_var).SetType( - framework::proto::VarType::LOD_TENSOR); + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + auto input_var_n = op_desc.Input("Param")[0]; + auto in_var_type = block->FindRecursiveOrCreateVar(input_var_n).GetType(); + PADDLE_ENFORCE(in_var_type == framework::proto::VarType::SELECTED_ROWS || + in_var_type == framework::proto::VarType::LOD_TENSOR, + "The input Var's type should be LoDtensor or SelectedRows," + " but the received var(%s)'s type is %s", + input_var_n, in_var_type); + + for (auto &out_var_n : op_desc.Output("ParamOut")) { + auto &out_var = block->FindRecursiveOrCreateVar(out_var_n); + if (out_var.GetType() != in_var_type) { + out_var.SetType(in_var_type); } } } diff --git a/paddle/fluid/operators/sgd_op.cu b/paddle/fluid/operators/sgd_op.cu index 243609075713305a90dc162991166ba24d54e835..d3f4eba3b24ec1ac0328ef270256cdf3abe499db 100644 --- a/paddle/fluid/operators/sgd_op.cu +++ b/paddle/fluid/operators/sgd_op.cu @@ -56,6 +56,12 @@ template class SGDOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + auto* param = ctx.Input("Param"); auto* param_out = ctx.Output("ParamOut"); auto* learning_rate = ctx.Input("LearningRate"); diff --git a/paddle/fluid/operators/sum_op.h b/paddle/fluid/operators/sum_op.h index 34403c7a7aa717cca470be2931009e219e00e3ae..11987c61aebaad00f8a71f1b909c83c44ddc8b0e 100644 --- a/paddle/fluid/operators/sum_op.h +++ b/paddle/fluid/operators/sum_op.h @@ -43,17 +43,31 @@ class SumKernel : public framework::OpKernel { out->mutable_data(context.GetPlace()); } auto result = EigenVector::Flatten(*out); + auto &place = + *context.template device_context().eigen_device(); + int start = in_place ? 1 : 0; if (!in_place) { - math::SetConstant constant_functor; - constant_functor(context.template device_context(), out, - 0.0); + if ((in_num >= 2) && in_vars[0]->IsType() && + in_vars[1]->IsType()) { + auto &in_0 = in_vars[0]->Get(); + auto &in_1 = in_vars[1]->Get(); + if (in_0.numel() && in_1.numel()) { + auto in_0_e = EigenVector::Flatten(in_0); + auto in_1_e = EigenVector::Flatten(in_1); + result.device(place) = in_0_e + in_1_e; + start = 2; + } + } + if (start != 2) { + math::SetConstant constant_functor; + constant_functor(context.template device_context(), + out, 0.0); + } } math::SelectedRowsAddToTensor functor; - auto &place = - *context.template device_context().eigen_device(); // If in_place, just skip the first tensor - for (size_t i = in_place ? 1 : 0; i < in_num; i++) { + for (size_t i = start; i < in_num; i++) { if (in_vars[i]->IsType()) { auto &in_t = in_vars[i]->Get(); if (in_t.numel() == 0) { diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.cc b/paddle/fluid/operators/truncated_gaussian_random_op.cc index d854e2803975543b51c50ea2bc173322d3c3ca5e..1e8708f2648d7dd3c10319bd0a4be193d2458d53 100644 --- a/paddle/fluid/operators/truncated_gaussian_random_op.cc +++ b/paddle/fluid/operators/truncated_gaussian_random_op.cc @@ -148,7 +148,7 @@ struct TruncatedNormal { T operator()(T value) const { auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value; - return (std::sqrt(2.0) * Erfinv(2 * p - 1) + mean) * std; + return std::sqrt(2.0) * Erfinv(2 * p - 1) * std + mean; } }; diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.cu b/paddle/fluid/operators/truncated_gaussian_random_op.cu index ad2a9021bfe344d838dff2040b3fb9371274e218..5a3510babe4d57b9e80f0e7898df98033834ca15 100644 --- a/paddle/fluid/operators/truncated_gaussian_random_op.cu +++ b/paddle/fluid/operators/truncated_gaussian_random_op.cu @@ -42,7 +42,7 @@ struct TruncatedNormal { rng.discard(n); T value = dist(rng); auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value; - return (std::sqrt(2.0) * erfinvf(2 * p - 1) + mean) * std; + return std::sqrt(2.0) * erfinvf(2 * p - 1) * std + mean; } }; @@ -52,6 +52,7 @@ class GPUTruncatedGaussianRandomKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); + unsigned int seed = static_cast(context.Attr("seed")); if (seed == 0) { std::random_device rd; diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index 763bb403588d13c15271d26b09813dddf3a5dd8c..aa907595cb7cf165974caa69fe8eb0370471732d 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -23,14 +23,14 @@ namespace operators { template class CPUUniformRandomKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - framework::Tensor* tensor = nullptr; + void Compute(const framework::ExecutionContext &ctx) const override { + framework::Tensor *tensor = nullptr; auto out_var = ctx.OutputVar("Out"); if (out_var->IsType()) { tensor = out_var->GetMutable(); } else if (out_var->IsType()) { auto shape = ctx.Attr>("shape"); - auto* selected_rows = out_var->GetMutable(); + auto *selected_rows = out_var->GetMutable(); tensor = selected_rows->mutable_value(); tensor->Resize(framework::make_ddim(shape)); selected_rows->mutable_rows()->reserve(shape[0]); @@ -39,7 +39,7 @@ class CPUUniformRandomKernel : public framework::OpKernel { "uniform_random_op's output only" "supports SelectedRows and LoDTensor"); } - T* data = tensor->mutable_data(ctx.GetPlace()); + T *data = tensor->mutable_data(ctx.GetPlace()); unsigned int seed = static_cast(ctx.Attr("seed")); std::minstd_rand engine; if (seed == 0) { @@ -60,14 +60,14 @@ class UniformRandomOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of UniformRandomOp should not be null."); PADDLE_ENFORCE( ctx->Attrs().Get("min") < ctx->Attrs().Get("max"), "uniform_random's min must less then max"); - auto& shape = ctx->Attrs().Get>("shape"); + auto &shape = ctx->Attrs().Get>("shape"); std::vector temp; temp.reserve(shape.size()); for (auto dim : shape) { @@ -78,7 +78,7 @@ class UniformRandomOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { + const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( static_cast(ctx.Attr("dtype")), ctx.GetPlace()); @@ -112,17 +112,17 @@ uniform distribution. The random result is in set [min, max]. class UniformRandomOpVarTypeInference : public framework::VarTypeInference { public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { auto out_var_name = op_desc.Output("Out").front(); - if (block->FindRecursiveOrCreateVar(out_var_name).GetType() == - framework::proto::VarType::SELECTED_ROWS) { - block->FindRecursiveOrCreateVar(out_var_name) - .SetType(framework::proto::VarType::SELECTED_ROWS); - } else { - block->FindRecursiveOrCreateVar(out_var_name) - .SetType(framework::proto::VarType::LOD_TENSOR); + auto var_data_type = static_cast( + boost::get(op_desc.GetAttr("dtype"))); + + auto out_var = block->FindRecursiveOrCreateVar(out_var_name); + if (out_var.GetType() != framework::proto::VarType::SELECTED_ROWS) { + out_var.SetType(framework::proto::VarType::LOD_TENSOR); } + out_var.SetDataType(var_data_type); } }; diff --git a/paddle/fluid/platform/cpu_info.cc b/paddle/fluid/platform/cpu_info.cc index f12070acf8ba1fc1353fcffa95642a79f3597774..d466f28d1ea0a8327f8d7a45c3e55c5aacd61544 100644 --- a/paddle/fluid/platform/cpu_info.cc +++ b/paddle/fluid/platform/cpu_info.cc @@ -135,7 +135,7 @@ bool MayIUse(const cpu_isa_t cpu_isa) { return cpu.has(Cpu::tAVX); case avx2: return cpu.has(Cpu::tAVX2); - case avx512_common: + case avx512f: return cpu.has(Cpu::tAVX512F); case avx512_core: return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) && diff --git a/paddle/fluid/platform/cpu_info.h b/paddle/fluid/platform/cpu_info.h index e2221414e1a6c29b53b1bc1f64da4e984a98f25f..fd31ef77b46d5b5b641983a0421da31914c87c18 100644 --- a/paddle/fluid/platform/cpu_info.h +++ b/paddle/fluid/platform/cpu_info.h @@ -45,7 +45,7 @@ typedef enum { sse42, avx, avx2, - avx512_common, + avx512f, avx512_core, avx512_core_vnni, avx512_mic, diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index e026ff703dafab1e6ca04de2e01f90c875649f46..c5d2bd7d70d64cf3b3f8546345b4e23789d6acd2 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -35,6 +35,16 @@ platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) { return it->second.get(); } +const std::vector +DeviceContextPool::GetAllDeviceContexts() const { + std::vector all_device_ctx; + all_device_ctx.reserve(device_contexts_.size()); + for (auto& dev_ctx : device_contexts_) { + all_device_ctx.emplace_back(dev_ctx.second.get()); + } + return all_device_ctx; +} + DeviceContextPool::DeviceContextPool( const std::vector& places) { PADDLE_ENFORCE_GT(places.size(), 0); @@ -214,9 +224,9 @@ class CudnnHolder { CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place), cudnn_holder_(nullptr) { CUDADeviceGuard guard(place_.device); - compute_capability = GetCUDAComputeCapability(place_.device); - multi_process = GetCUDAMultiProcessors(place_.device); - max_threads_per_mp = GetCUDAMaxThreadsPerMultiProcessor(place_.device); + compute_capability_ = GetCUDAComputeCapability(place_.device); + multi_process_ = GetCUDAMultiProcessors(place_.device); + max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device); PADDLE_ENFORCE(cudaStreamCreate(&stream_)); eigen_stream_.reset(new EigenCudaStreamDevice()); eigen_stream_->Reinitialize(&stream_, place); @@ -227,6 +237,16 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace place) cudnn_holder_.reset(new CudnnHolder(&stream_, place)); } + driver_version_ = GetCUDADriverVersion(place_.device); + runtime_version_ = GetCUDARuntimeVersion(place_.device); + + LOG(INFO) << "device: " << place_.device + << ", CUDA Capability: " << compute_capability_ + << ", Driver Version: " << driver_version_ / 1000 << "." + << (driver_version_ % 100) / 10 + << ", Runtime Version: " << runtime_version_ / 1000 << "." + << (runtime_version_ % 100) / 10; + callback_manager_.reset(new StreamCallbackManager(stream_)); } @@ -248,11 +268,11 @@ void CUDADeviceContext::Wait() const { } int CUDADeviceContext::GetComputeCapability() const { - return compute_capability; + return compute_capability_; } int CUDADeviceContext::GetMaxPhysicalThreadCount() const { - return multi_process * max_threads_per_mp; + return multi_process_ * max_threads_per_mp_; } Eigen::GpuDevice* CUDADeviceContext::eigen_device() const { diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h index 79539195157d74d4d757edee5e008cbb76c93ee2..999bbe00f1659881050cb0dc89570b74b201aca7 100644 --- a/paddle/fluid/platform/device_context.h +++ b/paddle/fluid/platform/device_context.h @@ -135,9 +135,11 @@ class CUDADeviceContext : public DeviceContext { cudaStream_t stream_; cublasHandle_t cublas_handle_; - int compute_capability; - int multi_process; - int max_threads_per_mp; + int compute_capability_; + int runtime_version_; + int driver_version_; + int multi_process_; + int max_threads_per_mp_; mutable std::mutex mtx_; @@ -215,6 +217,9 @@ class DeviceContextPool { /*! \brief Return handle of single device context. */ platform::DeviceContext* Get(const platform::Place& place); + /*! \brief Return all the device contexts. */ + const std::vector GetAllDeviceContexts() const; + template const typename DefaultDeviceContextType::TYPE* GetByPlace( const Place& place) { diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h index f04395a8ac00f33501008aa12f22773ddda9b138..a251bfcd9914422cb6300adbbcdef3dfa79f441c 100644 --- a/paddle/fluid/platform/enforce.h +++ b/paddle/fluid/platform/enforce.h @@ -130,6 +130,13 @@ struct EOFException : public std::exception { #define UNLIKELY(condition) (condition == 0) #endif +#if !defined(_WIN32) +#define LIKELY(condition) __builtin_expect(static_cast(condition), 1) +#else +// there is no equivalent intrinsics in msvc. +#define LIKELY(condition) (condition != 0) +#endif + template inline typename std::enable_if::type throw_on_error( bool stat, const Args&... args) { diff --git a/paddle/fluid/platform/gpu_info.cc b/paddle/fluid/platform/gpu_info.cc index f599e7fbc886a60394ae4690e4160275b55b8596..8fff9844db738dbd6508569a8aaeed044e445e5f 100644 --- a/paddle/fluid/platform/gpu_info.cc +++ b/paddle/fluid/platform/gpu_info.cc @@ -46,6 +46,24 @@ int GetCUDAComputeCapability(int id) { return device_prop.major * 10 + device_prop.minor; } +int GetCUDARuntimeVersion(int id) { + PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count"); + int runtime_version = 0; + PADDLE_ENFORCE(cudaRuntimeGetVersion(&runtime_version), + "cudaRuntimeGetVersion failed in " + "paddle::platform::cudaRuntimeGetVersion"); + return runtime_version; +} + +int GetCUDADriverVersion(int id) { + PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count"); + int driver_version = 0; + PADDLE_ENFORCE(cudaDriverGetVersion(&driver_version), + "cudaDriverGetVersion failed in " + "paddle::platform::GetCUDADriverVersion"); + return driver_version; +} + int GetCUDAMultiProcessors(int id) { PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count"); int count; diff --git a/paddle/fluid/platform/gpu_info.h b/paddle/fluid/platform/gpu_info.h index f4640d3eaa2165c35e8e14690d83e9e7e7168c0b..be44158431ff80a41f7fdf4dfd4d070667f2ac63 100644 --- a/paddle/fluid/platform/gpu_info.h +++ b/paddle/fluid/platform/gpu_info.h @@ -29,6 +29,12 @@ int GetCUDADeviceCount(); //! Get the compute capability of the ith GPU (format: major * 10 + minor) int GetCUDAComputeCapability(int i); +//! Get the runtime version of the ith GPU +int GetCUDARuntimeVersion(int id); + +//! Get the driver version of the ith GPU +int GetCUDADriverVersion(int id); + //! Get the MultiProcessors of the ith GPU. int GetCUDAMultiProcessors(int i); diff --git a/paddle/fluid/platform/init.cc b/paddle/fluid/platform/init.cc index 3d5c4ac2dc0ccd74674cc978dc4fb093ad180f9a..8612b88ef7a6c8c7a282eb8a6830687833d0aa9d 100644 --- a/paddle/fluid/platform/init.cc +++ b/paddle/fluid/platform/init.cc @@ -119,7 +119,7 @@ void InitDevices(bool init_p2p, const std::vector devices) { platform::SetNumThreads(FLAGS_paddle_num_threads); #endif - if (platform::jit::MayIUse(platform::jit::avx512_common)) { + if (platform::jit::MayIUse(platform::jit::avx512f)) { #ifndef __AVX512F__ LOG(WARNING) << "AVX512F is available, Please re-compile on local machine"; #endif diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 652a6ec7a4e2e823b28f39b449570cd375e88e18..da46a1abe12258b47b2fd4afb5f146daf15e026d 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -30,6 +30,8 @@ limitations under the License. */ #include "paddle/fluid/platform/device_tracer.h" #include "paddle/fluid/string/printf.h" +DEFINE_bool(enable_rpc_profiler, false, "Enable rpc profiler or not."); + namespace paddle { namespace platform { @@ -193,6 +195,13 @@ RecordEvent::~RecordEvent() { PopEvent(name_, dev_ctx_); } +RecordRPCEvent::RecordRPCEvent(const std::string& name, + const DeviceContext* dev_ctx) { + if (FLAGS_enable_rpc_profiler) { + event_.reset(new platform::RecordEvent(name, dev_ctx)); + } +} + RecordBlock::RecordBlock(int block_id) : is_enabled_(false), start_ns_(PosixInNsec()) { std::lock_guard l(profiler_mu); @@ -276,7 +285,7 @@ struct EventItem { // Print results void PrintProfiler(const std::vector>& events_table, const std::string& sorted_domain, const size_t name_width, - const size_t data_width, double total) { + const size_t data_width, bool merge_thread) { // Output header information std::cout << "\n------------------------->" << " Profiling Report " @@ -292,6 +301,10 @@ void PrintProfiler(const std::vector>& events_table, PADDLE_THROW("Invalid profiler state", g_state); } + if (merge_thread) { + std::cout << "Note! This Report merge all thread info into one." + << std::endl; + } std::cout << "Place: " << place << std::endl; std::cout << "Time unit: ms" << std::endl; std::cout << "Sorted by " << sorted_domain @@ -312,8 +325,7 @@ void PrintProfiler(const std::vector>& events_table, << std::setw(data_width) << event_item.min_time << std::setw(data_width) << event_item.max_time << std::setw(data_width) << event_item.ave_time - << std::setw(data_width) << event_item.total_time / total - << std::endl; + << std::setw(data_width) << event_item.ratio << std::endl; } } std::cout << std::endl; @@ -321,8 +333,10 @@ void PrintProfiler(const std::vector>& events_table, // Parse the event list and output the profiling report void ParseEvents(const std::vector>& events, + bool merge_thread, EventSortingKey sorted_by = EventSortingKey::kDefault) { if (g_state == ProfilerState::kDisabled) return; + if (merge_thread && events.size() < 2) return; std::string sorted_domain; std::function sorted_func; @@ -361,34 +375,55 @@ void ParseEvents(const std::vector>& events, sorted_domain = "event first end time"; } + const std::vector>* analyze_events; + std::vector> merged_events_list; + if (merge_thread) { + std::vector merged_events; + for (size_t i = 0; i < events.size(); ++i) { + for (size_t j = 0; j < events[i].size(); ++j) { + merged_events.push_back(events[i][j]); + } + } + merged_events_list.push_back(merged_events); + analyze_events = &merged_events_list; + } else { + analyze_events = &events; + } + std::vector> events_table; size_t max_name_width = 0; - double total = 0.; // the total time - for (size_t i = 0; i < events.size(); i++) { + for (size_t i = 0; i < (*analyze_events).size(); i++) { + double total = 0.; // the total time in one thread std::list pushed_events; std::vector event_items; std::unordered_map event_idx; - for (size_t j = 0; j < events[i].size(); j++) { - if (events[i][j].type() == EventType::kPushRange) { - pushed_events.push_back(events[i][j]); - } else if (events[i][j].type() == EventType::kPopRange) { + for (size_t j = 0; j < (*analyze_events)[i].size(); j++) { + if ((*analyze_events)[i][j].type() == EventType::kPushRange) { + pushed_events.push_back((*analyze_events)[i][j]); + } else if ((*analyze_events)[i][j].type() == EventType::kPopRange) { std::list::reverse_iterator rit = pushed_events.rbegin(); while (rit != pushed_events.rend() && - rit->name() != events[i][j].name()) { + rit->name() != (*analyze_events)[i][j].name()) { ++rit; } if (rit != pushed_events.rend()) { double event_time = (g_state == ProfilerState::kCUDA || g_state == ProfilerState::kAll) - ? rit->CudaElapsedMs(events[i][j]) - : rit->CpuElapsedMs(events[i][j]); + ? rit->CudaElapsedMs((*analyze_events)[i][j]) + : rit->CpuElapsedMs((*analyze_events)[i][j]); total += event_time; - std::string event_name = - "thread" + std::to_string(rit->thread_id()) + "::" + rit->name(); - max_name_width = std::max(max_name_width, event_name.size()); + std::string event_name; + if (merge_thread) { + event_name = rit->name(); + max_name_width = std::max(max_name_width, event_name.size()); + } else { + event_name = "thread" + std::to_string(rit->thread_id()) + "::" + + rit->name(); + max_name_width = std::max(max_name_width, event_name.size()); + } if (event_idx.find(event_name) == event_idx.end()) { event_idx[event_name] = event_items.size(); @@ -413,7 +448,7 @@ void ParseEvents(const std::vector>& events, pushed_events.erase((++rit).base()); } else { LOG(WARNING) << "Cannot find the push marker of event \'" - << events[i][j].name() + << (*analyze_events)[i][j].name() << "\', which will be ignored in profiling report."; } } @@ -421,6 +456,7 @@ void ParseEvents(const std::vector>& events, // average time for (auto& item : event_items) { item.ave_time = item.total_time / item.calls; + item.ratio = item.total_time / total; } // sort if (sorted_by != EventSortingKey::kDefault) { @@ -438,7 +474,8 @@ void ParseEvents(const std::vector>& events, } // Print report - PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12, total); + PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12, + merge_thread); } void DisableProfiler(EventSortingKey sorted_key, @@ -449,7 +486,8 @@ void DisableProfiler(EventSortingKey sorted_key, Mark("_stop_profiler_", nullptr); std::vector> all_events = GetAllEvents(); - ParseEvents(all_events, sorted_key); + ParseEvents(all_events, true, sorted_key); + ParseEvents(all_events, false, sorted_key); ResetProfiler(); DeviceTracer* tracer = GetDeviceTracer(); if (tracer->IsEnabled()) { diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index 38630686f7cf3c669373f941d989adf11ba6cfe6..e8eae874afa3d17f0d3374eef457cdbacb3f8424 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -71,6 +71,7 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx); #if !defined(_WIN32) struct RecordEvent { + // dev_ctx can be set to nullptr if device is cpu. RecordEvent(const std::string& name, const DeviceContext* dev_ctx); ~RecordEvent(); @@ -86,6 +87,16 @@ struct RecordEvent { std::string full_name_; }; +class RecordRPCEvent { + public: + // dev_ctx can be set to nullptr if device is cpu. + RecordRPCEvent(const std::string& name, const DeviceContext* dev_ctx); + ~RecordRPCEvent() {} + + private: + std::unique_ptr event_; +}; + struct RecordBlock { explicit RecordBlock(int block_id); ~RecordBlock(); diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 295af1c5837d70c32b522cc47c8c3e12d5bd61c7..339a7c98c6a2bba2cd46790cecc169ef447c63ce 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -57,6 +57,10 @@ limitations under the License. */ #include "pybind11/stl.h" +DEFINE_bool(reader_queue_speed_test_mode, false, + "If set true, the queue.pop will only get data from queue but not " + "remove the data from queue for speed testing"); + // disable auto conversion to list in Python PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray); @@ -157,7 +161,50 @@ PYBIND11_PLUGIN(core) { .def("_get_double_element", TensorGetElement) .def("_dtype", [](Tensor &self) { return ToDataType(self.type()); }); - py::class_(m, "LoDTensor") + py::class_(m, "LoDTensor", R"DOC( + LoDTensor is a Tensor with optional LoD information. + + np.array(lod_tensor) can convert LoDTensor to numpy array. + lod_tensor.lod() can retrieve the LoD information. + + LoD is short for Level of Details and is usually used for varied sequence + length. You can skip the following comment if you don't need optional LoD. + + For example: + A LoDTensor X can look like the example below. It contains 2 sequences. + The first has length 2 and the second has length 3, as described by x.lod. + + The first tensor dimension 5=2+3 is calculated from LoD if it's available. + It means the total number of sequence element. In X, each element has 2 + columns, hence [5, 2]. + + x.lod = [[2, 3]] + x.data = [[1, 2], [3, 4], + [5, 6], [7, 8], [9, 10]] + x.shape = [5, 2] + + LoD can have multiple levels (for example, a paragraph can have multiple + sentences and a sentence can have multiple words). In the following + LodTensor Y, the lod_level is 2. It means there are 2 sequence, the + first sequence length is 2 (has 2 sub-sequences), the second one's + length is 1. The first sequence's 2 sub-sequences have length 2 and 2, + respectively. And the second sequence's 1 sub-sequence has length 3. + + y.lod = [[2 1], [2 2 3]] + y.shape = [2+2+3, ...] + + Note: + In above description, LoD is length-based. In Paddle internal + implementation, lod is offset-based. Hence, internally, + y.lod is represented as [[0, 2, 3], [0, 2, 4, 7]] (length-based + equivlent would be [[2-0, 3-2], [2-0, 4-2, 7-4]]). + + Sometimes LoD is called recursive_sequence_length to be more + self-explanatory. In this case, it must be length-based. Due to history + reasons. when LoD is called lod in public API, it might be offset-based. + Users should be careful about it. + + )DOC") .def_buffer( [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); }) .def("__init__", @@ -337,7 +384,8 @@ All parameter, weight, gradient are variables in Paddle. return make_ddim(shape); }); auto *holder = var.GetMutable(); - holder->InitOnce(capacity, dims); + holder->InitOnce(capacity, dims, + FLAGS_reader_queue_speed_test_mode); return holder->GetQueue(); }, py::return_value_policy::copy); @@ -620,26 +668,58 @@ All parameter, weight, gradient are variables in Paddle. // -- python binds for parallel executor. py::class_ pe(m, "ParallelExecutor"); - py::class_ exec_strategy(pe, "ExecutionStrategy"); + py::class_ exec_strategy(pe, "ExecutionStrategy", R"DOC( + ExecutionStrategy allows the user to more preciously control how to run + the program in ParallelExecutor by setting the property. + + Examples: + .. code-block:: python + + exec_strategy = fluid.ExecutionStrategy() + exec_strategy.num_threads = 4 + + train_exe = fluid.ParallelExecutor(use_cuda=True, + loss_name=loss.name, + exec_strategy=exec_strategy) + + train_loss, = train_exe.run([loss.name], feed=feed_dict) + + )DOC"); + exec_strategy.def(py::init()) .def_property( "num_threads", [](const ExecutionStrategy &self) { return self.num_threads_; }, [](ExecutionStrategy &self, size_t num_threads) { self.num_threads_ = num_threads; - }) + }, + R"DOC(The type is INT, num_threads represents the size of thread pool that + used to run the operators of the current program in ParallelExecutor. + If :math:`num\_threads=1`, all the operators will execute one by one, + but the order maybe difference between iterations. + If it is not set, it will be set in ParallelExecutor according to the + device type and device count, for GPU, :math:`num\_threads=device\_count*4`, for CPU, + :math:`num\_threads=CPU\_NUM*4`, the explanation of:math:`CPU\_NUM` is in ParallelExecutor. + if it is not set, ParallelExecutor will get the cpu count by calling + `multiprocessing.cpu_count()`. Default 0.)DOC") .def_property( "use_cuda", [](const ExecutionStrategy &self) { return self.use_cuda_; }, [](ExecutionStrategy &self, bool use_cuda) { self.use_cuda_ = use_cuda; - }) + }) // FIXME(chengduo): Doesn't add doc for 'use_cuda', use_cuda may + // make user confuse, because ParallelExecutor has a parameter named + // 'use_cuda' too, in current implementation, ParallelExecutor's + // 'use_cuda' will rewrite ExecutionStrategy's 'use_cuda'. .def_property( "allow_op_delay", [](const ExecutionStrategy &self) { return self.allow_op_delay_; }, [](ExecutionStrategy &self, bool allow_op_delay) { self.allow_op_delay_ = allow_op_delay; - }) + }, + R"DOC(The type is BOOL, allow_op_delay represents whether to delay the + communication operators to run, it may make the execution faster. + Note that in some models, allow_op_delay may cause program hang. Default False.)DOC") .def_property( "num_iteration_per_drop_scope", [](const ExecutionStrategy &self) { @@ -647,7 +727,19 @@ All parameter, weight, gradient are variables in Paddle. }, [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) { self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope; - }); + }, + R"DOC(The type is INT, num_iteration_per_drop_scope indicates how + many iterations to clean up the temp variables which + is generated during execution. It may make the execution faster, + because the temp variable's shape maybe the same between two iterations. Default 100. + + NOTES: + 1. If you fetch data when calling the 'run', the ParallelExecutor + will clean up the temp variables at the end of the current iteration. + 2. In some NLP model, it may cause the GPU memory is insufficient, + in this case, you should reduce `num_iteration_per_drop_scope`. + )DOC"); + exec_strategy.def_property( "use_experimental_executor", [](const ExecutionStrategy &self) { @@ -658,7 +750,22 @@ All parameter, weight, gradient are variables in Paddle. : ExecutionStrategy::kDefault; }); - py::class_ build_strategy(pe, "BuildStrategy"); + py::class_ build_strategy(pe, "BuildStrategy", R"DOC( + BuildStrategy allows the user to more preciously control how to + build the SSA Graph in ParallelExecutor by setting the property. + + Examples: + .. code-block:: python + + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + + train_exe = fluid.ParallelExecutor(use_cuda=True, + loss_name=loss.name, + build_strategy=build_strategy) + + train_loss, = train_exe.run([loss.name], feed=feed_dict) +)DOC"); py::enum_(build_strategy, "ReduceStrategy") .value("Reduce", BuildStrategy::ReduceStrategy::kReduce) @@ -676,31 +783,51 @@ All parameter, weight, gradient are variables in Paddle. [](const BuildStrategy &self) { return self.reduce_; }, [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) { self.reduce_ = strategy; - }) + }, + R"DOC(The type is STR, there are two reduce strategies in ParallelExecutor, + 'AllReduce' and 'Reduce'. If you want that all the parameters' + optimization are done on all devices independently, you should choose 'AllReduce'; + if you choose 'Reduce', all the parameters' optimization will be evenly distributed + to different devices, and then broadcast the optimized parameter to other devices. + In some models, `Reduce` is faster. Default 'AllReduce'. )DOC") .def_property( "gradient_scale_strategy", [](const BuildStrategy &self) { return self.gradient_scale_; }, [](BuildStrategy &self, BuildStrategy::GradientScaleStrategy strategy) { self.gradient_scale_ = strategy; - }) + }, + R"DOC(The type is STR, there are three ways of defining :math:`loss@grad` in + ParallelExecutor, 'CoeffNumDevice', 'One' and 'Customized'. By default, + ParallelExecutor sets the :math:`loss@grad` according to the number of devices. + If you want to customize :math:`loss@grad`, you can choose 'Customized'. + Default 'CoeffNumDevice'.)DOC") .def_property( "debug_graphviz_path", [](const BuildStrategy &self) { return self.debug_graphviz_path_; }, [](BuildStrategy &self, const std::string &path) { self.debug_graphviz_path_ = path; - }) + }, + R"DOC(The type is STR, debug_graphviz_path indicate the path that + writing the SSA Graph to file in the form of graphviz, you. + It is useful for debugging. Default "")DOC") .def_property( "enable_data_balance", [](const BuildStrategy &self) { return self.enable_data_balance_; }, - [](BuildStrategy &self, bool b) { self.enable_data_balance_ = b; }) - .def_property("fuse_elewise_add_act_ops", - [](const BuildStrategy &self) { - return self.fuse_elewise_add_act_ops_; - }, - [](BuildStrategy &self, bool b) { - self.fuse_elewise_add_act_ops_ = b; - }) + [](BuildStrategy &self, bool b) { + self.enable_data_balance_ = b; + }) // FIXME(chengudo): enable_data_balance seems not important + .def_property( + "fuse_elewise_add_act_ops", + [](const BuildStrategy &self) { + return self.fuse_elewise_add_act_ops_; + }, + [](BuildStrategy &self, bool b) { + self.fuse_elewise_add_act_ops_ = b; + }, + R"DOC(The type is BOOL, fuse_elewise_add_act_ops indicate whether + to fuse elementwise_add_op and activation_op, + it may make the execution faster. Default False)DOC") .def("_create_passes_from_strategy", [](BuildStrategy &self) -> std::shared_ptr { return self.CreatePassesFromStrategy(); diff --git a/paddle/fluid/train/demo/README.md b/paddle/fluid/train/demo/README.md index 41b01d33828f750f67bba5f82cb7ed6fe4d4ea0a..191da20669e185d819ec5eed55427461cc0b10e4 100644 --- a/paddle/fluid/train/demo/README.md +++ b/paddle/fluid/train/demo/README.md @@ -15,7 +15,7 @@ cmake .. -DFLUID_INSTALL_DIR=$PADDLE_LIB \ -DWITH_MKL=OFF \ -DWITH_MKLDNN=OFF make -j8 -make -j8 inference_lib_dist +make -j8 fluid_lib_dist ``` ### step 2. generate program desc diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index e133323ae420ba68d90215767ab940aed744acd6..85493c10549c290330ed09b9f28accb7a980de6a 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -390,7 +390,9 @@ function run_mac_test() { Running unit tests ... ======================================== EOF - + #remove proxy here to fix dist error on mac + export http_proxy= + export https_proxy= # TODO: jiabin need to refine this part when these tests fixed on mac ctest --output-on-failure -j $1 # make install should also be test when unittest @@ -648,42 +650,47 @@ function gen_capi_package() { fi } -function gen_fluid_inference_lib() { +function gen_fluid_lib() { mkdir -p ${PADDLE_ROOT}/build cd ${PADDLE_ROOT}/build if [[ ${WITH_C_API:-OFF} == "OFF" && ${WITH_INFERENCE:-ON} == "ON" ]] ; then cat <>> with program.lr_schedule_guard(): >>> lr = lr * decay """ + + tmp_role = self._current_role + tmp_var = self._op_role_var + OpRole = core.op_proto_and_checker_maker.OpRole self._current_role = OpRole.LRSched # TODO(typhoonzero): how to set target learning rate var self._op_role_var = [] yield - self._op_role_var = [] - self._current_role = OpRole.Forward + self._op_role_var = tmp_var + self._current_role = tmp_role def __str__(self): """ diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 25fde782b7ec6aae089d99f212f867e75e4352ac..dcd5a064a85cbfeedf41b7dc122cacba987e6570 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -30,8 +30,7 @@ from ..unique_name import generate as unique_name __all__ = [ 'data', 'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer', - 'random_data_generator', 'py_reader', 'create_py_reader_by_data', - 'Preprocessor', 'load' + 'random_data_generator', 'py_reader', 'Preprocessor', 'load' ] @@ -56,7 +55,11 @@ def data(name, Args: name(str): The name/alias of the function shape(list): Tuple declaring the shape. - append_batch_size(bool): Whether or not to append the data as a batch. + append_batch_size(bool): + 1. If true, it prepends -1 to the shape. + For example if shape=[1], the resulting shape is [-1, 1]. + 2. If shape contains -1, such as shape=[1, -1], + append_batch_size will be enforced to be be False (ineffective). dtype(int|float): The type of data : float32, float_16, int etc type(VarType): The output type. By default it is LOD_TENSOR. lod_level(int): The LoD Level. 0 means the input data is not a sequence. @@ -471,158 +474,6 @@ def random_data_generator(low, high, shapes, lod_levels, for_parallel=True): return monkey_patch_reader_methods(main_prog_var) -def _py_reader(capacity, - shapes, - dtypes, - lod_levels=None, - name=None, - use_double_buffer=True, - feed_list=None): - - if feed_list is not None: - if not isinstance(feed_list, list): - raise TypeError("feed_list should be a list of Variable" - " instead of " + str(type(feed_list))) - lod_levels = [] - dtypes = [] - shape_concat = [] - ranks = [] - shapes = [] - - for data in feed_list: - dtypes.append(data.dtype) - shape_concat.extend(data.shape) - ranks.append(len(data.shape)) - shapes.append(data.shape) - lod_levels.append(data.lod_level) - else: - dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] - shape_concat = [] - ranks = [] - - for shape in shapes: - shape_concat.extend(shape) - ranks.append(len(shape)) - - if lod_levels is None: - lod_levels = [0] * len(shapes) - - if name is None: - queue_name = unique_name('lod_tensor_blocking_queue') - reader_name = unique_name('create_py_reader') - double_buffer_name = unique_name('double_buffer') - else: - queue_name = "_".join([name, "queue"]) - reader_name = "_".join([name, "reader"]) - double_buffer_name = "_".join([name, "double_buffer"]) - - var = global_scope().var(queue_name) - feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes) - - startup_blk = default_startup_program().current_block() - startup_var = startup_blk.create_var(name=reader_name) - startup_blk.append_op( - type='create_py_reader', - inputs={'blocking_queue': [queue_name]}, - outputs={'Out': [startup_var]}, - attrs={ - 'shape_concat': shape_concat, - 'lod_levels': lod_levels, - 'ranks': ranks - }) - - startup_var.desc.set_dtypes(dtypes) - startup_var.persistable = True - - main_prog_var = _copy_reader_var_(default_main_program().current_block(), - startup_var) - - reader = monkey_patch_reader_methods(main_prog_var) - if use_double_buffer: - double_buffer_reader = double_buffer(reader, name=double_buffer_name) - # we return a double buffer reader. However, the reset method comes from - # py_reader. - double_buffer_reader.reset = reader.reset - reader = double_buffer_reader - - # monkey patch py_reader special methods - reader.queue = feed_queue - current_reset_method = reader.reset - reader.thread = None - reader.tensor_provider = None - reader.exited = False - - def start_provide_thread(func): - def __provider_thread__(): - for tensors in func(): - array = core.LoDTensorArray() - for item in tensors: - if not isinstance(item, core.LoDTensor): - tmp = core.LoDTensor() - tmp.set(item, core.CPUPlace()) - item = tmp - - array.append(item) - - if reader.exited: - break - feed_queue.push(array) - if reader.exited: - break - feed_queue.close() - - reader.thread = threading.Thread(target=__provider_thread__) - reader.thread.daemon = True - reader.thread.start() - - def __set_tensor_provider__(func): - reader.tensor_provider = func - - def __set_paddle_reader__(paddle_reader): - with program_guard(Program(), Program()): - actual_feed_list = feed_list - if actual_feed_list is None: - actual_feed_list = [] - counter = 0 - for dtype, shape, lod_level in zip(dtypes, shapes, lod_levels): - name = str(counter) - actual_feed_list.append( - data( - name=name, - dtype=dtype, - shape=shape, - lod_level=lod_level)) - counter += 1 - - feeder = DataFeeder( - feed_list=actual_feed_list, place=core.CPUPlace()) - paddle_reader = feeder.decorate_reader( - paddle_reader, multi_devices=False) - - def __tensor_provider__(): - for slots in paddle_reader(): - yield [slots[str(idx)] for idx in six.moves.xrange(counter)] - - __set_tensor_provider__(__tensor_provider__) - - def __reset__(): - current_reset_method() - if reader.thread is not None and reader.tensor_provider is not None: - reader.exited = True - reader.thread.join() - reader.exited = False - - def __start__(): - start_provide_thread(reader.tensor_provider) - - reader.reset = __reset__ - reader.decorate_tensor_provider = __set_tensor_provider__ - reader.decorate_paddle_reader = __set_paddle_reader__ - reader.start = __start__ - - return reader - - def py_reader(capacity, shapes, dtypes, @@ -747,72 +598,128 @@ def py_reader(capacity, >>> except fluid.core.EOFException: >>> test_reader.reset() """ - return _py_reader( - capacity=capacity, - shapes=shapes, - dtypes=dtypes, - lod_levels=lod_levels, - name=name, - use_double_buffer=use_double_buffer) + dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] + shape_concat = [] + ranks = [] + for shape in shapes: + shape_concat.extend(shape) + ranks.append(len(shape)) -def create_py_reader_by_data(capacity, - feed_list, - name=None, - use_double_buffer=True): - """ - Create a Python reader for data feeding in Python + if lod_levels is None: + lod_levels = [0] * len(shapes) - This layer returns a Reader Variable. + if name is None: + queue_name = unique_name('lod_tensor_blocking_queue') + reader_name = unique_name('create_py_reader') + double_buffer_name = unique_name('double_buffer') + else: + queue_name = "_".join([name, "queue"]) + reader_name = "_".join([name, "reader"]) + double_buffer_name = "_".join([name, "double_buffer"]) - Works much like py_reader except that it's input is feed_list - instead of shapes, dtypes and lod_levels + var = global_scope().var(queue_name) + feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes) - Args: - capacity(int): The buffer capacity maintained by :code:`py_reader`. - feed_list(list(Variable)): The data feed list. - name(basestring): The prefix Python queue name and Reader name. None will - be generated automatically. - use_double_buffer(bool): Whether use double buffer or not. + startup_blk = default_startup_program().current_block() + startup_var = startup_blk.create_var(name=reader_name) + startup_blk.append_op( + type='create_py_reader', + inputs={'blocking_queue': [queue_name]}, + outputs={'Out': [startup_var]}, + attrs={ + 'shape_concat': shape_concat, + 'lod_levels': lod_levels, + 'ranks': ranks + }) - Returns: - Variable: A Reader from which we can get feeding data. + startup_var.desc.set_dtypes(dtypes) + startup_var.persistable = True - Examples: + main_prog_var = _copy_reader_var_(default_main_program().current_block(), + startup_var) - 1. The basic usage of :code:`py_reader` is as follows: + reader = monkey_patch_reader_methods(main_prog_var) + if use_double_buffer: + double_buffer_reader = double_buffer(reader, name=double_buffer_name) + # we return a double buffer reader. However, the reset method comes from + # py_reader. + double_buffer_reader.reset = reader.reset + reader = double_buffer_reader - >>> import paddle.fluid as fluid - >>> import paddle.dataset.mnist as mnist - >>> - >>> image = fluid.layers.data(name='image', shape=[3,224,224], dtypes='float32') - >>> label = fluid.layers.data(name='label', shape=[1], dtypes='int64') - >>> reader = fluid.layers.create_py_reader_by_data(capacity=64, feed_list=[image, label]) - >>> reader.decorate_paddle_reader( - >>> paddle.reader.shuffle(paddle.batch(mnist.train()) - >>> - >>> img, label = fluid.layers.read_file(reader) - >>> loss = network(img, label) # some network definition - >>> - >>> fluid.Executor(fluid.CUDAPlace(0)).run(fluid.default_startup_program()) - >>> - >>> exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name) - >>> for epoch_id in range(10): - >>> reader.start() - >>> try: - >>> while True: - >>> exe.run(fetch_list=[loss.name]) - >>> except fluid.core.EOFException: - >>> reader.reset() - """ - return _py_reader( - capacity=capacity, - shapes=None, - dtypes=None, - lod_levels=None, - name=name, - use_double_buffer=use_double_buffer, - feed_list=feed_list) + # monkey patch py_reader special methods + reader.queue = feed_queue + current_reset_method = reader.reset + reader.thread = None + reader.tensor_provider = None + reader.exited = False + + def start_provide_thread(func): + def __provider_thread__(): + for tensors in func(): + array = core.LoDTensorArray() + for item in tensors: + if not isinstance(item, core.LoDTensor): + tmp = core.LoDTensor() + tmp.set(item, core.CPUPlace()) + item = tmp + + array.append(item) + + if reader.exited: + break + feed_queue.push(array) + if reader.exited: + break + feed_queue.close() + + reader.thread = threading.Thread(target=__provider_thread__) + reader.thread.daemon = True + reader.thread.start() + + def __set_tensor_provider__(func): + reader.tensor_provider = func + + def __set_paddle_reader__(paddle_reader): + with program_guard(Program(), Program()): + feed_list = [] + counter = 0 + for dtype, shape, lod_level in zip(dtypes, shapes, lod_levels): + name = str(counter) + feed_list.append( + data( + name=name, + dtype=dtype, + shape=shape, + lod_level=lod_level)) + counter += 1 + + feeder = DataFeeder(feed_list=feed_list, place=core.CPUPlace()) + paddle_reader = feeder.decorate_reader( + paddle_reader, multi_devices=False) + + def __tensor_provider__(): + for slots in paddle_reader(): + yield [slots[str(idx)] for idx in six.moves.xrange(counter)] + + __set_tensor_provider__(__tensor_provider__) + + def __reset__(): + current_reset_method() + if reader.thread is not None and reader.tensor_provider is not None: + reader.exited = True + reader.thread.join() + reader.exited = False + + def __start__(): + start_provide_thread(reader.tensor_provider) + + reader.reset = __reset__ + reader.decorate_tensor_provider = __set_tensor_provider__ + reader.decorate_paddle_reader = __set_paddle_reader__ + reader.start = __start__ + + return reader def open_files(filenames, diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 8c0ef7a82421ffc04bf669e6850e075226c09d27..538035de1a7a062a91cd48a8b4c7c11d5352d6c1 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -56,6 +56,7 @@ __all__ = [ 'sequence_expand', 'sequence_expand_as', 'sequence_pad', + 'sequence_unpad', 'lstm_unit', 'reduce_sum', 'reduce_mean', @@ -64,6 +65,7 @@ __all__ = [ 'reduce_prod', 'sequence_first_step', 'sequence_last_step', + 'sequence_slice', 'dropout', 'split', 'ctc_greedy_decoder', @@ -94,6 +96,7 @@ __all__ = [ 'pad_constant_like', 'label_smooth', 'roi_pool', + 'roi_align', 'dice_loss', 'image_resize', 'image_resize_short', @@ -107,6 +110,7 @@ __all__ = [ 'log', 'crop', 'rank_loss', + 'margin_rank_loss', 'elu', 'relu6', 'pow', @@ -150,6 +154,7 @@ __all__ = [ 'mul', 'sigmoid_cross_entropy_with_logits', 'maxout', + 'affine_channel', ] @@ -351,7 +356,6 @@ def dynamic_lstm(input, c_0(Variable): The initial cell state is an optional input, default is zero. This is a tensor with shape (N x D), where N is the batch size. `h_0` and `c_0` can be NULL but only at the same time. - param_attr(ParamAttr|None): The parameter attribute for the learnable hidden-hidden weights. @@ -359,6 +363,11 @@ def dynamic_lstm(input, W_{fh}, W_{oh}`} - The shape is (D x 4D), where D is the hidden size. + + If it is set to None or one attribute of ParamAttr, + dynamic_lstm will create ParamAttr as param_attr. + If the Initializer of the param_attr is not set, the + parameter is initialized with Xavier. Default: None. bias_attr (ParamAttr|None): The bias attribute for the learnable bias weights, which contains two parts, input-hidden bias weights and peephole connections weights if @@ -371,6 +380,11 @@ def dynamic_lstm(input, - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \ W_{fc}, W_{oc}`}. - The shape is (1 x 7D). + + If it is set to None or one attribute of ParamAttr, + dynamic_lstm will create ParamAttr as bias_attr. + If the Initializer of the bias_attr is not set, + the bias is initialized zero. Default: None. use_peepholes (bool): ${use_peepholes_comment} is_reverse (bool): ${is_reverse_comment} gate_activation (str): ${gate_activation_comment} @@ -389,11 +403,11 @@ def dynamic_lstm(input, hidden_dim = 512 forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4, - act=None, bias_attr=None) + bias_attr=False) forward, _ = fluid.layers.dynamic_lstm( input=forward_proj, size=hidden_dim * 4, use_peepholes=False) """ - + assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp." helper = LayerHelper('lstm', **locals()) size = size // 4 weight = helper.create_parameter( @@ -528,6 +542,11 @@ def dynamic_lstmp(input, size. - Projection weight = {:math:`W_{rh}`}. - The shape of projection weight is (D x P). + + If it is set to None or one attribute of ParamAttr, + dynamic_lstm will create ParamAttr as param_attr. + If the Initializer of the param_attr is not set, the + parameter is initialized with Xavier. Default: None. bias_attr(ParamAttr|None): The bias attribute for the learnable bias weights, which contains two parts, input-hidden bias weights and peephole connections weights if @@ -540,6 +559,11 @@ def dynamic_lstmp(input, - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \ W_{fc}, W_{oc}`}. - The shape is (1 x 7D). + + If it is set to None or one attribute of ParamAttr, + dynamic_lstm will create ParamAttr as bias_attr. + If the Initializer of the bias_attr is not set, + the bias is initialized zero. Default: None. use_peepholes(bool): Whether to enable diagonal/peephole connections, default `True`. is_reverse(bool): Whether to compute reversed LSTM, default `False`. @@ -584,6 +608,7 @@ def dynamic_lstmp(input, proj_activation="tanh") """ + assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp." helper = LayerHelper('lstmp', **locals()) size = size // 4 weight = helper.create_parameter( @@ -1265,7 +1290,8 @@ def sequence_conv(input, padding=None, bias_attr=None, param_attr=None, - act=None): + act=None, + name=None): """ This function creates the op for sequence_conv, using the inputs and other convolutional configurations for the filters and stride as given @@ -1277,9 +1303,19 @@ def sequence_conv(input, filter_size (int): the filter size (H and W). filter_stride (int): stride of the filter. padding (bool): if True, add paddings. - bias_attr (ParamAttr|None): attributes for bias - param_attr (ParamAttr|None): attributes for parameter - act (str): the activation type + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of sequence_conv. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, sequence_conv + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of sequence_conv. If it is set to None or one attribute of ParamAttr, sequence_conv + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + act (str): Activation type, if it is set to None, activation is not appended. + Default: None. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Variable: output of sequence_conv @@ -1308,7 +1344,7 @@ def sequence_conv(input, return helper.append_activation(pre_act) -def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False): +def sequence_softmax(input, use_cudnn=False, name=None): """ This function computes the softmax activation among all time-steps for each sequence. The dimension of each time-step should be 1. Thus, the shape of @@ -1328,10 +1364,10 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False): Args: input (Variable): The input variable which is a LoDTensor. - bias_attr (ParamAttr|None): attributes for bias - param_attr (ParamAttr|None): attributes for parameter use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \ - library is installed. Default: False + library is installed. Default: False. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Variable: output of sequence_softmax @@ -1355,7 +1391,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False): return softmax_out -def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None): +def softmax(input, use_cudnn=True, name=None): """ The input of the softmax operator is a tensor of any rank. The output tensor has the same shape as the input. @@ -1382,10 +1418,10 @@ def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None): Args: input (Variable): The input variable. - bias_attr (ParamAttr): attributes for bias - param_attr (ParamAttr): attributes for parameter use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \ - library is installed. + library is installed. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Variable: output of softmax @@ -1491,14 +1527,23 @@ def conv2d(input, convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only - connected to the second half of the input channels. Default: groups=1 - param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None - bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None + connected to the second half of the input channels. Default: groups=1. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv2d. If it is set to None or one attribute of ParamAttr, conv2d + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with :math:`Normal(0.0, std)`, + and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv2d + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True - act (str): Activation type. Default: None + act (str): Activation type, if it is set to None, activation is not appended. + Default: None name (str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + will be named automatically. Default: None Returns: Variable: The tensor variable storing the convolution and \ @@ -1516,7 +1561,7 @@ def conv2d(input, """ num_channels = input.shape[1] - + assert param_attr is not False, "param_attr should not be False here." l_type = 'conv2d' if (num_channels == groups and num_filters % num_channels == 0 and not use_cudnn): @@ -1544,7 +1589,8 @@ def conv2d(input, filter_shape = [num_filters, int(num_filter_channels)] + filter_size def _get_default_param_initializer(): - std = (2.0 / (filter_size[0]**2 * num_channels))**0.5 + filter_elem_num = filter_size[0] * filter_size[1] * num_channels + std = (2.0 / filter_elem_num)**0.5 return Normal(0.0, std, 0) filter_param = helper.create_parameter( @@ -1655,13 +1701,22 @@ def conv3d(input, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels. Default: groups=1 - param_attr (ParamAttr): The parameters to the Conv3d Layer. Default: None - bias_attr (ParamAttr): Bias parameter for the Conv3d layer. Default: None + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv3d. If it is set to None or one attribute of ParamAttr, conv3d + will create ParamAttr as param_attr. If it is set to None, the parameter + is initialized with :math:`Normal(0.0, std)`, and the :math:`std` is + :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv3d + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True - act (str): Activation type. Default: None + act (str): Activation type, if it is set to None, activation is not appended. + Default: None. name (str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + will be named automatically. Default: None. Returns: Variable: The tensor variable storing the convolution and \ @@ -1679,7 +1734,7 @@ def conv3d(input, """ l_type = 'conv3d' - + assert param_attr is not False, "param_attr should not be False here." helper = LayerHelper(l_type, **locals()) dtype = helper.input_dtype() @@ -1704,7 +1759,9 @@ def conv3d(input, filter_shape = [num_filters, num_filter_channels] + filter_size def _get_default_param_initializer(): - std = (2.0 / (filter_size[0]**3 * num_channels))**0.5 + filter_elem_num = filter_size[0] * filter_size[1] * filter_size[ + 2] * num_channels + std = (2.0 / filter_elem_num)**0.5 return Normal(0.0, std, 0) filter_param = helper.create_parameter( @@ -1901,6 +1958,76 @@ def sequence_last_step(input): return sequence_pool(input=input, pool_type="last") +def sequence_slice(input, offset, length, name=None): + """ + **Sequence Slice Layer** + + The layer crops a subsequence from given sequence with given start + offset and subsequence length. + + It only supports sequence data (LoDTensor with lod_level equal to 1). + + .. code-block:: text + + - Case: + + Given the input Variable **input**: + + input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]], + input.lod = [[3, 2]], + input.dims = (5, 2), + + with offset.data = [[0], [1]] and length.data = [[2], [1]], + + the output Variable will be + + out.data = [[a1, a2], [b1, b2], [e1, e2]], + out.lod = [[2, 1]], + out.dims = (3, 2). + + NOTE: The first dimension size of **input**, **offset** and **length** + should be equal. The **offset** should start from 0. + + Args: + input(Variable): The input Variable which consists of the complete + sequences. + offset(Variable): The offset to slice each sequence. + length(Variable): The length of each subsequence. + name(str|None): A name for this layer(optional). If set None, the + layer will be named automatically. + + Returns: + Variable: The output subsequences. + + Examples: + + .. code-block:: python + + import numpy as np + seqs = fluid.layers.data(name='x', shape=[10, 5], + dtype='float32', lod_level=1) + offset = fluid.layers.assign(input=np.array([[0, 1]]).astype("int32")) + length = fluid.layers.assign(input=np.array([[2, 1]]).astype("int32")) + subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset, + length=length) + """ + helper = LayerHelper("sequence_slice", **locals()) + dtype = helper.input_dtype() + out = helper.create_tmp_variable(dtype) + + offset.stop_gradient = True + length.stop_gradient = True + + helper.append_op( + type="sequence_slice", + inputs={"X": input, + "Offset": offset, + "Length": length}, + outputs={"Out": out}) + + return out + + @templatedoc() def pool2d(input, pool_size=-1, @@ -2106,8 +2233,14 @@ def batch_norm(input, is_test(bool, Default False): Used for training or training. momentum(float, Default 0.9): epsilon(float, Default 1e-05): - param_attr(ParamAttr): The parameter attribute for Parameter `scale`. - bias_attr(ParamAttr): The parameter attribute for Parameter `bias`. + param_attr(ParamAttr|None): The parameter attribute for Parameter `scale` + of batch_norm. If it is set to None or one attribute of ParamAttr, batch_norm + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr(ParamAttr|None): The parameter attribute for the bias of batch_norm. + If it is set to None or one attribute of ParamAttr, batch_norm + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. data_layout(string, default NCHW): NCHW|NHWC in_place(bool, Default False): Make the input and output of batch norm reuse memory. name(string, Default None): A name for this layer(optional). If set None, the layer @@ -2127,6 +2260,7 @@ def batch_norm(input, hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w') hidden2 = fluid.layers.batch_norm(input=hidden1) """ + assert bias_attr is not False, "bias_attr should not be False in batch_norm." helper = LayerHelper('batch_norm', **locals()) dtype = helper.input_dtype() @@ -2243,19 +2377,28 @@ def layer_norm(input, Args: input(Variable): The input tensor variable. scale(bool): Whether to learn the adaptive gain :math:`g` after - normalization. + normalization. Default True. shift(bool): Whether to learn the adaptive bias :math:`b` after - normalization. - begin_norm_axis(bool): The normalization will be performed along + normalization. Default True. + begin_norm_axis(int): The normalization will be performed along dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`. + Default 1. epsilon(float): The small value added to the variance to prevent - division by zero. + division by zero. Default 1e-05. param_attr(ParamAttr|None): The parameter attribute for the learnable - gain :math:`g`. + gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is + omitted. If :attr:`scale` is True and :attr:`param_attr` is None, + a default :code:`ParamAttr` would be added as scale. The + :attr:`param_attr` is initialized as 1 if it is added. Default None. bias_attr(ParamAttr|None): The parameter attribute for the learnable - bias :math:`b`. + bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is + omitted. If :attr:`shift` is True and :attr:`param_attr` is None, + a default :code:`ParamAttr` would be added as bias. The + :attr:`bias_attr` is initialized as 0 if it is added. Default None. act(str): Activation to be applied to the output of layer normalizaiton. - name (str): The name of this layer. It is optional. + Default None. + name(str): The name of this layer. It is optional. Default None, and a + unique name would be generated automatically. Returns: ${y_comment} @@ -2396,15 +2539,22 @@ def conv2d_transpose(input, when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels. - Default: groups=1 - param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. - Default: None - bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None + Default: groups = 1. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv2d_transpose. If it is set to None or one attribute of ParamAttr, conv2d_transpose + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d_transpose. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv2d_transpose + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn - library is installed. Default: True - act(str): Activation type. Default: None + library is installed. Default: True. + act (str): Activation type, if it is set to None, activation is not appended. + Default: None. name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + will be named automatically. Default: True. Returns: Variable: The tensor variable storing the convolution transpose result. @@ -2419,7 +2569,7 @@ def conv2d_transpose(input, data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32') conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3) """ - + assert param_attr is not False, "param_attr should not be False in conv2d_transpose." input_channel = input.shape[1] op_type = 'conv2d_transpose' @@ -2455,6 +2605,7 @@ def conv2d_transpose(input, else: filter_size = utils.convert_to_list(filter_size, 2, 'conv2d_transpose.filter_size') + if output_size is None: output_size = [] elif isinstance(output_size, list) or isinstance(output_size, int): @@ -2464,6 +2615,7 @@ def conv2d_transpose(input, padding = utils.convert_to_list(padding, 2, 'padding') groups = 1 if groups is None else groups filter_shape = [input_channel, num_filters // groups] + filter_size + img_filter = helper.create_parameter( dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) @@ -2576,12 +2728,19 @@ def conv3d_transpose(input, first half of the input channels, while the second half of the filters is only connected to the second half of the input channels. Default: groups=1 - param_attr(ParamAttr): The parameters to the Conv3d_transpose Layer. - Default: None - bias_attr(ParamAttr): Bias parameter for the Conv3d layer. Default: None + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv3d_transpose. If it is set to None or one attribute of ParamAttr, conv3d_transpose + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d_transpose. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv3d_transpose + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True - act(str): Activation type. Default: None + act (str): Activation type, if it is set to None, activation is not appended. + Default: None. name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. @@ -2598,6 +2757,7 @@ def conv3d_transpose(input, data = fluid.layers.data(name='data', shape=[3, 12, 32, 32], dtype='float32') conv3d_transpose = fluid.layers.conv3d_transpose(input=data, num_filters=2, filter_size=3) """ + assert param_attr is not False, "param_attr should not be False in conv3d_transpose." l_type = "conv3d_transpose" helper = LayerHelper(l_type, **locals()) if not isinstance(input, Variable): @@ -2792,7 +2952,7 @@ def sequence_expand_as(x, y, name=None): @templatedoc() -def sequence_pad(x, pad_value, maxlen=None): +def sequence_pad(x, pad_value, maxlen=None, name=None): """ ${comment} @@ -2806,7 +2966,9 @@ def sequence_pad(x, pad_value, maxlen=None): None or any positive int. When it is None, all sequences will be padded up to the length of the longest one among them; when it a certain positive value, it must be greater than the length of the - longest original sequence." + longest original sequence. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The padded sequence batch and the original lengths before @@ -2843,6 +3005,66 @@ def sequence_pad(x, pad_value, maxlen=None): return out, length +def sequence_unpad(x, length, name=None): + """ + **Sequence Unpad Layer** + + This layer removes the padding data in the input sequences and convert + them into sequences with actual length as output, identitied by lod + information. + + .. code-block:: text + + Example: + + Given input Variable **x**: + x.data = [[ 1.0, 2.0, 3.0, 4.0, 5.0], + [ 6.0, 7.0, 8.0, 9.0, 10.0], + [11.0, 12.0, 13.0, 14.0, 15.0]], + + in which there are 3 sequences padded to length 5, and the acutal length + specified by input Variable **length**: + + length.data = [[2], [3], [4]], + + after unpadding, the output Variable will be: + + out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]] + out.lod = [[2, 3, 4]] + + Args: + x(Variable): Input Variable which contains the padded sequences with + equal length. + length(Variable): The Variable that specifies the actual ength of + sequences after unpadding. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The Variable contains the unpadded sequences. + + Examples: + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[10, 5], dtype='float32') + len = fluid.layers.data(name='length', shape=[1], dtype='int64') + out = fluid.layers.sequence_unpad(x=x, length=len) + """ + + helper = LayerHelper('sequence_unpad', input=x, **locals()) + dtype = helper.input_dtype() + out = helper.create_tmp_variable(dtype) + + length.stop_gradient = True + + helper.append_op( + type='sequence_unpad', + inputs={'X': x, + 'Length': length}, + outputs={'Out': out}) + return out + + def beam_search(pre_ids, pre_scores, ids, @@ -3054,10 +3276,18 @@ def lstm_unit(x_t, cell_t_prev (Variable): The cell value of lstm unit, a 2-D tensor with shape M x S, M for batch size and S for size of lstm unit. forget_bias (float): The forget bias of lstm unit. - param_attr (ParamAttr): The attributes of parameter weights, used to set - initializer, name etc. - bias_attr (ParamAttr): The attributes of bias weights, if not False, - bias weights will be created and be set to default value. + param_attr(ParamAttr|None): The parameter attribute for the learnable + hidden-hidden weights. + If it is set to None or one attribute of ParamAttr, + lstm_unit will create ParamAttr as param_attr. + If the Initializer of the param_attr is not set, the + parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|None): The bias attribute for the learnable bias + weights. If it is set to False, no bias will be added + to the output units. If it is set to None or one attribute of ParamAttr, + lstm_unit will create ParamAttr as bias_attr. + If the Initializer of the bias_attr is not set, + the bias is initialized zero. Default: None. name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. @@ -3971,7 +4201,8 @@ def nce(input, sample_weight=None, param_attr=None, bias_attr=None, - num_neg_samples=None): + num_neg_samples=None, + name=None): """ ${comment} @@ -3982,9 +4213,18 @@ def nce(input, sample_weight (Variable|None): A Variable of shape [batch_size, 1] storing a weight for each sample. The default weight for each sample is 1.0. - param_attr (ParamAttr|None): attributes for parameter - bias_attr (ParamAttr|None): attributes for bias + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of nce. If it is set to None or one attribute of ParamAttr, nce + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of nce. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, nce + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. num_neg_samples (int): ${num_neg_samples_comment} + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Variable: The output nce loss. @@ -4017,19 +4257,28 @@ def nce(input, """ helper = LayerHelper('nce', **locals()) assert isinstance(input, Variable) - dim = input.shape[1] assert isinstance(label, Variable) + + dim = input.shape[1] num_true_class = label.shape[1] w = helper.create_parameter( attr=helper.param_attr, shape=[num_total_classes, dim], is_bias=False, dtype=input.dtype) - b = helper.create_parameter( - attr=helper.bias_attr, - shape=[num_total_classes, 1], - is_bias=True, - dtype=input.dtype) + inputs = { + 'Input': input, + 'Label': label, + 'Weight': w, + 'SampleWeight': sample_weight if sample_weight is not None else [] + } + if helper.bias_attr: + b = helper.create_parameter( + attr=helper.bias_attr, + shape=[num_total_classes, 1], + is_bias=True, + dtype=input.dtype) + inputs['Bias'] = b cost = helper.create_tmp_variable(dtype=input.dtype) sample_logits = helper.create_tmp_variable(dtype=input.dtype) sample_labels = helper.create_tmp_variable(dtype=label.dtype) @@ -4046,13 +4295,7 @@ def nce(input, helper.append_op( type='nce', - inputs={ - 'Input': input, - 'Label': label, - 'Weight': w, - 'Bias': b, - 'SampleWeight': sample_weight if sample_weight is not None else [] - }, + inputs=inputs, outputs={ 'Cost': cost, 'SampleLogits': sample_logits, @@ -4062,7 +4305,12 @@ def nce(input, return cost / (num_neg_samples + 1) -def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): +def hsigmoid(input, + label, + num_classes, + param_attr=None, + bias_attr=None, + name=None): """ The hierarchical sigmoid operator is used to accelerate the training process of language model. This operator organizes the classes into a @@ -4083,11 +4331,17 @@ def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): label (Variable): The tensor variable contains labels of training data. It's a tensor with shape is :math:`[N \\times 1]`. num_classes: (int), The number of classes, must not be less than 2. - param_attr (ParamAttr|list of ParamAttr, default None): The parameter - attribute for learnable parameters/weights of this layer. - bias_attr (ParamAttr|list of ParamAttr, default None): The parameter - attribute for the bias of this layer. If it is set to False, no - bias will be applied. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of hsigmoid. If it is set to None or one attribute of ParamAttr, hsigmoid + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of hsigmoid. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, hsigmoid + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Out: (Tensor) The cost of hierarchical sigmoid operator. the shape is [N, 1] @@ -5177,6 +5431,54 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): return pool_out +@templatedoc() +def roi_align(input, + rois, + pooled_height=1, + pooled_width=1, + spatial_scale=1.0, + sampling_ratio=-1, + name=None): + """ + ${comment} + + Args: + input (Variable): ${x_comment} + rois (Variable): ROIs (Regions of Interest) to pool over. + pooled_height (integer): ${pooled_height_comment} Default: 1 + pooled_width (integer): ${pooled_width_comment} Default: 1 + spatial_scale (float): ${spatial_scale_comment} Default: 1.0 + sampling_ratio(intger): ${sampling_ratio_comment} Default: -1 + + Returns: + Variable: ${out_comment}. + Examples: + .. code-block:: python + + align_out = fluid.layers.roi_align(input=x, + rois=rois, + pooled_height=7, + pooled_width=7, + spatial_scale=0.5, + sampling_ratio=-1) + """ + helper = LayerHelper('roi_align', **locals()) + dtype = helper.input_dtype() + align_out = helper.create_tmp_variable(dtype) + helper.append_op( + type="roi_align", + inputs={"X": input, + "ROIs": rois}, + outputs={"Out": align_out}, + attrs={ + "pooled_height": pooled_height, + "pooled_width": pooled_width, + "spatial_scale": spatial_scale, + "sampling_ratio": sampling_ratio + }) + return align_out + + def dice_loss(input, label, epsilon=0.00001): """ Dice loss for comparing the similarity of two batch of data, @@ -5827,6 +6129,54 @@ def rank_loss(label, left, right, name=None): return out +def margin_rank_loss(label, left, right, margin=0.1, name=None): + """ + Margin Ranking Loss Layer for ranking problem, + which compares left score and right score passed in. + The ranking loss can be defined as following equation: + + .. math:: + + rank\_loss &= max(0, -label * (left - right) + margin) + + Args: + label (Variable): Indicates whether the left is ranked higher than the right or not. + left (Variable): Ranking score for left. + right (Variable): Ranking score for right. + margin (float): Indicates the given margin. + name (str|None): A name for this layer (optional). If set None, the layer + will be named automatically. + Returns: + Variable: The ranking loss. + Raises: + ValueError: Any of label, left, and right is not a Variable. + Examples: + .. code-block:: python + label = fluid.layers.data(name="label", shape=[4, 1], dtype="float32") + left = fluid.layers.data(name="left", shape=[4, 1], dtype="float32") + right = fluid.layers.data(name="right", shape=[4, 1], dtype="float32") + out = fluid.layers.margin_rank_loss(label, left, right) + """ + helper = LayerHelper('margin_rank_loss', **locals()) + if not isinstance(label, Variable): + raise ValueError("The label should be a Variable.") + if not isinstance(left, Variable): + raise ValueError("The left should be a Variable.") + if not isinstance(right, Variable): + raise ValueError("The right should be a Variable.") + out = helper.create_tmp_variable(left.dtype) + act = helper.create_tmp_variable(left.dtype) + helper.append_op( + type='margin_rank_loss', + inputs={"Label": label, + "X1": left, + "X2": right}, + outputs={'Out': out, + 'Activated': act}, + attrs={'margin': margin}) + return out + + def pad2d(input, paddings=[0, 0, 0, 0], mode='constant', @@ -6290,6 +6640,7 @@ def sequence_enumerate(input, win_size, pad_value=0, name=None): outputs={'Out': out}, attrs={'win_size': win_size, 'pad_value': pad_value}) + return out def sequence_mask(x, maxlen=None, dtype='int64', name=None): @@ -7084,3 +7435,44 @@ def maxout(x, groups, name=None): attrs={"groups": groups}, outputs={"Out": out}) return out + + +def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None): + """ + Applies a separate affine transformation to each channel of the input. + Useful for replacing spatial batch norm with its equivalent fixed + transformation. The input also can be 2D tensor and applies a affine + transformation in second dimension. + + Args: + x (Variable): Feature map input can be a 4D tensor with order NCHW + or NHWC. It also can be a 2D tensor and the affine transformation + is applied in the second dimension. + scale (Variable): 1D input of shape (C), the c-th element is the scale + factor of the affine transformation for the c-th channel of + the input. + bias (Variable): 1D input of shape (C), the c-th element is the bias + of the affine transformation for the c-th channel of the input. + data_layout (string, default NCHW): NCHW or NHWC. If input is 2D + tensor, you can ignore data_layout. + name (str, default None): The name of this layer. + + Returns: + out (Variable): A tensor of the same shape and data layout with x. + """ + helper = LayerHelper("affine_channel", **locals()) + + if name is None: + out = helper.create_tmp_variable(dtype=x.dtype) + else: + out = helper.create_variable( + name=name, dtype=x.dtype, persistable=False) + + helper.append_op( + type="affine_channel", + inputs={"X": x, + 'Scale': scale, + 'Bias': bias}, + attrs={"data_layout": data_layout}, + outputs={"Out": out}) + return out diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 9a8300524d8784fae598635796888382b1adbccf..1ff40a26f2f24e2ff06719972489b0c1e5d140c3 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -14,6 +14,8 @@ from __future__ import print_function from .layer_function_generator import generate_layer_fn, generate_layer_fn_noattr +from .. import core +from ..framework import convert_np_dtype_to_dtype_ __activations_noattr__ = [ 'sigmoid', @@ -58,8 +60,11 @@ _uniform_random_ = generate_layer_fn('uniform_random') def uniform_random(shape, dtype=None, min=None, max=None, seed=None): + locals_var = locals().keys() + if not isinstance(dtype, core.VarDesc.VarType): + dtype = convert_np_dtype_to_dtype_(dtype) kwargs = dict() - for name in locals(): + for name in locals_var: val = locals()[name] if val is not None: kwargs[name] = val @@ -78,8 +83,9 @@ _hard_shrink_ = generate_layer_fn('hard_shrink') def hard_shrink(x, threshold=None): + locals_var = locals().keys() kwargs = dict() - for name in locals(): + for name in locals_var: val = locals()[name] if val is not None: kwargs[name] = val @@ -99,12 +105,12 @@ _cum_sum_ = generate_layer_fn('cumsum') def cumsum(x, axis=None, exclusive=None, reverse=None): + locals_var = locals().keys() kwargs = dict() - for name in locals(): + for name in locals_var: val = locals()[name] if val is not None: kwargs[name] = val - return _cum_sum_(**kwargs) @@ -121,8 +127,9 @@ _thresholded_relu_ = generate_layer_fn('thresholded_relu') def thresholded_relu(x, threshold=None): + locals_var = locals().keys() kwargs = dict() - for name in locals(): + for name in locals_var: val = locals()[name] if val is not None: kwargs[name] = val diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index 44b92af7acc012f89b271c74b026d18e1a4075f8..9c6a2112a69637aad4baff9a0296330620e3c52f 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -100,7 +100,7 @@ def create_global_var(shape, force_cpu=False, name=None): """ - Create a new variable in the global block(block 0). + Create a new tensor variable with value in the global block(block 0). Args: shape(list[int]): shape of the variable diff --git a/python/paddle/fluid/nets.py b/python/paddle/fluid/nets.py index 1dabad54f5b976e0fcabf6918d3bc6ece4eed384..00d33b36fcc3266bf7f08020052d28172665e53e 100644 --- a/python/paddle/fluid/nets.py +++ b/python/paddle/fluid/nets.py @@ -64,23 +64,33 @@ def simple_img_conv_pool(input, average-pooling. Default :math:`max`. global_pooling (bool): Whether to use the global pooling. If global_pooling = true, pool_size and pool_padding while be ignored. Default False - conv_stride (int|list|tuple): The stride size of the Conv2d Layer. If stride is a + conv_stride (int|list|tuple): The stride size of the conv2d Layer. If stride is a list or tuple, it must contain two integers, (conv_stride_H, conv_stride_W). Otherwise, the conv_stride_H = conv_stride_W = conv_stride. Default: conv_stride = 1. - conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is + conv_padding (int|list|tuple): The padding size of the conv2d Layer. If padding is a list or tuple, it must contain two integers, (conv_padding_H, conv_padding_W). Otherwise, the conv_padding_H = conv_padding_W = conv_padding. Default: conv_padding = 0. - conv_dilation (int|list|tuple): The dilation size of the Conv2d Layer. If dilation is + conv_dilation (int|list|tuple): The dilation size of the conv2d Layer. If dilation is a list or tuple, it must contain two integers, (conv_dilation_H, conv_dilation_W). Otherwise, the conv_dilation_H = conv_dilation_W = conv_dilation. Default: conv_dilation = 1. - conv_groups (int): The groups number of the Conv2d Layer. According to grouped + conv_groups (int): The groups number of the conv2d Layer. According to grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only - connected to the second half of the input channels. Default: groups=1 - param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None - bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None - act (str): Activation type for Conv2d. Default: None + connected to the second half of the input channels. Default: groups=1. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv2d. If it is set to None or one attribute of ParamAttr, conv2d + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with :math:`Normal(0.0, std)`, + and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. + Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv2d + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. + act (str): Activation type for conv2d, if it is set to None, activation is not + appended. Default: None. use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 1b9571f6d3a6a69d1ac35f6be74b80eaa2ce6251..17af44afdde5cdbec082d473457ef01974695bc6 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -15,7 +15,7 @@ from __future__ import print_function import re from collections import defaultdict -from paddle.fluid.framework import Program, Variable, name_scope +from paddle.fluid.framework import Program, Variable, name_scope, default_main_program from . import framework from . import layers from .backward import append_backward @@ -111,7 +111,8 @@ class Optimizer(object): if param_lr == 1.0: return self._global_learning_rate() else: - return self._global_learning_rate() * param_lr + with default_main_program()._lr_schedule_guard(): + return self._global_learning_rate() * param_lr def _create_accumulators(self, block, parameters): """Create all accumulators needed by the parameters @@ -659,6 +660,9 @@ class AdamaxOptimizer(Optimizer): optimizer = fluid.optimizer.Adamax(learning_rate=0.2) optimizer.minimize(cost) + + Notes: + Currently, AdamaxOptimizer doesn't support sparse parameter optimization. """ _moment_acc_str = "moment" _inf_norm_acc_str = "inf_norm" @@ -778,6 +782,9 @@ class DecayedAdagradOptimizer(Optimizer): optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2) optimizer.minimize(cost) + + Notes: + Currently, DecayedAdagradOptimizer doesn't support sparse parameter optimization. """ _moment_acc_str = "moment" @@ -858,6 +865,9 @@ class AdadeltaOptimizer(Optimizer): optimizer = fluid.optimizer.Adadelta( learning_rate=0.0003, epsilon=1.0e-6, rho=0.95) _, params_grads = optimizer.minimize(cost) + + Notes: + Currently, AdadeltaOptimizer doesn't support sparse parameter optimization. """ _avg_squared_grad_acc_str = "_avg_squared_grad" @@ -1126,6 +1136,9 @@ class FtrlOptimizer(Optimizer): optimizer = fluid.optimizer.Ftrl(0.0001) _, params_grads = optimizer.minimize(cost) + + Notes: + Currently, FtrlOptimizer doesn't support sparse parameter optimization. """ _squared_acc_str = "squared" diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 57d272cbfb948840679e80e8db40379c57603113..3f4dd5eb712e738bbee8f93c062375033b8ab2f6 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -31,15 +31,32 @@ BuildStrategy = core.ParallelExecutor.BuildStrategy class ParallelExecutor(object): """ - ParallelExecutor can run program in parallel. + ParallelExecutor is designed for data parallelism, which focuses on distributing + the data across different nodes and every node operates on the data in parallel. + If you use ParallelExecutor to run the current program on GPU, the node means GPU + device, and ParallelExecutor will get the available GPU device automatically on + the current machine. If you use ParallelExecutor to run the current program on CPU, + the node means the CPU device, and you can specify the CPU device number by adding + 'CPU_NUM' environment variable, for example 'CPU_NUM=4', if the environment variable + is not found, ParallelExecutor will call `multiprocessing.cpu_count` to get the number + of CPUs in the system. Args: use_cuda (bool): Whether to use CUDA or not. loss_name (str): The loss name must set in training. Default None. main_program (Program): The program that need to run, if not provided, then default_main_program will be used. Default None. - share_vars_from(ParallelExecutor): If provied, it will share variables + share_vars_from(ParallelExecutor): If provide, it will share variables from the specified ParallelExecutor. Default None. + exec_strategy(ExecutionStrategy): exec_strategy is used to control how to run + the program in ParallelExecutor, for example how many threads are used to + execute the program, how many iterations to clean up the temp variables + which is generated during execution. For more information, please refer + to fluid.ExecutionStrategy. Default None. + build_strategy(BuildStrategy): build_strategy is used to control how to + build the SSA Graph in ParallelExecutor by setting the property, + for example reduce_strategy, gradient_scale_strategy. For more information, + please refer to fluid.BuildStrategy. Default None. num_trainers(int): If greater than 1, NCCL will be initialized with multiple rank of nodes, each node should have same number of GPUs. Distributed training will be enabled then. Default 1. diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py index a4336e955f21b0b09bf3dadbd437855c06745860..97644df007117be35ccd4d66495d8ca0ec7b0e88 100644 --- a/python/paddle/fluid/regularizer.py +++ b/python/paddle/fluid/regularizer.py @@ -237,6 +237,7 @@ class L1DecayRegularizer(WeightDecayRegularizer): 'Ids': idx}, outputs={'Out': decay}, attrs={'is_sparse': True}) + param = decay # Append sign op block.append_op( diff --git a/python/paddle/fluid/tests/CMakeLists.txt b/python/paddle/fluid/tests/CMakeLists.txt index 1885dda44ab5eaeca6a4f54e4b84379c71ec3167..d6568cd38e714bf9eb9d34da8a1c6a5cdb6677e3 100644 --- a/python/paddle/fluid/tests/CMakeLists.txt +++ b/python/paddle/fluid/tests/CMakeLists.txt @@ -1,4 +1,9 @@ -set(PYTHON_TESTS_DIR ${CMAKE_CURRENT_BINARY_DIR} CACHE PATH "python tests directory") +if(NOT APPLE) + set(PYTHON_TESTS_DIR ${CMAKE_CURRENT_BINARY_DIR} CACHE PATH "python tests directory") +else() + set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) +endif(NOT APPLE) + file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") diff --git a/python/paddle/fluid/tests/unittests/dist_simnet_bow.py b/python/paddle/fluid/tests/unittests/dist_simnet_bow.py index 6456d1b53a129db04ace7ff4413a3d76e922ccde..fac5e037a46715d146e354825f09ee8ccc4f3d70 100644 --- a/python/paddle/fluid/tests/unittests/dist_simnet_bow.py +++ b/python/paddle/fluid/tests/unittests/dist_simnet_bow.py @@ -81,7 +81,10 @@ def get_optimizer(): return optimizer -def train_network(batch_size, is_distributed=False, is_sparse=False): +def train_network(batch_size, + is_distributed=False, + is_sparse=False, + is_self_contained_lr=False): # query q = fluid.layers.data( name="query_ids", shape=[1], dtype="int64", lod_level=1) @@ -93,7 +96,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False): param_attr=fluid.ParamAttr( initializer=fluid.initializer.Constant(value=0.01), name="__emb__", - learning_rate=emb_lr), + learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.01), + name="__emb__"), is_sparse=is_sparse) ## vsum q_sum = fluid.layers.sequence_pool(input=q_emb, pool_type='sum') @@ -119,7 +124,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False): param_attr=fluid.ParamAttr( initializer=fluid.initializer.Constant(value=0.01), name="__emb__", - learning_rate=emb_lr), + learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.01), + name="__emb__"), is_sparse=is_sparse) ## vsum pt_sum = fluid.layers.sequence_pool(input=pt_emb, pool_type='sum') @@ -144,7 +151,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False): param_attr=fluid.ParamAttr( initializer=fluid.initializer.Constant(value=0.01), name="__emb__", - learning_rate=emb_lr), + learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.01), + name="__emb__"), is_sparse=is_sparse) ## vsum nt_sum = fluid.layers.sequence_pool(input=nt_emb, pool_type='sum') @@ -220,7 +229,10 @@ class TestDistSimnetBow2x2(TestDistRunnerBase): def get_model(self, batch_size=2): # Train program avg_cost, acc, predict = \ - train_network(batch_size, bool(int(os.environ["IS_DISTRIBUTED"])), bool(int(os.environ["IS_SPARSE"]))) + train_network(batch_size, + bool(int(os.environ["IS_DISTRIBUTED"])), + bool(int(os.environ["IS_SPARSE"])), + bool(int(os.environ["IS_SELF_CONTAINED_LR"]))) inference_program = fluid.default_main_program().clone() diff --git a/python/paddle/fluid/tests/unittests/test_affine_channel_op.py b/python/paddle/fluid/tests/unittests/test_affine_channel_op.py new file mode 100644 index 0000000000000000000000000000000000000000..2c9a063e6ee75371e0d05e1ff6964753017881a1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_affine_channel_op.py @@ -0,0 +1,106 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import numpy as np +from op_test import OpTest +import paddle.fluid.core as core + + +def affine_channel(x, scale, bias, layout): + C = x.shape[1] if layout == 'NCHW' else x.shape[-1] + if len(x.shape) == 4: + new_shape = (1, C, 1, 1) if layout == 'NCHW' else (1, 1, 1, C) + else: + new_shape = (1, C) + scale = scale.reshape(new_shape) + bias = bias.reshape(new_shape) + return x * scale + bias + + +class TestAffineChannelOp(OpTest): + def setUp(self): + self.op_type = "affine_channel" + self.init_test_case() + + x = np.random.random(self.shape).astype("float32") + scale = np.random.random(self.C).astype("float32") + bias = np.random.random(self.C).astype("float32") + + y = affine_channel(x, scale, bias, self.layout) + + self.inputs = {'X': x, 'Scale': scale, 'Bias': bias} + self.attrs = {'data_layout': self.layout} + self.outputs = {'Out': y} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X', 'Scale', 'Bias'], 'Out') + + def test_check_grad_stopgrad_dx(self): + self.check_grad(['Scale', 'Bias'], 'Out', no_grad_set=set('X')) + + def test_check_grad_stopgrad_dscale_dbias(self): + self.check_grad(['X'], 'Out', no_grad_set=set(['Scale', 'Bias'])) + + def init_test_case(self): + self.shape = [2, 32, 14, 14] + self.C = 32 + self.layout = 'NCHW' + + +class TestAffineChannelNHWC(TestAffineChannelOp): + def init_test_case(self): + self.shape = [2, 14, 14, 32] + self.C = 32 + self.layout = 'NHWC' + + +class TestAffineChannel2D(TestAffineChannelOp): + def init_test_case(self): + self.shape = [16, 64] + self.C = 64 + self.layout = 'NCHW' + + +class TestAffineChannelNCHWLargeShape(TestAffineChannelOp): + def init_test_case(self): + self.shape = [64, 128, 112, 112] + self.C = 128 + self.layout = 'NCHW' + + # since the gradient check is very slow in large shape, so skip check_grad + def test_check_grad(self): + pass + + def test_check_grad_stopgrad_dx(self): + pass + + def test_check_grad_stopgrad_dscale_dbias(self): + pass + + +class TestAffineChannelNCHWLargeShape(TestAffineChannelNCHWLargeShape): + def init_test_case(self): + self.shape = [64, 112, 112, 512] + self.C = 512 + self.layout = 'NHWC' + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py b/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py index 6103c3aafc0bb154194314830c5c8c5d89460cfe..46433d78252219fe02c3c4b5ddfc525bda177f18 100644 --- a/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py @@ -18,6 +18,9 @@ import unittest import numpy as np from op_test import OpTest +import paddle.fluid as fluid +import paddle.fluid.core as core + class TestClipByNormOp(OpTest): def setUp(self): @@ -62,5 +65,59 @@ class TestCase3(TestClipByNormOp): self.max_norm = 1.0 +class TestClipByNormOpWithSelectedRows(OpTest): + def check_with_place(self, place): + self.config_test_case() + scope = core.Scope() + + # set input + x_selected_rows = scope.var('X').get_selected_rows() + x_selected_rows.set_rows(self.grad_rows) + x_tensor = x_selected_rows.get_tensor() + x_np = np.random.random(self.grad_shape).astype("float32") + x_np[np.abs(x_np) < self.max_relative_error] = 0.5 + x_tensor.set(x_np, place) + + # set output + out_selected_rows = scope.var('Out').get_selected_rows() + + # run clip_by_norm_op + clip_by_norm_op = fluid.op.Operator( + "clip_by_norm", max_norm=self.max_norm, X='X', Out='Out') + clip_by_norm_op.run(scope, place) + + # check output + self.assertEqual(out_selected_rows.rows(), self.grad_clipped_rows) + out_tensor = out_selected_rows.get_tensor() + y_np = np.zeros(self.grad_clipped_shape) + y_np[0] = np.sum(x_np[0:2]) + y_np[1] = x_np[2] + y_np[2] = x_np[3] + norm = np.sqrt(np.sum(np.square(y_np))) + if norm > self.max_norm: + output = self.max_norm * y_np / norm + else: + output = y_np + self.assertTrue( + np.allclose( + np.array(out_tensor), output, atol=1e-5, equal_nan=False)) + + def test_clip_by_norm_with_selected_ros(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + + for place in places: + self.check_with_place(place) + + def config_test_case(self): + self.max_norm = 1.0 + self.max_relative_error = 0.006 + self.grad_shape = (4, 1) + self.grad_clipped_shape = (3, 1) + self.grad_rows = [0, 0, 1, 2] + self.grad_clipped_rows = [0, 1, 2] + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py b/python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py index e971f29db42a7c1a2394505a8ece3d2fd6b347e9..a0b6879f99e80a9710ee76f981769299a066b85b 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py +++ b/python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py @@ -25,7 +25,11 @@ class TestDistSimnetBowDense2x2(TestDistBase): self._enforce_place = "CPU" def test_simnet_bow(self): - need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '0'} + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '0', + 'IS_SELF_CONTAINED_LR': '1' + } self.check_with_place( "dist_simnet_bow.py", delta=1e-5, @@ -39,7 +43,11 @@ class TestDistSimnetBow2x2DenseAsync(TestDistBase): self._enforce_place = "CPU" def test_simnet_bow(self): - need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '0'} + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '0', + 'IS_SELF_CONTAINED_LR': '1' + } self.check_with_place( "dist_simnet_bow.py", delta=100, @@ -53,7 +61,11 @@ class TestDistSimnetBowSparse2x2(TestDistBase): self._enforce_place = "CPU" def test_simnet_bow(self): - need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '1'} + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '1' + } self.check_with_place( "dist_simnet_bow.py", delta=1e-5, @@ -67,7 +79,11 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase): self._enforce_place = "CPU" def test_simnet_bow(self): - need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '1'} + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '1' + } self.check_with_place( "dist_simnet_bow.py", delta=100, @@ -75,5 +91,61 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase): need_envs=need_envs) +# FIXME(tangwei): Learningrate variable is not created on pserver. +""" +class TestDistSimnetBow2x2LookupTableSync(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._enforce_place = "CPU" + + def test_simnet_bow(self): + need_envs = { + "IS_DISTRIBUTED": '1', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '1' + } + self.check_with_place( + "dist_simnet_bow.py", + delta=1e-5, + check_error_log=True, + need_envs=need_envs) + + +class TestDistSimnetBow2x2LookupTableAsync(TestDistBase): + def _setup_config(self): + self._sync_mode = False + self._enforce_place = "CPU" + + def test_simnet_bow(self): + need_envs = { + "IS_DISTRIBUTED": '1', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '1' + } + self.check_with_place( + "dist_simnet_bow.py", + delta=100, + check_error_log=False, + need_envs=need_envs) + + +class TestDistSimnetBow2x2LookupTableNotContainLRSync(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._enforce_place = "CPU" + + def test_simnet_bow(self): + need_envs = { + "IS_DISTRIBUTED": '1', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '0' + } + self.check_with_place( + "dist_simnet_bow.py", + delta=1e-5, + check_error_log=False, + need_envs=need_envs) +""" + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fused_embedding_fc_lstm_op.py b/python/paddle/fluid/tests/unittests/test_fused_embedding_fc_lstm_op.py new file mode 100644 index 0000000000000000000000000000000000000000..70ca521d3387ac11cd41d8496b4d094667232d4c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fused_embedding_fc_lstm_op.py @@ -0,0 +1,218 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import numpy as np +from op_test import OpTest +from test_lstm_op import lstm, ACTIVATION + + +def fc(x, w, b): + return np.dot(x, w) + b + + +def fused_embedded_fc_lstm( + ids, # T x 1 + lod, # 1 x N + embeddings=None, # Dict_size x M + wx=None, # M x 4D + bx=None, # 1 x 4D + h0=None, # N x D + c0=None, # N x D + w_h=None, # D x 4D + w_b=None, # 1 x 4D + w_c=None, # 1 x 3D + is_reverse=False, + act_gate=None, + act_cell=None, + act_cand=None): + # Make a lookup for embeddings and pass result into lstm reference + T = ids.shape[0] + M = embeddings.shape[1] + x = embeddings[ids].reshape([T, M]) + return lstm( + fc(x, wx, bx), lod, h0, c0, w_h, w_b, w_c, is_reverse, act_gate, + act_cell, act_cand) + + +class TestFusionLSTMOp(OpTest): + def set_conf(self): + pass + + def setUp(self): + self.op_type = 'fused_embedding_fc_lstm' + self.lod = [[2, 3, 5, 4]] + self.M = 8 # Embedding size + self.D = 16 # Hidden size + self.dict_size = 18 + self.has_initial_state = False + self.use_peepholes = False + self.is_reverse = False + self.act_gate = 'sigmoid' + self.act_cell = 'tanh' + self.act_cand = 'tanh' + self.set_conf() + + T = sum(self.lod[0]) + bs = len(self.lod[0]) + + # this is the weight of fc + wx = np.random.normal(size=(self.M, 4 * self.D)).astype('float32') + # this is the bias of fc + bx = np.random.normal(size=(1, 4 * self.D)).astype('float32') + + if self.use_peepholes: + b = np.random.normal(size=(1, 7 * self.D)).astype('float32') + else: + b = np.random.normal(size=(1, 4 * self.D)).astype('float32') + w_b = np.copy(b[:, 0:4 * self.D]) + w_c = b[:, 4 * self.D:] if self.use_peepholes else None + + # low is 0 , high is voc_size - 1 + ids = np.random.randint( + low=0, high=self.dict_size - 1, size=(T, 1)).astype("int64") + # embeddings as they were trained , so each entry is of M size + embeddings = np.random.random( + (self.dict_size, self.M)).astype("float32") + + # multiply embeddings via Weights + fc_embeddings = np.dot(embeddings, wx) + + # bias should be manually added into the bias of this fused embedding fc LSTM + b[0, 0:4 * self.D] += bx[0, :] + combined_biases = b[:, 0:4 * self.D] + # So let broadcast it , so they can be added + ones = np.ones([self.dict_size, 1]) + broadcasted_biases = np.dot(ones, combined_biases) + # Sum biases with Wx*embeddings + fc_embeddings += broadcasted_biases + + if self.has_initial_state: + h0 = np.random.normal(size=(bs, self.D)).astype('float32') + c0 = np.random.normal(size=(bs, self.D)).astype('float32') + else: + h0 = np.zeros((bs, self.D)).astype('float32') + c0 = np.zeros((bs, self.D)).astype('float32') + + wh = np.random.normal(size=(self.D, 4 * self.D)).astype('float32') + + h, c = fused_embedded_fc_lstm( + ids, self.lod, embeddings, wx, bx, h0, c0, wh, w_b, w_c, + self.is_reverse, ACTIVATION[self.act_gate], + ACTIVATION[self.act_cell], ACTIVATION[self.act_cand]) + + self.inputs = { + 'Ids': (ids, self.lod), + 'Embeddings': fc_embeddings, + 'WeightH': wh, + 'Bias': b + } + + if self.has_initial_state: + self.inputs['H0'] = h0 + self.inputs['C0'] = c0 + + self.outputs = { + 'Hidden': (h, self.lod), + 'Cell': (c, self.lod), + } + self.attrs = { + 'use_peepholes': self.use_peepholes, + 'is_reverse': self.is_reverse, + 'gate_activation': self.act_gate, + 'cell_activation': self.act_cell, + 'candidate_activation': self.act_cand + } + + def test_check_output(self): + for use_seq in {True, False}: + self.attrs['use_seq'] = use_seq + self.check_output() + + +class TestFusionLSTMOpInit(TestFusionLSTMOp): + def set_conf(self): + self.has_initial_state = True + + +class TestFusionLSTMOpReverse(TestFusionLSTMOp): + def set_conf(self): + self.is_reverse = True + + +class TestFusionLSTMOpInitReverse(TestFusionLSTMOp): + def set_conf(self): + self.has_initial_state = True + self.is_reverse = True + + +class TestFusionLSTMOpMD1(TestFusionLSTMOp): + def set_conf(self): + self.M = 36 + self.D = 8 + + +class TestFusionLSTMOpMD2(TestFusionLSTMOp): + def set_conf(self): + self.M = 8 + self.D = 8 + + +class TestFusionLSTMOpMD3(TestFusionLSTMOp): + def set_conf(self): + self.M = 15 + self.D = 3 + + +class TestFusionLSTMOpBS1(TestFusionLSTMOp): + def set_conf(self): + self.lod = [[3]] + self.D = 16 + + +class TestFusionLSTMOpPeepholes(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + + +class TestFusionLSTMOpPeepholesInit(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + self.has_initial_state = True + + +class TestFusionLSTMOpPeepholesReverse(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + self.is_reverse = True + + +class TestFusionLSTMOpPeepholesInitReverse(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + self.has_initial_state = True + self.is_reverse = True + + +class TestFusionLSTMOpPeepholesBS1(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + self.lod = [[2]] + self.D = 8 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fusion_seqconv_eltadd_relu_op.py b/python/paddle/fluid/tests/unittests/test_fusion_seqconv_eltadd_relu_op.py new file mode 100644 index 0000000000000000000000000000000000000000..ba6f1415b1c832eb688443953866652e3458b172 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fusion_seqconv_eltadd_relu_op.py @@ -0,0 +1,94 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import numpy as np +import random +from op_test import OpTest +from test_seq_conv import seqconv + + +class TestSeqConvEltAddRelu(OpTest): + def set_conf(self): + pass + + def setUp(self): + self.op_type = 'fusion_seqconv_eltadd_relu' + self.lod = [[6, 4]] + self.in_fea_size = 16 + self.out_fea_size = 8 + self.context_length = 4 + self.context_stride = 1 + self.context_start = 0 + self.set_conf() + + assert self.context_stride == 1 + + T = sum(self.lod[0]) + x = np.random.uniform(-1, 1, [T, self.in_fea_size]).astype('float32') + w = np.random.uniform( + -1, 1, [self.in_fea_size * self.context_length, + self.out_fea_size]).astype('float32') + b = np.random.uniform(-2, 1, [1, self.out_fea_size]).astype('float32') + out = seqconv(x, self.lod, w, self.context_length, self.context_start) + out = np.maximum(out + b, 0) + + self.inputs = {'X': (x, self.lod), 'Filter': w, 'Bias': b} + self.attrs = { + 'contextStart': self.context_start, + 'contextLength': self.context_length, + 'contextStride': self.context_stride + } + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + +class TestSeqConvEltAddReluBS1(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[10]] + + +class TestSeqConvEltAddReluBS1Case2(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[2]] + + +class TestSeqConvEltAddReluCase1(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[3, 5, 1, 6]] + self.context_length = 3 + self.context_start = -2 + + +class TestSeqConvEltAddReluCase2(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[10, 1, 2, 4, 1, 5, 6]] + self.in_fea_size = 2 + self.context_length = 4 + self.context_start = -1 + + +class TestSeqConvEltAddReluCase3(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[10, 1, 2, 4, 1, 5, 6]] + self.context_length = 5 + self.context_start = -4 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 1d8d0b55f0c5d7cffa01a100847bdf48b6d7023d..50de468dba803d0a2a0c129ad04aac8a3822cdbc 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -194,6 +194,14 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(layers.sequence_expand(x=x, y=y, ref_level=1)) print(str(program)) + def test_sequence_unpad(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[10, 5], dtype='float32') + length = layers.data(name='length', shape=[1], dtype='int64') + self.assertIsNotNone(layers.sequence_unpad(x=x, length=length)) + print(str(program)) + def test_lstm_unit(self): program = Program() with program_guard(program): @@ -406,6 +414,19 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(out) print(str(program)) + def test_sequence_slice(self): + program = Program() + with program_guard(program): + import numpy as np + seqs = layers.data( + name='x', shape=[10, 5], dtype='float32', lod_level=1) + offset = layers.assign(input=np.array([[0, 1]]).astype('int32')) + length = layers.assign(input=np.array([[2, 1]]).astype('int32')) + out = layers.sequence_slice( + input=seqs, offset=offset, length=length) + self.assertIsNotNone(out) + print(str(program)) + def test_lod_reset(self): program = Program() with program_guard(program): @@ -444,6 +465,16 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(output) print(str(program)) + def test_roi_align(self): + program = Program() + with program_guard(program): + x = layers.data(name="x", shape=[256, 30, 30], dtype="float32") + rois = layers.data( + name="rois", shape=[4], dtype="float32", lod_level=1) + output = layers.roi_align(x, rois, 14, 14, 0.5, 2) + self.assertIsNotNone(output) + print(str(program)) + def test_resize_bilinear(self): program = Program() with program_guard(program): diff --git a/python/paddle/fluid/tests/unittests/test_momentum_op.py b/python/paddle/fluid/tests/unittests/test_momentum_op.py index 7137fd0fdb7c503492107da684b95989037eb872..a3d89610b40ff9bd5002e843f8667ada87e67981 100644 --- a/python/paddle/fluid/tests/unittests/test_momentum_op.py +++ b/python/paddle/fluid/tests/unittests/test_momentum_op.py @@ -16,6 +16,8 @@ from __future__ import print_function import unittest import numpy as np +import paddle.fluid.core as core +from paddle.fluid.op import Operator from op_test import OpTest @@ -88,5 +90,97 @@ class TestMomentumOp2(OpTest): self.check_output() +class TestSparseMomentumOp(unittest.TestCase): + def setUp(self): + self.use_nesterov = False + + def check_with_place(self, place): + self.init_kernel() + scope = core.Scope() + # create and initialize Grad Variable + height = 10 + rows = [0, 4, 7] + row_numel = 12 + mu = 1.0 + use_nesterov = self.use_nesterov + + # create and initialize Param Variable + param = scope.var('Param').get_tensor() + param_array = np.full((height, row_numel), 5.0).astype("float32") + param.set(param_array, place) + param_out = scope.var("ParamOut").get_tensor() + param_out_array = np.full((height, row_numel), 0.0).astype("float32") + param_out.set(param_out_array, place) + + grad_selected_rows = scope.var('Grad').get_selected_rows() + grad_selected_rows.set_height(height) + grad_selected_rows.set_rows(rows) + grad_np_array = np.ones((len(rows), row_numel)).astype("float32") + grad_np_array[0, 0] = 2.0 + grad_np_array[2, 8] = 4.0 + grad_tensor = grad_selected_rows.get_tensor() + grad_tensor.set(grad_np_array, place) + + velocity = scope.var('Velocity').get_tensor() + velocity_np_array = np.ones((height, row_numel)).astype("float32") + velocity.set(velocity_np_array, place) + velocity_out = scope.var('VelocityOut').get_tensor() + velocity_out_np_array = np.full((height, row_numel), + 0.0).astype("float32") + velocity_out.set(velocity_out_np_array, place) + + # create and initialize LeraningRate Variable + lr = scope.var('LearningRate').get_tensor() + lr_array = np.full((1), 2.0).astype("float32") + lr.set(lr_array, place) + + # create and run operator + op = Operator( + "momentum", + Param='Param', + Grad='Grad', + Velocity='Velocity', + ParamOut='ParamOut', + VelocityOut='VelocityOut', + LearningRate='LearningRate', + mu=mu, + use_nesterov=use_nesterov) + op.run(scope, place) + + # get and compare result + param_out_np_array = np.array(param_out) + velocity_out_np_array = np.array(velocity_out) + + # TODO(dzh): add a more suitable general numpy interface + # for sparse update. + _grad_np_array = np.full((height, row_numel), 0.0).astype("float32") + for i in range(len(rows)): + _grad_np_array[rows[i]] = grad_np_array[i] + _velocity_out = mu * velocity_np_array + _grad_np_array + _param = param_array + if use_nesterov: + _param_out = _param - (_grad_np_array + _velocity_out * mu + ) * lr_array + else: + _param_out = _param - lr_array * _velocity_out + self.assertTrue((_velocity_out == velocity_out_np_array).all()) + self.assertTrue((_param_out == param_out_np_array).all()) + + def init_kernel(self): + pass + + def test_sparse_momentum(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + for place in places: + self.check_with_place(place) + + +class TestSparseMomentumOp2(TestSparseMomentumOp): + def init_kernel(self): + self.use_nesterov = True + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py index dfedf8190f75ec26532f281338f076ca0c7d83af..7f266056a9d98be1a6f67473be65a74957f943e9 100644 --- a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py +++ b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py @@ -37,7 +37,7 @@ def PolygonBoxRestore(input): indexes = indexes.repeat( [batch_size], axis=0) # [batch_size, geo_channels/2, 2, h, w] return indexes.reshape( - input.shape) - input # [batch_size, geo_channels, h, w] + input.shape) * 4 - input # [batch_size, geo_channels, h, w] class TestPolygonBoxRestoreOp(OpTest): diff --git a/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py b/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py index b85b94c939f76a3ba67a301682d9d356b481e0c9..b7fad9b3a60632adb564e1d155a3d935706b467f 100644 --- a/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py +++ b/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py @@ -53,22 +53,13 @@ def simple_fc_net(in_size, hidden_sizes, batch_size, queue_capacity, - use_double_buffer=False, - use_feed_list=True): - if use_feed_list: - data = fluid.layers.data(name="data", dtype='float32', shape=[in_size]) - label = fluid.layers.data(name='label', dtype='int64', shape=[1]) - reader = fluid.layers.create_py_reader_by_data( - capacity=queue_capacity, - use_double_buffer=False, - feed_list=[data, label]) - else: - reader = fluid.layers.py_reader( - capacity=queue_capacity, - shapes=[[-1, in_size], [-1, 1]], - lod_levels=[0, 0], - dtypes=['float32', 'int64'], - use_double_buffer=False) + use_double_buffer=False): + reader = fluid.layers.py_reader( + capacity=queue_capacity, + shapes=[[-1, in_size], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64'], + use_double_buffer=False) feed_queue = reader.queue reader = fluid.layers.batch(reader, batch_size=batch_size) if use_double_buffer: @@ -109,16 +100,14 @@ class TestPyReaderUsingExecutor(unittest.TestCase): if core.is_compiled_with_cuda() else [False]): for use_parallel_executor in [False, True]: for use_double_buffer in [False, True]: - for use_feed_list in [False, True]: - print('Test Parameters:'), - print({ - 'use_cuda': use_cuda, - 'use_parallel_executor': use_parallel_executor, - 'use_double_buffer': use_double_buffer, - 'use_feed_list': use_feed_list - }) - self.main(use_cuda, use_parallel_executor, - use_double_buffer, use_feed_list) + print('Test Parameters:'), + print({ + 'use_cuda': use_cuda, + 'use_parallel_executor': use_parallel_executor, + 'use_double_buffer': use_double_buffer + }) + self.main(use_cuda, use_parallel_executor, + use_double_buffer) def random_reader(self): def reader(): @@ -154,14 +143,12 @@ class TestPyReaderUsingExecutor(unittest.TestCase): def main(self, use_cuda=True, use_parallel_executor=False, - use_double_buffer=False, - use_feed_list=False): + use_double_buffer=False): assert not use_cuda or use_cuda and core.is_compiled_with_cuda() self.use_cuda = use_cuda self.use_parallel_executor = use_parallel_executor self.use_double_buffer = use_double_buffer - self.use_feed_list = use_feed_list startup_program = fluid.Program() main_program = fluid.Program() @@ -173,8 +160,7 @@ class TestPyReaderUsingExecutor(unittest.TestCase): hidden_sizes=self.hidden_sizes, batch_size=self.batch_size, queue_capacity=self.queue_capacity, - use_double_buffer=self.use_double_buffer, - use_feed_list=self.use_feed_list) + use_double_buffer=self.use_double_buffer) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() diff --git a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py index 70848e4e2239e2be160bb0c1a28a5aecd01a87dc..eb12bc741767340a3e7e3580a8b95065d4267693 100644 --- a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py +++ b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py @@ -19,33 +19,76 @@ import unittest import numpy as np import paddle.fluid.core as core from paddle.fluid.op import Operator +import paddle.fluid as fluid + + +def create_selected_rows_and_tensor(scope, place, height, row_num, + embedding_size): + sr = scope.var("@selected_rows@").get_selected_rows() + tensor = scope.var("grad").get_tensor() + + rows = np.random.random_integers( + low=0, high=height - 1, size=[row_num, ]).astype('int64') + sr_val = np.random.random(size=[row_num, embedding_size]).astype('float32') + + sr.set_height(height) + sr.set_rows(rows) + sr.get_tensor().set(sr_val, place) + + tensor_val = np.zeros(shape=[height, embedding_size], dtype='float32') + for i in range(row_num): + row = rows[i] + tensor_val[row, :] = tensor_val[row, :] + sr_val[i, :] + + tensor.set(tensor_val, place) + return tensor_val, sr_val class TestBase(unittest.TestCase): - def setup(self, centered, epsilon=1e-6): + def setup(self, + place, + is_sparse, + centered, + size, + row_num=None, + epsilon=1e-6): np.random.seed(5) # fix seed + self.scope = fluid.global_scope() + self.place = place + self.param_name = "param" - self.param = np.random.random((123, 321)).astype("float32") + self.param = np.random.random(size).astype("float32") self.mean_square_name = "mean_square" - self.mean_square = np.random.random((123, 321)).astype("float32") + self.mean_square = np.random.uniform( + low=1, high=2, size=size).astype("float32") self.mean_grad_name = "mean_grad" - self.mean_grad = np.random.random((123, 321)).astype("float32") + self.mean_grad = np.random.random(size).astype("float32") self.lr_name = "lr" self.learning_rate = np.array([0.01]).astype("float32") self.grad_name = "grad" - self.grad = np.random.random((123, 321)).astype("float32") + + self.is_sparse = is_sparse + if self.is_sparse: + self.grad_sr_name = "@selected_rows@" + self.grad, self.grad_sr = create_selected_rows_and_tensor( + self.scope, place, size[0], row_num, size[1]) + else: + self.grad = np.random.random(size).astype("float32") + grad_tensor = self.scope.var(self.grad_name).get_tensor() + grad_tensor.set(self.grad, place) self.moment_name = "moment" - self.moment = np.zeros((123, 321)).astype("float32") + self.moment = np.random.uniform( + low=0, high=1, size=size).astype("float32") self.epsilon = epsilon self.decay = 0.9 - self.momentum = 0.0 + self.momentum = 0.1 self.centered = centered self.ms_out = self.decay * self.mean_square + (1 - self.decay @@ -61,118 +104,122 @@ class TestBase(unittest.TestCase): self.param_out = self.param - self.moment_out - def check(self, - actual_t, - expect_t, - place, - out_name, - atol=1e-5, - equal_nan=False): - self.assertTrue( - np.allclose( - actual_t, expect_t, atol=atol, equal_nan=equal_nan), - "Output (" + out_name + ") has diff at " + str(place) + "\nExpect " - + str(expect_t) + "\n" + "But Got" + str(actual_t)) - - -class TestRmspropOp(TestBase): - def check_with_place(self, place, centered, epsilon): - self.setup(centered, epsilon) - scope = core.Scope() - # create and initialize Param Variable - param = scope.var(self.param_name).get_tensor() - param.set(self.param, place) + self.param_tensor = self.scope.var(self.param_name).get_tensor() + self.param_tensor.set(self.param, place) - mean_square = scope.var(self.mean_square_name).get_tensor() - mean_square.set(self.mean_square, place) + self.mean_square_tensor = self.scope.var( + self.mean_square_name).get_tensor() + self.mean_square_tensor.set(self.mean_square, place) - lr = scope.var(self.lr_name).get_tensor() + lr = self.scope.var(self.lr_name).get_tensor() lr.set(self.learning_rate, place) - grad = scope.var(self.grad_name).get_tensor() - grad.set(self.grad, place) + self.moment_tensor = self.scope.var(self.moment_name).get_tensor() + self.moment_tensor.set(self.moment, place) - moment = scope.var(self.moment_name).get_tensor() - moment.set(self.moment, place) + if self.centered: + self.mean_grad_tensor = self.scope.var( + self.mean_grad_name).get_tensor() + self.mean_grad_tensor.set(self.mean_grad, place) - # create and run sgd operator + def check(self, actual_t, expect_t, place, out_name, atol=1e-5): + self.assertTrue( + np.allclose( + actual_t, expect_t, atol=atol), + "Output (" + out_name + ") has diff at " + str(place) + "\nExpect " + + str(expect_t) + "\n" + "But Got" + str(actual_t)) - if self.centered: - mean_grad = scope.var(self.mean_grad_name).get_tensor() - mean_grad.set(self.mean_grad, place) - - rmsprop_op = Operator( - "rmsprop", - Param=self.param_name, - Grad=self.grad_name, - MeanSquare=self.mean_square_name, - MeanGrad=self.mean_grad_name, - Moment=self.moment_name, - LearningRate=self.lr_name, - ParamOut=self.param_name, - MeanSquareOut=self.mean_square_name, - MomentOut=self.moment_name, - MeanGradOut=self.mean_grad_name, - epsilon=self.epsilon, - decay=self.decay, - momentum=self.momentum, - centered=True) - else: - rmsprop_op = Operator( - "rmsprop", - Param=self.param_name, - Grad=self.grad_name, - MeanSquare=self.mean_square_name, - Moment=self.moment_name, - LearningRate=self.lr_name, - ParamOut=self.param_name, - MeanSquareOut=self.mean_square_name, - MomentOut=self.moment_name, - epsilon=self.epsilon, - decay=self.decay, - momentum=self.momentum, - centered=False) - - rmsprop_op.run(scope, place) - - atol = 1e-5 - equal_nan = False + +class TestRmspropOp(TestBase): + def check_with_place(self, + place, + is_sparse, + centered, + size, + row_num=None, + epsilon=1e-6): + self.setup(place, is_sparse, centered, size, row_num, epsilon) + self.run_and_check() + + def run_and_check(self): + grad_name = self.grad_sr_name if self.is_sparse else self.grad_name + + kwargs = { + 'Param': self.param_name, + 'Grad': grad_name, + 'MeanSquare': self.mean_square_name, + 'Moment': self.moment_name, + 'LearningRate': self.lr_name, + 'ParamOut': self.param_name, + 'MeanSquareOut': self.mean_square_name, + 'MomentOut': self.moment_name, + 'epsilon': self.epsilon, + 'decay': self.decay, + 'momentum': self.momentum, + 'centered': self.centered + } if self.centered: - atol = 1e-3 - equal_nan = True + kwargs['MeanGrad'] = self.mean_grad_name + kwargs['MeanGradOut'] = self.mean_grad_name + + rmsprop_op = Operator('rmsprop', **kwargs) + atol = 1e-6 + + rmsprop_op.run(self.scope, self.place) self.check( - np.array(mean_square), self.ms_out, place, self.mean_square_name) + np.array(self.mean_square_tensor), + self.ms_out, + self.place, + self.mean_square_name, + atol=atol) self.check( - np.array(moment), + np.array(self.moment_tensor), self.moment_out, - place, + self.place, self.moment_name, - atol=atol, - equal_nan=equal_nan) + atol=atol) self.check( - np.array(param), + np.array(self.param_tensor), self.param_out, - place, + self.place, self.param_name, - atol=atol, - equal_nan=equal_nan) + atol=atol) if self.centered: self.check( - np.array(mean_grad), self.mg_out, place, self.mean_grad_name) + np.array(self.mean_grad_tensor), self.mg_out, self.place, + self.mean_grad_name) def test_rmsprop(self): places = [core.CPUPlace()] if core.is_compiled_with_cuda(): places.append(core.CUDAPlace(0)) + + size = (128, 320) for place in places: - self.check_with_place(place, False, 1e-6) - self.check_with_place(place, False, 1e-10) - self.check_with_place(place, True, 1e-6) - self.check_with_place(place, True, 1e-10) + for centered in [False, True]: + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, is_sparse=False, centered=centered, size=size) + + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, + is_sparse=True, + centered=centered, + row_num=512, + size=size) + + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, + is_sparse=True, + centered=centered, + row_num=60, + size=size) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_roi_align_op.py b/python/paddle/fluid/tests/unittests/test_roi_align_op.py new file mode 100644 index 0000000000000000000000000000000000000000..1a252ea547e4d93d83f64fa9cdb3605eeef0a3cf --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_roi_align_op.py @@ -0,0 +1,170 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import numpy as np +import math +import sys +from op_test import OpTest + + +class TestROIAlignOp(OpTest): + def set_data(self): + self.init_test_case() + self.make_rois() + self.calc_roi_align() + self.inputs = {'X': self.x, 'ROIs': (self.rois[:, 1:5], self.rois_lod)} + self.attrs = { + 'spatial_scale': self.spatial_scale, + 'pooled_height': self.pooled_height, + 'pooled_width': self.pooled_width, + 'sampling_ratio': self.sampling_ratio + } + + self.outputs = {'Out': self.out_data} + + def init_test_case(self): + self.batch_size = 3 + self.channels = 3 + self.height = 8 + self.width = 6 + + # n, c, h, w + self.x_dim = (self.batch_size, self.channels, self.height, self.width) + + self.spatial_scale = 1.0 / 2.0 + self.pooled_height = 2 + self.pooled_width = 2 + self.sampling_ratio = -1 + + self.x = np.random.random(self.x_dim).astype('float32') + + def pre_calc(self, x_i, roi_xmin, roi_ymin, roi_bin_grid_h, roi_bin_grid_w, + bin_size_h, bin_size_w): + count = roi_bin_grid_h * roi_bin_grid_w + bilinear_pos = np.zeros( + [self.channels, self.pooled_height, self.pooled_width, count, 4], + np.float32) + bilinear_w = np.zeros( + [self.pooled_height, self.pooled_width, count, 4], np.float32) + for ph in range(self.pooled_width): + for pw in range(self.pooled_height): + c = 0 + for iy in range(roi_bin_grid_h): + y = roi_ymin + ph * bin_size_h + (iy + 0.5) * \ + bin_size_h / roi_bin_grid_h + for ix in range(roi_bin_grid_w): + x = roi_xmin + pw * bin_size_w + (ix + 0.5) * \ + bin_size_w / roi_bin_grid_w + if y < -1.0 or y > self.height or \ + x < -1.0 or x > self.width: + continue + if y <= 0: + y = 0 + if x <= 0: + x = 0 + y_low = int(y) + x_low = int(x) + if y_low >= self.height - 1: + y = y_high = y_low = self.height - 1 + else: + y_high = y_low + 1 + if x_low >= self.width - 1: + x = x_high = x_low = self.width - 1 + else: + x_high = x_low + 1 + ly = y - y_low + lx = x - x_low + hy = 1 - ly + hx = 1 - lx + for ch in range(self.channels): + bilinear_pos[ch, ph, pw, c, 0] = x_i[ch, y_low, + x_low] + bilinear_pos[ch, ph, pw, c, 1] = x_i[ch, y_low, + x_high] + bilinear_pos[ch, ph, pw, c, 2] = x_i[ch, y_high, + x_low] + bilinear_pos[ch, ph, pw, c, 3] = x_i[ch, y_high, + x_high] + bilinear_w[ph, pw, c, 0] = hy * hx + bilinear_w[ph, pw, c, 1] = hy * lx + bilinear_w[ph, pw, c, 2] = ly * hx + bilinear_w[ph, pw, c, 3] = ly * lx + c = c + 1 + return bilinear_pos, bilinear_w + + def calc_roi_align(self): + self.out_data = np.zeros( + (self.rois_num, self.channels, self.pooled_height, + self.pooled_width)).astype('float32') + + for i in range(self.rois_num): + roi = self.rois[i] + roi_batch_id = int(roi[0]) + x_i = self.x[roi_batch_id] + roi_xmin = roi[1] * self.spatial_scale + roi_ymin = roi[2] * self.spatial_scale + roi_xmax = roi[3] * self.spatial_scale + roi_ymax = roi[4] * self.spatial_scale + roi_width = max(roi_xmax - roi_xmin, 1) + roi_height = max(roi_ymax - roi_ymin, 1) + bin_size_h = float(roi_height) / float(self.pooled_height) + bin_size_w = float(roi_width) / float(self.pooled_width) + roi_bin_grid_h = self.sampling_ratio if self.sampling_ratio > 0 else \ + math.ceil(roi_height / self.pooled_height) + roi_bin_grid_w = self.sampling_ratio if self.sampling_ratio > 0 else \ + math.ceil(roi_width / self.pooled_width) + count = int(roi_bin_grid_h * roi_bin_grid_w) + pre_size = count * self.pooled_width * self.pooled_height + bilinear_pos, bilinear_w = self.pre_calc(x_i, roi_xmin, roi_ymin, + int(roi_bin_grid_h), + int(roi_bin_grid_w), + bin_size_h, bin_size_w) + for ch in range(self.channels): + align_per_bin = (bilinear_pos[ch] * bilinear_w).sum(axis=-1) + output_val = align_per_bin.mean(axis=-1) + self.out_data[i, ch, :, :] = output_val + + def make_rois(self): + rois = [] + self.rois_lod = [[]] + for bno in range(self.batch_size): + self.rois_lod[0].append(bno + 1) + for i in range(bno + 1): + x1 = np.random.random_integers( + 0, self.width // self.spatial_scale - self.pooled_width) + y1 = np.random.random_integers( + 0, self.height // self.spatial_scale - self.pooled_height) + + x2 = np.random.random_integers(x1 + self.pooled_width, + self.width // self.spatial_scale) + y2 = np.random.random_integers( + y1 + self.pooled_height, self.height // self.spatial_scale) + + roi = [bno, x1, y1, x2, y2] + rois.append(roi) + self.rois_num = len(rois) + self.rois = np.array(rois).astype("float32") + + def setUp(self): + self.op_type = "roi_align" + self.set_data() + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') diff --git a/python/paddle/fluid/tests/unittests/test_seq_conv.py b/python/paddle/fluid/tests/unittests/test_seq_conv.py index dcc86382e5286f354c4f2e81ead598f12c75b2c1..2285e9496768aea6f48fb7796536e8344839d862 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_conv.py +++ b/python/paddle/fluid/tests/unittests/test_seq_conv.py @@ -20,6 +20,53 @@ import random from op_test import OpTest +def seqconv(x, + lod, + filter, + context_length, + context_start, + padding_trainable=False, + padding_data=None): + [T, M] = x.shape + col = np.zeros((T, context_length * M)).astype('float32') + offset = [0] + for seq_len in lod[0]: + offset.append(offset[-1] + seq_len) + begin_pad = np.max([0, -context_start]) + for i in range(len(offset) - 1): + for j in range(context_length): + in_begin = offset[i] + context_start + j + in_end = offset[i + 1] + context_start + j + out_begin = offset[i] + out_end = offset[i + 1] + if in_begin < offset[i]: + pad_size = np.min( + [offset[i] - in_begin, offset[i + 1] - offset[i]]) + if padding_trainable: + sub_w = padding_data[j:j + pad_size, :] + col[offset[i]:offset[i] + pad_size, j * M:(j + 1) * + M] = sub_w + out_begin = offset[i] + pad_size + in_begin = offset[i] + + if in_end > offset[i + 1]: + pad_size = np.min( + [in_end - offset[i + 1], offset[i + 1] - offset[i]]) + if padding_trainable: + sub_w = padding_data[begin_pad + context_start + j - + pad_size:begin_pad + context_start + + j, :] + col[offset[i + 1] - pad_size:offset[i + 1], j * M:(j + 1) * + M] = sub_w + in_end = offset[i + 1] + out_end = offset[i + 1] - pad_size + if in_end <= in_begin: + continue + in_sub = x[in_begin:in_end, :] + col[out_begin:out_end, j * M:(j + 1) * M] += in_sub + return np.dot(col, filter) + + class TestSeqProject(OpTest): def setUp(self): self.init_test_case() @@ -66,57 +113,9 @@ class TestSeqProject(OpTest): 'paddingTrainable': self.padding_trainable, 'contextStride': self.context_stride } - out = np.zeros( - (self.input_size[0], self.output_represention)).astype('float32') + out = seqconv(x, self.lod, w, self.context_length, self.context_start, + self.padding_trainable, self.pad_data) self.outputs = {'Out': out} - self.compute() - - def compute(self): - x, lod = self.inputs['X'] - filter = self.inputs['Filter'] - pading_data = self.pad_data - out = np.zeros((self.input_size[0], self.context_length * - self.input_size[1])).astype('float32') - offset = [0] - for seq_len in lod[0]: - offset.append(offset[-1] + seq_len) - begin_pad = np.max([0, -self.context_start]) - - for i in range(len(offset) - 1): - for j in range(self.context_length): - in_begin = offset[i] + self.context_start + j - in_end = offset[i + 1] + self.context_start + j - out_begin = offset[i] - out_end = offset[i + 1] - if in_begin < offset[i]: - pad_size = np.min( - [offset[i] - in_begin, offset[i + 1] - offset[i]]) - if self.padding_trainable: - sub_w = pading_data[j:j + pad_size, :] - out[offset[i]:offset[i] + pad_size, j * self.input_size[ - 1]:(j + 1) * self.input_size[1]] = sub_w - out_begin = offset[i] + pad_size - in_begin = offset[i] - - if in_end > offset[i + 1]: - pad_size = np.min( - [in_end - offset[i + 1], offset[i + 1] - offset[i]]) - if self.padding_trainable: - sub_w = pading_data[begin_pad + self.context_start + j - - pad_size:begin_pad + - self.context_start + j, :] - out[offset[i + 1] - pad_size:offset[i + 1], j * self. - input_size[1]:(j + 1) * self.input_size[1]] = sub_w - in_end = offset[i + 1] - out_end = offset[i + 1] - pad_size - if in_end <= in_begin: - continue - - in_sub = x[in_begin:in_end, :] - out[out_begin:out_end, j * self.input_size[1]:(j + 1) * - self.input_size[1]] += in_sub - - np.dot(out, filter, out=self.outputs['Out']) def test_check_output(self): self.check_output() diff --git a/python/paddle/fluid/tests/unittests/test_sequence_unpad_op.py b/python/paddle/fluid/tests/unittests/test_sequence_unpad_op.py new file mode 100644 index 0000000000000000000000000000000000000000..673b0ea180464b8b8f6f5c6e76d5c5c80f347d25 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_sequence_unpad_op.py @@ -0,0 +1,75 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import six +import numpy as np +from op_test import OpTest + + +class TestSequenceUnpadOp(OpTest): + def init(self): + self.length = [2, 3, 4] + self.x_shape = (3, 5) + self.dtype = "float32" + + def compute(self): + assert len(self.length) == self.x_shape[0] + x = np.random.random(self.x_shape).astype(self.dtype) + out_lod = [self.length] + + out = x[0, 0:self.length[0]] + for i in six.moves.xrange(1, x.shape[0]): + out = np.append(out, x[i, 0:self.length[i]], axis=0) + + out_shape = (sum(self.length), ) + if len(self.x_shape) == 2: + out_shape = out_shape + (1, ) + else: + out_shape = out_shape + self.x_shape[2:] + + self.inputs = { + 'X': x, + 'Length': np.array(self.length).astype('int64').reshape(-1, 1) + } + self.outputs = {'Out': (out.reshape(out_shape), out_lod)} + + def setUp(self): + self.op_type = 'sequence_unpad' + self.init() + self.compute() + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + +class TestSequenceUnpadOp2(TestSequenceUnpadOp): + def init(self): + self.length = [2, 3, 4] + self.x_shape = (3, 5, 4, 3) + self.dtype = "float32" + + +class TestSequenceUnpadOp3(TestSequenceUnpadOp): + def init(self): + self.length = [5, 2, 3, 4] + self.x_shape = (4, 5, 3, 3, 6) + self.dtype = "float64" + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index ecdbe27f4d90268d755a712e25289cfaf4715f29..2192139f8d5950286691a77333dd8ec35505b033 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -788,7 +788,8 @@ in a single call.") tuple: (main_program, startup_program), of type "Program" """ pserver_prog = self.get_pserver_program(endpoint) - pserver_startup = self.get_startup_program(endpoint) + pserver_startup = self.get_startup_program( + endpoint, pserver_program=pserver_prog) return pserver_prog, pserver_startup def get_startup_program(self, @@ -1118,6 +1119,7 @@ to transpile() call.") def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints): # 2. add split_ids_op and send_op to send gradient to pservers + # there should only be one table_name all_ops = program.global_block().ops table_grad_name = grad_var_name(self.table_name) @@ -1142,7 +1144,7 @@ to transpile() call.") if self.sync_mode else [] }, attrs={ - "sync_mode": self.sync_mode, + "sync_mode": not self.sync_mode, "epmap": pserver_endpoints, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE, OP_ROLE_VAR_ATTR_NAME: [ @@ -1188,7 +1190,15 @@ to transpile() call.") def _create_table_optimize_block(self, pserver_index, pserver_program, pre_block_idx, grad_to_block_id): # STEP: create table optimize block + table_opt_block = pserver_program._create_block(pre_block_idx) # create table param and grad var in pserver program + # create table optimize block in pserver program + table_opt_op = [ + op for op in self.optimize_ops + if 'Param' in op.input_names and op.input("Param")[0] == + self.table_name + ][0] + origin_param_var = self.origin_program.global_block().vars[ self.table_name] @@ -1204,19 +1214,16 @@ to transpile() call.") dtype=origin_param_var.dtype, type=core.VarDesc.VarType.SELECTED_ROWS, persistable=True) + # parameter must be selected rows param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS) grad_var = pserver_program.global_block()._clone_variable( self.origin_program.global_block().vars[grad_var_name( self.table_name)]) - # create table optimize block in pserver program - table_opt_op = [ - op for op in self.optimize_ops - if 'Param' in op.input_names and op.input("Param")[0] == - self.table_name - ][0] - table_opt_block = pserver_program._create_block(pre_block_idx) + lr_var = pserver_program.global_block()._clone_variable( + self.origin_program.global_block().vars[table_opt_op.input( + "LearningRate")[0]]) if self.sync_mode: # create grad vars in pserver program @@ -1248,8 +1255,6 @@ to transpile() call.") grad_var = pserver_program.global_block()._rename_var( origin_grad_name, splited_grad_name) - lr_var = pserver_program.global_block().vars[table_opt_op.input( - "LearningRate")[0]] inputs = { "Param": [param_var], "Grad": [grad_var], diff --git a/python/paddle/fluid/transpiler/inference_transpiler.py b/python/paddle/fluid/transpiler/inference_transpiler.py index c402535b27142e94af339a6c18401ba20bc6564d..5269bd94cec47a5262e2389c5b02f91edd5a7d17 100644 --- a/python/paddle/fluid/transpiler/inference_transpiler.py +++ b/python/paddle/fluid/transpiler/inference_transpiler.py @@ -74,7 +74,7 @@ class InferenceTranspiler(object): ''' Transpile the program fusing elementwise_add into conv for MKLDNN program. Elementwise add following convolution OP can be fused by adding - 'fuse_eltwise' attribute to convolution OP and replacing its output + 'fuse_residual_connection' attribute to convolution OP and replacing its output Tensor with second parameter of elementwise_add. The result of fuse is: - before: @@ -92,7 +92,8 @@ class InferenceTranspiler(object): if current_op.type in ['conv2d']: next_op = self.block.ops[i + 1] if next_op.type == 'elementwise_add': - self._fuse_conv_eltwise(current_op, next_op) + self._fuse_conv_eltwise(i, current_op, next_op) + self.block._remove_op(i + 1) # Remove old conv self.block._remove_op(i + 1) # Remove elementwise_add i = i + 1 self._adjust_input() @@ -444,7 +445,7 @@ class InferenceTranspiler(object): outputs={"Output": out_var}, attrs=attrs) - def _fuse_conv_eltwise(self, conv_op, eltwise_op): + def _fuse_conv_eltwise(self, index, conv_op, eltwise_op): ''' fuse the conv op with elementwise_add @@ -454,9 +455,30 @@ class InferenceTranspiler(object): :type eltwise_op: Operator ''' - conv_op._set_attr("fuse_eltwise", True) - self.input_map[conv_op.output("Output")[0]] = eltwise_op.input("Y")[0] - self.input_map[eltwise_op.output("Out")[0]] = eltwise_op.input("Y")[0] + eltwise_input = "X" + if eltwise_op.input("X")[0] == conv_op.output("Output")[0]: + eltwise_input = "Y" + + residual_var = self.block.vars[eltwise_op.input(eltwise_input)[0]] + out_var = self.block.vars[eltwise_op.output("Out")[0]] + filter_var = self.block.vars[conv_op.input("Filter")[0]] + in_var = self.block.vars[conv_op.input("Input")[0]] + bias_var = self.block.vars[conv_op.input("Bias")[0]] + + conv_op._set_attr("fuse_residual_connection", True) + attrs = {name: conv_op.attr(name) for name in conv_op.attr_names} + + self.block._insert_op( + index, + type="conv2d", + inputs={ + "Input": in_var, + "Filter": filter_var, + "Bias": bias_var, + "ResidualData": residual_var + }, + outputs={"Output": out_var}, + attrs=attrs) def _adjust_input(self): for i in range(len(self.block.ops)): diff --git a/python/paddle/utils/__init__.py b/python/paddle/utils/__init__.py index 15595d208583b567b8f768c8d7bd84986ca5a03f..5de6f966a038543ffffdf955251f587e3eb15cad 100644 --- a/python/paddle/utils/__init__.py +++ b/python/paddle/utils/__init__.py @@ -12,4 +12,5 @@ # See the License for the specific language governing permissions and # limitations under the License. -__all__ = ['dump_config'] +from plot import Ploter +__all__ = ['dump_config', 'Ploter'] diff --git a/python/paddle/utils/plot.py b/python/paddle/utils/plot.py new file mode 100644 index 0000000000000000000000000000000000000000..08889c0313fc24151cde6ca7b662d81eb53c9d7b --- /dev/null +++ b/python/paddle/utils/plot.py @@ -0,0 +1,115 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os + + +class PlotData(object): + def __init__(self): + self.step = [] + self.value = [] + + def append(self, step, value): + self.step.append(step) + self.value.append(value) + + def reset(self): + self.step = [] + self.value = [] + + +class Ploter(object): + """ + Plot input data in a 2D graph + + Args: + title: assign the title of input data. + step: x_axis of the data. + value: y_axis of the data. + """ + + def __init__(self, *args): + self.__args__ = args + self.__plot_data__ = {} + for title in args: + self.__plot_data__[title] = PlotData() + # demo in notebooks will use Ploter to plot figure, but when we convert + # the ipydb to py file for testing, the import of matplotlib will make the + # script crash. So we can use `export DISABLE_PLOT=True` to disable import + # these libs + self.__disable_plot__ = os.environ.get("DISABLE_PLOT") + if not self.__plot_is_disabled__(): + import matplotlib.pyplot as plt + from IPython import display + self.plt = plt + self.display = display + + def __plot_is_disabled__(self): + return self.__disable_plot__ == "True" + + def append(self, title, step, value): + """ + Feed data + + Args: + title: assign the group data to this subtitle. + step: the x_axis of data. + value: the y_axis of data. + + Examples: + .. code-block:: python + plot_curve = Ploter("Curve 1","Curve 2") + plot_curve.append(title="Curve 1",step=1,value=1) + """ + assert isinstance(title, basestring) + assert self.__plot_data__.has_key(title) + data = self.__plot_data__[title] + assert isinstance(data, PlotData) + data.append(step, value) + + def plot(self, path=None): + """ + Plot data in a 2D graph + + Args: + path: store the figure to this file path. Defaul None. + + Examples: + .. code-block:: python + plot_curve = Ploter() + plot_cure.plot() + """ + if self.__plot_is_disabled__(): + return + + titles = [] + for title in self.__args__: + data = self.__plot_data__[title] + assert isinstance(data, PlotData) + if len(data.step) > 0: + titles.append(title) + self.plt.plot(data.step, data.value) + self.plt.legend(titles, loc='upper left') + if path is None: + self.display.clear_output(wait=True) + self.display.display(self.plt.gcf()) + else: + self.plt.savefig(path) + self.plt.gcf().clear() + + def reset(self): + for key in self.__plot_data__: + data = self.__plot_data__[key] + assert isinstance(data, PlotData) + data.reset()