diff --git a/CMakeLists.txt b/CMakeLists.txt index d3379a663db4613e529cdba4ce22111765ff59cc..ed704585d8a6bf3befd9a549aa5a62a33fea3da9 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -62,13 +62,11 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF) option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF) option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF) -option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF) option(WITH_CONTRIB "Compile the third-party contributation" OFF) option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF) option(WITH_ANAKIN "Compile with Anakin library" OFF) option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE}) option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF) -option(WITH_INFERENCE "Compile fluid inference library" ON) option(ON_INFER "Turn on inference optimization." OFF) option(WITH_INFERENCE_API_TEST "Test fluid inference high-level api interface" OFF) option(WITH_SYSTEM_BLAS "Use system blas library" OFF) @@ -305,6 +303,9 @@ if(WITH_DOC) endif() if (ON_INFER) - message(WARNING "On inference mode, will take place some specific optimization.") + message(STATUS "On inference mode, will take place some specific optimization.") add_definitions(-DPADDLE_ON_INFERENCE) +else() + #TODO(luotao), combine this warning with `make inference_lib_dist` command. + message(WARNING "On inference mode, will take place some specific optimization. Turn on the ON_INFER flag when building inference_lib only.") endif() diff --git a/README.md b/README.md index 8ee67f66423df8bce27f70015be8752457cd9784..56d6c10c642787836abb55cb2974bda0b8d22da4 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://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) +[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/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/1.1/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 1.0.1](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0) +### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1) ### 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==1.0.1.post87 +pip install paddlepaddle-gpu==1.1.0.post87 # Linux GPU cuda8cudnn5 -pip install paddlepaddle-gpu==1.0.1.post85 +pip install paddlepaddle-gpu==1.1.0.post85 # For installation on other platform, refer to http://paddlepaddle.org/ ``` @@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==1.0.1.post85 ## Installation -It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) on our website. +It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website. ## 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. +We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and +[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/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/1.0/user_guides/howto/training/cluster_howto.html) +- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html) You can run distributed training jobs on MPI clusters. -- [Python API](http://paddlepaddle.org/documentation/api/zh/1.0/fluid.html) +- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html) Our new API enables much shorter programs. -- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.0/advanced_usage/development/contribute_to_paddle.html) +- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html) We appreciate your contributions! diff --git a/cmake/configure.cmake b/cmake/configure.cmake index e9852f00b1835adec31373f58ac538f9685251e0..7f5771e561f6cc419fc9b3094174645ac432546e 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -50,11 +50,7 @@ if(NOT WITH_PROFILER) endif(NOT WITH_PROFILER) if(NOT CMAKE_CROSSCOMPILING) - if(WITH_AVX AND AVX512F_FOUND) - set(SIMD_FLAG ${AVX512F_FLAG}) - elseif(WITH_AVX AND AVX2_FOUND) - set(SIMD_FLAG ${AVX2_FLAG}) - elseif(WITH_AVX AND AVX_FOUND) + if(WITH_AVX AND AVX_FOUND) set(SIMD_FLAG ${AVX_FLAG}) elseif(SSE3_FOUND) set(SIMD_FLAG ${SSE3_FLAG}) diff --git a/cmake/external/xxhash.cmake b/cmake/external/xxhash.cmake index 4deaab7545c20002fedcad1cca6df54fe9783eb0..c227e09719bd5f0e825f81fb96f78105aa10c79b 100644 --- a/cmake/external/xxhash.cmake +++ b/cmake/external/xxhash.cmake @@ -7,7 +7,11 @@ set(XXHASH_INCLUDE_DIR "${XXHASH_INSTALL_DIR}/include") IF(WITH_STATIC_LIB) SET(BUILD_CMD make lib) ELSE() - SET(BUILD_CMD sed -i "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib) + IF(APPLE) + SET(BUILD_CMD sed -i \"\" "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib) + ELSE(APPLE) + SET(BUILD_CMD sed -i "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib) + ENDIF(APPLE) ENDIF() ExternalProject_Add( diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index 1047b6f998a74e42114b9deab4f0e7ba1af36835..efdb093a7b28e19f3b2a774dd54f2e7f042e9ca7 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -14,9 +14,6 @@ # make package for paddle fluid shared and static library function(copy TARGET) - if (NOT ON_INFER) - message(WARNING "Turn on the ON_INFER flag when building inference_lib only.") - endif() set(options "") set(oneValueArgs "") set(multiValueArgs SRCS DSTS DEPS) diff --git a/cmake/simd.cmake b/cmake/simd.cmake index 3eacf4d86aa0385eddb690d72e85e3384929bb99..566dc75fda019eb66759eb403f60e16f18cffef1 100644 --- a/cmake/simd.cmake +++ b/cmake/simd.cmake @@ -89,7 +89,9 @@ CHECK_CXX_SOURCE_RUNS(" #include int main() { - __m512i a = _mm512_undefined_epi32(); + __m512i a = _mm512_set_epi32 (-1, 2, -3, 4, -1, 2, -3, 4, + 13, -5, 6, -7, 9, 2, -6, 3); + __m512i result = _mm512_abs_epi32 (a); return 0; }" AVX512F_FOUND) diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index 6653244507742b33d9524a7a0e4a5b2b575d358a..6b665a9effba4bef083d007c0c74f2f4c79e647e 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -24,6 +24,7 @@ if(NOT WITH_FLUID_ONLY) endif() add_subdirectory(testing) +set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests CACHE INTERNAL "python tests directory") if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API) add_subdirectory(fluid) endif() diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index 2b8b82e74fc49d454b5331460acbffd0e9404fb5..b6b7af951093e4d721e5d0c99e7bb818c67af749 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -64,11 +64,11 @@ paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', ' 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_pool ArgSpec(args=['input', 'pool_type', 'is_test'], varargs=None, keywords=None, defaults=(False,)) 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.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)) +paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)) 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)) paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.conv2d_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)) @@ -103,7 +103,7 @@ paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 's 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) paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None)) -paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100)) +paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode'], varargs=None, keywords=None, defaults=(False, -100, False)) paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1)) @@ -174,9 +174,13 @@ 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_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', '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.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None)) +paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None)) +paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(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) diff --git a/paddle/fluid/CMakeLists.txt b/paddle/fluid/CMakeLists.txt index 48b36df6499e59fe742766b5f81fd30a9fb8b900..7d48f0057140cf021a21ea7e304b7e38cc8b9ec2 100644 --- a/paddle/fluid/CMakeLists.txt +++ b/paddle/fluid/CMakeLists.txt @@ -9,8 +9,6 @@ add_subdirectory(pybind) add_subdirectory(recordio) endif(NOT WIN32) -if(WITH_INFERENCE) - # NOTE: please add subdirectory inference at last. - add_subdirectory(inference) - add_subdirectory(train) -endif() +# NOTE: please add subdirectory inference at last. +add_subdirectory(inference) +add_subdirectory(train) diff --git a/paddle/fluid/framework/attribute.cc b/paddle/fluid/framework/attribute.cc index 0dcecb62dba971b48c4f11c0ef47494be40eeea0..fabf2abfc803b8838edb48aa01ab8896799c97ac 100644 --- a/paddle/fluid/framework/attribute.cc +++ b/paddle/fluid/framework/attribute.cc @@ -64,6 +64,13 @@ Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) { case proto::AttrType::LONG: { return attr_desc.l(); } + case proto::AttrType::LONGS: { + std::vector val(attr_desc.longs_size()); + for (int i = 0; i < attr_desc.longs_size(); ++i) { + val[i] = attr_desc.longs(i); + } + return val; + } default: PADDLE_THROW("Unsupport attr type %d", attr_desc.type()); } diff --git a/paddle/fluid/framework/attribute.h b/paddle/fluid/framework/attribute.h index 14ca3e96209ed17f12e87fda8506806514698977..d9c76881b7e98d0b7cd29024b98c8f7720398c66 100644 --- a/paddle/fluid/framework/attribute.h +++ b/paddle/fluid/framework/attribute.h @@ -26,6 +26,113 @@ limitations under the License. */ namespace paddle { namespace framework { + +template +struct ExtractAttribute { + explicit ExtractAttribute(const std::string& attr_name) + : attr_name_(attr_name) {} + + T* operator()(Attribute& attr) const { + T* attr_value = nullptr; + try { + attr_value = &boost::get(attr); + } catch (boost::bad_get& bad_get) { + PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s", + attr_name_, paddle::platform::demangle(typeid(T).name()), + paddle::platform::demangle(attr.type().name())); + } + return attr_value; + } + + const std::string& attr_name_; +}; + +// special handle bool +// FIXME(yuyang18): Currently we cast bool into int in python binding. It is +// hard to change the logic there. In another way, we should correct handle +// if the user set `some_flag=1`. +// +// FIX ME anytime if there is a better solution. +template <> +struct ExtractAttribute { + explicit ExtractAttribute(const std::string& attr_name) + : attr_name_(attr_name) {} + + bool* operator()(Attribute& attr) const { + if (attr.type() == typeid(int)) { // NOLINT + int val = boost::get(attr); + attr = static_cast(val); + } else if (attr.type() == typeid(float)) { // NOLINT + float val = boost::get(attr); + attr = static_cast(val); + } + bool* attr_value = nullptr; + try { + attr_value = &boost::get(attr); + } catch (boost::bad_get& bad_get) { + PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s", + attr_name_, paddle::platform::demangle(attr.type().name())); + } + return attr_value; + } + + const std::string& attr_name_; +}; + +template <> +struct ExtractAttribute { + explicit ExtractAttribute(const std::string& attr_name) + : attr_name_(attr_name) {} + + int64_t* operator()(Attribute& attr) const { + if (attr.type() == typeid(int)) { // NOLINT + int val = boost::get(attr); + attr = static_cast(val); + } else if (attr.type() == typeid(float)) { // NOLINT + int val = boost::get(attr); + attr = static_cast(val); + } + int64_t* attr_value = nullptr; + try { + attr_value = &boost::get(attr); + } catch (boost::bad_get& bad_get) { + PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s", + attr_name_, paddle::platform::demangle(attr.type().name())); + } + return attr_value; + } + + const std::string& attr_name_; +}; + +template <> +struct ExtractAttribute> { + explicit ExtractAttribute(const std::string& attr_name) + : attr_name_(attr_name) {} + + std::vector* operator()(Attribute& attr) const { + if (attr.type() == typeid(std::vector)) { // NOLINT + std::vector val = boost::get>(attr); + std::vector vec(val.begin(), val.end()); + attr = vec; + } else if (attr.type() == typeid(std::vector)) { // NOLINT + std::vector val = boost::get>(attr); + std::vector vec(val.begin(), val.end()); + attr = vec; + } + std::vector* attr_value = nullptr; + try { + attr_value = &boost::get>(attr); + } catch (boost::bad_get& bad_get) { + PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s", + attr_name_, paddle::platform::demangle(attr.type().name())); + } + return attr_value; + } + + const std::string& attr_name_; +}; + template inline proto::AttrType AttrTypeID() { Attribute tmp = T(); @@ -42,7 +149,11 @@ class AttrReader { inline const T& Get(const std::string& name) const { PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap", name); - return boost::get(attrs_.at(name)); + + Attribute& attr = const_cast(attrs_.at(name)); + ExtractAttribute extract_attr(name); + T* attr_value = extract_attr(attr); + return *attr_value; } private: @@ -82,7 +193,7 @@ class DefaultValueSetter { public: explicit DefaultValueSetter(T default_value) : default_value_(default_value) {} - void operator()(T& value) const { value = default_value_; } + void operator()(T& value) const { value = default_value_; } // NOLINT private: T default_value_; @@ -117,84 +228,6 @@ class EnumInContainer { std::unordered_set container_; }; -template -struct ExtractAttribute { - explicit ExtractAttribute(const std::string& attr_name) - : attr_name_(attr_name) {} - - T* operator()(Attribute& attr) const { - T* attr_value = nullptr; - try { - attr_value = &boost::get(attr); - } catch (boost::bad_get& bad_get) { - PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s", - attr_name_, paddle::platform::demangle(typeid(T).name()), - paddle::platform::demangle(attr.type().name())); - } - return attr_value; - } - - const std::string& attr_name_; -}; - -// special handle bool -// FIXME(yuyang18): Currently we cast bool into int in python binding. It is -// hard to change the logic there. In another way, we should correct handle -// if the user set `some_flag=1`. -// -// FIX ME anytime if there is a better solution. -template <> -struct ExtractAttribute { - explicit ExtractAttribute(const std::string& attr_name) - : attr_name_(attr_name) {} - - bool* operator()(Attribute& attr) const { - if (attr.type() == typeid(int)) { // NOLINT - int val = boost::get(attr); - attr = static_cast(val); - } else if (attr.type() == typeid(float)) { // NOLINT - float val = boost::get(attr); - attr = static_cast(val); - } - bool* attr_value = nullptr; - try { - attr_value = &boost::get(attr); - } catch (boost::bad_get& bad_get) { - PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s", - attr_name_, paddle::platform::demangle(attr.type().name())); - } - return attr_value; - } - - const std::string& attr_name_; -}; - -template <> -struct ExtractAttribute { - explicit ExtractAttribute(const std::string& attr_name) - : attr_name_(attr_name) {} - - int64_t* operator()(Attribute& attr) const { - if (attr.type() == typeid(int)) { // NOLINT - int val = boost::get(attr); - attr = static_cast(val); - } else if (attr.type() == typeid(float)) { // NOLINT - int val = boost::get(attr); - attr = static_cast(val); - } - int64_t* attr_value = nullptr; - try { - attr_value = &boost::get(attr); - } catch (boost::bad_get& bad_get) { - PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s", - attr_name_, paddle::platform::demangle(attr.type().name())); - } - return attr_value; - } - - const std::string& attr_name_; -}; - // check whether a certain attribute fit its limits // an attribute can have more than one limits template @@ -235,7 +268,7 @@ class TypedAttrChecker { return *this; } - void operator()(AttributeMap& attr_map) const { + void operator()(AttributeMap& attr_map) const { // NOLINT if (!attr_map.count(attr_name_)) { // user do not set this attr PADDLE_ENFORCE(!default_value_setter_.empty(), @@ -271,7 +304,7 @@ class OpAttrChecker { return *(checker.target>()); } - void Check(AttributeMap& attr_map) const { + void Check(AttributeMap& attr_map) const { // NOLINT for (const auto& checker : attr_checkers_) { checker(attr_map); } diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index 17188ac5f301102ae79c6ace676b84ee66e28801..d6b5ad4570c1d8402dedb8596cc75d9eae5a91c7 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -1,5 +1,6 @@ cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto node) cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context lod_tensor) +cc_library(op_graph_view SRCS op_graph_view.cc DEPS op_handle_base) cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry) @@ -30,20 +31,25 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope) -if(WITH_GPU) +cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper) + +if (WITH_GPU) cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass) endif() +cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass) + cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle) -if(WITH_GPU) - cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass) -else() - cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto) +set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass) +if (WITH_GPU) + list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass) endif() +cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS}) + cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context) @@ -56,6 +62,7 @@ cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_execu # device_context reduce_op_handle ) cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context) +cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fused_broadcast_op_handle) cc_library(build_strategy SRCS build_strategy.cc DEPS graph_viz_pass multi_devices_graph_pass diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.cc b/paddle/fluid/framework/details/all_reduce_op_handle.cc index 7c5f5bd80a937bf1a1c891155764833d7b21c5c2..b8690156763e4037811245b8016982710445e6a2 100644 --- a/paddle/fluid/framework/details/all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc @@ -34,7 +34,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, nccl_ctxs_(ctxs) { if (nccl_ctxs_) { for (auto &p : places_) { - this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p); + this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p)); } } } @@ -46,7 +46,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, #endif void AllReduceOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); if (NoDummyInputSize() == 1) { return; // No need to all reduce when GPU count = 1; @@ -127,7 +127,7 @@ void AllReduceOpHandle::RunImpl() { *local_scopes_[i]->FindVar(kLocalExecScopeName)->Get(); auto &p = places_[i]; auto *var = scope.FindVar(out_var_handles[i]->name_); - auto *dev_ctx = dev_ctxes_[p]; + auto *dev_ctx = dev_ctxes_.at(p); RunAndRecordEvent(p, [&trg, var, dev_ctx, p] { auto &tensor_gpu = *var->GetMutable(); diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index 5b5a10e22776bee5c61a55c163c1732692551e36..7f0d06c892541a2697a4ed083f6f4c0fc774a2a4 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -59,6 +59,10 @@ void BroadcastOpHandle::BroadcastOneVar( var_scopes.at(in_var_handle.scope_idx_)->FindVar(in_var_handle.name_); PADDLE_ENFORCE_NOT_NULL(in_var); Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var); + if (UNLIKELY(!in_tensor.IsInitialized())) { + VLOG(3) << "in var " << in_var_handle.name_ << "not inited, return!"; + return; + } InitOutputValue(in_var_handle, out_var_handles); diff --git a/paddle/fluid/framework/details/broadcast_op_handle.h b/paddle/fluid/framework/details/broadcast_op_handle.h index 020d351e891c7afab37c59c0ff8d8e5e7ba184f2..72180fac864256ddda076c57e50ab1083c113d32 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.h +++ b/paddle/fluid/framework/details/broadcast_op_handle.h @@ -44,7 +44,8 @@ struct BroadcastOpHandle : public OpHandleBase { nccl_ctxs_(nccl_ctxs) { if (nccl_ctxs_) { for (auto &p_ctx : nccl_ctxs_->contexts_) { - dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get(); + this->SetDeviceContext(platform::CUDAPlace(p_ctx.first), + p_ctx.second.ctx_.get()); } } } diff --git a/paddle/fluid/framework/details/broadcast_op_handle_test.cc b/paddle/fluid/framework/details/broadcast_op_handle_test.cc index ab7412a19fbd13fa39dbae9af528d158cc9ddbd0..650de5a48de6b1fdab120cdeda563a169fd1a1c1 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle_test.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle_test.cc @@ -12,232 +12,12 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/details/broadcast_op_handle.h" -#include "gtest/gtest.h" - -#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/framework/details/broadcast_op_handle_test.h" namespace paddle { namespace framework { namespace details { -namespace f = paddle::framework; -namespace p = paddle::platform; - -// test data amount -const f::DDim kDims = {20, 20}; - -struct TestBroadcastOpHandle { - std::vector> ctxs_; - std::vector local_scopes_; - std::vector param_scopes_; - Scope g_scope_; - std::unique_ptr op_handle_; - std::vector> vars_; - std::vector gpu_list_; - bool use_gpu_; -#ifdef PADDLE_WITH_CUDA - std::unique_ptr nccl_ctxs_; -#endif - - void WaitAll() { - for (size_t j = 0; j < ctxs_.size(); ++j) { - ctxs_[j]->Wait(); - } -#ifdef PADDLE_WITH_CUDA - if (nccl_ctxs_) { - nccl_ctxs_->WaitAll(); - } -#endif - } - - void InitCtxOnGpu(bool use_gpu) { - use_gpu_ = use_gpu; - if (use_gpu_) { -#ifdef PADDLE_WITH_CUDA - int count = p::GetCUDADeviceCount(); - if (count <= 1) { - LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA " - "device count is " - << count; - exit(0); - } - for (int i = 0; i < count; ++i) { - auto p = p::CUDAPlace(i); - gpu_list_.push_back(p); - ctxs_.emplace_back(new p::CUDADeviceContext(p)); - } - nccl_ctxs_.reset(new platform::NCCLContextMap(gpu_list_)); -#else - PADDLE_THROW("CUDA is not support."); -#endif - } else { - int count = 8; - for (int i = 0; i < count; ++i) { - auto p = p::CPUPlace(); - gpu_list_.push_back(p); - ctxs_.emplace_back(new p::CPUDeviceContext(p)); - } -#ifdef PADDLE_WITH_CUDA - nccl_ctxs_.reset(nullptr); -#endif - } - } - - void InitBroadcastOp(size_t input_scope_idx) { - for (size_t j = 0; j < gpu_list_.size(); ++j) { - local_scopes_.push_back(&(g_scope_.NewScope())); - Scope& local_scope = local_scopes_.back()->NewScope(); - *local_scopes_.back() - ->Var(details::kLocalExecScopeName) - ->GetMutable() = &local_scope; - local_scope.Var("out"); - param_scopes_.emplace_back(&local_scope); - } - param_scopes_[input_scope_idx]->Var("input"); - - std::unique_ptr n = - ir::CreateNodeForTest("node0", ir::Node::Type::kOperation); - if (use_gpu_) { -#ifdef PADDLE_WITH_CUDA - op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_, - nccl_ctxs_.get())); -#else - PADDLE_THROW("CUDA is not support."); -#endif - } else { -#ifdef PADDLE_WITH_CUDA - op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_, - nccl_ctxs_.get())); -#else - op_handle_.reset( - new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_)); -#endif - } - - std::unique_ptr v = - ir::CreateNodeForTest("node1", ir::Node::Type::kVariable); - auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input", - gpu_list_[input_scope_idx]); - vars_.emplace_back(in_var_handle); - op_handle_->AddInput(in_var_handle); - - // add dummy var - - std::unique_ptr v2 = - ir::CreateNodeForTest("node2", ir::Node::Type::kVariable); - vars_.emplace_back(new DummyVarHandle(v2.get())); - DummyVarHandle* dummy_var_handle = - static_cast(vars_.back().get()); - dummy_var_handle->ClearGeneratedOp(); - op_handle_->AddInput(dummy_var_handle); - - for (size_t j = 0; j < gpu_list_.size(); ++j) { - if (!use_gpu_) { - op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); - } - std::unique_ptr v3 = - ir::CreateNodeForTest("node3", ir::Node::Type::kVariable); - VarHandle* out_var_handle = - new VarHandle(v3.get(), 2, j, "out", gpu_list_[j]); - vars_.emplace_back(out_var_handle); - op_handle_->AddOutput(out_var_handle); - } - - // add dummy var - std::unique_ptr v4 = - ir::CreateNodeForTest("node4", ir::Node::Type::kVariable); - vars_.emplace_back(new DummyVarHandle(v4.get())); - DummyVarHandle* out_dummy_var_handle = - static_cast(vars_.back().get()); - out_dummy_var_handle->ClearGeneratedOp(); - op_handle_->AddOutput(out_dummy_var_handle); - } - - void TestBroadcastLodTensor(size_t input_scope_idx) { - auto in_var = param_scopes_[input_scope_idx]->FindVar("input"); - PADDLE_ENFORCE_NOT_NULL(in_var); - auto in_lod_tensor = in_var->GetMutable(); - in_lod_tensor->mutable_data(kDims, gpu_list_[input_scope_idx]); - - std::vector send_vector(static_cast(f::product(kDims))); - for (size_t k = 0; k < send_vector.size(); ++k) { - send_vector[k] = k; - } - f::LoD lod{{0, 10, 20}}; - paddle::framework::TensorFromVector( - send_vector, *(ctxs_[input_scope_idx]), in_lod_tensor); - in_lod_tensor->set_lod(lod); - in_lod_tensor->Resize(kDims); - - op_handle_->Run(false); - - WaitAll(); - - p::CPUPlace cpu_place; - for (size_t j = 0; j < gpu_list_.size(); ++j) { - auto out_var = param_scopes_[j]->FindVar("out"); - PADDLE_ENFORCE_NOT_NULL(out_var); - auto out_tensor = out_var->Get(); - PADDLE_ENFORCE_EQ(out_tensor.lod(), lod, "lod is not equal."); - - f::Tensor result_tensor; - f::TensorCopySync(out_tensor, cpu_place, &result_tensor); - float* ct = result_tensor.mutable_data(cpu_place); - - for (int64_t i = 0; i < f::product(kDims); ++i) { - ASSERT_NEAR(ct[i], send_vector[i], 1e-5); - } - } - } - - void TestBroadcastSelectedRows(size_t input_scope_idx) { - auto in_var = param_scopes_[input_scope_idx]->FindVar("input"); - PADDLE_ENFORCE_NOT_NULL(in_var); - auto in_selected_rows = in_var->GetMutable(); - auto value = in_selected_rows->mutable_value(); - value->mutable_data(kDims, gpu_list_[input_scope_idx]); - int height = static_cast(kDims[0]) * 2; - std::vector rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1, - 2, 4, 6, 3, 1, 1, 1, 1, 3, 7}; - in_selected_rows->set_height(height); - in_selected_rows->set_rows(rows); - - std::vector send_vector(static_cast(f::product(kDims))); - for (size_t k = 0; k < send_vector.size(); ++k) { - send_vector[k] = k; - } - paddle::framework::TensorFromVector( - send_vector, *(ctxs_[input_scope_idx]), value); - - op_handle_->Run(false); - - WaitAll(); - - p::CPUPlace cpu_place; - for (size_t j = 0; j < gpu_list_.size(); ++j) { - auto out_var = param_scopes_[j]->FindVar("out"); - PADDLE_ENFORCE_NOT_NULL(out_var); - auto& out_select_rows = out_var->Get(); - auto rt = out_select_rows.value(); - - PADDLE_ENFORCE_EQ(out_select_rows.height(), height, - "height is not equal."); - for (size_t k = 0; k < out_select_rows.rows().size(); ++k) { - PADDLE_ENFORCE_EQ(out_select_rows.rows()[k], rows[k]); - } - - f::Tensor result_tensor; - f::TensorCopySync(rt, cpu_place, &result_tensor); - float* ct = result_tensor.data(); - - for (int64_t i = 0; i < f::product(kDims); ++i) { - ASSERT_NEAR(ct[i], send_vector[i], 1e-5); - } - } - } -}; - TEST(BroadcastTester, TestCPUBroadcastTestLodTensor) { TestBroadcastOpHandle test_op; size_t input_scope_idx = 0; diff --git a/paddle/fluid/framework/details/broadcast_op_handle_test.h b/paddle/fluid/framework/details/broadcast_op_handle_test.h new file mode 100644 index 0000000000000000000000000000000000000000..1a2a9ac328c4a9b89bfb89106af81b9fb3ed3028 --- /dev/null +++ b/paddle/fluid/framework/details/broadcast_op_handle_test.h @@ -0,0 +1,271 @@ +// 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 "gtest/gtest.h" +#include "paddle/fluid/framework/details/broadcast_op_handle.h" + +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace framework { +namespace details { + +namespace f = paddle::framework; +namespace p = paddle::platform; + +// test data amount +const f::DDim kDims = {20, 20}; + +struct TestBroadcastOpHandle { + std::vector> ctxs_; + std::vector local_scopes_; + std::vector param_scopes_; + Scope g_scope_; + std::unique_ptr op_handle_; + std::vector> vars_; + std::vector place_list_; + bool use_gpu_; +#ifdef PADDLE_WITH_CUDA + std::unique_ptr nccl_ctxs_; +#endif + + void WaitAll() { + for (size_t j = 0; j < ctxs_.size(); ++j) { + ctxs_[j]->Wait(); + } +#ifdef PADDLE_WITH_CUDA + if (nccl_ctxs_) { + nccl_ctxs_->WaitAll(); + } +#endif + } + + void InitCtxOnGpu(bool use_gpu) { + use_gpu_ = use_gpu; + if (use_gpu_) { +#ifdef PADDLE_WITH_CUDA + int count = p::GetCUDADeviceCount(); + if (count <= 1) { + LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA " + "device count is " + << count; + exit(0); + } + for (int i = 0; i < count; ++i) { + auto p = p::CUDAPlace(i); + place_list_.push_back(p); + ctxs_.emplace_back(new p::CUDADeviceContext(p)); + } + nccl_ctxs_.reset(new platform::NCCLContextMap(place_list_)); +#else + PADDLE_THROW("CUDA is not support."); +#endif + } else { + int count = 8; + for (int i = 0; i < count; ++i) { + auto p = p::CPUPlace(); + place_list_.push_back(p); + ctxs_.emplace_back(new p::CPUDeviceContext(p)); + } +#ifdef PADDLE_WITH_CUDA + nccl_ctxs_.reset(nullptr); +#endif + } + } + + void InitBroadcastOp(size_t input_scope_idx) { + for (size_t j = 0; j < place_list_.size(); ++j) { + local_scopes_.push_back(&(g_scope_.NewScope())); + Scope& local_scope = local_scopes_.back()->NewScope(); + *local_scopes_.back() + ->Var(details::kLocalExecScopeName) + ->GetMutable() = &local_scope; + local_scope.Var("out"); + param_scopes_.emplace_back(&local_scope); + } + param_scopes_[input_scope_idx]->Var("input"); + + std::unique_ptr n = + ir::CreateNodeForTest("node0", ir::Node::Type::kOperation); + if (use_gpu_) { +#ifdef PADDLE_WITH_CUDA + op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, + place_list_, nccl_ctxs_.get())); +#else + PADDLE_THROW("CUDA is not support."); +#endif + } else { +#ifdef PADDLE_WITH_CUDA + op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, + place_list_, nccl_ctxs_.get())); +#else + op_handle_.reset( + new BroadcastOpHandle(n.get(), local_scopes_, place_list_)); +#endif + } + + std::unique_ptr v = + ir::CreateNodeForTest("node1", ir::Node::Type::kVariable); + auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input", + place_list_[input_scope_idx]); + vars_.emplace_back(in_var_handle); + op_handle_->AddInput(in_var_handle); + + // add dummy var + + std::unique_ptr v2 = + ir::CreateNodeForTest("node2", ir::Node::Type::kVariable); + vars_.emplace_back(new DummyVarHandle(v2.get())); + DummyVarHandle* dummy_var_handle = + static_cast(vars_.back().get()); + dummy_var_handle->ClearGeneratedOp(); + op_handle_->AddInput(dummy_var_handle); + + for (size_t j = 0; j < place_list_.size(); ++j) { + if (!use_gpu_) { + op_handle_->SetDeviceContext(place_list_[j], ctxs_[j].get()); + } + std::unique_ptr v3 = + ir::CreateNodeForTest("node3", ir::Node::Type::kVariable); + VarHandle* out_var_handle = + new VarHandle(v3.get(), 2, j, "out", place_list_[j]); + vars_.emplace_back(out_var_handle); + op_handle_->AddOutput(out_var_handle); + } + + // add dummy var + std::unique_ptr v4 = + ir::CreateNodeForTest("node4", ir::Node::Type::kVariable); + vars_.emplace_back(new DummyVarHandle(v4.get())); + DummyVarHandle* out_dummy_var_handle = + static_cast(vars_.back().get()); + out_dummy_var_handle->ClearGeneratedOp(); + op_handle_->AddOutput(out_dummy_var_handle); + } + + std::vector InitLoDTensor(const std::string& varname, + size_t input_scope_idx, const f::LoD& lod, + float val_scalar = 0.0) { + auto var = param_scopes_[input_scope_idx]->FindVar(varname); + + PADDLE_ENFORCE_NOT_NULL(var); + auto lod_tensor = var->GetMutable(); + std::vector send_vector(static_cast(f::product(kDims))); + for (size_t k = 0; k < send_vector.size(); ++k) { + send_vector[k] = k + val_scalar; + } + paddle::framework::TensorFromVector( + send_vector, *(ctxs_[input_scope_idx]), lod_tensor); + lod_tensor->set_lod(lod); + lod_tensor->Resize(kDims); + return send_vector; + } + + std::vector InitSelectedRows(const std::string& varname, + size_t input_scope_idx, + const std::vector& rows, + int height, float value_scalar = 0.0) { + std::vector send_vector(static_cast(f::product(kDims))); + for (size_t k = 0; k < send_vector.size(); ++k) { + send_vector[k] = k + value_scalar; + } + + auto var = param_scopes_[input_scope_idx]->FindVar(varname); + PADDLE_ENFORCE_NOT_NULL(var); + auto selected_rows = var->GetMutable(); + auto value = selected_rows->mutable_value(); + value->mutable_data(kDims, place_list_[input_scope_idx]); + selected_rows->set_height(height); + selected_rows->set_rows(rows); + + paddle::framework::TensorFromVector( + send_vector, *(ctxs_[input_scope_idx]), value); + + return send_vector; + } + + void SelectedRowsEqual(const std::string& varname, int input_scope_idx, + const std::vector& send_vector, + const std::vector& rows, int height) { + auto var = param_scopes_[input_scope_idx]->FindVar(varname); + PADDLE_ENFORCE_NOT_NULL(var); + auto& selected_rows = var->Get(); + auto rt = selected_rows.value(); + PADDLE_ENFORCE_EQ(selected_rows.height(), height, "height is not equal."); + + for (size_t k = 0; k < selected_rows.rows().size(); ++k) { + PADDLE_ENFORCE_EQ(selected_rows.rows()[k], rows[k]); + } + + p::CPUPlace cpu_place; + f::Tensor result_tensor; + f::TensorCopySync(rt, cpu_place, &result_tensor); + float* ct = result_tensor.data(); + + for (int64_t i = 0; i < f::product(kDims); ++i) { + ASSERT_NEAR(ct[i], send_vector[i], 1e-5); + } + } + + void LoDTensorEqual(const std::string& varname, + const std::vector& send_vec, const f::LoD& lod, + framework::Scope* scope) { + p::CPUPlace cpu_place; + auto var = scope->FindVar(varname); + PADDLE_ENFORCE_NOT_NULL(var); + auto tensor = var->Get(); + PADDLE_ENFORCE_EQ(tensor.lod(), lod, "lod is not equal."); + f::Tensor result_tensor; + f::TensorCopySync(tensor, cpu_place, &result_tensor); + float* ct = result_tensor.mutable_data(cpu_place); + for (int64_t k = 0; k < f::product(kDims); ++k) { + ASSERT_NEAR(ct[k], send_vec[k], 1e-5); + } + } + + void TestBroadcastLodTensor(size_t input_scope_idx) { + f::LoD lod{{0, 10, 20}}; + auto send_vector = InitLoDTensor("input", input_scope_idx, lod); + + op_handle_->Run(false); + + WaitAll(); + for (size_t j = 0; j < place_list_.size(); ++j) { + LoDTensorEqual("out", send_vector, lod, param_scopes_[j]); + } + } + + void TestBroadcastSelectedRows(size_t input_scope_idx) { + std::vector rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1, + 2, 4, 6, 3, 1, 1, 1, 1, 3, 7}; + int height = static_cast(kDims[0] * 2); + auto send_vector = InitSelectedRows("input", input_scope_idx, rows, height); + + op_handle_->Run(false); + + WaitAll(); + for (size_t j = 0; j < place_list_.size(); ++j) { + SelectedRowsEqual("out", input_scope_idx, send_vector, rows, height); + } + } +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index fefd27fc86fb8dce3311fa580d90f518906dd862..48f94a1f05614d4b797562ac67cdb9828fd0456e 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -16,6 +16,7 @@ limitations under the License. */ #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h" #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h" +#include "paddle/fluid/framework/details/sequential_execution_pass.h" #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph_viz_pass.h" @@ -27,6 +28,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { public: explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy) : ir::PassBuilder(), strategy_(strategy) { + if (strategy_.enable_sequential_execution_) { + AppendPass("sequential_execution_pass"); + } + // Add a graph viz pass to record a graph. if (!strategy_.debug_graphviz_path_.empty()) { auto viz_pass = AppendPass("graph_viz_pass"); @@ -64,6 +69,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { // Verify that the graph is correct for multi-device executor. AppendPass("multi_devices_check_pass"); + + if (strategy_.remove_unnecessary_lock_) { + AppendPass("modify_op_lock_and_record_event_pass"); + } } private: @@ -110,6 +119,11 @@ std::unique_ptr BuildStrategy::Apply( pass->Erase("nccl_ctxs"); pass->SetNotOwned("nccl_ctxs", nctx); #endif + } else if (pass->Type() == "sequential_execution_pass") { + pass->Erase(kAllOpDescs); + pass->Set>( + kAllOpDescs, + new std::vector(main_program.Block(0).AllOps())); } graph = pass->Apply(std::move(graph)); } @@ -125,3 +139,5 @@ USE_PASS(multi_batch_merge_pass); USE_PASS(multi_devices_pass); USE_PASS(multi_devices_check_pass); USE_PASS(multi_devices_print_pass); +USE_PASS(sequential_execution_pass); +USE_PASS(modify_op_lock_and_record_event_pass); diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h index f3ffaf6ecd7c4dd99c40fe58ba88c0cbdc14bde7..6c7b54db8f610aa34cd51dcbc13063290cae3ac0 100644 --- a/paddle/fluid/framework/details/build_strategy.h +++ b/paddle/fluid/framework/details/build_strategy.h @@ -69,8 +69,12 @@ struct BuildStrategy { bool enable_data_balance_{false}; + bool enable_sequential_execution_{false}; + bool fuse_broadcast_op_{false}; + bool remove_unnecessary_lock_{false}; + // User normally doesn't need to call this API. // The PassBuilder allows for more customized insert, remove of passes // from python side. diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc index b6282debdb4eb6b1f29c39e54ac4f3e2296838da..7ad1e40c600c6e70cea822fac777ff20163078e6 100644 --- a/paddle/fluid/framework/details/computation_op_handle.cc +++ b/paddle/fluid/framework/details/computation_op_handle.cc @@ -29,15 +29,21 @@ ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope, void ComputationOpHandle::RunImpl() { WaitInputVarGenerated(place_); - this->RunAndRecordEvent([this] { + auto run_func = [this]() { op_->Run(*scope_->FindVar(kLocalExecScopeName)->Get(), place_); - }); + }; + + if (is_lock_and_record_event_free_) { + run_func(); + } else { + this->RunAndRecordEvent(run_func); + } } bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) { bool need_wait = in_var && in_var->GeneratedOp() && - in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_[place_]; + in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_); return need_wait; } diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h index e98f1ab148db083ac63a1afd43e334fbfae62539..662a91d6b4dfcfed563fdf2e46c22f83f90b40af 100644 --- a/paddle/fluid/framework/details/computation_op_handle.h +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -36,6 +36,8 @@ struct ComputationOpHandle : public OpHandleBase { const platform::Place &GetPlace() const { return place_; } + void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; } + protected: void RunImpl() override; @@ -45,6 +47,7 @@ struct ComputationOpHandle : public OpHandleBase { std::unique_ptr op_; Scope *scope_; platform::Place place_; + bool is_lock_and_record_event_free_{false}; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/data_balance_op_handle.cc b/paddle/fluid/framework/details/data_balance_op_handle.cc index 525d24322442ef4dd6e8c24212af61c908959b87..0b772f9b63e2cfb78175f5e0d7011db8e6a5ec20 100644 --- a/paddle/fluid/framework/details/data_balance_op_handle.cc +++ b/paddle/fluid/framework/details/data_balance_op_handle.cc @@ -28,7 +28,7 @@ DataBalanceOpHandle::DataBalanceOpHandle( : OpHandleBase(node), local_scopes_(local_scopes), places_(places) { if (ctxs) { for (auto &p : places_) { - this->dev_ctxes_[p] = ctxs->DevCtx(p); + this->SetDeviceContext(p, ctxs->DevCtx(p)); } } } @@ -89,8 +89,8 @@ void DataBalanceOpHandle::RunImpl() { PADDLE_ENFORCE_GT(places_.size(), 1, "Data balance can only be enabled when the number of " "places to run larger than 1."); - auto in_var_handles = DynamicCast(inputs_); - auto out_var_handles = DynamicCast(outputs_); + auto in_var_handles = DynamicCast(this->Inputs()); + auto out_var_handles = DynamicCast(this->Outputs()); PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0); PADDLE_ENFORCE_EQ( in_var_handles.size(), out_var_handles.size(), diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc index 6e22fedf1c39428528c00cce4c9a4460dfb95cb3..98fc390e72fab3701538fd6f974460fa5114fdb0 100644 --- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc @@ -92,13 +92,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( size_t num_complete = 0; remaining_ = 0; - BlockingQueue complete_q; + auto complete_q = std::make_shared>(); for (auto op : bootstrap_ops_) { - RunOpAsync(op_deps.get(), op, &complete_q); + RunOpAsync(op_deps.get(), op, complete_q); } while (num_complete != op_deps->size()) { - size_t num_comp = complete_q.Pop(); + size_t num_comp = complete_q->Pop(); if (num_comp == -1UL) { int remaining = 0; while (true) { @@ -107,7 +107,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( break; } for (int i = 0; i < remaining; ++i) { - complete_q.Pop(); + complete_q->Pop(); } } exception_.ReThrow(); @@ -120,7 +120,8 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( } void FastThreadedSSAGraphExecutor::RunOpAsync( std::unordered_map> *op_deps, - OpHandleBase *op, BlockingQueue *complete_q) { + OpHandleBase *op, + const std::shared_ptr> &complete_q) { ++remaining_; this->pool_.enqueue([=] { OpHandleBase *op_to_run = op; @@ -144,7 +145,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync( if (op_to_run == nullptr) { op_to_run = pending_op; } else { - this->RunOpAsync(op_deps, pending_op, complete_q); + RunOpAsync(op_deps, pending_op, complete_q); } } } @@ -156,8 +157,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync( } void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() { atomic_op_deps_ = pool_.enqueue([&] { - std::unordered_map> *op_deps = - new std::unordered_map>; + auto *op_deps = new std::unordered_map>; for (auto &pair : op_deps_) { (*op_deps)[pair.first] = pair.second; } diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h index dad3a231cba6402f57ba654a9ac5fb520b9c8f04..8b8382447105c8caa36963214684d6ee9fa15200 100644 --- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h @@ -50,7 +50,8 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor { std::atomic remaining_; void RunOpAsync(std::unordered_map> *op_deps, - OpHandleBase *op, BlockingQueue *complete_q); + OpHandleBase *op, + const std::shared_ptr> &complete_q); void PrepareAtomicOpDeps(); diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc b/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..0f12bd2b4e857648342aeb5ad33b6c0fe01c9c73 --- /dev/null +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc @@ -0,0 +1,165 @@ +// 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/details/fused_broadcast_op_handle.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/details/broadcast_op_handle_test.h" + +namespace paddle { +namespace framework { +namespace details { + +struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { + std::vector out_varnames_; + + void InitFusedBroadcastOp(std::vector input_scope_idxes) { + // initialize scope and var + for (size_t i = 0; i < place_list_.size(); ++i) { + local_scopes_.push_back(&(g_scope_.NewScope())); + Scope& local_scope = local_scopes_.back()->NewScope(); + *local_scopes_.back() + ->Var(details::kLocalExecScopeName) + ->GetMutable() = &local_scope; + for (size_t j = 0; j < input_scope_idxes.size(); ++j) { + local_scope.Var("out_var" + j); + if (i == j) local_scope.Var("in_var" + j); + } + param_scopes_.emplace_back(&local_scope); + } + + // create op handle node + std::unique_ptr n = + ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation); + if (use_gpu_) { +#ifdef PADDLE_WITH_CUDA + op_handle_.reset(new FusedBroadcastOpHandle( + n.get(), local_scopes_, place_list_, nccl_ctxs_.get())); +#else + PADDLE_THROW("CUDA is not supported."); +#endif + } else { +#ifdef PADDLE_WITH_CUDA + op_handle_.reset(new FusedBroadcastOpHandle( + n.get(), local_scopes_, place_list_, nccl_ctxs_.get())); +#else + op_handle_.reset( + new FusedBroadcastOpHandle(n.get(), local_scopes_, place_list_)); +#endif + } + + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + // add input var handle + std::unique_ptr in_node = + ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable); + VarHandle* in_var_handle = + new VarHandle(in_node.get(), 1, input_scope_idxes[i], "in_var" + i, + place_list_[input_scope_idxes[i]]); + vars_.emplace_back(in_var_handle); + op_handle_->AddInput(in_var_handle); + + // add output var handle + for (size_t j = 0; j < place_list_.size(); ++j) { + std::unique_ptr out_node = + ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable); + VarHandle* out_var_handle = + new VarHandle(out_node.get(), 2, j, "out_var" + i, place_list_[j]); + vars_.emplace_back(out_var_handle); + op_handle_->AddOutput(out_var_handle); + } + } + } + + void TestFusedBroadcastLoDTensor(std::vector input_scope_idxes) { + std::vector> send_vec; + f::LoD lod{{0, 10, 20}}; + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + const std::string varname("in_var" + i); + float val_scalar = static_cast(i); + send_vec.push_back( + InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar)); + } + + op_handle_->Run(false); + + WaitAll(); + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + const std::string& varname("out_var" + i); + for (size_t j = 0; j < place_list_.size(); ++j) { + LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]); + } + } + } + + void TestFusedBroadcastSelectedRows(std::vector input_scope_idxes) { + std::vector> send_vector; + std::vector rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1, + 2, 4, 6, 3, 1, 1, 1, 1, 3, 7}; + int height = static_cast(kDims[0] * 2); + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + const std::string varname("in_var" + i); + float val_scalar = static_cast(i); + send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i], + rows, height, val_scalar)); + } + + op_handle_->Run(false); + + WaitAll(); + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + const std::string& varname("out_var" + i); + for (size_t j = 0; j < place_list_.size(); ++j) { + SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows, + height); + } + } + } +}; + +TEST(FusedBroadcastTester, CPULodTensor) { + TestFusedBroadcastOpHandle test_op; + std::vector input_scope_idxes = {0, 1}; + test_op.InitCtxOnGpu(false); + test_op.InitFusedBroadcastOp(input_scope_idxes); + test_op.TestFusedBroadcastLoDTensor(input_scope_idxes); +} + +TEST(FusedBroadcastTester, CPUSelectedRows) { + TestFusedBroadcastOpHandle test_op; + std::vector input_scope_idxes = {0, 1}; + test_op.InitCtxOnGpu(false); + test_op.InitFusedBroadcastOp(input_scope_idxes); + test_op.TestFusedBroadcastSelectedRows(input_scope_idxes); +} + +#ifdef PADDLE_WITH_CUDA +TEST(FusedBroadcastTester, GPULodTensor) { + TestFusedBroadcastOpHandle test_op; + std::vector input_scope_idxes = {0, 1}; + test_op.InitCtxOnGpu(true); + test_op.InitFusedBroadcastOp(input_scope_idxes); + test_op.TestFusedBroadcastLoDTensor(input_scope_idxes); +} + +TEST(FusedBroadcastTester, GPUSelectedRows) { + TestFusedBroadcastOpHandle test_op; + std::vector input_scope_idxes = {0, 1}; + test_op.InitCtxOnGpu(true); + test_op.InitFusedBroadcastOp(input_scope_idxes); + test_op.TestFusedBroadcastSelectedRows(input_scope_idxes); +} +#endif + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/gather_op_handle.cc b/paddle/fluid/framework/details/gather_op_handle.cc index 9aae19fc73de4387186da47c55710c94d53f1b88..ca4633c5a8f22fc9f7319b06aa766f9fe37dc68c 100644 --- a/paddle/fluid/framework/details/gather_op_handle.cc +++ b/paddle/fluid/framework/details/gather_op_handle.cc @@ -36,7 +36,7 @@ void GatherOpHandle::RunImpl() { VarHandle *out_var_handle; { - auto out_var_handles = DynamicCast(outputs_); + auto out_var_handles = DynamicCast(this->Outputs()); PADDLE_ENFORCE_EQ(out_var_handles.size(), 1, "The number of output should be one."); out_var_handle = out_var_handles.front(); @@ -99,7 +99,7 @@ void GatherOpHandle::RunImpl() { Tensor *out_tensor = out_value->mutable_value(); // copy - auto dev_ctx = dev_ctxes_[out_var_handle->place_]; + auto dev_ctx = dev_ctxes_.at(out_var_handle->place_); RunAndRecordEvent(out_var_handle->place_, [in_tensors, out_tensor, &dev_ctx, t_out_p] { int s = 0, e = 0; diff --git a/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..169ce3ae7ca497e40d99b1c16633e35e1e4f1009 --- /dev/null +++ b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc @@ -0,0 +1,59 @@ +// 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/details/modify_op_lock_and_record_event_pass.h" +#include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/details/op_graph_view.h" + +namespace paddle { +namespace framework { +namespace details { + +static bool IsLockAndRecordEventFreeComputationOpHandle( + ComputationOpHandle *op, const OpGraphView &graph_view) { + if (!platform::is_gpu_place(op->GetPlace())) return false; + for (auto &pending_op : graph_view.PendingOps(op)) { + auto *tmp = dynamic_cast(pending_op); + if (tmp == nullptr || !(tmp->GetPlace() == op->GetPlace())) { + return false; + } + } + return true; +} + +std::unique_ptr ModifyOpLockAndRecordEventPass::ApplyImpl( + std::unique_ptr ir_graph) const { + auto &all_ops = ir_graph->Get(kGraphOps); + OpGraphView graph_view(all_ops); + for (auto &op : all_ops) { + auto *compute_op = dynamic_cast(op.get()); + if (compute_op == nullptr) continue; + bool is_lock_and_record_event_free = + IsLockAndRecordEventFreeComputationOpHandle(compute_op, graph_view); + compute_op->SetLockAndRecordEventFree(is_lock_and_record_event_free); + if (is_lock_and_record_event_free) { + VLOG(10) << "Set is_lock_and_record_event_free be true in op " + << compute_op->DebugString(); + } + } + return ir_graph; +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(modify_op_lock_and_record_event_pass, + paddle::framework::details::ModifyOpLockAndRecordEventPass); diff --git a/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..b54e1b318be95e1e0abf6830f8c918895df02718 --- /dev/null +++ b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h @@ -0,0 +1,32 @@ +// 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/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +class ModifyOpLockAndRecordEventPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc index f2d5b182e577714d6138e99932af637a711cc9f2..f3819887a196a7c8bf35897467bb9d68b428094e 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc @@ -722,7 +722,8 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result, } if (node->Op()->Type() == "split_byref" || - node->Op()->Type() == "split_selected_rows") { + node->Op()->Type() == "split_selected_rows" || + node->Op()->Type() == "split_ids") { // TODO(paddle-dev): getting the first var is not safe. op_dev_id = GetVarDeviceID(*result, input_var_names[0]); if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) { diff --git a/paddle/fluid/framework/details/op_graph_view.cc b/paddle/fluid/framework/details/op_graph_view.cc new file mode 100644 index 0000000000000000000000000000000000000000..65dafd376f7c687410270e35f105ff595fe78f59 --- /dev/null +++ b/paddle/fluid/framework/details/op_graph_view.cc @@ -0,0 +1,77 @@ +// 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/details/op_graph_view.h" +#include +#include + +namespace paddle { +namespace framework { +namespace details { + +OpGraphView::OpGraphView( + const std::vector> &ops) { + Build(ops); +} + +void OpGraphView::Build(const std::vector> &ops) { + for (auto &op : ops) { + preceding_ops_[op.get()]; + pending_ops_[op.get()]; + for (auto &var : op->Outputs()) { + for (auto &pending_op : var->PendingOps()) { + preceding_ops_[pending_op].insert(op.get()); + pending_ops_[op.get()].insert(pending_op); + } + } + } + PADDLE_ENFORCE( + preceding_ops_.size() == ops.size() && pending_ops_.size() == ops.size(), + "There are duplicate ops in graph."); +} + +size_t OpGraphView::OpNumber() const { return preceding_ops_.size(); } + +std::unordered_set OpGraphView::AllOps() const { + std::unordered_set ret; + for (auto &pair : preceding_ops_) { + ret.insert(pair.first); + } + return ret; +} + +bool OpGraphView::HasOp(OpHandleBase *op) const { + return preceding_ops_.count(op) != 0; +} + +void OpGraphView::EnforceHasOp(OpHandleBase *op) const { + PADDLE_ENFORCE(HasOp(op), "Cannot find op %s in OpGraphView", + op == nullptr ? "nullptr" : op->DebugString()); +} + +const std::unordered_set &OpGraphView::PrecedingOps( + OpHandleBase *op) const { + EnforceHasOp(op); + return preceding_ops_.at(op); +} + +const std::unordered_set &OpGraphView::PendingOps( + OpHandleBase *op) const { + EnforceHasOp(op); + return pending_ops_.at(op); +} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/op_graph_view.h b/paddle/fluid/framework/details/op_graph_view.h new file mode 100644 index 0000000000000000000000000000000000000000..398c019be00a6ff5f5b39fdcbe97339341b1685b --- /dev/null +++ b/paddle/fluid/framework/details/op_graph_view.h @@ -0,0 +1,54 @@ +// 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 +#include +#include "paddle/fluid/framework/details/op_handle_base.h" + +namespace paddle { +namespace framework { +namespace details { + +class OpGraphView { + public: + explicit OpGraphView(const std::vector> &ops); + + size_t OpNumber() const; + + std::unordered_set AllOps() const; + + const std::unordered_set &PrecedingOps( + OpHandleBase *op) const; + + const std::unordered_set &PendingOps(OpHandleBase *op) const; + + bool HasOp(OpHandleBase *op) const; + + private: + void Build(const std::vector> &ops); + void EnforceHasOp(OpHandleBase *op) const; + + std::unordered_map> + preceding_ops_; + std::unordered_map> + pending_ops_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc index 3812f0abf1b7069525c4420054c61c01c908acfe..4822627ac3b65972f41d9a23d9fe3dba3de3f97d 100644 --- a/paddle/fluid/framework/details/op_handle_base.cc +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -103,7 +103,7 @@ void OpHandleBase::WaitInputVarGenerated() { void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) { for (auto *in : inputs_) { if (NeedWait(in)) { - in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[place]); + in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(place)); } } } diff --git a/paddle/fluid/framework/details/reduce_op_handle.cc b/paddle/fluid/framework/details/reduce_op_handle.cc index 7fc06f234d42a992328c0b6164f17945d8075c28..4503123eac810917cabcf1e62cff98552ed2f742 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.cc +++ b/paddle/fluid/framework/details/reduce_op_handle.cc @@ -27,7 +27,7 @@ namespace framework { namespace details { void ReduceOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); if (places_.size() == 1) return; // the input and output may have dummy var. diff --git a/paddle/fluid/framework/details/reduce_op_handle.h b/paddle/fluid/framework/details/reduce_op_handle.h index a6289b055f97b7b0e57928358d84117b33cf2df8..999828ae457ba43541da06088ce7c25331fd05ec 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.h +++ b/paddle/fluid/framework/details/reduce_op_handle.h @@ -46,7 +46,8 @@ struct ReduceOpHandle : public OpHandleBase { nccl_ctxs_(nccl_ctxs) { if (nccl_ctxs_) { for (auto &p_ctx : nccl_ctxs_->contexts_) { - dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get(); + this->SetDeviceContext(platform::CUDAPlace(p_ctx.first), + p_ctx.second.ctx_.get()); } } } diff --git a/paddle/fluid/framework/details/reference_count_op_handle.h b/paddle/fluid/framework/details/reference_count_op_handle.h index fc479a4c4a1e7d5c824d3c202e0cccf743dd52c9..cc4ccfbdfc720284e683a8f3f59a4aa57a3a9eb1 100644 --- a/paddle/fluid/framework/details/reference_count_op_handle.h +++ b/paddle/fluid/framework/details/reference_count_op_handle.h @@ -51,7 +51,7 @@ class ReferenceCountOpHandle : public OpHandleBase { dev_ctx_ = static_cast( platform::DeviceContextPool::Instance().Get(place)); if (IsStreamGarabageCollector()) { - PADDLE_ENFORCE(cudaSetDevice(place.device)); + platform::SetDeviceId(place.device); PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming)); } @@ -61,7 +61,7 @@ class ReferenceCountOpHandle : public OpHandleBase { ~ReferenceCountOpHandle() { if (IsStreamGarabageCollector()) { auto gpu_place = boost::get(dev_ctx_->GetPlace()); - PADDLE_ENFORCE(cudaSetDevice(gpu_place.device)); + platform::SetDeviceId(gpu_place.device); PADDLE_ENFORCE(cudaEventDestroy(event_)); } } diff --git a/paddle/fluid/framework/details/reference_count_pass.cc b/paddle/fluid/framework/details/reference_count_pass.cc index 2d1f688d64ece3322e253b0c070264b9eb73d678..0b994ced7f751f056fec076e3dea8d14d0bed991 100644 --- a/paddle/fluid/framework/details/reference_count_pass.cc +++ b/paddle/fluid/framework/details/reference_count_pass.cc @@ -43,6 +43,23 @@ static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) { return nullptr; } +static void AddDependencyBetween(OpHandleBase *in, OpHandleBase *out, + ir::Graph *graph) { + auto it = std::find_if( + in->Outputs().begin(), in->Outputs().end(), [](VarHandleBase *var) { + return dynamic_cast(var) != nullptr; + }); + + if (it != in->Outputs().end()) { + out->AddInput(*it); + } else { + auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); + graph->Get(kGraphDepVars).emplace(dep_var); + in->AddOutput(dep_var); + out->AddInput(dep_var); + } +} + std::unique_ptr ReferenceCountPass::ApplyImpl( std::unique_ptr graph) const { auto &ref_cnts = Get(kGlobalReferenceCount); @@ -133,12 +150,7 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( auto *ref_cnt_handle = new ReferenceCountOpHandle( ref_cnt_node, next_compute_op->GetScope(), place, {var_name}, gcs[place.device].get(), cur_ref_cnts[place.device].get()); - if (next_compute_op->Outputs().empty()) { - auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); - next_compute_op->AddOutput(dep_var); - graph->Get(kGraphDepVars).emplace(dep_var); - } - ref_cnt_handle->AddInput(next_compute_op->Outputs().front()); + AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get()); compute_ref_cnt_map[next_compute_op].reset(ref_cnt_handle); } } @@ -160,12 +172,7 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( auto *ref_cnt_handle = new ReferenceCountOpHandle( ref_cnt_node, compute_op->GetScope(), place, in_var_names, gcs[place.device].get(), cur_ref_cnts[place.device].get()); - if (compute_op->Outputs().empty()) { - auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); - compute_op->AddOutput(dep_var); - graph->Get(kGraphDepVars).emplace(dep_var); - } - ref_cnt_handle->AddInput(compute_op->Outputs().front()); + AddDependencyBetween(compute_op, ref_cnt_handle, graph.get()); compute_ref_cnt_map[compute_op].reset(ref_cnt_handle); } diff --git a/paddle/fluid/framework/details/rpc_op_handle.cc b/paddle/fluid/framework/details/rpc_op_handle.cc index f44b374edb29228dff5a8bf003d945291f166d49..dfa6c1ade1a024bb9087144d0e96fa5b0417f06a 100644 --- a/paddle/fluid/framework/details/rpc_op_handle.cc +++ b/paddle/fluid/framework/details/rpc_op_handle.cc @@ -29,22 +29,19 @@ RPCOpHandle::RPCOpHandle(ir::Node *node, const framework::OpDesc &op_desc, place_(place) {} void RPCOpHandle::RunImpl() { - // TODO(wuyi): need further analysis whether wait VarDummyHandle. - // Wait input done for (auto *in : inputs_) { auto &p = static_cast(in)->place_; - // FIXME(Yancey1989): need a better solution instead of use DebugString() - if (ir::IsControlDepVar(*in->Node())) { // HACK + if (ir::IsControlDepVar(*in->Node())) { continue; } if (in->GeneratedOp()) { - in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[p]); + in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(p)); } } - auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); - // FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead - // lock. - op_->Run(*tmp_scope, place_); + this->RunAndRecordEvent([this] { + op_->Run(*local_scope_->FindVar(kLocalExecScopeName)->Get(), + place_); + }); } std::string RPCOpHandle::Name() const { return name_; } diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc index ba243979b34aa1f683de707525403becaf0a1c00..ef1626599795a553e654fe5d3ed74ef3a3a67d78 100644 --- a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc @@ -27,7 +27,7 @@ ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev, coeff_(static_cast(1.0 / num_dev)), scope_(scope), place_(place) { - dev_ctxes_[place_] = dev_ctx; + this->SetDeviceContext(place_, dev_ctx); } ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {} @@ -46,9 +46,9 @@ void ScaleLossGradOpHandle::RunImpl() { } else { #ifdef PADDLE_WITH_CUDA this->RunAndRecordEvent([&] { - auto stream = - static_cast(this->dev_ctxes_[place_]) - ->stream(); + auto stream = static_cast( + this->dev_ctxes_.at(place_)) + ->stream(); memory::Copy(boost::get(place_), tmp, platform::CPUPlace(), &coeff_, sizeof(float), stream); VLOG(10) << place_ << "RUN Scale loss grad op"; diff --git a/paddle/fluid/framework/details/sequential_execution_pass.cc b/paddle/fluid/framework/details/sequential_execution_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..cc2c8bfef9f9f54c2e499467df0d22ce3f69d6b8 --- /dev/null +++ b/paddle/fluid/framework/details/sequential_execution_pass.cc @@ -0,0 +1,109 @@ +// 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/details/sequential_execution_pass.h" +#include +#include +#include +#include +#include "paddle/fluid/framework/op_proto_maker.h" + +namespace paddle { +namespace framework { +namespace details { + +static bool IsSameOpDesc(OpDesc *op1, OpDesc *op2) { + return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && + op1->Outputs() == op2->Outputs(); +} + +std::unique_ptr SequentialExecutionPass::ApplyImpl( + std::unique_ptr graph) const { + // FIXME(zjl): Insert dependencies between some distributed ops may cause + // the multi_devices_graph_pass fails. So we skip these ops here. + // Indeed, maybe we should not insert dependencies between these ops + // casually, which may cause deadlock easily. + // We should add more skipped distributed ops when found errors in + // multi_devices_graph_pass + static std::unordered_set skip_dist_ops{ + "send", "recv", "send_barrier", "fetch_barrier"}; + + auto &ops = Get>(kAllOpDescs); + std::vector op_node_list; + op_node_list.reserve(ops.size()); + + std::unordered_map op_deps; + std::unordered_map> pending_ops; + std::unordered_set ready_ops; + + for (ir::Node *node : graph->Nodes()) { + if (!node->IsOp()) continue; + std::unordered_set preceding_ops; + for (auto *in : node->inputs) { + PADDLE_ENFORCE(in->IsVar(), + "Preceding Node of Op Nodes must be Var Node"); + if (in->inputs.empty()) continue; + PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp(), + "Preceding Op Node of Var Node must be unique"); + preceding_ops.insert(in->inputs[0]); + pending_ops[in->inputs[0]].insert(node); + } + op_deps[node] = preceding_ops.size(); + if (preceding_ops.empty()) { + ready_ops.insert(node); + } + } + + for (auto *op_desc : ops) { + ir::Node *found_node = nullptr; + for (auto *node : ready_ops) { + if (IsSameOpDesc(op_desc, node->Op())) { + PADDLE_ENFORCE(found_node == nullptr, + "Found multiple op_desc in graph: %s", op_desc->Type()); + found_node = node; + } + } + + PADDLE_ENFORCE_NOT_NULL(found_node, "Cannot find op_desc in graph: %s", + op_desc->Type()); + for (auto *pending_op : pending_ops[found_node]) { + if (--op_deps.at(pending_op) == 0) { + ready_ops.insert(pending_op); + } + } + ready_ops.erase(found_node); + if (skip_dist_ops.count(op_desc->Type()) == 0) { + op_node_list.push_back(found_node); + } + } + + for (size_t i = 1; i < op_node_list.size(); ++i) { + auto *dep_var = graph->CreateControlDepVar(); + op_node_list[i]->inputs.push_back(dep_var); + op_node_list[i - 1]->outputs.push_back(dep_var); + dep_var->outputs.push_back(op_node_list[i]); + dep_var->inputs.push_back(op_node_list[i - 1]); + VLOG(10) << "Add dependencies between " << op_node_list[i - 1]->Name() + << " and " << op_node_list[i]->Name(); + } + return graph; +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(sequential_execution_pass, + paddle::framework::details::SequentialExecutionPass) + .RequirePassAttr(paddle::framework::details::kAllOpDescs); diff --git a/paddle/fluid/framework/details/sequential_execution_pass.h b/paddle/fluid/framework/details/sequential_execution_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..a04c08bc2eb3bae797d648b30a22a5fee7ba0eaa --- /dev/null +++ b/paddle/fluid/framework/details/sequential_execution_pass.h @@ -0,0 +1,34 @@ +// 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/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +constexpr char kAllOpDescs[] = "all_op_descs"; + +class SequentialExecutionPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 31beef3ae829d72570ee7c879dac71ed600cd216..dc63effd1b7c8fe5bb3fc91058eb855e552d3926 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -39,7 +39,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr)); std::unordered_map pending_ops; std::unordered_set pending_vars; - BlockingQueue ready_vars; + auto ready_vars = std::make_shared>(); std::unordered_set ready_ops; // For ops (e.g. nccl_all_reduce) that need to coordinate multiple // streams from multiple GPUs, it's faster to buffer them and schedule @@ -51,12 +51,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( for (auto &var_map : graph_->Get(details::kGraphVars)) { for (auto &name_pair : var_map) { for (auto &version_pair : name_pair.second) { - InsertPendingVar(&pending_vars, &ready_vars, version_pair.get()); + InsertPendingVar(&pending_vars, ready_vars.get(), version_pair.get()); } } } for (auto &var : graph_->Get(details::kGraphDepVars)) { - InsertPendingVar(&pending_vars, &ready_vars, var.get()); + InsertPendingVar(&pending_vars, ready_vars.get(), var.get()); } for (auto &op : graph_->Get(details::kGraphOps)) { @@ -73,12 +73,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( FeedFetchList fetch_data(fetch_tensors.size()); InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops, - &pending_vars, &ready_vars, &fetch_data); + &pending_vars, ready_vars.get(), &fetch_data); auto run_all_ops = [&](std::unordered_set &set) { for (auto *op : set) { running_ops_++; - RunOp(&ready_vars, op); + RunOp(ready_vars, op); } set.clear(); }; @@ -87,7 +87,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( run_op_futures_.clear(); exception_holder_.Clear(); event.reset(nullptr); - // Step 3. Execution while (!pending_vars.empty()) { // 1. Run All Ready ops @@ -103,7 +102,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // 2. Find ready variable bool timeout; - auto cur_ready_vars = ready_vars.PopAll(1, &timeout); + auto cur_ready_vars = ready_vars->PopAll(1, &timeout); if (timeout) { if (exception_holder_.IsCaught()) { @@ -133,7 +132,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( } } PADDLE_ENFORCE(ready_ops.empty()); - // Wait FetchOps. ClearFetchOp(graph_.get(), &fetch_ops); @@ -206,7 +204,8 @@ void ThreadedSSAGraphExecutor::InsertPendingVar( } void ThreadedSSAGraphExecutor::RunOp( - BlockingQueue *ready_var_q, details::OpHandleBase *op) { + const std::shared_ptr> &ready_var_q, + details::OpHandleBase *op) { auto op_run = [ready_var_q, op, this] { try { if (VLOG_IS_ON(10)) { diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index 512f8a4ca5a9b82a395dde11722b8db44ea5ec27..dbb0b498d995a897b109bd4ef98521b2193276ed 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -51,7 +51,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { ~ThreadedSSAGraphExecutor() {} private: - void RunOp(BlockingQueue *ready_var_q, + void RunOp(const std::shared_ptr> &ready_var_q, details::OpHandleBase *op); private: diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 9095119efce49a5a789f70fb1e677a8d02e5b5fb..af47f0e5ab9d700b6b4c05f716a2b997ef469e6d 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -85,8 +85,10 @@ Executor::Executor(const platform::Place& place) : place_(place) {} void Executor::Close() { #ifdef PADDLE_WITH_DISTRIBUTE + // TODO(typhoonzero): complete message will need to use real trainer_id, + // except 0. ::paddle::operators::distributed::RPCClient::GetInstance< - ::paddle::operators::distributed::GRPCClient>() + ::paddle::operators::distributed::GRPCClient>(0) ->SendComplete(); #endif } diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto index c99406799ba5f664c4b9f80e0567b293e4ffea51..efdabffb9b33ddf007c13008d0f3afb7a3961eda 100644 --- a/paddle/fluid/framework/framework.proto +++ b/paddle/fluid/framework/framework.proto @@ -35,6 +35,7 @@ enum AttrType { BLOCK = 8; LONG = 9; BLOCKS = 10; + LONGS = 11; } // OpDesc describes an instance of a C++ framework::OperatorBase @@ -55,6 +56,7 @@ message OpDesc { optional int32 block_idx = 12; optional int64 l = 13; repeated int32 blocks_idx = 14; + repeated int64 longs = 15; }; message Var { diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index ce006b7a3fbc16f3c9149933390969b14a46b484..28231a53bad50fe9f19cfe3e73c3dc09aa3762cf 100644 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -41,6 +41,7 @@ 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(depthwise_conv_mkldnn_pass base) pass_library(conv_bias_mkldnn_fuse_pass inference) pass_library(conv_relu_mkldnn_fuse_pass inference) pass_library(conv_elementwise_add_mkldnn_fuse_pass inference) @@ -59,6 +60,7 @@ cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector) cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto) if (WITH_MKLDNN) + cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass) 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/conv_relu_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h index b5de0d548713772e7ad41cfb6d8b3e9460683efb..fe585bd7c41bb32ae00462e989ab4c0051fc89a8 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h +++ b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h @@ -31,7 +31,8 @@ class ConvReLUFusePass : public FusePassBase { virtual ~ConvReLUFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir 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 8f4bab25ed4919881baf19a961a52aa229e06a8f..19248b4dfee1da81d18cd2effac08ba68dde80fb 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 @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h" #include +#include "paddle/fluid/framework/op_proto_maker.h" namespace paddle { namespace framework { @@ -36,6 +37,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name, op->SetInput("X", inputs); } op->SetOutput("Out", outputs); + op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(OpRole::kForward)); } // a->OP0->b diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..19056e18aa892dbc83dfbf7305b6ad8b6b6bc51c --- /dev/null +++ b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc @@ -0,0 +1,58 @@ +/* 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/depthwise_conv_mkldnn_pass.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +#define GET_NODE(id, pattern) \ + PADDLE_ENFORCE(subgraph.count(pattern.RetrieveNode(#id)), \ + "pattern has no Node called %s", #id); \ + auto* id = subgraph.at(pattern.RetrieveNode(#id)); \ + PADDLE_ENFORCE_NOT_NULL(id, "subgraph has no node %s", #id); + +std::unique_ptr DepthwiseConvMKLDNNPass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init("depthwise_conv_mkldnn_pass", graph.get()); + GraphPatternDetector gpd; + + auto* pattern = gpd.mutable_pattern(); + pattern->NewNode("depthwise_conv") + ->assert_is_op("depthwise_conv2d") + ->assert_op_attr("use_mkldnn", true); + + int found_depthwise_conv_mkldnn_count = 0; + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(3) << "handle DepthwiseConvMKLDNN fuse"; + GET_NODE(depthwise_conv, (*pattern)); + depthwise_conv->Op()->SetType("conv2d"); + found_depthwise_conv_mkldnn_count++; + }; + + gpd(graph.get(), handler); + AddStatis(found_depthwise_conv_mkldnn_count); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(depthwise_conv_mkldnn_pass, + paddle::framework::ir::DepthwiseConvMKLDNNPass); diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..8ca6a7325186401c26eb7f9375cf83b7b97cc1c9 --- /dev/null +++ b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h @@ -0,0 +1,34 @@ +/* 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/fuse_pass_base.h" + +namespace paddle { +namespace framework { +namespace ir { + +class DepthwiseConvMKLDNNPass : public FusePassBase { + public: + virtual ~DepthwiseConvMKLDNNPass() {} + + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..09d0b15f46a7e50afb6aea46383013ce6a6c6118 --- /dev/null +++ b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc @@ -0,0 +1,123 @@ +// 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/depthwise_conv_mkldnn_pass.h" + +#include + +namespace paddle { +namespace framework { +namespace ir { + +void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name, + const std::vector& inputs, + const std::vector& outputs, bool use_mkldnn = false) { + auto* op = prog->MutableBlock(0)->AppendOp(); + op->SetType(type); + 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]}); + op->SetOutput("Out", outputs); +} + +// (a, weights, bias)->depthwise conv mkldnn->b +// (b, weights2, bias2)->depthwise conv no mkldnn->c +// (c, weights3, bias3)->conv mkldnn->d +// (d, weights3, bias3)->conv no mkldnn->e +ProgramDesc BuildProgramDesc() { + ProgramDesc prog; + for (auto& v : std::vector( + {"a", "b", "c", "d", "e", "weights", "bias", "weights2", "bias2", + "weights3", "bias3", "weights4", "bias4"})) { + auto* var = prog.MutableBlock(0)->Var(v); + var->SetType(proto::VarType::SELECTED_ROWS); + if (v == "weights" || v == "bias" || v == "weights2" || v == "bias2" || + v == "weights3" || v == "bias3" || v == "weights4" || v == "bias4") { + var->SetPersistable(true); + } + } + + // depthwise conv with MKL-DNN + SetOp(&prog, "depthwise_conv2d", "conv1", + std::vector({"a", "weights", "bias"}), + std::vector({"b"}), true); + // depthwise conv without MKL-DNN + SetOp(&prog, "depthwise_conv2d", "conv2", + std::vector({"b", "weights2", "bias2"}), + std::vector({"c"}), false); + // conv with MKL-DNN + SetOp(&prog, "conv2d", "conv3", + std::vector({"c", "weights3", "bias3"}), + std::vector({"d"}), true); + // conv without MKL-dNN + SetOp(&prog, "conv2d", "conv4", + std::vector({"d", "weights4", "bias4"}), + std::vector({"e"}), false); + + return prog; +} + +TEST(DepthwiseConvMKLDNNPass, basic) { + auto prog = BuildProgramDesc(); + + std::unique_ptr graph(new ir::Graph(prog)); + + auto pass = PassRegistry::Instance().Get("depthwise_conv_mkldnn_pass"); + + struct counters { + int mkldnn_depthwise_conv_nodes; + int other_depthwise_conv_nodes; + int mkldnn_conv_nodes; + int other_conv_nodes; + }; + + counters before{1, 1, 1, 1}; + + graph = pass->Apply(std::move(graph)); + + // initialize counters before loop + counters after{0, 0, 0, 0}; + + for (auto* node : graph->Nodes()) { + if (node->IsOp()) { + auto* op = node->Op(); + if (op->Type() == "conv2d") { + if (boost::get(op->GetAttr("use_mkldnn"))) + after.mkldnn_conv_nodes++; + else + after.other_conv_nodes++; + } else if (op->Type() == "depthwise_conv2d") { + if (boost::get(op->GetAttr("use_mkldnn"))) + after.mkldnn_depthwise_conv_nodes++; + else + after.other_depthwise_conv_nodes++; + } + } + } + + EXPECT_EQ(after.other_depthwise_conv_nodes, + before.other_depthwise_conv_nodes); + EXPECT_EQ(after.other_conv_nodes, before.other_conv_nodes); + EXPECT_EQ(after.mkldnn_depthwise_conv_nodes, + before.mkldnn_depthwise_conv_nodes - 1); + EXPECT_EQ(after.mkldnn_conv_nodes, before.mkldnn_conv_nodes + 1); +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +USE_PASS(depthwise_conv_mkldnn_pass); diff --git a/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc b/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc index 06286a109d01af638e74e06ccc83e2a5500663ea..2db7d95cae1c8c59691fd642e2462e92ed58814f 100644 --- a/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc +++ b/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/ir/fc_fuse_pass.h" #include +#include "paddle/fluid/framework/op_proto_maker.h" namespace paddle { namespace framework { @@ -32,6 +33,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, op->SetInput("X", inputs); } op->SetOutput("Out", outputs); + op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(OpRole::kForward)); } // a->OP0->b diff --git a/paddle/fluid/framework/ir/graph.cc b/paddle/fluid/framework/ir/graph.cc index 265a128e95e6205159de67d87d5cb8ca1ad7d475..132159b8b272f311060a39b58919c26822bf50ee 100644 --- a/paddle/fluid/framework/ir/graph.cc +++ b/paddle/fluid/framework/ir/graph.cc @@ -23,8 +23,67 @@ limitations under the License. */ namespace paddle { namespace framework { namespace ir { +namespace { + +void CheckProgram(const ProgramDesc &program) { +#define _INT(role) static_cast(role) + + std::map visit; + for (OpDesc *op : program.Block(0).AllOps()) { + // For backward compatibility, some program doesn't have role added. + if (!op->HasAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) continue; + int role_id = + boost::get(op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())); + visit[role_id] = true; + switch (role_id) { + case _INT(OpRole::kForward): + if (visit.find(_INT(OpRole::kBackward)) != visit.end()) { + LOG(ERROR) + << "Cannot add backward operator before forward operator %s." + << op->Type(); + } + break; + case _INT(OpRole::kBackward): + case _INT(OpRole::kBackward) | _INT(OpRole::kLoss): + PADDLE_ENFORCE( + visit.find(_INT(OpRole::kOptimize)) == visit.end(), + "Cannot add backward operator %s after optimize operator.", + op->Type()); + break; + case _INT(OpRole::kForward) | _INT(OpRole::kLoss): + PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward) | + _INT(OpRole::kLoss)) == visit.end(), + "Cannot add backward|loss operator before " + "forward|loss operator %s.", + op->Type()); + PADDLE_ENFORCE( + visit.find(_INT(OpRole::kOptimize)) == visit.end(), + "Cannot add forward|loss operator %s after optimize operator.", + op->Type()); + break; + case _INT(OpRole::kOptimize): + case _INT(OpRole::kOptimize) | _INT(OpRole::kLRSched): + PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward)) != visit.end(), + "Optimize operators %s must follow backward operator.", + op->Type()); + break; + case _INT(OpRole::kLRSched): + case _INT(OpRole::kDist): + case _INT(OpRole::kRPC): + case _INT(OpRole::kNotSpecified): + break; + default: + LOG(FATAL) << "Unknown operator role. Don't add new role because " + "you don't know what you are doing."; + } + } + +#undef _INT +} +} // namespace Graph::Graph(const ProgramDesc &program) : program_(program) { + CheckProgram(program_); // Make the nodes id start from 0. Node::ResetId(); auto var_nodes = InitFromProgram(program_); diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index 29b604afbfcfc2bac67e447db8cd4c671c036dbe..b20d70132256bd5df7411c46ff4eb246b1f14ba8 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -259,6 +259,15 @@ GraphPatternDetector::DetectPatterns() { return result; } +bool GraphItemCMP(const std::pair &a, + const std::pair &b) { + if (a.first != b.first) { + return a.first < b.first; + } else { + return a.second < b.second; + } +} + // TODO(Superjomn) enhance the function as it marks unique unique as duplicates // see https://github.com/PaddlePaddle/Paddle/issues/13550 void GraphPatternDetector::UniquePatterns( @@ -267,12 +276,16 @@ void GraphPatternDetector::UniquePatterns( std::vector result; std::unordered_set set; + std::hash hasher; for (auto &g : *subgraphs) { - size_t key = 0; - for (auto &item : g) { - key ^= std::hash{}(item.first); - key ^= std::hash{}(item.second); + // Sort the items in the sub-graph, and transform to a string key. + std::vector> sorted_keys(g.begin(), g.end()); + std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP); + std::stringstream ss; + for (auto &item : sorted_keys) { + ss << item.first << ":" << item.second; } + auto key = hasher(ss.str()); if (!set.count(key)) { result.emplace_back(g); set.insert(key); diff --git a/paddle/fluid/framework/lod_tensor.cc b/paddle/fluid/framework/lod_tensor.cc index 1e7da9a69c7cbf8c13306656599a759515802b76..669d08c70c9b7453264806b346a6c9eb211cfd4a 100644 --- a/paddle/fluid/framework/lod_tensor.cc +++ b/paddle/fluid/framework/lod_tensor.cc @@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor( PADDLE_ENFORCE_EQ(new_lod.size(), lod.size()); for (size_t j = 0; j < lod.size(); ++j) { auto &sub_lod = new_lod[j]; - auto &offset = sub_lod.back(); + size_t offset = sub_lod.back(); for (size_t k = 1; k < lod[j].size(); ++k) { sub_lod.push_back(lod[j][k] + offset); } diff --git a/paddle/fluid/framework/lod_tensor_array.h b/paddle/fluid/framework/lod_tensor_array.h index 0ad6a709008406257d6c0a220bce38bb24e188cd..36a5c3c5d601390beedaf37ceb98ee2c63ecf5a6 100644 --- a/paddle/fluid/framework/lod_tensor_array.h +++ b/paddle/fluid/framework/lod_tensor_array.h @@ -19,81 +19,7 @@ limitations under the License. */ namespace paddle { namespace framework { -// NOTE The vector can't be replaced with the class LoDTensorArray -// directly, because there are many vector used accross the project, -// and some of them are treated as LoDTensorArray. -#if !defined(PADDLE_ON_INFERENCE) - using LoDTensorArray = std::vector; -#else // !PADDLE_ON_INFERENCE - -#pragma message "LoDTensorArray is replaced with the inference one." -/* - * A LoDTensorArray which will not deallocate buffer when resized, fix the data - * diff in inference, and more performance friendly in the concurrency - * scenerios. - */ -class LoDTensorArray { - public: - LoDTensorArray() = default; - - using iterator = std::vector::iterator; - using const_iterator = std::vector::const_iterator; - - const_iterator begin() const { return array_.begin(); } - const_iterator end() const { return array_.begin() + size_; } - iterator begin() { return array_.begin(); } - iterator end() { return array_.begin() + size_; } - - void push_back(const LoDTensor& x) { - if (size_ < array_.size()) { - array_[size_++] = x; - } else { - array_.push_back(x); - ++size_; - } - } - void resize(size_t size) { - if (array_.size() < size) { - array_.resize(size); - } - size_ = size; - } - - void emplace_back() { array_.emplace_back(); } - - void emplace_back(LoDTensor&& x) { array_.emplace_back(std::move(x)); } - - LoDTensor& back() { return array_.back(); } - - size_t space() const { return array_.size(); } - - void reserve(size_t size) { - // Naive warning to tell user this array might be to large. The memory and - // buffer used by this TensorArray will not be deleted during the training - // and inference phase, so attention not to make it expand too long. - if (size > 800UL) { - LOG(WARNING) << "TensorArray has more than 800 items"; - } - array_.reserve(size); - } - - bool empty() const { return size_ == 0UL; } - void clear() { size_ = 0UL; } - - LoDTensor& operator[](size_t id) { return array_[id]; } - const LoDTensor& operator[](size_t id) const { return array_[id]; } - LoDTensor& at(size_t id) { return array_.at(id); } - const LoDTensor& at(size_t id) const { return array_.at(id); } - - size_t size() const { return size_; } - - private: - size_t size_{0}; - std::vector array_; -}; -#endif // !PADDLE_ON_INFERENCE - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index c293cf92b4f3d530109c76850df184af9cad7399..8ece618f3f72552fedcffab3e03ebb30476b7cab 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -419,8 +419,15 @@ struct SetAttrDescVisitor : public boost::static_visitor { } VectorToRepeated(blocks_idx, attr_->mutable_blocks_idx()); } + void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); } + void operator()(int64_t v) const { attr_->set_l(v); } + + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_longs()); + } + void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); } }; diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h index 678c14a44b5c259ddc6f914b3fbb5b50e649993c..4c59c73d8779eceb267ad532aabccabbd54b0df2 100644 --- a/paddle/fluid/framework/op_proto_maker.h +++ b/paddle/fluid/framework/op_proto_maker.h @@ -33,7 +33,7 @@ enum class OpRole { // used for distributed training. kDist = 0x0008, // Tag all learning rate scheduler operators. - kLRSched = 0x0016, + kLRSched = 0x0010, kLoss = 0x0100, // The default value of op's role. This should be only used for unittests and diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index 14fcde2fe3b1c3acfc0994e9cd37a784c57826d7..45fc36c70633204dbfadbd10757c08b009d2cc74 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -354,18 +354,18 @@ void OperatorBase::GenerateTemporaryNames() { } } -static bool VarIsTensor(const Variable* var) { - return var->IsType() || var->IsType(); +static bool VarIsTensor(const Variable& var) { + return var.IsType() || var.IsType(); } -static const Tensor* GetTensorFromVar(Variable* var) { - if (var->IsType()) { - return var->GetMutable(); - } else if (var->IsType()) { - return var->GetMutable()->mutable_value(); +const Tensor* GetTensorFromVar(const Variable& var) { + if (var.IsType()) { + return static_cast(&(var.Get())); + } else if (var.IsType()) { + return &(var.Get().value()); } else { PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.", - var->Type().name()); + var.Type().name()); } } @@ -415,8 +415,7 @@ bool ExecutionContext::HasOutput(const std::string& name) const { template <> const Tensor* ExecutionContext::Input(const std::string& name) const { auto* var = InputVar(name); - return var == nullptr ? nullptr - : GetTensorFromVar(const_cast(var)); + return var == nullptr ? nullptr : GetTensorFromVar(*var); } template <> @@ -428,7 +427,7 @@ const std::vector ExecutionContext::MultiInput( std::transform(names.begin(), names.end(), std::back_inserter(res), [&](const std::string& sub_name) { auto var = scope_.FindVar(sub_name); - return var == nullptr ? nullptr : GetTensorFromVar(var); + return var == nullptr ? nullptr : GetTensorFromVar(*var); }); return res; } @@ -770,8 +769,10 @@ void OperatorWithKernel::TransferInplaceVarsBack( for (auto& var_name : inplace_vars) { VLOG(3) << "share inplace var " + var_name + " back to it's original scope"; auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name)); - auto* transformed_tensor = - GetTensorFromVar(transfer_scope.FindVar(var_name)); + auto* var = transfer_scope.FindVar(var_name); + PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr", + var_name); + auto* transformed_tensor = GetTensorFromVar(*var); original_tensor->ShareDataWith(*transformed_tensor); } } @@ -784,11 +785,11 @@ Scope* OperatorWithKernel::TryTransferData( for (auto& var_name : var_name_item.second) { auto* var = scope.FindVar(var_name); // Only tensor can be tranfer to another device. - if (var == nullptr || !VarIsTensor(var)) { + if (var == nullptr || !VarIsTensor(*var)) { continue; } - auto* tensor_in = GetTensorFromVar(var); + auto* tensor_in = GetTensorFromVar(*var); if (!tensor_in->IsInitialized()) { continue; } diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 626b50edfd39424473be33e9f8baec5970471477..96ad3205235b921a7cf60ed674a8350f74d18509 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -63,6 +63,7 @@ inline std::string GradVarName(const std::string& var_name) { } proto::VarType::Type GetDataTypeOfVar(const Variable* var); +const Tensor* GetTensorFromVar(const Variable& var); class OperatorBase; class ExecutionContext; diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index cffb96bedf7638ee52856f21662437085035eca6..a45b9ec7a20ac3629d182f009b735d4d82fb5dc2 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -187,6 +187,10 @@ void ParallelExecutor::BCastParamsToDevices( } auto &main_tensor = main_var->Get(); + if (!main_tensor.IsInitialized()) { + VLOG(3) << "one in var not inited, return!"; + continue; + } auto &dims = main_tensor.dims(); if (paddle::platform::is_gpu_place(main_tensor.place())) { #ifdef PADDLE_WITH_CUDA @@ -299,10 +303,8 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes( } ParallelExecutor::~ParallelExecutor() { - const auto dev_ctxs = - platform::DeviceContextPool::Instance().GetAllDeviceContexts(); - for (auto &dev_ctx : dev_ctxs) { - dev_ctx->Wait(); + for (auto &p : member_->places_) { + platform::DeviceContextPool::Instance().Get(p)->Wait(); } if (member_->own_local_scope_) { diff --git a/paddle/fluid/framework/tensor_test.cc b/paddle/fluid/framework/tensor_test.cc index cb2061c06a429d8e8116001a4aa4e8c46ea13428..a0a9a573603ceb6b577529101cb331adbc81337a 100644 --- a/paddle/fluid/framework/tensor_test.cc +++ b/paddle/fluid/framework/tensor_test.cc @@ -75,6 +75,19 @@ TEST(Tensor, MutableData) { platform::CPUPlace()); EXPECT_EQ(p1, p2); } + // Not sure if it's desired, but currently, Tensor type can be changed. + { + framework::Tensor src_tensor; + int8_t* p1 = src_tensor.mutable_data(framework::make_ddim({1}), + platform::CPUPlace()); + EXPECT_NE(p1, nullptr); + *p1 = 1; + + uint8_t* p2 = src_tensor.mutable_data(framework::make_ddim({1}), + platform::CPUPlace()); + EXPECT_NE(p2, nullptr); + EXPECT_EQ(static_cast(p2[0]), 1); + } #ifdef PADDLE_WITH_CUDA { diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index 057482b76dbe76acdb72f9e9e7c28e9892b9d588..c54e897251b8bf403c04587ae0dd2c2e69e1e659 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -154,6 +154,12 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, 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); + } else if (platform::is_cuda_pinned_place(src_place) && + platform::is_gpu_place(dst_place)) { + auto src_pinned_place = boost::get(src_place); + auto dst_gpu_place = boost::get(dst_place); + memory::Copy(dst_gpu_place, dst_ptr, src_pinned_place, src_ptr, size, + nullptr); } #endif } diff --git a/paddle/fluid/framework/threadpool.cc b/paddle/fluid/framework/threadpool.cc index a588cb417aebe94bd4aeda02b1bc8ba07a04b960..fcec955360f1c681a62929e904d5736854a8ffad 100644 --- a/paddle/fluid/framework/threadpool.cc +++ b/paddle/fluid/framework/threadpool.cc @@ -57,10 +57,10 @@ ThreadPool::ThreadPool(int num_threads) : running_(true) { ThreadPool::~ThreadPool() { { // notify all threads to stop running - std::lock_guard l(mutex_); + std::unique_lock l(mutex_); running_ = false; - scheduled_.notify_all(); } + scheduled_.notify_all(); for (auto& t : threads_) { t->join(); @@ -70,19 +70,25 @@ ThreadPool::~ThreadPool() { void ThreadPool::TaskLoop() { while (true) { - std::unique_lock lock(mutex_); + Task task; - scheduled_.wait( - lock, [this] { return !this->tasks_.empty() || !this->running_; }); + { + std::unique_lock lock(mutex_); + scheduled_.wait( + lock, [this] { return !this->tasks_.empty() || !this->running_; }); - if (!running_ || tasks_.empty()) { - return; - } + if (!running_ && tasks_.empty()) { + return; + } + + if (tasks_.empty()) { + PADDLE_THROW("This thread has no task to Run"); + } - // pop a task from the task queue - auto task = std::move(tasks_.front()); - tasks_.pop(); - lock.unlock(); + // pop a task from the task queue + task = std::move(tasks_.front()); + tasks_.pop(); + } // run the task task(); diff --git a/paddle/fluid/framework/threadpool.h b/paddle/fluid/framework/threadpool.h index 0687e628aaa4fb7b2e67938fa09a319c8bb35aff..7a51d18fbbf65f68725aa86a6a0ce4d15dff5673 100644 --- a/paddle/fluid/framework/threadpool.h +++ b/paddle/fluid/framework/threadpool.h @@ -58,7 +58,7 @@ class ThreadPool { ~ThreadPool(); // Run pushes a function to the task queue and returns a std::future - // object. To wait for the completion of the task, call + // object. To wait for the completion of the task, call // std::future::wait(). template std::future Run(Callback fn) { @@ -69,7 +69,6 @@ class ThreadPool { template std::future> RunAndGetException( Callback fn) { - std::unique_lock lock(mutex_); Task task([fn]() -> std::unique_ptr { try { fn(); @@ -84,7 +83,13 @@ class ThreadPool { return nullptr; }); std::future> f = task.get_future(); - tasks_.push(std::move(task)); + { + std::unique_lock lock(mutex_); + if (!running_) { + PADDLE_THROW("enqueue on stopped ThreadPool"); + } + tasks_.push(std::move(task)); + } scheduled_.notify_one(); return f; } diff --git a/paddle/fluid/framework/type_defs.h b/paddle/fluid/framework/type_defs.h index e099e40f121ff13657e563eb608feecbca0551be..2de6233a9e0d320ec9a06d547db3575eb61925c0 100644 --- a/paddle/fluid/framework/type_defs.h +++ b/paddle/fluid/framework/type_defs.h @@ -36,7 +36,7 @@ using Attribute = boost::variant, std::vector, std::vector, bool, std::vector, BlockDesc*, int64_t, - std::vector>; + std::vector, std::vector>; using AttributeMap = std::unordered_map; diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index dbbe8bcba69a1d87e21c8eae18834fb708e8b1e4..e5678cf607a8ff3763e79c1f321a81c021846fb1 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -1,3 +1,6 @@ +if(WITH_TESTING) + include(tests/test.cmake) # some generic cmake funtion for inference +endif() # analysis and tensorrt must be added before creating static library, # otherwise, there would be undefined reference to them in static library. add_subdirectory(analysis) diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt index d4d2fd4634f9e11f3f002e11e177c332ced49885..0354f9e6e9588af601210b8a71ae98c1f90d62f0 100644 --- a/paddle/fluid/inference/analysis/CMakeLists.txt +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -20,22 +20,17 @@ cc_test(test_node SRCS node_tester.cc DEPS analysis) cc_test(test_dot SRCS dot_tester.cc DEPS analysis) cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid) -function (inference_analysis_test TARGET) - if(WITH_TESTING) - set(options "") - set(oneValueArgs "") - set(multiValueArgs SRCS ARGS EXTRA_DEPS) - cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - set(mem_opt "") - if(WITH_GPU) - set(mem_opt "--fraction_of_gpu_memory_to_use=0.5") - endif() - cc_test(${TARGET} - SRCS "${analysis_test_SRCS}" - DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS} - ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt} ${analysis_test_ARGS}) - set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec) - endif(WITH_TESTING) +function(inference_analysis_test TARGET) + if(WITH_TESTING) + set(options "") + set(oneValueArgs "") + set(multiValueArgs SRCS ARGS EXTRA_DEPS) + cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + inference_base_test(${TARGET} + SRCS ${analysis_test_SRCS} + DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS} + ARGS --inference_model_dir=${WORD2VEC_MODEL_DIR} ${analysis_test_ARGS}) + endif() endfunction(inference_analysis_test) inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api) diff --git a/paddle/fluid/inference/analysis/analyzer.h b/paddle/fluid/inference/analysis/analyzer.h index 7114f5222c5904bb4422b9e67ad035b85bbb770c..3af1d572dfd81197797dd7e57d87ba12c2f3548e 100644 --- a/paddle/fluid/inference/analysis/analyzer.h +++ b/paddle/fluid/inference/analysis/analyzer.h @@ -79,6 +79,7 @@ class Analyzer : public OrderedRegistry { "conv_bn_fuse_pass", // "conv_eltwiseadd_bn_fuse_pass", // #ifdef PADDLE_WITH_MKLDNN + "depthwise_conv_mkldnn_pass", // "conv_bias_mkldnn_fuse_pass", // "conv_relu_mkldnn_fuse_pass", // "conv_elementwise_add_mkldnn_fuse_pass", // diff --git a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc index 1682011c3d8cc9927a4b026b370671798cace625..50ce20621fb289023ecccf7bb39d98169765d5ee 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/framework/op_proto_maker.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/inference/analysis/ut_helper.h" @@ -130,6 +131,8 @@ void SetOp(framework::ProgramDesc* prog, const std::string& type, op->SetType(type); op->SetInput("Xs", inputs); op->SetOutput("Xs", outputs); + op->SetAttr(framework::OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(framework::OpRole::kForward)); } TEST(DataFlowGraph, Build_IR_Graph) { diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt index e2027b7cb4d584ffcc48624d2c01e65a61829975..49a9ebe3ddec1e4fd59ae1155a706859e249d25c 100644 --- a/paddle/fluid/inference/api/CMakeLists.txt +++ b/paddle/fluid/inference/api/CMakeLists.txt @@ -17,39 +17,12 @@ if(APPLE) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") endif(APPLE) - -set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB} - ) +set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB}) if(WITH_GPU AND TENSORRT_FOUND) set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor) endif() -function(inference_api_test TARGET_NAME) - if (WITH_TESTING) - set(options "") - set(oneValueArgs SRC) - set(multiValueArgs ARGS) - cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - - if (WITH_GPU) - cc_test(${TARGET_NAME} - SRCS ${inference_test_SRC} - DEPS "${inference_deps}" - ARGS --dirname=${PYTHON_TESTS_DIR}/book/ --fraction_of_gpu_memory_to_use=0.15) - else() - cc_test(${TARGET_NAME} - SRCS ${inference_test_SRC} - DEPS "${inference_deps}" - ARGS --dirname=${PYTHON_TESTS_DIR}/book/) - endif() - if(inference_test_ARGS) - set_tests_properties(${TARGET_NAME} - PROPERTIES DEPENDS "${inference_test_ARGS}") - endif() - endif(WITH_TESTING) -endfunction(inference_api_test) - cc_library(reset_tensor_array SRCS details/reset_tensor_array.cc DEPS lod_tensor scope) cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS reset_tensor_array lod_tensor scope) cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor) @@ -59,10 +32,11 @@ cc_test(test_paddle_inference_api SRCS api_tester.cc DEPS paddle_inference_api) -inference_api_test(test_api_impl SRC api_impl_tester.cc - ARGS test_word2vec test_image_classification) - -set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) +if(WITH_TESTING) + inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps} + ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book) + set_tests_properties(test_api_impl PROPERTIES DEPENDS test_image_classification) +endif() cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} paddle_inference_api ARGS --dirname=${PYTHON_TESTS_DIR}/book) @@ -70,8 +44,10 @@ if(WITH_GPU AND TENSORRT_FOUND) cc_library(paddle_inference_tensorrt_subgraph_engine SRCS api_tensorrt_subgraph_engine.cc DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api tensorrt_converter zero_copy_tensor_dummy) - -inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec) + if(WITH_TESTING) + inference_base_test(test_api_tensorrt_subgraph_engine SRCS api_tensorrt_subgraph_engine_tester.cc DEPS ${inference_deps} + ARGS --dirname=${WORD2VEC_MODEL_DIR}) + endif() endif() if (WITH_ANAKIN AND WITH_MKL) # only needed in CI diff --git a/paddle/fluid/inference/api/api_impl_tester.cc b/paddle/fluid/inference/api/api_impl_tester.cc index b7b8ee6ea08fe907f3f052ae1118f782ac853ca7..5152b8670ddb206f0927c03149684af4a096df42 100644 --- a/paddle/fluid/inference/api/api_impl_tester.cc +++ b/paddle/fluid/inference/api/api_impl_tester.cc @@ -27,7 +27,9 @@ limitations under the License. */ #define ACC_DIFF 1e-3 #endif -DEFINE_string(dirname, "", "Directory of the inference model."); +DEFINE_string(word2vec_dirname, "", + "Directory of the word2vec inference model."); +DEFINE_string(book_dirname, "", "Directory of the book inference model."); namespace paddle { @@ -49,7 +51,7 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { NativeConfig GetConfig() { NativeConfig config; - config.model_dir = FLAGS_dirname + "/word2vec.inference.model"; + config.model_dir = FLAGS_word2vec_dirname; LOG(INFO) << "dirname " << config.model_dir; config.fraction_of_gpu_memory = 0.15; #ifdef PADDLE_WITH_CUDA @@ -116,7 +118,7 @@ void MainImageClassification(bool use_gpu) { NativeConfig config = GetConfig(); config.use_gpu = use_gpu; config.model_dir = - FLAGS_dirname + "/image_classification_resnet.inference.model"; + FLAGS_book_dirname + "/image_classification_resnet.inference.model"; const bool is_combined = false; std::vector> feed_target_shapes = @@ -187,7 +189,7 @@ void MainThreadsWord2Vec(bool use_gpu) { std::vector threads; for (int tid = 0; tid < num_jobs; ++tid) { threads.emplace_back([&, tid]() { - auto predictor = main_predictor->Clone(); + auto predictor = CreatePaddlePredictor(config); auto& local_inputs = paddle_tensor_feeds[tid]; std::vector local_outputs; ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs)); @@ -220,7 +222,7 @@ void MainThreadsImageClassification(bool use_gpu) { NativeConfig config = GetConfig(); config.use_gpu = use_gpu; config.model_dir = - FLAGS_dirname + "/image_classification_resnet.inference.model"; + FLAGS_book_dirname + "/image_classification_resnet.inference.model"; auto main_predictor = CreatePaddlePredictor(config); std::vector jobs(num_jobs); @@ -245,7 +247,7 @@ void MainThreadsImageClassification(bool use_gpu) { std::vector threads; for (int tid = 0; tid < num_jobs; ++tid) { threads.emplace_back([&, tid]() { - auto predictor = main_predictor->Clone(); + auto predictor = CreatePaddlePredictor(config); auto& local_inputs = paddle_tensor_feeds[tid]; std::vector local_outputs; ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs)); @@ -271,7 +273,7 @@ TEST(inference_api_native, word2vec_cpu_threads) { MainThreadsWord2Vec(false /*use_gpu*/); } TEST(inference_api_native, image_classification_cpu) { - MainThreadsImageClassification(false /*use_gpu*/); + MainImageClassification(false /*use_gpu*/); } TEST(inference_api_native, image_classification_cpu_threads) { MainThreadsImageClassification(false /*use_gpu*/); @@ -279,15 +281,17 @@ TEST(inference_api_native, image_classification_cpu_threads) { #ifdef PADDLE_WITH_CUDA TEST(inference_api_native, word2vec_gpu) { MainWord2Vec(true /*use_gpu*/); } -TEST(inference_api_native, word2vec_gpu_threads) { - MainThreadsWord2Vec(true /*use_gpu*/); -} +// Turn off temporarily for the unstable result. +// TEST(inference_api_native, word2vec_gpu_threads) { +// MainThreadsWord2Vec(true /*use_gpu*/); +// } TEST(inference_api_native, image_classification_gpu) { - MainThreadsImageClassification(true /*use_gpu*/); -} -TEST(inference_api_native, image_classification_gpu_threads) { - MainThreadsImageClassification(true /*use_gpu*/); + MainImageClassification(true /*use_gpu*/); } +// Turn off temporarily for the unstable result. +// TEST(inference_api_native, image_classification_gpu_threads) { +// MainThreadsImageClassification(true /*use_gpu*/); +// } #endif 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 702158ea3bcab854eece3ccd40724d92efcbae67..89c9a65cb06ba565f0e0cbdb9b6031c6adbcb64e 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc @@ -29,13 +29,13 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) { //# 1. Create PaddlePredictor with a config. NativeConfig config0; - config0.model_dir = FLAGS_dirname + "word2vec.inference.model"; + config0.model_dir = FLAGS_dirname; config0.use_gpu = true; config0.fraction_of_gpu_memory = 0.3; config0.device = 0; MixedRTConfig config1; - config1.model_dir = FLAGS_dirname + "word2vec.inference.model"; + config1.model_dir = FLAGS_dirname; config1.use_gpu = true; config1.fraction_of_gpu_memory = 0.3; config1.device = 0; diff --git a/paddle/fluid/inference/api/demo_ci/run.sh b/paddle/fluid/inference/api/demo_ci/run.sh index 340e84d9312c20e2d10eb4c0a13066511d5d2eb5..ff718077c1ba6b10fe87aac10d84f96a23ad6bba 100755 --- a/paddle/fluid/inference/api/demo_ci/run.sh +++ b/paddle/fluid/inference/api/demo_ci/run.sh @@ -60,10 +60,9 @@ for WITH_STATIC_LIB in ON OFF; do -DWITH_MKL=$TURN_ON_MKL \ -DDEMO_NAME=simple_on_word2vec \ -DWITH_GPU=$TEST_GPU_CPU \ - -DWITH_STATIC_LIB=$WITH_STATIC_LIB \ - -DON_INFER=ON + -DWITH_STATIC_LIB=$WITH_STATIC_LIB make -j - word2vec_model=${PADDLE_ROOT}'/build/python/paddle/fluid/tests/book/word2vec.inference.model' + word2vec_model=$DATA_DIR'/word2vec/word2vec.inference.model' if [ -d $word2vec_model ]; then for use_gpu in $use_gpu_list; do ./simple_on_word2vec \ @@ -81,8 +80,7 @@ for WITH_STATIC_LIB in ON OFF; do -DWITH_MKL=$TURN_ON_MKL \ -DDEMO_NAME=vis_demo \ -DWITH_GPU=$TEST_GPU_CPU \ - -DWITH_STATIC_LIB=$WITH_STATIC_LIB \ - -DON_INFER=ON + -DWITH_STATIC_LIB=$WITH_STATIC_LIB make -j for use_gpu in $use_gpu_list; do for vis_demo_name in $vis_demo_list; do @@ -108,8 +106,7 @@ for WITH_STATIC_LIB in ON OFF; do -DWITH_STATIC_LIB=$WITH_STATIC_LIB \ -DUSE_TENSORRT=$USE_TENSORRT \ -DTENSORRT_INCLUDE_DIR=$TENSORRT_INCLUDE_DIR \ - -DTENSORRT_LIB_DIR=$TENSORRT_LIB_DIR \ - -DON_INFER=ON + -DTENSORRT_LIB_DIR=$TENSORRT_LIB_DIR make -j ./trt_mobilenet_demo \ --modeldir=$DATA_DIR/mobilenet/model \ diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index 0a37d3968c39d2c244bbd82161afddf6330e421d..7bcf2dd1eeb17e802c5647df31945284ae08fa95 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -18,6 +18,21 @@ namespace paddle { namespace inference { namespace tensorrt { +bool to_skip_merging_optimize(TensorRTEngine* engine_, + const std::vector& filters, + const std::vector& strides, + const std::vector& paddings, + std::string input_name) { + if (engine_->itensor_quote_num[input_name] > 0) { + return true; + } + if (filters[0] == 1 && filters[1] == 1 && strides[0] == 1 && + strides[1] == 1 && paddings[0] == 0 && paddings[1] == 0) + engine_->itensor_quote_num[input_name] += 1; + + return false; +} + class Conv2dOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, @@ -31,6 +46,7 @@ class Conv2dOpConverter : public OpConverter { PADDLE_ENFORCE_EQ(op_desc.Output("Output").size(), 1); auto* X = engine_->GetITensor(op_desc.Input("Input").front()); + // Declare weights auto* Y_v = scope.FindVar(op_desc.Input("Filter").front()); PADDLE_ENFORCE_NOT_NULL(Y_v); @@ -83,7 +99,10 @@ class Conv2dOpConverter : public OpConverter { std::move(weight_tensor); layer->getOutput(0)->setName(output_name.c_str()); engine_->SetITensor(output_name, layer->getOutput(0)); - if (test_mode) { + + if (test_mode || + to_skip_merging_optimize(engine_, {filter_h, filter_w}, strides, + paddings, op_desc.Input("Input").front())) { engine_->DeclareOutput(output_name); } } diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index 14e9e14d33d637ee68e37593cc48721e5169499f..9e0f95844761db7571c5313726d34685a9aa66b2 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -133,6 +133,10 @@ void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer *layer, int offset, buffer_sizes_[name] = 0; } +bool TensorRTEngine::HasDeclared(const std::string &name) { + return buffer_sizes_.count(name) > 0; +} + void TensorRTEngine::DeclareOutput(const std::string &name) { PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate output name %s", name); diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index bd3ba4cea6551a7f6651e311e2649de191a6faa1..d9d3827321127631c0af6e5cfd2dfdd640cee146 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -91,6 +91,8 @@ class TensorRTEngine : public EngineBase { const std::string& name); // Set the itensor_map_[name] as the network's output, and set its name. void DeclareOutput(const std::string& name); + // Check if the ITensor has been declared + bool HasDeclared(const std::string& name); // GPU memory address for an ITensor with specific name. One can operate on // these memory directly for acceleration, for example, output the converted @@ -132,6 +134,16 @@ class TensorRTEngine : public EngineBase { std::unordered_map> weight_map; + // TODO: (NHZLX) + // In the normal case, the paddle-trt exists bug when runing the googlenet. + // When there are more than two convolutions of 1 * 1 with the same input, the + // paddle-tensorrt will do the merging optimization, which fuse those conv + // into + // one conv, and then trigger bug. So, We should use strategy to avoid this + // optimization for the time being. This bug will be fixed in the future. + std::unordered_map + itensor_quote_num; + private: // the max batch size int max_batch_; diff --git a/paddle/fluid/inference/tests/api/CMakeLists.txt b/paddle/fluid/inference/tests/api/CMakeLists.txt index c3dd1f433691e1c96e9f38ef7b595befad26408f..2ca84c80058b35840aff5d072cdc99ecf5165f8e 100644 --- a/paddle/fluid/inference/tests/api/CMakeLists.txt +++ b/paddle/fluid/inference/tests/api/CMakeLists.txt @@ -1,17 +1,9 @@ -set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com") -set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING - "A path setting inference demo download directories.") set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor) -function (inference_download install_dir url filename) - message(STATUS "Download inference test stuff from ${url}/${filename}") - execute_process(COMMAND bash -c "mkdir -p ${install_dir}") - execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}") - message(STATUS "finish downloading ${filename}") -endfunction() -function (inference_download_and_uncompress install_dir url filename) - inference_download(${install_dir} ${url} ${filename}) - execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}") +function(download_model install_dir model_name) + if (NOT EXISTS ${install_dir}) + inference_download_and_uncompress(${install_dir} ${INFERENCE_URL} ${model_name}) + endif() endfunction() function(download_model_and_data install_dir model_name data_name) @@ -27,6 +19,13 @@ function(inference_analysis_api_test target install_dir filename) ARGS --infer_model=${install_dir}/model --infer_data=${install_dir}/data.txt) endfunction() +function(inference_analysis_api_test_with_fake_data target install_dir filename model_name) + download_model(${install_dir} ${model_name}) + inference_analysis_test(${target} SRCS ${filename} + EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} + ARGS --infer_model=${install_dir}/model) +endfunction() + # RNN1 if(NOT APPLE) set(RNN1_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn1") @@ -43,6 +42,15 @@ set(RNN2_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn2") download_model_and_data(${RNN2_INSTALL_DIR} "rnn2_model.tar.gz" "rnn2_data.txt.tar.gz") inference_analysis_api_test(test_analyzer_rnn2 ${RNN2_INSTALL_DIR} analyzer_rnn2_tester.cc) +# DAM +set(DAM_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/dam") +download_model_and_data(${DAM_INSTALL_DIR} "DAM_model.tar.gz" "DAM_data.txt.tar.gz") +inference_analysis_test(test_analyzer_dam SRCS analyzer_dam_tester.cc + EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} ARGS + --infer_model=${DAM_INSTALL_DIR}/model + --infer_data=${DAM_INSTALL_DIR}/data.txt + --use_analysis=0) + # chinese_ner set(CHINESE_NER_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/chinese_ner") download_model_and_data(${CHINESE_NER_INSTALL_DIR} "chinese_ner_model.tar.gz" "chinese_ner-data.txt.tar.gz") @@ -66,17 +74,13 @@ inference_analysis_api_test(test_analyzer_seq_conv1 ${SEQ_CONV1_INSTALL_DIR} ana # ocr set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr") if (NOT EXISTS ${OCR_INSTALL_DIR}) - inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz") + inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz") endif() inference_analysis_api_test(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc) # resnet50 -set(RESNET50_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/resnet50") -if (NOT EXISTS ${RESNET50_INSTALL_DIR}) - inference_download_and_uncompress(${RESNET50_INSTALL_DIR} ${INFERENCE_URL} "resnet50_model.tar.gz") -endif() -inference_analysis_test(test_analyzer_resnet50 SRCS analyzer_resnet50_tester.cc - EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} ARGS --infer_model=${RESNET50_INSTALL_DIR}/model) +inference_analysis_api_test_with_fake_data(test_analyzer_resnet50 + "${INFERENCE_DEMO_INSTALL_DIR}/resnet50" analyzer_resnet50_tester.cc "resnet50_model.tar.gz") # anakin if (WITH_ANAKIN AND WITH_MKL) # only needed in CI diff --git a/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc b/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..ceac5dc7e14365c77cf1cbbbc16e4bf3ebfced73 --- /dev/null +++ b/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc @@ -0,0 +1,224 @@ +// 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/tests/api/tester_helper.h" + +namespace paddle { +namespace inference { +using contrib::AnalysisConfig; +#define MAX_TURN_NUM 9 +#define MAX_TURN_LEN 50 +static std::vector result_data; + +struct DataRecord { + std::vector> + turns[MAX_TURN_NUM]; // turns data : MAX_TURN_NUM + std::vector> + turns_mask[MAX_TURN_NUM]; // turns mask data : MAX_TURN_NUM + std::vector> response; // response data : 1 + std::vector> response_mask; // response mask data : 1 + size_t batch_iter{0}; + size_t batch_size{1}; + size_t num_samples; // total number of samples + DataRecord() = default; + explicit DataRecord(const std::string &path, int batch_size = 1) + : batch_size(batch_size) { + Load(path); + } + DataRecord NextBatch() { + DataRecord data; + size_t batch_end = batch_iter + batch_size; + // NOTE skip the final batch, if no enough data is provided. + if (batch_end <= response.size()) { + for (int i = 0; i < MAX_TURN_NUM; ++i) { + data.turns[i].assign(turns[i].begin() + batch_iter, + turns[i].begin() + batch_end); + } + for (int i = 0; i < MAX_TURN_NUM; ++i) { + data.turns_mask[i].assign(turns_mask[i].begin() + batch_iter, + turns_mask[i].begin() + batch_end); + } + data.response.assign(response.begin() + batch_iter, + response.begin() + batch_end); + data.response_mask.assign(response_mask.begin() + batch_iter, + response_mask.begin() + batch_end); + CHECK(!data.response.empty()); + CHECK(!data.response_mask.empty()); + CHECK_EQ(data.response.size(), data.response_mask.size()); + } + batch_iter += batch_size; + return data; + } + void Load(const std::string &path) { + std::ifstream file(path); + std::string line; + size_t num_lines = 0; + result_data.clear(); + while (std::getline(file, line)) { + num_lines++; + std::vector data; + split(line, ',', &data); + CHECK_EQ(data.size(), 2 * MAX_TURN_NUM + 3); + // load turn data + std::vector turns_tmp[MAX_TURN_NUM]; + for (int i = 0; i < MAX_TURN_NUM; ++i) { + split_to_int64(data[i], ' ', &turns_tmp[i]); + turns[i].push_back(std::move(turns_tmp[i])); + } + // load turn_mask data + std::vector turns_mask_tmp[MAX_TURN_NUM]; + for (int i = 0; i < MAX_TURN_NUM; ++i) { + split_to_float(data[MAX_TURN_NUM + i], ' ', &turns_mask_tmp[i]); + turns_mask[i].push_back(std::move(turns_mask_tmp[i])); + } + // load response data + std::vector response_tmp; + split_to_int64(data[2 * MAX_TURN_NUM], ' ', &response_tmp); + response.push_back(std::move(response_tmp)); + // load response_mask data + std::vector response_mask_tmp; + split_to_float(data[2 * MAX_TURN_NUM + 1], ' ', &response_mask_tmp); + response_mask.push_back(std::move(response_mask_tmp)); + // load result data + float result_tmp; + result_tmp = std::stof(data[2 * MAX_TURN_NUM + 2]); + result_data.push_back(result_tmp); + } + num_samples = num_lines; + } +}; + +void PrepareInputs(std::vector *input_slots, DataRecord *data, + int batch_size) { + PaddleTensor turns_tensor[MAX_TURN_NUM]; + PaddleTensor turns_mask_tensor[MAX_TURN_NUM]; + PaddleTensor response_tensor; + PaddleTensor response_mask_tensor; + std::string turn_pre = "turn_"; + std::string turn_mask_pre = "turn_mask_"; + + auto one_batch = data->NextBatch(); + int size = one_batch.response[0].size(); + CHECK_EQ(size, MAX_TURN_LEN); + // turn tensor assignment + for (int i = 0; i < MAX_TURN_NUM; ++i) { + turns_tensor[i].name = turn_pre + std::to_string(i); + turns_tensor[i].shape.assign({batch_size, size, 1}); + turns_tensor[i].dtype = PaddleDType::INT64; + TensorAssignData(&turns_tensor[i], one_batch.turns[i]); + } + // turn mask tensor assignment + for (int i = 0; i < MAX_TURN_NUM; ++i) { + turns_mask_tensor[i].name = turn_mask_pre + std::to_string(i); + turns_mask_tensor[i].shape.assign({batch_size, size, 1}); + turns_mask_tensor[i].dtype = PaddleDType::FLOAT32; + TensorAssignData(&turns_mask_tensor[i], one_batch.turns_mask[i]); + } + // response tensor assignment + response_tensor.name = "response"; + response_tensor.shape.assign({batch_size, size, 1}); + response_tensor.dtype = PaddleDType::INT64; + TensorAssignData(&response_tensor, one_batch.response); + // response mask tensor assignment + response_mask_tensor.name = "response_mask"; + response_mask_tensor.shape.assign({batch_size, size, 1}); + response_mask_tensor.dtype = PaddleDType::FLOAT32; + TensorAssignData(&response_mask_tensor, one_batch.response_mask); + + // Set inputs. + for (int i = 0; i < MAX_TURN_NUM; ++i) { + input_slots->push_back(std::move(turns_tensor[i])); + } + for (int i = 0; i < MAX_TURN_NUM; ++i) { + input_slots->push_back(std::move(turns_mask_tensor[i])); + } + input_slots->push_back(std::move(response_tensor)); + input_slots->push_back(std::move(response_mask_tensor)); +} + +void SetConfig(contrib::AnalysisConfig *cfg) { + cfg->prog_file = FLAGS_infer_model + "/__model__"; + cfg->param_file = FLAGS_infer_model + "/param"; + cfg->use_gpu = false; + cfg->device = 0; + cfg->specify_input_name = true; + cfg->enable_ir_optim = true; +} + +void SetInput(std::vector> *inputs) { + DataRecord data(FLAGS_infer_data, FLAGS_batch_size); + std::vector input_slots; + int test_batch_num = + FLAGS_test_all_data ? data.num_samples / FLAGS_batch_size : 1; + LOG(INFO) << "The number of samples to be test: " + << test_batch_num * FLAGS_batch_size; + for (int bid = 0; bid < test_batch_num; ++bid) { + input_slots.clear(); + PrepareInputs(&input_slots, &data, FLAGS_batch_size); + (*inputs).emplace_back(input_slots); + } +} + +// Easy for profiling independently. +TEST(Analyzer_dam, profile) { + contrib::AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector outputs; + std::vector> input_slots_all; + SetInput(&input_slots_all); + TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads); + + if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { + 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], result_data[i], 1e-3); + } + } +} + +// Check the fuse status +TEST(Analyzer_dam, fuse_statis) { + contrib::AnalysisConfig cfg; + SetConfig(&cfg); + + if (FLAGS_use_analysis) { + int num_ops; + auto predictor = CreatePaddlePredictor(cfg); + auto fuse_statis = GetFuseStatis( + static_cast(predictor.get()), &num_ops); + ASSERT_TRUE(fuse_statis.count("fc_fuse")); + EXPECT_EQ(fuse_statis.at("fc_fuse"), 317); + EXPECT_EQ(num_ops, 2020); + } +} + +// Compare result of NativeConfig and AnalysisConfig +TEST(Analyzer_dam, compare) { + contrib::AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + + if (FLAGS_use_analysis) { + CompareNativeAndAnalysis(cfg, input_slots_all); + } +} + +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc b/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc index 577b97e271aacab5d6740de7c8bc00bc87ae54dd..d91f7c314d0a936da6f5b0c41920c905af5cd0ee 100644 --- a/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc @@ -20,7 +20,6 @@ using contrib::AnalysisConfig; struct DataRecord { std::vector> word_data_all, mention_data_all; - std::vector> rnn_word_datas, rnn_mention_datas; std::vector lod; // two inputs have the same lod info. size_t batch_iter{0}; size_t batch_size{1}; @@ -45,8 +44,6 @@ struct DataRecord { CHECK(!data.mention_data_all.empty()); CHECK_EQ(data.word_data_all.size(), data.mention_data_all.size()); for (size_t j = 0; j < data.word_data_all.size(); j++) { - data.rnn_word_datas.push_back(data.word_data_all[j]); - data.rnn_mention_datas.push_back(data.mention_data_all[j]); // calculate lod data.lod.push_back(data.lod.back() + data.word_data_all[j].size()); } @@ -87,8 +84,8 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data, lod_mention_tensor.shape.assign({size, 1}); lod_mention_tensor.lod.assign({one_batch.lod}); // assign data - TensorAssignData(&lod_word_tensor, one_batch.rnn_word_datas); - TensorAssignData(&lod_mention_tensor, one_batch.rnn_mention_datas); + TensorAssignData(&lod_word_tensor, one_batch.word_data_all); + TensorAssignData(&lod_mention_tensor, one_batch.mention_data_all); // Set inputs. input_slots->assign({lod_word_tensor, lod_mention_tensor}); for (auto &tensor : *input_slots) { diff --git a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc index c2151eea0823f80feb17b014c1f739d2a15ae862..e5c8dfd22a006d5271248c5b083aab2c22417502 100644 --- a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc @@ -30,25 +30,7 @@ void SetConfig(AnalysisConfig *cfg) { } void SetInput(std::vector> *inputs) { - PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data."); - - PaddleTensor input; - // channel=3, height/width=318 - std::vector shape({FLAGS_batch_size, 3, 318, 318}); - input.shape = shape; - input.dtype = PaddleDType::FLOAT32; - - // fill input data, for profile easily, do not use random data here. - size_t size = FLAGS_batch_size * 3 * 318 * 318; - input.data.Resize(size * sizeof(float)); - float *input_data = static_cast(input.data.data()); - for (size_t i = 0; i < size; i++) { - *(input_data + i) = static_cast(i) / size; - } - - std::vector input_slots; - input_slots.assign({input}); - (*inputs).emplace_back(input_slots); + SetFakeImageInput(inputs, FLAGS_infer_model); } // Easy for profiling independently. @@ -61,13 +43,6 @@ void profile(bool use_mkldnn = false) { std::vector> input_slots_all; SetInput(&input_slots_all); TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads); - - if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { - PADDLE_ENFORCE_EQ(outputs.size(), 1UL); - size_t size = GetSize(outputs[0]); - // output is a 512-dimension feature - EXPECT_EQ(size, 512 * FLAGS_batch_size); - } } TEST(Analyzer_resnet50, profile) { profile(); } @@ -83,8 +58,7 @@ TEST(Analyzer_resnet50, fuse_statis) { auto predictor = CreatePaddlePredictor(cfg); auto fuse_statis = GetFuseStatis( static_cast(predictor.get()), &num_ops); - ASSERT_TRUE(fuse_statis.count("fc_fuse")); - EXPECT_EQ(fuse_statis.at("fc_fuse"), 1); + LOG(INFO) << "num_ops: " << num_ops; } // Compare result of NativeConfig and AnalysisConfig diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h index 19c3f532d5dcb7588793fa21fa179f6b48649103..8c5888d8da7b33eeca77311c10dd818648e8e524 100644 --- a/paddle/fluid/inference/tests/api/tester_helper.h +++ b/paddle/fluid/inference/tests/api/tester_helper.h @@ -25,6 +25,7 @@ #include "paddle/fluid/inference/api/analysis_predictor.h" #include "paddle/fluid/inference/api/helper.h" #include "paddle/fluid/inference/api/paddle_inference_pass.h" +#include "paddle/fluid/inference/tests/test_helper.h" #include "paddle/fluid/platform/profiler.h" DEFINE_string(infer_model, "", "model path"); @@ -105,6 +106,34 @@ std::unordered_map GetFuseStatis(PaddlePredictor *predictor, return fuse_statis; } +void SetFakeImageInput(std::vector> *inputs, + const std::string &dirname) { + // Set fake_image_data + PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data."); + std::vector> feed_target_shapes = + GetFeedTargetShapes(dirname, true, "model", "params"); + int dim1 = feed_target_shapes[0][1]; + int dim2 = feed_target_shapes[0][2]; + int dim3 = feed_target_shapes[0][3]; + + PaddleTensor input; + std::vector shape({FLAGS_batch_size, dim1, dim2, dim3}); + input.shape = shape; + input.dtype = PaddleDType::FLOAT32; + + // fill input data, for profile easily, do not use random data here. + size_t size = FLAGS_batch_size * dim1 * dim2 * dim3; + input.data.Resize(size * sizeof(float)); + float *input_data = static_cast(input.data.data()); + for (size_t i = 0; i < size; i++) { + *(input_data + i) = static_cast(i) / size; + } + + std::vector input_slots; + input_slots.assign({input}); + (*inputs).emplace_back(input_slots); +} + void TestOneThreadPrediction( const AnalysisConfig &config, const std::vector> &inputs, diff --git a/paddle/fluid/inference/tests/api/trt_models_tester.cc b/paddle/fluid/inference/tests/api/trt_models_tester.cc index 91111f2af56065bbf57ba3a41bddd55ecced1060..75840a9c437d956da4f542a38b2532ea20ee96c5 100644 --- a/paddle/fluid/inference/tests/api/trt_models_tester.cc +++ b/paddle/fluid/inference/tests/api/trt_models_tester.cc @@ -93,11 +93,16 @@ void CompareTensorRTWithFluid(int batch_size, std::string model_dirname) { } } -TEST(trt_models_test, main) { - std::vector infer_models = {"mobilenet", "resnet50", - "resnext50"}; - for (auto &model_dir : infer_models) { - CompareTensorRTWithFluid(1, FLAGS_dirname + "/" + model_dir); - } +TEST(trt_models_test, mobilenet) { + CompareTensorRTWithFluid(1, FLAGS_dirname + "/mobilenet"); +} + +TEST(trt_models_test, resnet50) { + CompareTensorRTWithFluid(1, FLAGS_dirname + "/resnet50"); } + +TEST(trt_models_test, resnext50) { + CompareTensorRTWithFluid(1, FLAGS_dirname + "/resnext50"); +} + } // namespace paddle diff --git a/paddle/fluid/inference/tests/test.cmake b/paddle/fluid/inference/tests/test.cmake new file mode 100644 index 0000000000000000000000000000000000000000..ab3a30ce6bba14a7d5ec700a159d90031e6b5dc7 --- /dev/null +++ b/paddle/fluid/inference/tests/test.cmake @@ -0,0 +1,31 @@ +set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url") +set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING + "A path setting inference demo download directories.") +function (inference_download install_dir url filename) + message(STATUS "Download inference test stuff from ${url}/${filename}") + execute_process(COMMAND bash -c "mkdir -p ${install_dir}") + execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}") + message(STATUS "finish downloading ${filename}") +endfunction() + +function (inference_download_and_uncompress install_dir url filename) + inference_download(${install_dir} ${url} ${filename}) + execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}") +endfunction() + +set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec") +if (NOT EXISTS ${WORD2VEC_INSTALL_DIR}) + inference_download_and_uncompress(${WORD2VEC_INSTALL_DIR} ${INFERENCE_URL} "word2vec.inference.model.tar.gz") +endif() +set(WORD2VEC_MODEL_DIR "${WORD2VEC_INSTALL_DIR}/word2vec.inference.model") + +function (inference_base_test TARGET) + set(options "") + set(oneValueArgs "") + set(multiValueArgs SRCS ARGS DEPS) + cmake_parse_arguments(base_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + if(WITH_GPU) + set(mem_opt "--fraction_of_gpu_memory_to_use=0.5") + endif() + cc_test(${TARGET} SRCS ${base_test_SRCS} DEPS ${base_test_DEPS} ARGS ${mem_opt} ${base_test_ARGS}) +endfunction() diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 94f0550df57e79fa68c135f5c9c4b7effe6ac156..2118fcfd4bb1589947617e462f09971fcc090b98 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -18,7 +18,6 @@ limitations under the License. */ #include #include -#include "paddle/fluid/framework/ir/graph_to_program_pass.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/inference/io.h" #include "paddle/fluid/platform/profiler.h" @@ -94,15 +93,15 @@ void CheckError(const paddle::framework::LoDTensor& output1, std::unique_ptr InitProgram( paddle::framework::Executor* executor, paddle::framework::Scope* scope, - const std::string& dirname, const bool is_combined = false) { + const std::string& dirname, const bool is_combined = false, + const std::string& prog_filename = "__model_combined__", + const std::string& param_filename = "__params_combined__") { std::unique_ptr inference_program; if (is_combined) { // All parameters are saved in a single file. // Hard-coding the file names of program and parameters in unittest. // The file names should be consistent with that used in Python API // `fluid.io.save_inference_model`. - std::string prog_filename = "__model_combined__"; - std::string param_filename = "__params_combined__"; inference_program = paddle::inference::Load(executor, scope, dirname + "/" + prog_filename, dirname + "/" + param_filename); @@ -115,12 +114,15 @@ std::unique_ptr InitProgram( } std::vector> GetFeedTargetShapes( - const std::string& dirname, const bool is_combined = false) { + const std::string& dirname, const bool is_combined = false, + const std::string& prog_filename = "__model_combined__", + const std::string& param_filename = "__params_combined__") { auto place = paddle::platform::CPUPlace(); auto executor = paddle::framework::Executor(place); auto* scope = new paddle::framework::Scope(); - auto inference_program = InitProgram(&executor, scope, dirname, is_combined); + auto inference_program = InitProgram(&executor, scope, dirname, is_combined, + prog_filename, param_filename); auto& global_block = inference_program->Block(0); const std::vector& feed_target_names = @@ -136,15 +138,6 @@ std::vector> GetFeedTargetShapes( return feed_target_shapes; } -void Compile(paddle::framework::ProgramDesc* program) { - std::unique_ptr g( - new paddle::framework::ir::Graph(*program)); - auto pass = paddle::framework::ir::PassRegistry::Instance().Get( - "graph_to_program_pass"); - pass->SetNotOwned("program", program); - pass->Apply(std::move(g)); -} - template void TestInference(const std::string& dirname, const std::vector& cpu_feeds, @@ -182,7 +175,6 @@ void TestInference(const std::string& dirname, paddle::platform::DeviceContextPool::Instance().Get(place)); inference_program = InitProgram(&executor, scope, dirname, is_combined); } - Compile(inference_program.get()); // Disable the profiler and print the timing information paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, @@ -261,5 +253,3 @@ void TestInference(const std::string& dirname, delete scope; } - -USE_PASS(graph_to_program_pass); diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index bd7813cec31be0c2e19afc33429d0fad432b6606..242f72e2cade63c89ad9643c69043f5804749d46 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -301,6 +301,7 @@ 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(crf_decoding_op DEPS jit_kernel) op_library(fusion_lstm_op DEPS jit_kernel) if (WITH_GPU) op_library(conv_op DEPS vol2col depthwise_conv im2col) diff --git a/paddle/fluid/operators/activation_op.cu b/paddle/fluid/operators/activation_op.cu index 27487b396ccf63d962defa6b270063ccb409164e..d3a7ceed466a9b5e4d773f1531d198adff97eac2 100644 --- a/paddle/fluid/operators/activation_op.cu +++ b/paddle/fluid/operators/activation_op.cu @@ -26,6 +26,8 @@ namespace plat = paddle::platform; act_type##_grad, ops::ActivationGradKernel>, \ ops::ActivationGradKernel>); + ops::grad_functor>, \ + ops::ActivationGradKernel>); FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL); diff --git a/paddle/fluid/operators/activation_op.h b/paddle/fluid/operators/activation_op.h index 2e31d1c9c708225135e27c93ba94722794c4b282..0747469e0f4c4fe6a323a499c720a54d1e278e09 100644 --- a/paddle/fluid/operators/activation_op.h +++ b/paddle/fluid/operators/activation_op.h @@ -333,8 +333,7 @@ struct SqrtGradFunctor : public BaseActivationFunctor { template void operator()(Device d, X x, Out out, dOut dout, dX dx) const { - const Out out_conj = Eigen::numext::conj(out); - dx.device(d) = static_cast(0.5) * dout / out_conj; + dx.device(d) = static_cast(0.5) * dout / out; } }; @@ -740,7 +739,7 @@ struct PowGradFunctor : public BaseActivationFunctor { typename dX> void operator()(Device d, X x, Out out, dOut dout, dX dx) const { dx.device(d) = dout * static_cast(factor) * - x.pow(static_cast(factor - static_cast(1))); + x.pow(static_cast(factor) - static_cast(1)); } }; diff --git a/paddle/fluid/operators/add_position_encoding_op.cc b/paddle/fluid/operators/add_position_encoding_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..8127e554bed1aae7a5ce8837bcadf1b7f13f1ac2 --- /dev/null +++ b/paddle/fluid/operators/add_position_encoding_op.cc @@ -0,0 +1,97 @@ +/* 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/add_position_encoding_op.h" + +namespace paddle { +namespace operators { + +class AddPositionEncodingOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "X(Input) of add_position_encoding_op should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Out"), + "Out(Output) of add_position_encoding_op should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class AddPositionEncodingOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "X(Input) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("Out"), "Out must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Out@GRAD must not be null."); + + auto out_dims = ctx->GetInputDim("Out"); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), out_dims); + } + } +}; + +class AddPositionEncodingOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "Input of AddPositionEncoding operator"); + AddOutput("Out", "Output of AddPositionEncoding operator"); + AddAttr("alpha", "The scale of Original Embedding.") + .SetDefault(1.0f) + .AddCustomChecker([](const float& alpha) { + PADDLE_ENFORCE(alpha >= 0.0f, "'alpha' must be above 0.0."); + }); + AddAttr("beta", "The scale of Position Embedding.") + .SetDefault(1.0f) + .AddCustomChecker([](const float& beta) { + PADDLE_ENFORCE(beta >= 0.0f, "'beta' must be between 0.0."); + }); + AddComment(R"DOC( + Add Position Encoding Operator. + + The add position encoding calculates the output based on the input, alpha, beta. + The size of each dimension of the parameters checked in the infer-shape. + )DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +namespace plt = paddle::platform; + +REGISTER_OPERATOR(add_position_encoding, ops::AddPositionEncodingOp, + ops::AddPositionEncodingOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(add_position_encoding_grad, ops::AddPositionEncodingOpGrad); + +REGISTER_OP_CPU_KERNEL( + add_position_encoding, + ops::AddPositionEncodingKernel, + ops::AddPositionEncodingKernel); + +REGISTER_OP_CPU_KERNEL( + add_position_encoding_grad, + ops::AddPositionEncodingGradKernel, + ops::AddPositionEncodingGradKernel); diff --git a/paddle/fluid/operators/add_position_encoding_op.h b/paddle/fluid/operators/add_position_encoding_op.h new file mode 100644 index 0000000000000000000000000000000000000000..5f371235f160c416058e877dbba2d9fe89abf7db --- /dev/null +++ b/paddle/fluid/operators/add_position_encoding_op.h @@ -0,0 +1,105 @@ +/* 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/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/safe_ref.h" + +namespace paddle { +namespace operators { + +template +class AddPositionEncodingKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* X = context.Input("X"); + auto& x_lod = X->lod(); + auto* src_ptr = X->data(); + + auto* Out = context.Output("Out"); + auto* dst_ptr = Out->mutable_data(context.GetPlace()); + + float alpha = context.Attr("alpha"); + float beta = context.Attr("beta"); + + auto x_dim = X->dims(); + int batch_size = 0; + int max_seq_len = 0; + int enc_size = 0; + + if (x_lod.empty()) { + PADDLE_ENFORCE( + x_dim.size() == 3UL, + "The input X of Add Position Encoding should be 3-D Tensor!"); + batch_size = x_dim[0]; + max_seq_len = x_dim[1]; + enc_size = x_dim[2]; + } else { + PADDLE_ENFORCE( + x_dim.size() == 2UL, + "The input X of Add Position Encoding should be 2-D LoDTensor!"); + PADDLE_ENFORCE( + x_lod.size() == 1UL, + "The Add Position Encoding Op only supports lod_level == 1!"); + batch_size = x_lod[0].size() - 1; + max_seq_len = -1; + enc_size = x_dim[1]; + } + + PADDLE_ENFORCE(enc_size % 2 == 0, "Only support even encode size!"); + + const int half_size = enc_size / 2; + for (int i = 0; i < batch_size; ++i) { + const int max_length = + x_lod.empty() ? max_seq_len : x_lod[0][i + 1] - x_lod[0][i]; + for (int j = 0; j < max_length; ++j) { + for (int k = 0; k < half_size; ++k) { + const double val = (half_size > 1) + ? j / pow(10000.0, double(k) / (half_size - 1)) + : j / 10000.0; + dst_ptr[k] = src_ptr[k] * alpha + sin(val) * beta; + dst_ptr[half_size + k] = + src_ptr[half_size + k] * alpha + cos(val) * beta; + } + src_ptr += enc_size; + dst_ptr += enc_size; + } + } + } +}; + +template +class AddPositionEncodingGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* dOut = + context.Input(framework::GradVarName("Out")); + auto dout = framework::EigenVector::Flatten(*dOut); + + auto* dX = + context.Output(framework::GradVarName("X")); + dX->mutable_data(context.GetPlace()); + auto dx = framework::EigenVector::Flatten(*dX); + + float alpha = context.Attr("alpha"); + + auto* place = + context.template device_context().eigen_device(); + dx.device(*place) = dout * static_cast(alpha); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/affine_grid_cudnn_op.cu.cc b/paddle/fluid/operators/affine_grid_cudnn_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..ed71594ba5781590f3291d56c4ba1a4443003bd5 --- /dev/null +++ b/paddle/fluid/operators/affine_grid_cudnn_op.cu.cc @@ -0,0 +1,112 @@ +/* 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/op_registry.h" +#include "paddle/fluid/platform/cudnn_helper.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using ScopedSpatialTransformerDescriptor = + platform::ScopedSpatialTransformerDescriptor; + +template +class CUDNNAffineGridOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace."); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto* theta = ctx.Input("Theta"); + auto* output = ctx.Output("Output"); + const T* theta_data = theta->data(); + + int n = theta->dims()[0]; + auto size_attr = ctx.Attr>("output_shape"); + Tensor h_sizes; + int* h_size_data; + if (size_attr.size() == 0) { + auto* output_shape = ctx.Input("OutputShape"); + framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes); + h_size_data = h_sizes.data(); + } else { + h_size_data = h_sizes.mutable_data({4}, platform::CPUPlace()); + h_size_data[0] = n; + h_size_data[1] = size_attr[1]; + h_size_data[2] = size_attr[2]; + h_size_data[3] = size_attr[3]; + } + + T* output_data = output->mutable_data( + {n, h_size_data[2], h_size_data[3], 2}, ctx.GetPlace()); + ScopedSpatialTransformerDescriptor st_desc; + cudnnSpatialTransformerDescriptor_t cudnn_st_desc = + st_desc.descriptor(4, h_size_data); + + PADDLE_ENFORCE(platform::dynload::cudnnSpatialTfGridGeneratorForward( + handle, cudnn_st_desc, theta_data, output_data)); + } +}; + +template +class CUDNNAffineGridGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace."); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto output_grad = ctx.Input(framework::GradVarName("Output")); + auto theta_grad = ctx.Output(framework::GradVarName("Theta")); + + int n = output_grad->dims()[0]; + auto size_attr = ctx.Attr>("output_shape"); + Tensor h_sizes; + int* h_size_data; + if (size_attr.size() == 0) { + auto* output_shape = ctx.Input("OutputShape"); + framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes); + h_size_data = h_sizes.data(); + } else { + h_size_data = h_sizes.mutable_data({4}, platform::CPUPlace()); + h_size_data[0] = n; + h_size_data[1] = size_attr[1]; + h_size_data[2] = size_attr[2]; + h_size_data[3] = size_attr[3]; + } + + ScopedSpatialTransformerDescriptor st_desc; + cudnnSpatialTransformerDescriptor_t cudnn_st_desc = + st_desc.descriptor(4, h_size_data); + + const T* output_grad_data = output_grad->data(); + T* theta_grad_data = theta_grad->mutable_data(ctx.GetPlace()); + + PADDLE_ENFORCE(platform::dynload::cudnnSpatialTfGridGeneratorBackward( + handle, cudnn_st_desc, output_grad_data, theta_grad_data)); + } +}; + +} // namespace operators +} // namespace paddle + +namespace plat = paddle::platform; +REGISTER_OP_KERNEL(affine_grid, CUDNN, plat::CUDAPlace, + paddle::operators::CUDNNAffineGridOpKernel, + paddle::operators::CUDNNAffineGridOpKernel); +REGISTER_OP_KERNEL(affine_grid_grad, CUDNN, plat::CUDAPlace, + paddle::operators::CUDNNAffineGridGradOpKernel, + paddle::operators::CUDNNAffineGridGradOpKernel); diff --git a/paddle/fluid/operators/affine_grid_op.cc b/paddle/fluid/operators/affine_grid_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..6f7da445fc84fc1f14b01a633af0e886aec6f8ed --- /dev/null +++ b/paddle/fluid/operators/affine_grid_op.cc @@ -0,0 +1,231 @@ +/* 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/affine_grid_op.h" +#include +#include "paddle/fluid/framework/op_registry.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/cudnn_helper.h" +#endif + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +struct Linspace { + void operator()(T start, T end, int count, framework::Tensor* numbers, + const framework::ExecutionContext& ctx) { + T* number_data = numbers->mutable_data({count}, platform::CPUPlace()); + T slice = (end - start) / (T)(count - 1); + for (int i = 0; i < count; ++i) { + number_data[i] = start + (T)i * slice; + } + } +}; + +class AffineGridOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Theta"), + "Input(Theta) of AffineGridOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Output"), + "Output(Output) of AffineGridOp should not be null."); + auto theta_dims = ctx->GetInputDim("Theta"); + PADDLE_ENFORCE(theta_dims.size() == 3, + "AffineGrid's Input(Theta) should be 3-D tensor."); + + auto output_shape = ctx->Attrs().Get>("output_shape"); + if (output_shape.size() == 0) { + PADDLE_ENFORCE(ctx->HasInput("OutputShape"), + "Input(OutputShape) of AffineGridOp should not be null if " + "attr(output_shape) is not configured."); + auto output_shape_dims = ctx->GetInputDim("OutputShape"); + PADDLE_ENFORCE(output_shape_dims.size() == 1, + "AffineGrid's Input(OutputShape) should be 1-D tensor."); + } else { + PADDLE_ENFORCE(output_shape.size() == 4, + "The size of attr(output_shape) should be 4."); + } + + PADDLE_ENFORCE(theta_dims[1] == 2, "Input(theta) dims[1] should be 2."); + PADDLE_ENFORCE(theta_dims[2] == 3, "Input(theta) dims[2] should be 3."); + // N * H * W * 2 + ctx->SetOutputDim("Output", + framework::make_ddim({theta_dims[0], -1, -1, 2})); + ctx->ShareLoD("Theta", "Output"); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::LibraryType library{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_CUDA + if (platform::CanCUDNNBeUsed(ctx)) { + library = framework::LibraryType::kCUDNN; + } +#endif + auto data_type = framework::ToDataType(ctx.Input("Theta")->type()); + return framework::OpKernelType(data_type, ctx.GetPlace(), + framework::DataLayout::kAnyLayout, library); + } +}; + +class AffineGridOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput( + "Theta", + "(Tensor) A batch of affine transform parameters with shape [N, 2, 3]. " + "It is used to transform coordinate (x_0, y_0) to coordinate (x_1, " + "y_1)."); + AddInput("OutputShape", + "(Tensor) The shape of target image with format [N, C, H, W].") + .AsDispensable(); + AddOutput("Output", "(Tensor) Output Tensor with shape [N, H, W, 2]."); + AddAttr( + "use_cudnn", + "(bool, default false) Only used in cudnn kernel, need install cudnn") + .SetDefault(true); + AddAttr>( + "output_shape", + "The target output image shape with format [N, C, H, W].") + .SetDefault(std::vector()); + + AddComment(R"DOC( + It generates a grid of (x,y) coordinates using the parameters of the + affine transformation that correspond to a set of points where the input + feature map should be sampled to produce the transformed output feature map. + + Given: + Theta = [[[x_11, x_12, x_13] + [x_14, x_15, x_16]] + [[x_21, x_22, x_23] + [x_24, x_25, x_26]]] + + OutputShape = [2, 3, 5, 5] + + Step 1: + + Generate relative coordinates according to OutputShape. + The values of relative coordinates are in the interval between -1 and 1. + The shape of the relative coordinates is [2, H, W] as below: + + C = [[[-1. -1. -1. -1. -1. ] + [-0.5 -0.5 -0.5 -0.5 -0.5] + [ 0. 0. 0. 0. 0. ] + [ 0.5 0.5 0.5 0.5 0.5] + [ 1. 1. 1. 1. 1. ]] + [[-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ]]] + C[0] is the coordinates in height axis and C[1] is the coordinates in width axis. + + Step2: + Tanspose and reshape C to shape [H * W, 2] and append ones to last dimension. The we get: + C_ = [[-1. -1. 1. ] + [-0.5 -1. 1. ] + [ 0. -1. 1. ] + [ 0.5 -1. 1. ] + [ 1. -1. 1. ] + [-1. -0.5 1. ] + [-0.5 -0.5 1. ] + [ 0. -0.5 1. ] + [ 0.5 -0.5 1. ] + [ 1. -0.5 1. ] + [-1. 0. 1. ] + [-0.5 0. 1. ] + [ 0. 0. 1. ] + [ 0.5 0. 1. ] + [ 1. 0. 1. ] + [-1. 0.5 1. ] + [-0.5 0.5 1. ] + [ 0. 0.5 1. ] + [ 0.5 0.5 1. ] + [ 1. 0.5 1. ] + [-1. 1. 1. ] + [-0.5 1. 1. ] + [ 0. 1. 1. ] + [ 0.5 1. 1. ] + [ 1. 1. 1. ]] + Step3: + Compute output by equation $$Output[i] = C_ * Theta[i]^T$$ + )DOC"); + } +}; + +class AffineGridOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + auto theta_dims = ctx->GetInputDim("Theta"); + if (ctx->HasOutput(framework::GradVarName("Theta"))) { + ctx->SetOutputDim(framework::GradVarName("Theta"), theta_dims); + } + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_CUDA + if (platform::CanCUDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kCUDNN; + } +#endif + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Theta")->type()), + ctx.GetPlace(), framework::DataLayout::kAnyLayout, library_); + } +}; + +class AffineGridGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("affine_grid_grad"); + op->SetInput("Theta", Input("Theta")); + op->SetInput("OutputShape", Input("OutputShape")); + op->SetInput(framework::GradVarName("Output"), OutputGrad("Output")); + + op->SetAttrMap(Attrs()); + + op->SetOutput(framework::GradVarName("Theta"), InputGrad("Theta")); + return std::unique_ptr(op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(affine_grid, ops::AffineGridOp, ops::AffineGridOpMaker, + ops::AffineGridGradMaker); +REGISTER_OPERATOR(affine_grid_grad, ops::AffineGridOpGrad); + +REGISTER_OP_CPU_KERNEL( + affine_grid, + ops::AffineGridOpKernel, + ops::AffineGridOpKernel); +REGISTER_OP_CPU_KERNEL( + affine_grid_grad, + ops::AffineGridGradOpKernel, + ops::AffineGridGradOpKernel); diff --git a/paddle/fluid/operators/affine_grid_op.h b/paddle/fluid/operators/affine_grid_op.h new file mode 100644 index 0000000000000000000000000000000000000000..87d23831486e658374d4c011412fdef57be1b994 --- /dev/null +++ b/paddle/fluid/operators/affine_grid_op.h @@ -0,0 +1,174 @@ +/* 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/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenTensor = framework::EigenTensor; + +using Array1 = Eigen::DSizes; +using Array2 = Eigen::DSizes; +using Array3 = Eigen::DSizes; +using Array4 = Eigen::DSizes; + +/** + *Return a tensor with evenly spaced numbers over a specified interval. + */ +template +struct Linspace { + void operator()(T start, T end, int count, framework::Tensor* numbers, + const framework::ExecutionContext& ctx); +}; + +template +inline void GetIdxMap(int n, int h, int w, Tensor* grid, + const framework::ExecutionContext& ctx) { + auto& place = *ctx.template device_context().eigen_device(); + grid->mutable_data({n, h, w, 3}, ctx.GetPlace()); + auto grid_t = EigenTensor::From(*grid); + // Get indexes of height with shape [height, width, 1] + Tensor h_idx; + Linspace linspace; + linspace((T)-1, (T)1, h, &h_idx, ctx); + auto h_idx_t = EigenTensor::From(h_idx); + // Get indexes of width with shape [height, width, 1] + Tensor w_idx; + linspace((T)-1, (T)1, w, &w_idx, ctx); + auto w_idx_t = EigenTensor::From(w_idx); + // Get constant ones tensor with shape [height, width, 1] + Tensor ones; + ones.mutable_data({h, w, 1}, ctx.GetPlace()); + auto ones_t = EigenTensor::From(ones).setConstant((T)1); + // Get grid tensor with shape [n, h, w, 3] by concatenating h_idx, w_idx and + // ones + Tensor w_idx_map; + w_idx_map.mutable_data({h, w, 1}, ctx.GetPlace()); + auto w_idx_map_t = EigenTensor::From(w_idx_map); + Tensor h_idx_map; + h_idx_map.mutable_data({h, w, 1}, ctx.GetPlace()); + auto h_idx_map_t = EigenTensor::From(h_idx_map); + Tensor w_h_idx_map; + w_h_idx_map.mutable_data({h, w, 2}, ctx.GetPlace()); + auto w_h_idx_map_t = EigenTensor::From(w_h_idx_map); + Tensor w_h_one_idx_map; + w_h_one_idx_map.mutable_data({h, w, 3}, ctx.GetPlace()); + auto w_h_one_idx_map_t = EigenTensor::From(w_h_one_idx_map); + + w_idx_map_t.device(place) = w_idx_t.reshape(Array2(1, w)) + .broadcast(Array2(h, 1)) + .reshape(Array3(h, w, 1)); + + h_idx_map_t.device(place) = h_idx_t.reshape(Array2(1, h)) + .broadcast(Array2(w, 1)) + .shuffle(Array2(1, 0)) + .reshape(Array3(h, w, 1)); + + w_h_idx_map_t.device(place) = w_idx_map_t.concatenate(h_idx_map_t, 2); + w_h_one_idx_map_t.device(place) = w_h_idx_map_t.concatenate(ones_t, 2); + grid_t.device(place) = w_h_one_idx_map_t.reshape(Array4(1, h, w, 3)) + .broadcast(Array4(n, 1, 1, 1)); +} + +template +class AffineGridOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* theta = ctx.Input("Theta"); + int n = theta->dims()[0]; + auto size_attr = ctx.Attr>("output_shape"); + int h = 0; + int w = 0; + if (size_attr.size() == 0) { + auto* output_shape = ctx.Input("OutputShape"); + Tensor h_sizes; + framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes); + const int* h_size_data = h_sizes.data(); + h = h_size_data[2]; + w = h_size_data[3]; + } else { + h = size_attr[2]; + w = size_attr[3]; + } + auto* output = ctx.Output("Output"); + output->mutable_data({n, h, w, 2}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), output, + static_cast(0)); + Tensor grid; + GetIdxMap(n, h, w, &grid, ctx); + // output = grid * theta.T + // TODO(wanghaoshuang): Refine batched matrix multiply + auto blas = math::GetBlas(ctx); + for (int i = 0; i < n; ++i) { + Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3}); + Tensor sliced_theta = theta->Slice(i, i + 1).Resize({2, 3}); + Tensor sliced_out = output->Slice(i, i + 1).Resize({h * w, 2}); + blas.MatMul(sliced_grid, false, sliced_theta, true, T(1), &sliced_out, + T(0)); + } + } +}; + +template +class AffineGridGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto output_grad = ctx.Input(framework::GradVarName("Output")); + auto theta_grad = ctx.Output(framework::GradVarName("Theta")); + int n = output_grad->dims()[0]; + auto size_attr = ctx.Attr>("output_shape"); + int h = 0; + int w = 0; + if (size_attr.size() == 0) { + auto* output_shape = ctx.Input("OutputShape"); + Tensor h_sizes; + framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes); + const int* h_size_data = h_sizes.data(); + h = h_size_data[2]; + w = h_size_data[3]; + } else { + h = size_attr[2]; + w = size_attr[3]; + } + theta_grad->mutable_data({n, 2, 3}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), theta_grad, + static_cast(0)); + Tensor grid; + GetIdxMap(n, h, w, &grid, ctx); + // output = grid * theta.T + // TODO(wanghaoshuang): Refine batched matrix multiply + auto blas = math::GetBlas(ctx); + for (int i = 0; i < n; ++i) { + Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3}); + Tensor sliced_out_grad = output_grad->Slice(i, i + 1).Resize({h * w, 2}); + Tensor sliced_theta_grad = theta_grad->Slice(i, i + 1).Resize({2, 3}); + blas.MatMul(sliced_out_grad, true, sliced_grid, false, T(1), + &sliced_theta_grad, T(0)); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/batch_norm_op.cu.cc b/paddle/fluid/operators/batch_norm_op.cu.cc index ca6cd8669352fd5814f25a04433ca97fe4abe9ff..aaed335c905c0d80cd519afc5fecb06af73fcfe7 100644 --- a/paddle/fluid/operators/batch_norm_op.cu.cc +++ b/paddle/fluid/operators/batch_norm_op.cu.cc @@ -219,8 +219,8 @@ class BatchNormGradKernel auto *d_bias = ctx.Output(framework::GradVarName("Bias")); d_x->mutable_data(ctx.GetPlace()); - d_scale->mutable_data(ctx.GetPlace()); - d_bias->mutable_data(ctx.GetPlace()); + d_scale->mutable_data>(ctx.GetPlace()); + d_bias->mutable_data>(ctx.GetPlace()); auto &dev_ctx = ctx.template device_context(); if ((N * H * W * D) == 1) { @@ -272,8 +272,10 @@ class BatchNormGradKernel const auto *saved_mean = ctx.Input("SavedMean"); const auto *saved_var = ctx.Input("SavedVariance"); - const void *saved_mean_data = saved_mean->template data(); - const void *saved_var_data = saved_var->template data(); + const void *saved_mean_data = + saved_mean->template data>(); + const void *saved_var_data = + saved_var->template data>(); CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward( dev_ctx.cudnn_handle(), mode_, CudnnDataType::kOne(), @@ -281,10 +283,10 @@ class BatchNormGradKernel CudnnDataType::kZero(), data_desc_, x->template data(), data_desc_, d_y->template data(), data_desc_, d_x->template mutable_data(ctx.GetPlace()), bn_param_desc_, - scale->template data(), - d_scale->template mutable_data(ctx.GetPlace()), - d_bias->template mutable_data(ctx.GetPlace()), epsilon, - saved_mean_data, saved_var_data)); + scale->template data>(), + d_scale->template mutable_data>(ctx.GetPlace()), + d_bias->template mutable_data>(ctx.GetPlace()), + epsilon, saved_mean_data, saved_var_data)); // clean when exit. CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(data_desc_)); @@ -304,4 +306,5 @@ REGISTER_OP_CUDA_KERNEL( ops::BatchNormKernel); REGISTER_OP_CUDA_KERNEL( batch_norm_grad, ops::BatchNormGradKernel, - ops::BatchNormGradKernel); + ops::BatchNormGradKernel, + ops::BatchNormGradKernel); diff --git a/paddle/fluid/operators/checkpoint_notify_op.cc b/paddle/fluid/operators/checkpoint_notify_op.cc index 3a2527e407bb179c4873fa3ffe2e8f22fb47faf7..7c072cb071a5d1e6a0549cf6d9eff18fd2533edc 100644 --- a/paddle/fluid/operators/checkpoint_notify_op.cc +++ b/paddle/fluid/operators/checkpoint_notify_op.cc @@ -38,9 +38,10 @@ class CheckpointNotifyOp : public framework::OperatorBase { std::vector epmap = Attr>("epmap"); std::string dir = Attr("dir"); std::string lookup_table_name = Attr("lookup_table"); + int trainer_id = Attr("trainer_id"); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(trainer_id); for (size_t i = 0; i < epmap.size(); i++) { auto lookup_table_save_dir = string::Sprintf("%s/%s_%d", dir, lookup_table_name, i); @@ -63,6 +64,7 @@ class CheckpointNotifyOpMaker : public framework::OpProtoAndCheckerMaker { "dir", "(string, default '') indicate the folder checkpoint will use"); AddAttr("lookup_table", "(string, default '') the lookup table name"); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddComment(R"DOC( CheckpointNotify operator diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc index 4a7a6bcf7154d5680de751e3c933be46fb09fd74..76eda51ad414030074b69ee8d4f796c5c32d12f3 100644 --- a/paddle/fluid/operators/conv_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc @@ -143,9 +143,11 @@ class CUDNNConvOpKernel : public framework::OpKernel { cudnn_conv_desc, CUDNN_TENSOR_OP_MATH)); // Currently tensor core is only enabled using this algo algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM; + VLOG(5) << "use cudnn_tensor_op_math"; } else { CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( cudnn_conv_desc, CUDNN_DEFAULT_MATH)); + VLOG(5) << "NOT use cudnn_tensor_op_math"; } #endif @@ -160,6 +162,7 @@ class CUDNNConvOpKernel : public framework::OpKernel { // ------------------- cudnn conv forward --------------------- ScalingParamType alpha = 1.0f, beta = 0.0f; + auto workspace_handle = dev_ctx.cudnn_workspace_handle(); for (int i = 0; i < groups; i++) { auto cudnn_func = [&](void* cudnn_workspace) { CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward( @@ -168,7 +171,7 @@ class CUDNNConvOpKernel : public framework::OpKernel { cudnn_conv_desc, algo, cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_output_desc, output_data + i * group_offset_out)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } }; @@ -314,6 +317,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { // ------------------- cudnn conv backward data --------------------- ScalingParamType alpha = 1.0f, beta = 0.0f; + auto workspace_handle = dev_ctx.cudnn_workspace_handle(); if (input_grad) { T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset input_grad. @@ -327,7 +331,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { data_algo, cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc, input_grad_data + i * group_offset_in)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } // ------------------- cudnn conv backward filter --------------------- @@ -343,7 +347,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { filter_algo, cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_filter_desc, filter_grad_data + i * group_offset_filter)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } } @@ -359,7 +363,8 @@ REGISTER_OP_KERNEL(conv2d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel); REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvGradOpKernel, - paddle::operators::CUDNNConvGradOpKernel); + paddle::operators::CUDNNConvGradOpKernel, + paddle::operators::CUDNNConvGradOpKernel); REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel, diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc index 71a5db06fce07571944a0003c2eab629e1710b41..5a2ac66cd242d773d172243e07f213cc39a2841c 100644 --- a/paddle/fluid/operators/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/conv_mkldnn_op.cc @@ -15,6 +15,8 @@ #include "paddle/fluid/operators/conv_op.h" #include "paddle/fluid/platform/mkldnn_helper.h" +#include "paddle/fluid/framework/data_layout_transform.h" + namespace paddle { namespace operators { @@ -57,6 +59,11 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler { return conv_pd_->dst_primitive_desc().get_size(); } + mkldnn::memory::format GetDstFormat() const { + return static_cast( + conv_pd_->dst_primitive_desc().desc().data.format); + } + size_t GetDiffWeightsMemorySize() const { return conv_bwd_weights_pd_->diff_weights_primitive_desc().get_size(); } @@ -108,6 +115,20 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler { "@data-weights_mem_p", pipeline); } + std::shared_ptr AcquireResidualDataMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_residual_data_mem_p"); + } + + std::shared_ptr AcquireDstMemoryFromResidualDataMemory( + const std::shared_ptr& user_residual_memory_p, + void* dst_ptr, + std::vector& pipeline) { // NOLINT + return this->AcquireMemory(user_residual_memory_p, + this->AcquireDstMemoryFromPrimitive(dst_ptr), + "@residual_data_mem_p", pipeline); + } + std::shared_ptr AcquireDiffSrcMemoryFromDataPrimitive( void* ptr) { return this->AcquireMemoryFromPrimitive( @@ -386,7 +407,13 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto user_weights_memory_p = handler.AcquireWeightsMemory( user_weights_md, to_void_cast(filter_data)); - T* output_data = nullptr; + // create reorder primitive if the input format is not the preferred one + auto src_memory_p = + handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline); + auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive( + user_weights_memory_p, pipeline, is_test); + + std::shared_ptr dst_memory_p; if (fuse_residual_conn) { auto residual_param = ctx.Input("ResidualData"); @@ -399,20 +426,34 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { "Output and elementwise parameter need to have the " "same dimension sizes"); - output->ShareDataWith(*residual_param); - output_data = output->mutable_data(ctx.GetPlace()); + if (residual_param->format() != handler.GetDstFormat()) { + auto output_data = + output->mutable_data(ctx.GetPlace(), handler.GetDstMemorySize()); + auto residual_data_tz = + paddle::framework::vectorize2int(residual_param->dims()); + auto residual_data_type = + paddle::framework::ToMKLDNNDataType(residual_param->type()); + + auto user_residual_md = platform::MKLDNNMemDesc( + residual_data_tz, residual_data_type, residual_param->format()); + auto user_residual_memory_p = handler.AcquireResidualDataMemory( + user_residual_md, to_void_cast(residual_param_data)); + + dst_memory_p = handler.AcquireDstMemoryFromResidualDataMemory( + user_residual_memory_p, to_void_cast(output_data), pipeline); + } else { + output->ShareDataWith(*residual_param); + auto output_data = output->mutable_data(ctx.GetPlace()); + dst_memory_p = + handler.AcquireDstMemoryFromPrimitive(to_void_cast(output_data)); + } } else { - output_data = output->mutable_data(ctx.GetPlace(), - paddle::memory::Allocator::kDefault, - handler.GetDstMemorySize()); + auto output_data = output->mutable_data( + ctx.GetPlace(), paddle::memory::Allocator::kDefault, + handler.GetDstMemorySize()); + dst_memory_p = + handler.AcquireDstMemoryFromPrimitive(to_void_cast(output_data)); } - // create reorder primitive if the input format is not the preferred one - auto src_memory_p = - handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline); - auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive( - user_weights_memory_p, pipeline, is_test); - auto dst_memory_p = - handler.AcquireDstMemoryFromPrimitive(to_void_cast(output_data)); // create convolution op primitive std::shared_ptr conv_p; diff --git a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc index 73831611d01b8c5b8d2d9f7f15634a0094e4a608..f44094ca6b7b7f23f2e7593ad79e4e2a6f0d3070 100644 --- a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc @@ -104,6 +104,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { int output_offset = output->numel() / output->dims()[0] / groups; int filter_offset = filter->numel() / groups; T alpha = 1.0f, beta = 0.0f; + auto workspace_handle = dev_ctx.cudnn_workspace_handle(); for (int g = 0; g < groups; g++) { auto cudnn_func = [&](void* cudnn_workspace) { CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData( @@ -112,7 +113,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { algo, cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_output_desc, output_data + output_offset * g)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } }; @@ -208,6 +209,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { output_grad->numel() / output_grad->dims()[0] / groups; int filter_offset = filter->numel() / groups; T alpha = 1.0f, beta = 0.0f; + auto workspace_handle = dev_ctx.cudnn_workspace_handle(); if (input_grad) { T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset input_grad. @@ -220,7 +222,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc, input_grad_data + input_offset * g)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } @@ -238,7 +240,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_filter_desc, filter_grad_data + filter_offset * g)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } } diff --git a/paddle/fluid/operators/crf_decoding_op.h b/paddle/fluid/operators/crf_decoding_op.h index 8181897c3d3844bda5574e85a08b2af038fcd664..e9d2e84a434d7084c526a6e75363a65577197262 100644 --- a/paddle/fluid/operators/crf_decoding_op.h +++ b/paddle/fluid/operators/crf_decoding_op.h @@ -16,6 +16,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/jit_kernel.h" #include "paddle/fluid/operators/math/math_function.h" namespace paddle { @@ -69,9 +70,6 @@ class CRFDecodingOpKernel : public framework::OpKernel { auto emission_dims = emission_weights.dims(); const size_t seq_len = emission_dims[0]; const size_t tag_num = emission_dims[1]; - - const size_t state_trans_base_idx = 2; - const T* x = emission_weights.data(); const T* w = transition_weights.data(); int64_t* path = decoded_path->data(); @@ -84,221 +82,10 @@ class CRFDecodingOpKernel : public framework::OpKernel { Tensor track; int* track_value = track.mutable_data(emission_dims, platform::CPUPlace()); - -#ifdef __AVX__ -// It use the AVX or AVX512 instruction to deal the data as the vector of 8 or -// 16 elements per iteration. Then it can implement the parallel processing. -// Only optimize for float type. -#ifdef __AVX512F__ - size_t step_size = 16; -#else - size_t step_size = 8; -#endif - if (std::is_same::value && (tag_num >= step_size)) { - size_t steps = tag_num / step_size; - size_t remain = tag_num % step_size; - int last_offset = static_cast(remain) - static_cast(step_size); - - // Setup the alpha initial value. - size_t i_offset = 0; - for (size_t i = 0; i <= steps; ++i) { -#ifdef __AVX512F__ - // Declare the variable for the content of weights, input and alpha - // values. - __m512 w_content, x_content, alpha_content; - - // Load the relevant data into the variables from un-aligned address. - w_content = _mm512_loadu_ps((const float*)(w + i_offset)); - x_content = _mm512_loadu_ps((const float*)(x + i_offset)); - alpha_content = _mm512_add_ps(w_content, x_content); - - // Save the alpha value. - _mm512_storeu_ps(reinterpret_cast(alpha_value + i_offset), - alpha_content); -#else - // Declare the variable for the content of weights, input and alpha - // values. - __m256 w_content, x_content, alpha_content; - - // Load the relevant data into the variables from un-aligned address. - w_content = _mm256_loadu_ps((const float*)(w + i_offset)); - x_content = _mm256_loadu_ps((const float*)(x + i_offset)); - alpha_content = _mm256_add_ps(w_content, x_content); - - // Save the alpha value. - _mm256_storeu_ps(reinterpret_cast(alpha_value + i_offset), - alpha_content); -#endif - i_offset += step_size; - if (i == steps - 1) { - if (remain > 0) { - i_offset += last_offset; - } else { - break; - } - } - } - - // Use the column-major strategy to get the location of maximum score. - size_t seq_offset = 0; - for (size_t k = 1; k < seq_len; ++k) { - size_t j_offset = 0; - for (size_t j = 0; j <= steps; ++j) { -#ifdef __AVX512F__ - // Initialize the variables of maximum score and location. - __m512 max_score = _mm512_set1_ps(-std::numeric_limits::max()); - __m512i max_j = _mm512_setzero_si512(); -#else - // Initialize the variables of maximum score and location. - __m256 max_score = _mm256_set1_ps(-std::numeric_limits::max()); - __m256i max_j = _mm256_set1_epi32(0); -#endif - // Calculate the offset of transition_weights. - size_t trans_offset = state_trans_base_idx * tag_num + j_offset; - for (size_t i = 0; i < tag_num; ++i) { -#ifdef __AVX512F__ - // Initalize the content of alpha variable with related offset. - __m512 alpha_content = - _mm512_set1_ps(*(const float*)(alpha_value + seq_offset + i)); - // Obtain the content of weights from un-aligned address. - __m512 w_content = - _mm512_loadu_ps((const float*)(w + trans_offset)); - - __m512 score_v = _mm512_add_ps(alpha_content, w_content); - - __mmask16 mask = _mm512_cmp_ps_mask(score_v, max_score, _CMP_GT_OS); - - // According to the mask value, it update the index of the max_score - // location. - max_j = _mm512_mask_set1_epi32(max_j, mask, i); - - // Update the max_score value. - max_score = _mm512_max_ps(max_score, score_v); -#else - // Initalize the content of alpha variable with related offset. - __m256 alpha_content = _mm256_broadcast_ss( - (const float*)(alpha_value + seq_offset + i)); - // Obtain the content of weights from un-aligned address. - __m256 w_content = - _mm256_loadu_ps((const float*)(w + trans_offset)); - __m256 score_v = _mm256_add_ps(alpha_content, w_content); - - __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); - -#ifdef __AVX2__ - // According to the mask value, it update the index of the max_score - // location. - max_j = _mm256_or_si256( - _mm256_andnot_si256((__m256i)mask, max_j), - _mm256_and_si256((__m256i)mask, _mm256_set1_epi32(i))); -#else - __m128i lo_max_j = _mm256_extractf128_si256(max_j, 0); - __m128i hi_max_j = _mm256_extractf128_si256(max_j, 1); - __m128i lo_mask = _mm256_extractf128_si256((__m256i)mask, 0); - __m128i hi_mask = _mm256_extractf128_si256((__m256i)mask, 1); - - lo_max_j = _mm_andnot_si128(lo_mask, lo_max_j); - hi_max_j = _mm_andnot_si128(hi_mask, hi_max_j); - lo_mask = _mm_and_si128(lo_mask, _mm_set1_epi32(i)); - hi_mask = _mm_and_si128(hi_mask, _mm_set1_epi32(i)); - - lo_max_j = _mm_or_si128(lo_mask, lo_max_j); - hi_max_j = _mm_or_si128(hi_mask, hi_max_j); - - // According to the mask value, it update the index of the max_score - // location. - max_j = _mm256_insertf128_si256(max_j, lo_max_j, 0); - max_j = _mm256_insertf128_si256(max_j, hi_max_j, 1); -#endif - - // Update the max_score value. - max_score = _mm256_max_ps(max_score, score_v); -#endif - trans_offset += tag_num; - } - -#ifdef __AVX512F__ - // Update the alpha and track values. - __m512 x_content = _mm512_loadu_ps( - (const float*)(x + seq_offset + tag_num + j_offset)); - max_score = _mm512_add_ps(max_score, x_content); - _mm512_storeu_ps(reinterpret_cast(alpha_value + seq_offset + - tag_num + j_offset), - max_score); - _mm512_storeu_si512( - reinterpret_cast<__m512i*>(track_value + seq_offset + tag_num + - j_offset), - max_j); -#else - // Update the alpha and track values. - __m256 x_content = _mm256_loadu_ps( - (const float*)(x + seq_offset + tag_num + j_offset)); - max_score = _mm256_add_ps(max_score, x_content); - _mm256_storeu_ps(reinterpret_cast(alpha_value + seq_offset + - tag_num + j_offset), - max_score); - _mm256_storeu_si256( - reinterpret_cast<__m256i*>(track_value + seq_offset + tag_num + - j_offset), - max_j); -#endif - - // Calculate the offset of next step - j_offset += step_size; - if (j == steps - 1) { - if (remain > 0) { - j_offset += last_offset; - } else { - break; - } - } - } - - seq_offset += tag_num; - } - } else { - for (size_t i = 0; i < tag_num; ++i) alpha_value[i] = w[i] + x[i]; - - for (size_t k = 1; k < seq_len; ++k) { - for (size_t i = 0; i < tag_num; ++i) { - T max_score = -std::numeric_limits::max(); - int max_j = 0; - for (size_t j = 0; j < tag_num; ++j) { - T score = alpha_value[(k - 1) * tag_num + j] + - w[(j + state_trans_base_idx) * tag_num + i]; - if (score > max_score) { - max_score = score; - max_j = j; - } - } - - alpha_value[k * tag_num + i] = max_score + x[k * tag_num + i]; - track_value[k * tag_num + i] = max_j; - } - } - } -#else - for (size_t i = 0; i < tag_num; ++i) alpha_value[i] = w[i] + x[i]; - - for (size_t k = 1; k < seq_len; ++k) { - for (size_t i = 0; i < tag_num; ++i) { - T max_score = -std::numeric_limits::max(); - int max_j = 0; - for (size_t j = 0; j < tag_num; ++j) { - T score = alpha_value[(k - 1) * tag_num + j] + - w[(j + state_trans_base_idx) * tag_num + i]; - if (score > max_score) { - max_score = score; - max_j = j; - } - } - - alpha_value[k * tag_num + i] = max_score + x[k * tag_num + i]; - track_value[k * tag_num + i] = max_j; - } - } - -#endif + const auto& ker = math::jitkernel::KernelPool::Instance() + .template Get>( + static_cast(tag_num)); + ker->Compute(static_cast(seq_len), x, w, alpha_value, track_value); T max_score = -std::numeric_limits::max(); int max_i = 0; for (size_t i = 0; i < tag_num; ++i) { diff --git a/paddle/fluid/operators/cross_entropy_op.cu b/paddle/fluid/operators/cross_entropy_op.cu index 30dbd5bd3d39dd2992c3dd91364003bb7715a2eb..fcd34383a85f6984a8f27ce0625364f8fd5e31d6 100644 --- a/paddle/fluid/operators/cross_entropy_op.cu +++ b/paddle/fluid/operators/cross_entropy_op.cu @@ -13,12 +13,17 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/cross_entropy_op.h" +#include "paddle/fluid/platform/float16.h" +namespace plat = paddle::platform; namespace ops = paddle::operators; using CUDACtx = paddle::platform::CUDADeviceContext; REGISTER_OP_CUDA_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, - ops::CrossEntropyOpKernel); -REGISTER_OP_CUDA_KERNEL(cross_entropy_grad, - ops::CrossEntropyGradientOpKernel, - ops::CrossEntropyGradientOpKernel); + ops::CrossEntropyOpKernel, + ops::CrossEntropyOpKernel); + +REGISTER_OP_CUDA_KERNEL( + cross_entropy_grad, ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel); diff --git a/paddle/fluid/operators/delete_var_op.cc b/paddle/fluid/operators/delete_var_op.cc index d7a9bfbc437dbf4c723b9c87ff62ec6b62c38638..89416f7ab5d07ddac5b540b9bb361f831c1ef360 100644 --- a/paddle/fluid/operators/delete_var_op.cc +++ b/paddle/fluid/operators/delete_var_op.cc @@ -32,6 +32,11 @@ class DeleteVarOp : public framework::OperatorBase { } }; +class DeleteVarOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override {} +}; + class DeleteVarOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -48,4 +53,5 @@ It should not be configured by users directly. REGISTER_OPERATOR(delete_var, paddle::operators::DeleteVarOp, paddle::framework::EmptyGradOpMaker, - paddle::operators::DeleteVarOpInfoMaker); + paddle::operators::DeleteVarOpInfoMaker, + paddle::operators::DeleteVarOpShapeInference); diff --git a/paddle/fluid/operators/detection/generate_proposal_labels_op.cc b/paddle/fluid/operators/detection/generate_proposal_labels_op.cc index 339e63a2be13cec7b641b3a9eeb083480fc4b86e..fddd6884017c35112ba48f245759f5d846b55f9a 100644 --- a/paddle/fluid/operators/detection/generate_proposal_labels_op.cc +++ b/paddle/fluid/operators/detection/generate_proposal_labels_op.cc @@ -439,31 +439,88 @@ class GenerateProposalLabelsKernel : public framework::OpKernel { class GenerateProposalLabelsOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - // TODO(buxingyuan): Add Document - AddInput("RpnRois", "RpnRois."); - AddInput("GtClasses", "GtClasses."); - AddInput("IsCrowd", "IsCrowd."); - AddInput("GtBoxes", "GtBoxes."); - AddInput("ImInfo", "ImInfo."); - - AddOutput("Rois", "Rois."); - AddOutput("LabelsInt32", "LabelsInt32."); - AddOutput("BboxTargets", "BboxTargets."); - AddOutput("BboxInsideWeights", "BboxInsideWeights."); - AddOutput("BboxOutsideWeights", "BboxOutsideWeights."); - - AddAttr("batch_size_per_im", "batch_size_per_im"); - AddAttr("fg_fraction", "fg_fraction"); - AddAttr("fg_thresh", "fg_thresh"); - AddAttr("bg_thresh_hi", "bg_thresh_hi"); - AddAttr("bg_thresh_lo", "bg_thresh_lo"); - AddAttr>("bbox_reg_weights", "bbox_reg_weights"); - AddAttr("class_nums", "class_nums"); - AddAttr("use_random", "use_random").SetDefault(true); + AddInput( + "RpnRois", + "(LoDTensor), This input is a 2D LoDTensor with shape [N, 4]. " + "N is the number of the GenerateProposalOp's output, " + "each element is a bounding box with [xmin, ymin, xmax, ymax] format."); + AddInput("GtClasses", + "(LoDTensor), This input is a 2D LoDTensor with shape [M, 1]. " + "M is the number of groundtruth, " + "each element is a class label of groundtruth."); + AddInput( + "IsCrowd", + "(LoDTensor), This input is a 2D LoDTensor with shape [M, 1]. " + "M is the number of groundtruth, " + "each element is a flag indicates whether a groundtruth is crowd."); + AddInput( + "GtBoxes", + "(LoDTensor), This input is a 2D LoDTensor with shape [M, 4]. " + "M is the number of groundtruth, " + "each element is a bounding box with [xmin, ymin, xmax, ymax] format."); + AddInput("ImInfo", + "(Tensor), This input is a 2D Tensor with shape [B, 3]. " + "B is the number of input images, " + "each element consists of im_height, im_width, im_scale."); + + AddOutput( + "Rois", + "(LoDTensor), This output is a 2D LoDTensor with shape [P, 4]. " + "P usuall equal to batch_size_per_im * batch_size, " + "each element is a bounding box with [xmin, ymin, xmax, ymax] format."); + AddOutput("LabelsInt32", + "(LoDTensor), This output is a 2D LoDTensor with shape [P], " + "each element repersents a class label of a roi"); + AddOutput("BboxTargets", + "(LoDTensor), This output is a 2D LoDTensor with shape [P, 4 * " + "class_nums], " + "each element repersents a box label of a roi"); + AddOutput( + "BboxInsideWeights", + "(LoDTensor), This output is a 2D LoDTensor with shape [P, 4 * " + "class_nums], " + "each element indicates whether a box should contribute to loss."); + AddOutput( + "BboxOutsideWeights", + "(LoDTensor), This output is a 2D LoDTensor with shape [P, 4 * " + "class_nums], " + "each element indicates whether a box should contribute to loss."); + + AddAttr("batch_size_per_im", "Batch size of rois per images."); + AddAttr("fg_fraction", + "Foreground fraction in total batch_size_per_im."); + AddAttr( + "fg_thresh", + "Overlap threshold which is used to chose foreground sample."); + AddAttr("bg_thresh_hi", + "Overlap threshold upper bound which is used to chose " + "background sample."); + AddAttr("bg_thresh_lo", + "Overlap threshold lower bound which is used to chose " + "background sample."); + AddAttr>("bbox_reg_weights", "Box regression weights."); + AddAttr("class_nums", "Class number."); + AddAttr( + "use_random", + "Use random sampling to choose foreground and background boxes.") + .SetDefault(true); AddComment(R"DOC( -Generate Proposals Labels Operator. -)DOC"); +This operator can be, for given the GenerateProposalOp output bounding boxes and groundtruth, +to sample foreground boxes and background boxes, and compute loss target. + +RpnRois is the output boxes of RPN and was processed by generate_proposal_op, these boxes +were combined with groundtruth boxes and sampled according to batch_size_per_im and fg_fraction, +If an instance with a groundtruth overlap greater than fg_thresh, then it was considered as a foreground sample. +If an instance with a groundtruth overlap greater than bg_thresh_lo and lower than bg_thresh_hi, +then it was considered as a background sample. +After all foreground and background boxes are chosen (so called Rois), +then we apply random sampling to make sure +the number of foreground boxes is no more than batch_size_per_im * fg_fraction. + +For each box in Rois, we assign the classification (class label) and regression targets (box label) to it. +Finally BboxInsideWeights and BboxOutsideWeights are used to specify whether it would contribute to training loss. + )DOC"); } }; diff --git a/paddle/fluid/operators/distributed/grpc_client.cc b/paddle/fluid/operators/distributed/grpc_client.cc index f5d5627815c7320dad5051b0f7d95b8ec6703687..be5c20ad2e4b53e0ff98561b92543b03298381d9 100644 --- a/paddle/fluid/operators/distributed/grpc_client.cc +++ b/paddle/fluid/operators/distributed/grpc_client.cc @@ -79,7 +79,7 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, auto* var = p_scope->FindVar(var_name_val); ::grpc::ByteBuffer req; - SerializeToByteBuffer(var_name_val, var, *p_ctx, &req); + SerializeToByteBuffer(var_name_val, var, *p_ctx, &req, "", trainer_id_); VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; @@ -105,7 +105,10 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, void ProcGetResponse(const VarHandle& var_h, const ::grpc::ByteBuffer& ret_msg) { framework::Variable* outvar = nullptr; - DeserializeFromByteBuffer(ret_msg, *var_h.ctx(), var_h.scope(), &outvar); + // get response's trainer_id is not used + int trainer_id; + DeserializeFromByteBuffer(ret_msg, *var_h.ctx(), var_h.scope(), &outvar, + &trainer_id); } template @@ -135,6 +138,7 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep, // prepare input sendrecv::VariableMessage req; req.set_varname(var_name_val); + req.set_trainer_id(trainer_id_); ::grpc::ByteBuffer buf; RequestToByteBuffer(req, &buf); diff --git a/paddle/fluid/operators/distributed/grpc_serde.cc b/paddle/fluid/operators/distributed/grpc_serde.cc index 215405e694903b74630aa9bd34ec5bbfbb651633..f27b70a5a3dd2927b51a95af7bd1b84a6e232f86 100644 --- a/paddle/fluid/operators/distributed/grpc_serde.cc +++ b/paddle/fluid/operators/distributed/grpc_serde.cc @@ -41,13 +41,14 @@ static void SerializeDestroyCallback(void* payload) { void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, - ::grpc::ByteBuffer* msg, - const std::string& out_name) { + ::grpc::ByteBuffer* msg, const std::string& out_name, + const int trainer_id) { platform::RecordRPCEvent record_event("serial", &ctx); VarMsg request; TensorPayload* payload = nullptr; request.set_varname(name); + request.set_trainer_id(trainer_id); // Note: normally the profiler is enabled in 1 trainer, hence only // 1 trainer returns true for ShouldSendProfileState(). It tells PS // servers the trainer's profiling state so that PS can follow the @@ -141,11 +142,12 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, const framework::Scope* scope, - framework::Variable** var) { + framework::Variable** var, int* trainer_id) { 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(); + *trainer_id = resp.GetTrainerId(); } } // namespace distributed diff --git a/paddle/fluid/operators/distributed/grpc_serde.h b/paddle/fluid/operators/distributed/grpc_serde.h index 450c41dcd6b1bf9a33d3bbef3a1c94a2f83ff322..7ec489e961630747ba00e68ad3603cacbb1aa485 100644 --- a/paddle/fluid/operators/distributed/grpc_serde.h +++ b/paddle/fluid/operators/distributed/grpc_serde.h @@ -38,12 +38,13 @@ typedef void (*DestroyCallback)(void*); void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, ::grpc::ByteBuffer* msg, - const std::string& out_varname = std::string()); + const std::string& out_varname = std::string(), + const int trainer_id = 0); void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, const framework::Scope* scope, - framework::Variable** var); + framework::Variable** var, int* trainer_id); } // namespace distributed } // namespace operators diff --git a/paddle/fluid/operators/distributed/grpc_server.cc b/paddle/fluid/operators/distributed/grpc_server.cc index 8edb00276df3ade1b320fbf2873e8b54ff3e1464..eb9e36029c01fb38ca4438578190dd7895182ea1 100644 --- a/paddle/fluid/operators/distributed/grpc_server.cc +++ b/paddle/fluid/operators/distributed/grpc_server.cc @@ -102,9 +102,10 @@ class RequestSend final : public RequestBase { auto scope = request_->GetMutableLocalScope(); auto invar = request_->GetVar(); + int trainer_id = request_->GetTrainerId(); framework::Variable* outvar = nullptr; - request_handler_->Handle(varname, scope, invar, &outvar); + request_handler_->Handle(varname, scope, invar, &outvar, trainer_id); Finish(reply_, &responder_); } @@ -133,13 +134,14 @@ class RequestGet final : public RequestBase { void Process() override { // proc request. std::string varname = request_.varname(); + int trainer_id = request_.trainer_id(); VLOG(4) << "RequestGet " << varname; auto scope = request_handler_->scope(); auto invar = scope->FindVar(varname); framework::Variable* outvar = nullptr; - request_handler_->Handle(varname, scope, invar, &outvar); + request_handler_->Handle(varname, scope, invar, &outvar, trainer_id); if (outvar) { SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(), @@ -179,6 +181,7 @@ class RequestPrefetch final : public RequestBase { // prefetch process... std::string in_var_name = request_->Varname(); std::string out_var_name = request_->OutVarname(); + int trainer_id = request_->GetTrainerId(); VLOG(4) << "RequestPrefetch, in_var_name: " << in_var_name << " out_var_name: " << out_var_name; @@ -187,7 +190,8 @@ class RequestPrefetch final : public RequestBase { // out var must be created in local scope! framework::Variable* outvar = scope->Var(out_var_name); - request_handler_->Handle(in_var_name, scope, invar, &outvar, out_var_name); + request_handler_->Handle(in_var_name, scope, invar, &outvar, trainer_id, + out_var_name); SerializeToByteBuffer(out_var_name, outvar, *request_handler_->dev_ctx(), &reply_); @@ -225,12 +229,13 @@ class RequestCheckpointNotify final : public RequestBase { std::string checkpoint_notify = request_->Varname(); std::string checkpoint_dir = request_->OutVarname(); + int trainer_id = request_->GetTrainerId(); VLOG(4) << "RequestCheckpointNotify notify: " << checkpoint_notify << ", dir: " << checkpoint_dir; request_handler_->Handle(checkpoint_notify, scope, nullptr, nullptr, - checkpoint_dir); + trainer_id, checkpoint_dir); Finish(reply_, &responder_); } diff --git a/paddle/fluid/operators/distributed/grpc_variable_response.cc b/paddle/fluid/operators/distributed/grpc_variable_response.cc index 34d47f3ec0f3025109447b66078b724607d2953a..d6d219d4369ba785e5c369538d4a18dc682952c1 100644 --- a/paddle/fluid/operators/distributed/grpc_variable_response.cc +++ b/paddle/fluid/operators/distributed/grpc_variable_response.cc @@ -286,13 +286,21 @@ int GRPCVariableResponse::Parse(Source* source) { platform::EnableProfiler(platform::ProfilerState::kCPU); } else if (profiling == platform::kDisableProfiler && platform::IsProfileEnabled()) { - // TODO(panyx0718): Should we allow to customize file dir. platform::DisableProfiler( platform::EventSortingKey::kDefault, - string::Sprintf("/tmp/profile_ps_%lld", listener_id)); + string::Sprintf("%s_%lld", FLAGS_rpc_server_profile_path, + listener_id)); } break; } + case sendrecv::VariableMessage::kTrainerIdFieldNumber: { + uint64_t trainer_id = 0; + if (!input.ReadVarint64(&trainer_id)) { + return tag; + } + meta_.set_trainer_id(trainer_id); + break; + } default: { // Unknown tag, return unknown error. return -1; diff --git a/paddle/fluid/operators/distributed/request_handler.h b/paddle/fluid/operators/distributed/request_handler.h index 5be7095acd3c5ac6f880a8a26c246f60a93643b5..3c1db147098055e9974c9dc607266cdaf2e43dae 100644 --- a/paddle/fluid/operators/distributed/request_handler.h +++ b/paddle/fluid/operators/distributed/request_handler.h @@ -190,6 +190,7 @@ class RequestHandler { // } virtual bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") = 0; protected: diff --git a/paddle/fluid/operators/distributed/request_handler_impl.cc b/paddle/fluid/operators/distributed/request_handler_impl.cc index 849e412504eb9180b746db65fd4fa353ed0c05a1..025528fe70b8f4d353ab92f29b1bd71c77cf7850 100644 --- a/paddle/fluid/operators/distributed/request_handler_impl.cc +++ b/paddle/fluid/operators/distributed/request_handler_impl.cc @@ -36,6 +36,7 @@ bool RequestSendHandler::Handle(const std::string& varname, framework::Scope* scope, framework::Variable* invar, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name) { VLOG(4) << "RequestSendHandler:" << varname; @@ -50,7 +51,6 @@ bool RequestSendHandler::Handle(const std::string& varname, // Async if (!sync_mode_) { VLOG(3) << "async process var: " << varname; - rpc_server_->Profiler().OneStep(); try { executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(), scope); @@ -76,6 +76,7 @@ bool RequestGetHandler::Handle(const std::string& varname, framework::Scope* scope, framework::Variable* invar, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name) { VLOG(4) << "RequestGetHandler:" << varname; if (sync_mode_) { @@ -88,6 +89,19 @@ bool RequestGetHandler::Handle(const std::string& varname, } } else { if (varname != FETCH_BARRIER_MESSAGE && varname != COMPLETE_MESSAGE) { + if (enable_dc_asgd_) { + // NOTE: the format is determined by distributed_transpiler.py + std::string param_bak_name = + string::Sprintf("%s.trainer_%d_bak", varname, trainer_id); + VLOG(3) << "getting " << param_bak_name << " trainer_id " << trainer_id; + auto var = scope_->FindVar(varname); + auto t_orig = var->Get(); + auto param_bak = scope_->Var(param_bak_name); + auto t = param_bak->GetMutable(); + t->mutable_data(dev_ctx_->GetPlace(), t_orig.type()); + VLOG(3) << "copying " << varname << " to " << param_bak_name; + framework::TensorCopy(t_orig, dev_ctx_->GetPlace(), t); + } *outvar = scope_->FindVar(varname); } } @@ -98,6 +112,7 @@ bool RequestPrefetchHandler::Handle(const std::string& varname, framework::Scope* scope, framework::Variable* invar, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name) { VLOG(4) << "RequestPrefetchHandler " << varname; @@ -113,6 +128,7 @@ bool RequestCheckpointHandler::Handle(const std::string& varname, framework::Scope* scope, framework::Variable* invar, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name) { PADDLE_ENFORCE( checkpoint_notify_id != -1, diff --git a/paddle/fluid/operators/distributed/request_handler_impl.h b/paddle/fluid/operators/distributed/request_handler_impl.h index 8be5b21bb89a580f4091de19186fd2d7e5802478..c1afda9dd2445e492d8b93659c9ff13e6e1030b8 100644 --- a/paddle/fluid/operators/distributed/request_handler_impl.h +++ b/paddle/fluid/operators/distributed/request_handler_impl.h @@ -36,20 +36,34 @@ namespace distributed { class RequestSendHandler final : public RequestHandler { public: - explicit RequestSendHandler(bool sync_mode) : RequestHandler(sync_mode) {} + explicit RequestSendHandler(bool sync_mode, bool enable_dc_asgd = false) + : RequestHandler(sync_mode) { + enable_dc_asgd_ = enable_dc_asgd; + } virtual ~RequestSendHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") override; + + private: + bool enable_dc_asgd_; }; class RequestGetHandler final : public RequestHandler { public: - explicit RequestGetHandler(bool sync_mode) : RequestHandler(sync_mode) {} + explicit RequestGetHandler(bool sync_mode, bool enable_dc_asgd = false) + : RequestHandler(sync_mode) { + enable_dc_asgd_ = enable_dc_asgd; + } virtual ~RequestGetHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") override; + + private: + bool enable_dc_asgd_; }; class RequestPrefetchHandler final : public RequestHandler { @@ -58,6 +72,7 @@ class RequestPrefetchHandler final : public RequestHandler { virtual ~RequestPrefetchHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") override; }; @@ -70,6 +85,7 @@ class RequestCheckpointHandler final : public RequestHandler { virtual ~RequestCheckpointHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") override; private: diff --git a/paddle/fluid/operators/distributed/rpc_client.cc b/paddle/fluid/operators/distributed/rpc_client.cc index b5ec9fe5367beb97b3cc7298102deff1e8ca4ec9..390e9af0f38c920f39e8cd3e1b3d28fd89b239fe 100644 --- a/paddle/fluid/operators/distributed/rpc_client.cc +++ b/paddle/fluid/operators/distributed/rpc_client.cc @@ -24,6 +24,7 @@ namespace distributed { std::once_flag RPCClient::init_flag_; std::unique_ptr RPCClient::rpc_client_(nullptr); +int RPCClient::trainer_id_ = 0; } // namespace distributed } // namespace operators diff --git a/paddle/fluid/operators/distributed/rpc_client.h b/paddle/fluid/operators/distributed/rpc_client.h index 3539ee5e459d6dfe0b6510806464bcc6817910bb..1983802e49506c79041112ac87d429e4c084ddfd 100644 --- a/paddle/fluid/operators/distributed/rpc_client.h +++ b/paddle/fluid/operators/distributed/rpc_client.h @@ -72,14 +72,15 @@ class RPCClient { virtual bool Wait() = 0; template - static RPCClient* GetInstance() { - std::call_once(init_flag_, &RPCClient::Init); + static RPCClient* GetInstance(int trainer_id) { + std::call_once(init_flag_, &RPCClient::Init, trainer_id); return rpc_client_.get(); } // Init is called by GetInstance. template - static void Init() { + static void Init(int trainer_id) { + trainer_id_ = trainer_id; if (rpc_client_.get() == nullptr) { rpc_client_.reset(new T()); rpc_client_->InitImpl(); @@ -88,6 +89,8 @@ class RPCClient { protected: virtual void InitImpl() {} + // each trainer have exact one trainer id, it should be static + static int trainer_id_; private: static std::once_flag init_flag_; diff --git a/paddle/fluid/operators/distributed/rpc_server.cc b/paddle/fluid/operators/distributed/rpc_server.cc index 084480ae48b8b9267ade1a840f6a70519cb28e48..3e30ed4ac86bd2cb3f7c4301163e54a947c3d5b4 100644 --- a/paddle/fluid/operators/distributed/rpc_server.cc +++ b/paddle/fluid/operators/distributed/rpc_server.cc @@ -20,42 +20,10 @@ #include "paddle/fluid/operators/distributed/rpc_server.h" #include "paddle/fluid/platform/profiler.h" -DEFINE_int32(rpc_server_profile_period, 0, - "the period of listen_and_serv to do profile"); -DEFINE_string(rpc_server_profile_path, "/dev/null", - "the profile log file path"); - namespace paddle { namespace operators { namespace distributed { -RPCServerProfiler::RPCServerProfiler(int profile_period, - const std::string& profile_log_path) - : profile_period_(profile_period), profile_log_path_(profile_log_path) { - step_ = 0; -} - -void RPCServerProfiler::OneStep() { - PADDLE_ENFORCE_LE(step_, profile_period_, - "step_ should not be larger then " - "profile_period_"); - if (profile_period_ <= 0) { - return; - } - - if (step_ == 0) { - auto pf_state = paddle::platform::ProfilerState::kCPU; - paddle::platform::EnableProfiler(pf_state); - } - if (step_ == profile_period_) { - paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kTotal, - profile_log_path_); - step_ = 0; - } else { - step_++; - } -} - void RPCServer::ShutDown() { LOG(INFO) << "RPCServer ShutDown "; ShutDownImpl(); diff --git a/paddle/fluid/operators/distributed/rpc_server.h b/paddle/fluid/operators/distributed/rpc_server.h index f3e61e1575ced0b9ffbad23e6973121daca9751b..c78c5007a7f262f15305b6c284e8c4fbddef42a0 100644 --- a/paddle/fluid/operators/distributed/rpc_server.h +++ b/paddle/fluid/operators/distributed/rpc_server.h @@ -23,30 +23,14 @@ #include "paddle/fluid/operators/distributed/request_handler.h" -DECLARE_int32(rpc_server_profile_period); -DECLARE_string(rpc_server_profile_path); - namespace paddle { namespace operators { namespace distributed { -class RPCServerProfiler { - public: - RPCServerProfiler(int profile_period, const std::string& profile_log_path); - void OneStep(); - - private: - const int profile_period_; - std::string profile_log_path_; - int step_; -}; - class RPCServer { public: explicit RPCServer(const std::string& address, int client_num) : cur_cond_(0), - profiler_(FLAGS_rpc_server_profile_period, - FLAGS_rpc_server_profile_path), bind_address_(address), exit_flag_(false), selected_port_(0), @@ -86,7 +70,6 @@ class RPCServer { void Complete(); void ResetBarrierCounter(); - RPCServerProfiler& Profiler() { return profiler_; } bool NeedResetAllVars(); @@ -101,7 +84,6 @@ class RPCServer { std::unordered_map rpc_cond_map_; std::atomic cur_cond_; std::condition_variable rpc_cond_; - RPCServerProfiler profiler_; protected: std::string bind_address_; diff --git a/paddle/fluid/operators/distributed/rpc_server_test.cc b/paddle/fluid/operators/distributed/rpc_server_test.cc index d6176e1443d2a441af7878e5efe99796d486bb7a..c3dd459fc4e8c4bd304c09d7a3ed4f456c4dc69f 100644 --- a/paddle/fluid/operators/distributed/rpc_server_test.cc +++ b/paddle/fluid/operators/distributed/rpc_server_test.cc @@ -125,7 +125,7 @@ TEST(PREFETCH, CPU) { g_req_handler.reset(new distributed::RequestPrefetchHandler(true)); g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1)); distributed::RPCClient* client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(0); std::thread server_thread(StartServer, distributed::kRequestPrefetch); g_rpc_service->WaitServerReady(); @@ -165,7 +165,7 @@ TEST(COMPLETE, CPU) { g_req_handler.reset(new distributed::RequestSendHandler(true)); g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 2)); distributed::RPCClient* client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(0); PADDLE_ENFORCE(client != nullptr); std::thread server_thread(StartServer, distributed::kRequestSend); g_rpc_service->WaitServerReady(); diff --git a/paddle/fluid/operators/distributed/send_recv.proto.in b/paddle/fluid/operators/distributed/send_recv.proto.in index 8b0a09abe1d05dda10eda0030eb91cb9ca40683e..55820c980e8139625c1b589f9d2d68dfee74a212 100644 --- a/paddle/fluid/operators/distributed/send_recv.proto.in +++ b/paddle/fluid/operators/distributed/send_recv.proto.in @@ -79,6 +79,7 @@ message VariableMessage { // server stops profiling and generates a profile to /tmp/profile_ps_* // when profile switches from 1 to 2. int64 profile = 11; + int64 trainer_id = 12; } message VoidMessage {} diff --git a/paddle/fluid/operators/distributed/variable_response.cc b/paddle/fluid/operators/distributed/variable_response.cc index d24168745eec7a7c23f07e761a2129fd0adc004b..5b2be04e6a1656f79a8e2b3a6fa19c847e81b5cb 100644 --- a/paddle/fluid/operators/distributed/variable_response.cc +++ b/paddle/fluid/operators/distributed/variable_response.cc @@ -16,6 +16,9 @@ #include #include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +DEFINE_string(rpc_server_profile_path, "./profile_ps", + "the profile log file path"); + namespace paddle { namespace operators { namespace distributed { diff --git a/paddle/fluid/operators/distributed/variable_response.h b/paddle/fluid/operators/distributed/variable_response.h index 6aec52ca00f59a42ecca01da8df1680ce4eda432..4c7fcbbdfb305ce6b4fc9d1edd9738899b200ec6 100644 --- a/paddle/fluid/operators/distributed/variable_response.h +++ b/paddle/fluid/operators/distributed/variable_response.h @@ -27,6 +27,8 @@ #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/operators/distributed/send_recv.pb.h" +DECLARE_string(rpc_server_profile_path); + namespace paddle { namespace operators { namespace distributed { @@ -92,6 +94,8 @@ class VariableResponse { return scope_->FindVar(meta_.varname()); } + int GetTrainerId() { return static_cast(meta_.trainer_id()); } + protected: bool ReadRaw(::google::protobuf::io::CodedInputStream* input, const platform::DeviceContext& dev_ctx, platform::Place place, diff --git a/paddle/fluid/operators/elementwise_add_op.cu b/paddle/fluid/operators/elementwise_add_op.cu index dfff518f170b56d180b6883c363effb8dbd677b6..f9f5c66d34fa1d73db00173e493f9953b8579518 100644 --- a/paddle/fluid/operators/elementwise_add_op.cu +++ b/paddle/fluid/operators/elementwise_add_op.cu @@ -30,4 +30,5 @@ REGISTER_OP_CUDA_KERNEL( ops::ElementwiseAddGradKernel, ops::ElementwiseAddGradKernel, ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel); + ops::ElementwiseAddGradKernel, + ops::ElementwiseAddGradKernel); diff --git a/paddle/fluid/operators/elementwise_op_function.h b/paddle/fluid/operators/elementwise_op_function.h index 7c84a9d813948ab7347446872643c2e00823a5ad..93204216f947e5203863a3493005faa0c03ae4af 100644 --- a/paddle/fluid/operators/elementwise_op_function.h +++ b/paddle/fluid/operators/elementwise_op_function.h @@ -365,7 +365,7 @@ static __global__ void ElemwiseGradBroadcast1CUDAKernel( int j = blockIdx.x; int i = threadIdx.x; int tid = threadIdx.x; - T val = 0; + T val(0); do { int x_offset = i * w + j; @@ -433,7 +433,7 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel( int tid = threadIdx.x; int j = blockIdx.x; - T val = 0; + T val(0); int ttid = tid; while (true) { diff --git a/paddle/fluid/operators/fake_init_op.cc b/paddle/fluid/operators/fake_init_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..28ebdcb03ea83f3ec701106111a7cc5f0f7ed7dc --- /dev/null +++ b/paddle/fluid/operators/fake_init_op.cc @@ -0,0 +1,86 @@ +/* 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/data_type.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +class FakeInitInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of FakeInitOp should not be null."); + auto &shape = ctx->Attrs().Get>("shape"); + ctx->SetOutputDim("Out", framework::make_ddim(shape)); + } +}; + +class FakeInitOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &dev_place) const override { + framework::Tensor *tensor = nullptr; + + auto &out_var = *scope.FindVar(Output("Out")); + + if (out_var.IsType()) { + tensor = out_var.GetMutable(); + tensor->Resize(framework::make_ddim(Attr>("shape"))); + } else if (out_var.IsType()) { + tensor = out_var.GetMutable()->mutable_value(); + tensor->Resize(framework::make_ddim(Attr>("shape"))); + } else { + PADDLE_THROW( + "fake init op's output only" + "supports SelectedRows and LoDTensor"); + } + } +}; + +class FakeInitOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override {} +}; + +class FakeInitOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddAttr>("shape", + "(vector) The shape of the output"); + AddOutput("Out", + "(Tensor) Tensor of specified shape will be filled " + "with the specified value"); + AddComment(R"DOC( +FakeInit Operator. + +Init an variable but not alloc memory for it, it is used for init the +table parameter at trainer side in distributed lookup table. + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(fake_init, ops::FakeInitOp, ops::FakeInitInferShape, + ops::FakeInitOpMaker, paddle::framework::EmptyGradOpMaker, + ops::FakeInitOpVarTypeInference); diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc index 9d7ac7ab6194593747548fac3cefc8d4ed3058d8..8754856e140ed074782e6fccb8991571a12babab 100644 --- a/paddle/fluid/operators/fetch_barrier_op.cc +++ b/paddle/fluid/operators/fetch_barrier_op.cc @@ -37,7 +37,8 @@ class FetchBarrierOp : public framework::OperatorBase { const platform::Place& place) const override { std::vector eps = Attr>("endpoints"); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); @@ -61,6 +62,7 @@ This operator will send a send barrier signal to list_and_serv op, so that the Parameter Server would knew all variables have been sent. )DOC"); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr>("endpoints", "(string vector, default 127.0.0.1:6164)" "Server endpoints to send variables to.") diff --git a/paddle/fluid/operators/fill_constant_op.cc b/paddle/fluid/operators/fill_constant_op.cc index e04a68717b351ddb0be5a7e70aa9297e5eb0125f..252f313440296bd9e5eebf26f67b08bbe7decce8 100644 --- a/paddle/fluid/operators/fill_constant_op.cc +++ b/paddle/fluid/operators/fill_constant_op.cc @@ -24,7 +24,7 @@ class FillConstantInferShape : public framework::InferShapeBase { void operator()(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of FillConstantOp should not be null."); - auto &shape = ctx->Attrs().Get>("shape"); + auto &shape = ctx->Attrs().Get>("shape"); ctx->SetOutputDim("Out", framework::make_ddim(shape)); } }; @@ -47,10 +47,10 @@ class FillConstantOp : public framework::OperatorBase { if (out_var.IsType()) { tensor = out_var.GetMutable(); - tensor->Resize(framework::make_ddim(Attr>("shape"))); + tensor->Resize(framework::make_ddim(Attr>("shape"))); } else if (out_var.IsType()) { tensor = out_var.GetMutable()->mutable_value(); - tensor->Resize(framework::make_ddim(Attr>("shape"))); + tensor->Resize(framework::make_ddim(Attr>("shape"))); } else { PADDLE_THROW( "fill constant op's output only" @@ -83,7 +83,8 @@ class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker { "(int, default 5 (FP32)) " "Output data type") .SetDefault(framework::proto::VarType::FP32); - AddAttr>("shape", "(vector) The shape of the output"); + AddAttr>("shape", + "(vector) The shape of the output"); AddAttr("value", "(float, default 0) The value to be filled") .SetDefault(0.0f); AddAttr("force_cpu", diff --git a/paddle/fluid/operators/gather_op.cc b/paddle/fluid/operators/gather_op.cc index 089b541a0a61adb5efda6b2e027c913d5808dff0..f84ff206fffddef1030b7ed439e887bdfef342a6 100644 --- a/paddle/fluid/operators/gather_op.cc +++ b/paddle/fluid/operators/gather_op.cc @@ -102,7 +102,9 @@ REGISTER_OPERATOR(gather, ops::GatherOp, ops::GatherOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(gather_grad, ops::GatherGradOp); REGISTER_OP_CPU_KERNEL(gather, ops::GatherOpKernel, - ops::GatherOpKernel, ops::GatherOpKernel); + ops::GatherOpKernel, ops::GatherOpKernel, + ops::GatherOpKernel); REGISTER_OP_CPU_KERNEL(gather_grad, ops::GatherGradientOpKernel, + ops::GatherGradientOpKernel, ops::GatherGradientOpKernel, - ops::GatherGradientOpKernel); + ops::GatherGradientOpKernel); diff --git a/paddle/fluid/operators/gather_op.cu b/paddle/fluid/operators/gather_op.cu index 7e014dd1cb47ee0575308dc13ba7bc7617baebff..9f4aef08cd58e72ce344a640e6564b9e360ce169 100644 --- a/paddle/fluid/operators/gather_op.cu +++ b/paddle/fluid/operators/gather_op.cu @@ -61,5 +61,11 @@ class GatherGradOpCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel); -REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel, + ops::GatherOpCUDAKernel, + ops::GatherOpCUDAKernel, + ops::GatherOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel, + ops::GatherGradOpCUDAKernel, + ops::GatherGradOpCUDAKernel, + ops::GatherGradOpCUDAKernel); diff --git a/paddle/fluid/operators/gaussian_random_op.cc b/paddle/fluid/operators/gaussian_random_op.cc index 1488aab1926b5b4ba7bceed582700f5a11fc6c93..c70d5b8bc7569c38cbc003aca7c62dc503df11cf 100644 --- a/paddle/fluid/operators/gaussian_random_op.cc +++ b/paddle/fluid/operators/gaussian_random_op.cc @@ -52,7 +52,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of GaussianRandomOp should not be null."); - auto shape = ctx->Attrs().Get>("shape"); + auto shape = ctx->Attrs().Get>("shape"); std::vector temp; temp.reserve(shape.size()); for (auto dim : shape) { @@ -88,9 +88,9 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddOutput("Out", "Output matrix of gaussian random op"); - AddAttr>("shape", - "(vector) " - "The dimension of random tensor."); + AddAttr>("shape", + "(vector) " + "The dimension of random tensor."); AddAttr("mean", "(float, default 0.0) " "mean of random tensor.") diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc index 697c239e59d158428ae9ba9f7feded19637dff28..ef574ccdf48dcf6074a777bcb7667b114415674c 100644 --- a/paddle/fluid/operators/gen_nccl_id_op.cc +++ b/paddle/fluid/operators/gen_nccl_id_op.cc @@ -61,7 +61,7 @@ class GenNCCLIdOp : public framework::OperatorBase { std::vector endpoint_list = Attr>("endpoint_list"); distributed::RPCClient* client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(0); for (auto& ep : endpoint_list) { VLOG(3) << "sending nccl id to " << ep; diff --git a/paddle/fluid/operators/grid_sampler_cudnn_op.cu.cc b/paddle/fluid/operators/grid_sampler_cudnn_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..7cde7ca462fda9ae6ace7755af0a432afee28bba --- /dev/null +++ b/paddle/fluid/operators/grid_sampler_cudnn_op.cu.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. */ + +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/cudnn_helper.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; +using ScopedTensorDescriptor = platform::ScopedTensorDescriptor; +using DataLayout = platform::DataLayout; +using ScopedSpatialTransformerDescriptor = + platform::ScopedSpatialTransformerDescriptor; +template +using CudnnDataType = platform::CudnnDataType; + +template +class CUDNNGridSampleOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace"); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto* input = ctx.Input("X"); + auto* grid = ctx.Input("Grid"); + auto* output = ctx.Output("Output"); + + int n = input->dims()[0]; + int c = input->dims()[1]; + int h = input->dims()[2]; + int w = input->dims()[3]; + const int size[4] = {n, c, h, w}; + + const T* input_data = input->data(); + const T* grid_data = grid->data(); + T* output_data = output->mutable_data({n, c, h, w}, ctx.GetPlace()); + + ScopedSpatialTransformerDescriptor st_desc; + cudnnSpatialTransformerDescriptor_t cudnn_st_desc = + st_desc.descriptor(4, size); + + ScopedTensorDescriptor input_desc; + ScopedTensorDescriptor output_desc; + cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(input->dims())); + cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(output->dims())); + + CUDNN_ENFORCE(platform::dynload::cudnnSpatialTfSamplerForward( + handle, cudnn_st_desc, CudnnDataType::kOne(), cudnn_input_desc, + input_data, grid_data, CudnnDataType::kZero(), cudnn_output_desc, + output_data)); + } +}; + +template +class CUDNNGridSampleGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace"); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto* input = ctx.Input("X"); + auto* grid = ctx.Input("Grid"); + auto* output_grad = ctx.Input(framework::GradVarName("Output")); + auto* input_grad = ctx.Output(framework::GradVarName("X")); + auto* grid_grad = ctx.Output(framework::GradVarName("Grid")); + + auto output_grad_dims = output_grad->dims(); + const int n = output_grad_dims[0]; + const int c = output_grad_dims[1]; + const int h = output_grad_dims[2]; + const int w = output_grad_dims[3]; + const int size[4] = {n, c, h, w}; + + ScopedSpatialTransformerDescriptor st_dest; + cudnnSpatialTransformerDescriptor_t cudnn_st_dest = + st_dest.descriptor(4, size); + + const T* input_data = input->data(); + const T* grid_data = grid->data(); + const T* output_grad_data = output_grad->data(); + T* input_grad_data = + input_grad->mutable_data(output_grad_dims, ctx.GetPlace()); + T* grid_grad_data = + grid_grad->mutable_data({n, h, w, 2}, ctx.GetPlace()); + + ScopedTensorDescriptor input_desc; + ScopedTensorDescriptor input_grad_desc; + ScopedTensorDescriptor output_grad_desc; + cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(input->dims())); + cudnnTensorDescriptor_t cudnn_input_grad_desc = + input_grad_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(input_grad->dims())); + cudnnTensorDescriptor_t cudnn_output_grad_desc = + output_grad_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(output_grad->dims())); + + CUDNN_ENFORCE(platform::dynload::cudnnSpatialTfSamplerBackward( + handle, cudnn_st_dest, CudnnDataType::kOne(), cudnn_input_desc, + input_data, CudnnDataType::kZero(), cudnn_input_grad_desc, + input_grad_data, CudnnDataType::kOne(), cudnn_output_grad_desc, + output_grad_data, grid_data, CudnnDataType::kZero(), + grid_grad_data)); + } +}; + +} // namespace operators +} // namespace paddle + +namespace plat = paddle::platform; +REGISTER_OP_KERNEL(grid_sampler, CUDNN, plat::CUDAPlace, + paddle::operators::CUDNNGridSampleOpKernel, + paddle::operators::CUDNNGridSampleOpKernel); +REGISTER_OP_KERNEL(grid_sampler_grad, CUDNN, plat::CUDAPlace, + paddle::operators::CUDNNGridSampleGradOpKernel, + paddle::operators::CUDNNGridSampleGradOpKernel); diff --git a/paddle/fluid/operators/grid_sampler_op.cc b/paddle/fluid/operators/grid_sampler_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..e76eb6893b1f7b6a965682368860c02fa32f6330 --- /dev/null +++ b/paddle/fluid/operators/grid_sampler_op.cc @@ -0,0 +1,203 @@ +/* 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/grid_sampler_op.h" +#include "paddle/fluid/framework/op_registry.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/cudnn_helper.h" +#endif + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +class GridSampleOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of GridSampleOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grid"), + "Input(Grid) of GridSampleOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Output"), + "Output(Output) of GridSampleOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto grid_dims = ctx->GetInputDim("Grid"); + PADDLE_ENFORCE(x_dims.size() == 4, + "Input(X) of GridSampleOp should be 4-D Tensor."); + PADDLE_ENFORCE(grid_dims.size() == 4, + "Input(Grid) of GridSampleOp should be 4-D Tensor."); + PADDLE_ENFORCE(grid_dims[3] == 2, "Input(Grid) dims[3] should be 2."); + PADDLE_ENFORCE_EQ(grid_dims[0], x_dims[0], + "Input(X) and Input(Grid) dims[0] should be equal."); + PADDLE_ENFORCE_EQ( + grid_dims[1], x_dims[2], + "Input(X) dims[2] and Input(Grid) dims[1] should be equal."); + PADDLE_ENFORCE_EQ( + grid_dims[2], x_dims[3], + "Input(X) dims[3] and Input(Grid) dims[2] should be equal."); + + ctx->SetOutputDim("Output", x_dims); + ctx->ShareLoD("X", "Output"); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_CUDA + if (platform::CanCUDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kCUDNN; + } +#endif + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + framework::DataLayout::kAnyLayout, library_); + } +}; + +class GridSampleOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor) The input data of GridSampleOp, " + "This is a 4-D tensor with shape of [N, C, H, W]"); + AddInput( + "Grid", + "(Tensor) The input grid of GridSampleOp generated by AffineGridOp, " + "This is a 4-D tensor with shape of [N, H, W, 2] is the concatenation " + "of x and y coordinates with shape [N, H, W] in last dimention"); + AddOutput("Output", "(Tensor) Output tensor with shape [N, C, H, W]"); + AddAttr( + "use_cudnn", + "(bool, default true) Only used in cudnn kernel, need install cudnn") + .SetDefault(true); + + AddComment(R"DOC( + This operation samples input X by using bilinear interpolation based on + flow field grid, which is usually gennerated by affine_grid. The grid of + shape [N, H, W, 2] is the concatenation of (grid_x, grid_y) coordinates + with shape [N, H, W] each, where grid_x is indexing the 4th dimension + (in width dimension) of input data x and grid_y is indexng the 3rd + dimention (in height dimension), finally results is the bilinear + interpolation value of 4 nearest corner points. + + Step 1: + Get (x, y) grid coordinates and scale to [0, H-1/W-1]. + + grid_x = 0.5 * (grid[:, :, :, 0] + 1) * (W - 1) + grid_y = 0.5 * (grid[:, :, :, 1] + 1) * (H - 1) + + Step 2: + Indices input data X with grid (x, y) in each [H, W] area, and bilinear + interpolate point value by 4 nearest points. + + wn ------- y_n ------- en + | | | + | d_n | + | | | + x_w --d_w-- grid--d_e-- x_e + | | | + | d_s | + | | | + ws ------- y_s ------- wn + + x_w = floor(x) // west side x coord + x_e = x_w + 1 // east side x coord + y_n = floor(y) // north side y coord + y_s = y_s + 1 // south side y coord + + d_w = grid_x - x_w // distance to west side + d_e = x_e - grid_x // distance to east side + d_n = grid_y - y_n // distance to north side + d_s = y_s - grid_y // distance to south side + + wn = X[:, :, y_n, x_w] // north-west point value + en = X[:, :, y_n, x_e] // north-east point value + ws = X[:, :, y_s, x_w] // south-east point value + es = X[:, :, y_s, x_w] // north-east point value + + output = wn * d_e * d_s + en * d_w * d_s + + ws * d_e * d_n + es * d_w * d_n + )DOC"); + } +}; + +class GridSampleOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + auto input_dims = ctx->GetInputDim("X"); + auto grid_dims = ctx->GetInputDim("Grid"); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), input_dims); + } + if (ctx->HasOutput(framework::GradVarName("Grid"))) { + ctx->SetOutputDim(framework::GradVarName("Grid"), grid_dims); + } + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_CUDA + if (platform::CanCUDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kCUDNN; + } +#endif + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + framework::DataLayout::kAnyLayout, library_); + } +}; + +class GridSampleGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("grid_sampler_grad"); + op->SetInput("X", Input("X")); + op->SetInput("Grid", Input("Grid")); + op->SetInput(framework::GradVarName("Output"), OutputGrad("Output")); + + op->SetAttrMap(Attrs()); + + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + op->SetOutput(framework::GradVarName("Grid"), InputGrad("Grid")); + return std::unique_ptr(op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(grid_sampler, ops::GridSampleOp, ops::GridSampleOpMaker, + ops::GridSampleGradMaker); +REGISTER_OPERATOR(grid_sampler_grad, ops::GridSampleOpGrad); + +REGISTER_OP_CPU_KERNEL( + grid_sampler, + ops::GridSampleOpKernel, + ops::GridSampleOpKernel); +REGISTER_OP_CPU_KERNEL( + grid_sampler_grad, + ops::GridSampleGradOpKernel, + ops::GridSampleGradOpKernel); diff --git a/paddle/fluid/operators/grid_sampler_op.h b/paddle/fluid/operators/grid_sampler_op.h new file mode 100644 index 0000000000000000000000000000000000000000..0d5874fc0cc4b90bec141690b88f28a27443bd60 --- /dev/null +++ b/paddle/fluid/operators/grid_sampler_op.h @@ -0,0 +1,322 @@ +/* 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/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/gather.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/hostdevice.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenTensor = framework::EigenTensor; + +using Array3 = Eigen::DSizes; +using Array4 = Eigen::DSizes; + +template +static inline bool isInBound(T x, T y, T x_max, T y_max) { + if (x < 0 || x > x_max || y < 0 || y > y_max) { + return false; + } + return true; +} + +template +static void CalcGridLocations(const platform::CPUDeviceContext& ctx, + const Tensor& grid, Tensor* x_w, Tensor* x_e, + Tensor* y_n, Tensor* y_s, Tensor* d_w, + Tensor* d_e, Tensor* d_n, Tensor* d_s) { + auto& place = *ctx.eigen_device(); + const int n = grid.dims()[0]; + const int h = grid.dims()[1]; + const int w = grid.dims()[2]; + const T x_max = static_cast(w - 1); + const T y_max = static_cast(h - 1); + + // split grid with shape (n, h, w, 2) into (x, y) by the 3rd Dim + Tensor grid_x, grid_y; + T* grid_x_data = grid_x.mutable_data({n, h, w}, ctx.GetPlace()); + T* grid_y_data = grid_y.mutable_data({n, h, w}, ctx.GetPlace()); + const T* grid_data = grid.data(); + for (int i = 0; i < n * h * w; i++) { + grid_x_data[i] = grid_data[2 * i]; + grid_y_data[i] = grid_data[(2 * i) + 1]; + } + + Tensor ones; + ones.mutable_data({n, h, w}, ctx.GetPlace()); + auto ones_t = EigenTensor::From(ones).setConstant(1.0); + + // scale grid to [0, h-1/w-1] + auto grid_x_t = EigenTensor::From(grid_x); + auto grid_y_t = EigenTensor::From(grid_y); + grid_x_t.device(place) = 0.5 * ((grid_x_t + ones_t) * x_max); + grid_y_t.device(place) = 0.5 * ((grid_y_t + ones_t) * y_max); + + // calculate coords of 4 corner points + x_w->mutable_data({n, h, w}, ctx.GetPlace()); + x_e->mutable_data({n, h, w}, ctx.GetPlace()); + y_n->mutable_data({n, h, w}, ctx.GetPlace()); + y_s->mutable_data({n, h, w}, ctx.GetPlace()); + auto x_w_t = EigenTensor::From(*x_w); + auto x_e_t = EigenTensor::From(*x_e); + auto y_n_t = EigenTensor::From(*y_n); + auto y_s_t = EigenTensor::From(*y_s); + x_w_t.device(place) = grid_x_t.floor(); + x_e_t.device(place) = x_w_t + ones_t; + y_n_t.device(place) = grid_y_t.floor(); + y_s_t.device(place) = y_n_t + ones_t; + + // calculate distances to 4 sides + d_w->mutable_data({n, h, w}, ctx.GetPlace()); + d_e->mutable_data({n, h, w}, ctx.GetPlace()); + d_n->mutable_data({n, h, w}, ctx.GetPlace()); + d_s->mutable_data({n, h, w}, ctx.GetPlace()); + auto d_w_t = EigenTensor::From(*d_w); + auto d_e_t = EigenTensor::From(*d_e); + auto d_n_t = EigenTensor::From(*d_n); + auto d_s_t = EigenTensor::From(*d_s); + d_w_t.device(place) = grid_x_t - x_w_t; + d_e_t.device(place) = x_e_t - grid_x_t; + d_n_t.device(place) = grid_y_t - y_n_t; + d_s_t.device(place) = y_s_t - grid_y_t; +} + +template +static void GetGridPointValue(const Tensor& input, Tensor* output, + const Tensor& x, const Tensor& y) { + const int n = input.dims()[0]; + const int c = input.dims()[1]; + const int h = input.dims()[2]; + const int w = input.dims()[3]; + auto x_t = EigenTensor::From(x); + auto y_t = EigenTensor::From(y); + auto output_t = EigenTensor::From(*output).setConstant((T)0); + auto input_t = EigenTensor::From(input); + + for (int i = 0; i < n; i++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + if (isInBound(x_t(i, k, l), y_t(i, k, l), (T)(w - 1), (T)(h - 1))) { + for (int j = 0; j < c; j++) { + output_t(i, j, k, l) = + input_t(i, j, static_cast(round(y_t(i, k, l))), + static_cast(round(x_t(i, k, l)))); + } + } + } + } + } +} + +template +static void GatherOutputGradToInputGrad(const Tensor& output_grad, + Tensor* input_grad, const Tensor& x, + const Tensor& y, const Tensor& d1, + const Tensor& d2) { + const int n = output_grad.dims()[0]; + const int c = output_grad.dims()[1]; + const int h = output_grad.dims()[2]; + const int w = output_grad.dims()[3]; + auto x_t = EigenTensor::From(x); + auto y_t = EigenTensor::From(y); + auto d1_t = EigenTensor::From(d1); + auto d2_t = EigenTensor::From(d2); + auto input_grad_t = EigenTensor::From(*input_grad); + auto output_grad_t = EigenTensor::From(output_grad); + + for (int i = 0; i < n; i++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + if (isInBound(x_t(i, k, l), y_t(i, k, l), (T)(w - 1), (T)(h - 1))) { + for (int j = 0; j < c; j++) { + input_grad_t(i, j, static_cast(round(y_t(i, k, l))), + static_cast(round(x_t(i, k, l)))) += + output_grad_t(i, j, k, l) * d1_t(i, k, l) * d2_t(i, k, l); + } + } + } + } + } +} + +template +class GridSampleOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& place = *ctx.template device_context().eigen_device(); + auto* input = ctx.Input("X"); + auto* grid = ctx.Input("Grid"); + + const int n = input->dims()[0]; + const int c = input->dims()[1]; + const int h = input->dims()[2]; + const int w = input->dims()[3]; + + // calc locations and distances of 4 corner points + Tensor x_w, x_e, y_n, y_s; + Tensor d_w, d_e, d_n, d_s; + CalcGridLocations( + ctx.template device_context(), *grid, &x_w, + &x_e, &y_n, &y_s, &d_w, &d_e, &d_n, &d_s); + + auto* output = ctx.Output("Output"); + output->mutable_data({n, c, h, w}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), output, + static_cast(0)); + + // calc 4 corner points value + Tensor v_wn, v_en, v_ws, v_es; + v_wn.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_en.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_ws.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_es.mutable_data({n, c, h, w}, ctx.GetPlace()); + GetGridPointValue(*input, &v_wn, x_w, y_n); + GetGridPointValue(*input, &v_en, x_e, y_n); + GetGridPointValue(*input, &v_ws, x_w, y_s); + GetGridPointValue(*input, &v_es, x_e, y_s); + + auto d_w_t = EigenTensor::From(d_w); + auto d_e_t = EigenTensor::From(d_e); + auto d_n_t = EigenTensor::From(d_n); + auto d_s_t = EigenTensor::From(d_s); + auto d_w_scaled_t = + d_w_t.reshape(Array4(n, 1, h, w)).broadcast(Array4(1, c, 1, 1)); + auto d_e_scaled_t = + d_e_t.reshape(Array4(n, 1, h, w)).broadcast(Array4(1, c, 1, 1)); + auto d_n_scaled_t = + d_n_t.reshape(Array4(n, 1, h, w)).broadcast(Array4(1, c, 1, 1)); + auto d_s_scaled_t = + d_s_t.reshape(Array4(n, 1, h, w)).broadcast(Array4(1, c, 1, 1)); + auto v_wn_t = EigenTensor::From(v_wn); + auto v_en_t = EigenTensor::From(v_en); + auto v_ws_t = EigenTensor::From(v_ws); + auto v_es_t = EigenTensor::From(v_es); + auto output_t = EigenTensor::From(*output); + // bilinear interpolaetion by 4 corner points + output_t.device(place) = v_wn_t * d_e_scaled_t * d_s_scaled_t + + v_en_t * d_w_scaled_t * d_s_scaled_t + + v_ws_t * d_e_scaled_t * d_n_scaled_t + + v_es_t * d_w_scaled_t * d_n_scaled_t; + } +}; + +template +class GridSampleGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("X"); + auto* grid = ctx.Input("Grid"); + auto* output_grad = ctx.Input(framework::GradVarName("Output")); + + const int n = input->dims()[0]; + const int c = input->dims()[1]; + const int h = input->dims()[2]; + const int w = input->dims()[3]; + + auto* input_grad = ctx.Output(framework::GradVarName("X")); + input_grad->mutable_data({n, c, h, w}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), input_grad, + static_cast(0)); + auto* grid_grad = ctx.Output(framework::GradVarName("Grid")); + grid_grad->mutable_data({n, h, w, 2}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), grid_grad, + static_cast(0)); + + Tensor x_w, x_e, y_n, y_s; + Tensor d_w, d_e, d_n, d_s; + CalcGridLocations( + ctx.template device_context(), *grid, &x_w, + &x_e, &y_n, &y_s, &d_w, &d_e, &d_n, &d_s); + + // gather output grad value to input grad by corner point coords and weight + GatherOutputGradToInputGrad(*output_grad, input_grad, x_w, y_n, d_e, + d_s); + GatherOutputGradToInputGrad(*output_grad, input_grad, x_w, y_s, d_e, + d_n); + GatherOutputGradToInputGrad(*output_grad, input_grad, x_e, y_n, d_w, + d_s); + GatherOutputGradToInputGrad(*output_grad, input_grad, x_e, y_s, d_w, + d_n); + + // calc 4 corner points value + Tensor v_wn, v_en, v_ws, v_es; + v_wn.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_en.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_ws.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_es.mutable_data({n, c, h, w}, ctx.GetPlace()); + GetGridPointValue(*input, &v_wn, x_w, y_n); + GetGridPointValue(*input, &v_en, x_e, y_n); + GetGridPointValue(*input, &v_ws, x_w, y_s); + GetGridPointValue(*input, &v_es, x_e, y_s); + auto v_wn_t = EigenTensor::From(v_wn); + auto v_en_t = EigenTensor::From(v_en); + auto v_ws_t = EigenTensor::From(v_ws); + auto v_es_t = EigenTensor::From(v_es); + + auto d_w_t = EigenTensor::From(d_w); + auto d_e_t = EigenTensor::From(d_e); + auto d_n_t = EigenTensor::From(d_n); + auto d_s_t = EigenTensor::From(d_s); + + auto output_grad_t = EigenTensor::From(*output_grad); + + Tensor grid_grad_x, grid_grad_y; + grid_grad_x.mutable_data({n, h, w}, ctx.GetPlace()); + grid_grad_y.mutable_data({n, h, w}, ctx.GetPlace()); + auto grid_grad_x_t = EigenTensor::From(grid_grad_x).setConstant(0.0); + auto grid_grad_y_t = EigenTensor::From(grid_grad_y).setConstant(0.0); + for (int i = 0; i < n; i++) { + for (int j = 0; j < c; j++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + grid_grad_x_t(i, k, l) += + ((v_en_t(i, j, k, l) - v_wn_t(i, j, k, l)) * d_s_t(i, k, l) + + (v_es_t(i, j, k, l) - v_ws_t(i, j, k, l)) * d_n_t(i, k, l)) * + output_grad_t(i, j, k, l); + grid_grad_y_t(i, k, l) += + ((v_ws_t(i, j, k, l) - v_wn_t(i, j, k, l)) * d_e_t(i, k, l) + + (v_es_t(i, j, k, l) - v_en_t(i, j, k, l)) * d_w_t(i, k, l)) * + output_grad_t(i, j, k, l); + } + } + } + } + const T x_max = static_cast(w - 1); + const T y_max = static_cast(h - 1); + grid_grad_x_t = grid_grad_x_t * (x_max / (T)2); + grid_grad_y_t = grid_grad_y_t * (y_max / (T)2); + + // gather grid_grad [x, y] in 3rd Dim + T* grid_grad_data = grid_grad->data(); + T* grid_grad_x_data = grid_grad_x.data(); + T* grid_grad_y_data = grid_grad_y.data(); + for (int i = 0; i < n * h * w; i++) { + grid_grad_data[2 * i] = grid_grad_x_data[i]; + grid_grad_data[2 * i + 1] = grid_grad_y_data[i]; + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index 26f09c46c2224a4a46d302dff4b2ec594f0be103..1d8b1411cddf4fe16d2d00313c519cc173e1504d 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -27,6 +27,10 @@ limitations under the License. */ #include "paddle/fluid/operators/distributed/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" +DEFINE_int32(rpc_send_thread_num, 5, "number of threads for rpc send"); +DEFINE_int32(rpc_get_thread_num, 5, "number of threads for rpc get"); +DEFINE_int32(rpc_prefetch_thread_num, 5, "number of threads for rpc prefetch"); + namespace paddle { namespace operators { @@ -130,7 +134,6 @@ void ListenAndServOp::RunSyncLoop( rpc_service_->ResetBarrierCounter(); while (true) { - rpc_service_->Profiler().OneStep(); // Get from multiple trainers, we don't care about the order in which // the gradients arrives, just add suffix 0~n and merge the gradient. rpc_service_->SetCond(distributed::kRequestSend); @@ -214,23 +217,26 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, framework::ProgramDesc *program, framework::Scope *recv_scope) const { VLOG(2) << "RunAsyncLoop"; - // grad name to block id - std::unordered_map grad_to_block_id; - std::unordered_map id_to_grad; - auto grad_to_block_id_str = Attr>("grad_to_block_id"); - for (const auto &grad_and_id : grad_to_block_id_str) { + DoubleFindMap grad_to_block_id; + + auto append_block_maps = [](DoubleFindMap *out_map, + const std::string &grad_and_id) { std::vector pieces; split(grad_and_id, ':', &pieces); - VLOG(3) << "after split, grad = " << pieces[0] << ", id=" << pieces[1]; + VLOG(3) << "after split, key = " << pieces[0] << ", id=" << pieces[1]; PADDLE_ENFORCE_EQ(pieces.size(), 2); - PADDLE_ENFORCE_EQ(grad_to_block_id.count(pieces[0]), 0); + PADDLE_ENFORCE_EQ(out_map->count(pieces[0]), 0); int block_id = std::stoi(pieces[1]); - grad_to_block_id[pieces[0]] = block_id; - id_to_grad[block_id] = pieces[0]; + (*out_map)[pieces[0]] = block_id; + }; + + for (const auto &grad_and_id : grad_to_block_id_str) { + append_block_maps(&grad_to_block_id, grad_and_id); } + size_t num_blocks = program->Size(); PADDLE_ENFORCE_GE(num_blocks, 2, "server program should have at least 2 blocks"); @@ -240,15 +246,22 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, block_list.push_back(blkid); } auto optimize_prepared = executor->Prepare(*program, block_list); - // execute global block if needed - if (block_list[0] == 1 && id_to_grad.count(1) == 0) { + // execute global block if needed, block id 1 in the program is global + // block if it's not bind to a grad var for it's update. + if (block_list[0] == 1 && + grad_to_block_id.find_value(static_cast(1)) == + grad_to_block_id.end()) { executor->RunPreparedContext(optimize_prepared[0].get(), recv_scope); } std::unordered_map> - grad_to_prepared_ctx; + grad_to_prepared_ctx, param_to_prepared_ctx; for (size_t i = 0; i < block_list.size(); ++i) { - grad_to_prepared_ctx[id_to_grad[block_list[i]]] = optimize_prepared[i]; + auto blkid = block_list[i]; + auto it = grad_to_block_id.find_value(blkid); + if (it != grad_to_block_id.end()) { + grad_to_prepared_ctx[it->first] = optimize_prepared[i]; + } } request_send_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); @@ -311,6 +324,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, framework::Scope &recv_scope = scope.NewScope(); bool sync_mode = Attr("sync_mode"); + bool dc_sgd = Attr("dc_asgd"); auto fan_in = Attr("Fanin"); auto inputs = Inputs("X"); @@ -324,19 +338,24 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, rpc_service_.reset(new RPCSERVER_T(endpoint, fan_in)); - request_send_handler_.reset(new distributed::RequestSendHandler(sync_mode)); - request_get_handler_.reset(new distributed::RequestGetHandler(sync_mode)); + request_send_handler_.reset( + new distributed::RequestSendHandler(sync_mode, dc_sgd)); + request_get_handler_.reset( + new distributed::RequestGetHandler(sync_mode, dc_sgd)); request_prefetch_handler_.reset( new distributed::RequestPrefetchHandler(sync_mode)); request_checkpoint_handler_.reset(new distributed::RequestCheckpointHandler( sync_mode, checkpoint_block_id)); rpc_service_->RegisterRPC(distributed::kRequestSend, - request_send_handler_.get()); + request_send_handler_.get(), + FLAGS_rpc_send_thread_num); rpc_service_->RegisterRPC(distributed::kRequestGet, - request_get_handler_.get()); + request_get_handler_.get(), + FLAGS_rpc_get_thread_num); rpc_service_->RegisterRPC(distributed::kRequestPrefetch, - request_prefetch_handler_.get()); + request_prefetch_handler_.get(), + FLAGS_rpc_prefetch_thread_num); rpc_service_->RegisterRPC(distributed::kRequestCheckpoint, request_checkpoint_handler_.get()); @@ -436,6 +455,8 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { "a map from grad name to it's optimize block id") .SetDefault({}); AddAttr("sync_mode", "if works at sync_mode or not").SetDefault(true); + AddAttr("dc_asgd", "set to true will enable DC-ASGD training.") + .SetDefault(false); AddAttr>( kOptimizeBlocks, "Optimize blocks to run on server side.") .SetDefault({}); diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index 5f889793ab16249a4e06801090db087a089dbed1..9431978df836121baacc12ed4e1ee6b218cc7d7a 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -18,6 +18,7 @@ limitations under the License. */ #include #include #include +#include #include #include "paddle/fluid/framework/executor.h" @@ -37,6 +38,17 @@ constexpr char kCheckpointBlockId[] = "checkpint_block_id"; void RunServer(std::shared_ptr service); +template +class DoubleFindMap : public std::unordered_map { + public: + typename std::unordered_map::iterator find_value(TValue v) { + return std::find_if(this->begin(), this->end(), + [&v](const std::pair p) { + return p.second == v; + }); + } +}; + class ListenAndServOp : public framework::OperatorBase { public: ListenAndServOp(const std::string& type, diff --git a/paddle/fluid/operators/lookup_table_op.cc b/paddle/fluid/operators/lookup_table_op.cc index d7f6cd5ab0acd2b677a3e5bd51bbcffe82eb1e50..3226a727b1f5f6de9e97ce2068381be7c9b69ff3 100644 --- a/paddle/fluid/operators/lookup_table_op.cc +++ b/paddle/fluid/operators/lookup_table_op.cc @@ -121,7 +121,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("W")); + auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Out")); return framework::OpKernelType(data_type, ctx.device_context()); } }; diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt index 84ab2bec2500e6877c17e32da710b0855e002c77..d9bf2a8aafb9264aa62f3f7404746a53bbb4155f 100644 --- a/paddle/fluid/operators/math/CMakeLists.txt +++ b/paddle/fluid/operators/math/CMakeLists.txt @@ -75,7 +75,12 @@ if(WITH_GPU) endif() cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split) 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_rnn.cc - DEPS cpu_info cblas) + +set(JIT_KERNEL_SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc) +set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce) +if(WITH_XBYAK) + list(APPEND JIT_KERNEL_SRCS jit_gen.cc jit_code.cc) + list(APPEND JIT_KERNEL_DEPS xbyak) +endif() +cc_library(jit_kernel SRCS ${JIT_KERNEL_SRCS} DEPS ${JIT_KERNEL_DEPS}) cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel) diff --git a/paddle/fluid/operators/math/cross_entropy.cu b/paddle/fluid/operators/math/cross_entropy.cu index c92341ea55ea21773acba33665e267b2f1c25fe3..cb200ec8d6ea533d546f3e01a16a48c88b14f677 100644 --- a/paddle/fluid/operators/math/cross_entropy.cu +++ b/paddle/fluid/operators/math/cross_entropy.cu @@ -21,6 +21,16 @@ namespace operators { namespace math { namespace { + +__device__ __forceinline__ float real_log(float x) { return logf(x); } + +__device__ __forceinline__ double real_log(double x) { return log(x); } + +__device__ __forceinline__ platform::float16 real_log( + const platform::float16& val) { + return static_cast(logf(static_cast(val))); +} + template __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label, const int N, const int D, @@ -29,8 +39,8 @@ __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label, i += blockDim.x * gridDim.x) { PADDLE_ASSERT(label[i] >= 0 && label[i] < D || label[i] == ignore_index); Y[i] = ignore_index == label[i] - ? 0 - : -math::TolerableValue()(log(X[i * D + label[i]])); + ? static_cast(0) + : -math::TolerableValue()(real_log(X[i * D + label[i]])); } } @@ -38,12 +48,12 @@ template __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, const int class_num) { int tid = threadIdx.x; - T val = 0; + T val(0); int idx = blockIdx.x * class_num + tid; int end = blockIdx.x * class_num + class_num; for (; idx < end; idx += blockDim.x) { - val += math::TolerableValue()(std::log(X[idx])) * label[idx]; + val += math::TolerableValue()(real_log(X[idx])) * label[idx]; } val = paddle::platform::reduceSum(val, tid, blockDim.x); @@ -53,8 +63,6 @@ __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, } } // namespace -using Tensor = framework::Tensor; - template class CrossEntropyFunctor { public: @@ -89,6 +97,8 @@ class CrossEntropyFunctor { template class CrossEntropyFunctor; template class CrossEntropyFunctor; +template class CrossEntropyFunctor; } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/math/cross_entropy.h b/paddle/fluid/operators/math/cross_entropy.h index e8aeb5d0575ac0f6b8761e97896df73578e8a103..99a4935186e1e6f9e3bf36eb029ce3d230510117 100644 --- a/paddle/fluid/operators/math/cross_entropy.h +++ b/paddle/fluid/operators/math/cross_entropy.h @@ -13,6 +13,7 @@ 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/tensor.h" #include "paddle/fluid/platform/hostdevice.h" @@ -33,6 +34,26 @@ struct TolerableValue { } }; +// NOTE(dzh): float16 value clip behave different. +// 1. Our ValueClipping has a hardcore threshold 1e20 +// for float number. 1e20 will resulting in overflow in float16. +// 2. float16 should expose the the real number overflow to python. +// because mixed-training depends the inf/nan value to determine +// if the scale value will be adjusted. +// Also. In standard implementation of cross entropy, other +// framework not has the ValueClipping. +template <> +struct TolerableValue { + HOSTDEVICE platform::float16 operator()(const platform::float16& x) const { + if (platform::isfinite(x)) + return x; + else if (x > static_cast(0)) + return std::numeric_limits::max(); + else + return std::numeric_limits::min(); + } +}; + template class CrossEntropyFunctor { public: diff --git a/paddle/fluid/operators/math/jit_code.cc b/paddle/fluid/operators/math/jit_code.cc new file mode 100644 index 0000000000000000000000000000000000000000..9e2cc18c7a5e396be40b2336382f68a17f8a2bf9 --- /dev/null +++ b/paddle/fluid/operators/math/jit_code.cc @@ -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. */ + +#include "paddle/fluid/operators/math/jit_code.h" +#include "paddle/fluid/operators/math/jit_kernel.h" +#include "paddle/fluid/platform/cpu_info.h" + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace gen { + +using namespace platform::jit; // NOLINT + +bool VMulJitCode::init(int d) { + // It's not necessary to use avx512 since it would slow down the frequency + // and this kernel is not compute bound. + return MayIUse(avx); +} + +void VMulJitCode::generate() { + // do not need push stack, and do not need save avx512reg if do not use avx512 + int offset = 0; + for (int i = 0; i < num_ / AVX_FLOAT_BLOCK; ++i) { + vmovups(ymm_src1, ptr[param1 + offset]); + vmovups(ymm_src2, ptr[param2 + offset]); + vmulps(ymm_dst, ymm_src1, ymm_src2); + vmovups(ptr[param3 + offset], ymm_dst); + offset += sizeof(float) * AVX_FLOAT_BLOCK; + } + int rest = num_ % AVX_FLOAT_BLOCK; + if (rest >= 4) { + vmovups(xmm_src1, ptr[param1 + offset]); + vmovups(xmm_src2, ptr[param2 + offset]); + vmulps(xmm_dst, xmm_src1, xmm_src2); + vmovups(ptr[param3 + offset], xmm_dst); + offset += sizeof(float) * 4; + rest -= 4; + } + if (rest >= 2) { + vmovq(xmm_src1, ptr[param1 + offset]); + vmovq(xmm_src2, ptr[param2 + offset]); + vmulps(xmm_dst, xmm_src1, xmm_src2); + vmovq(ptr[param3 + offset], xmm_dst); + offset += sizeof(float) * 2; + rest -= 2; + } + if (rest > 0) { + vmovss(xmm_src1, ptr[param1 + offset]); + vmovss(xmm_src2, ptr[param2 + offset]); + vmulss(xmm_dst, xmm_src1, xmm_src2); + vmovss(ptr[param3 + offset], xmm_dst); + } + ret(); +} + +} // namespace gen +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_code.h b/paddle/fluid/operators/math/jit_code.h new file mode 100644 index 0000000000000000000000000000000000000000..6007b290815de0ceaa2226962c5273ae7da72e7e --- /dev/null +++ b/paddle/fluid/operators/math/jit_code.h @@ -0,0 +1,60 @@ +/* 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/operators/math/jit_gen.h" + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace gen { + +using reg64_t = const Xbyak::Reg64; +using reg32_t = const Xbyak::Reg32; +using xmm_t = const Xbyak::Xmm; +using ymm_t = const Xbyak::Ymm; +using zmm_t = const Xbyak::Zmm; +using Label = Xbyak::Label; + +class VMulJitCode : public JitCode { + public: + DECLARE_JIT_CODE(VMulJitCode); + explicit VMulJitCode(int d, size_t code_size = 256 * 1024, + void* code_ptr = nullptr) + : JitCode(code_size, code_ptr), num_(d) {} + static bool init(int d); + void generate() override; + + private: + int num_; + reg64_t param1{abi_param1}; + reg64_t param2{abi_param2}; + reg64_t param3{abi_param3}; + + xmm_t xmm_src1 = xmm_t(0); + xmm_t xmm_src2 = xmm_t(1); + xmm_t xmm_dst = xmm_t(2); + + ymm_t ymm_src1 = ymm_t(0); + ymm_t ymm_src2 = ymm_t(1); + ymm_t ymm_dst = ymm_t(2); +}; + +} // namespace gen +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_gen.cc b/paddle/fluid/operators/math/jit_gen.cc new file mode 100644 index 0000000000000000000000000000000000000000..6af39518ed926554c8c839bba701d3827923dba0 --- /dev/null +++ b/paddle/fluid/operators/math/jit_gen.cc @@ -0,0 +1,90 @@ +/* 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_gen.h" +#include +#include +#include +#include "paddle/fluid/platform/cpu_info.h" + +DEFINE_bool(dump_jitcode, false, "Whether to dump the jitcode to file"); + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace gen { + +constexpr Xbyak::Operand::Code g_abi_regs[] = { + Xbyak::Operand::RBX, Xbyak::Operand::RBP, Xbyak::Operand::R12, + Xbyak::Operand::R13, Xbyak::Operand::R14, Xbyak::Operand::R15}; + +constexpr int num_g_abi_regs = sizeof(g_abi_regs) / sizeof(g_abi_regs[0]); + +void JitCode::preCode() { + for (int i = 0; i < num_g_abi_regs; ++i) { + push(Xbyak::Reg64(g_abi_regs[i])); + } + if (platform::jit::MayIUse(platform::jit::avx512f)) { + mov(reg_EVEX_max_8b_offt, 2 * EVEX_max_8b_offt); + } +} + +void JitCode::postCode() { + for (int i = 0; i < num_g_abi_regs; ++i) { + pop(Xbyak::Reg64(g_abi_regs[num_g_abi_regs - 1 - i])); + } + ret(); +} + +void JitCode::dumpCode(const Xbyak::uint8 *code) const { + if (code) { + static int counter = 0; + std::ostringstream filename; + filename << "paddle_jitcode_" << name() << "." << counter << ".bin"; + counter++; + std::ofstream fout(filename.str(), std::ios::out); + if (fout.is_open()) { + fout.write(reinterpret_cast(code), getSize()); + fout.close(); + } + } +} + +Xbyak::Address JitCode::EVEX_compress_addr(Xbyak::Reg64 base, int offt, + bool bcast) { + int scale = 0; + if (EVEX_max_8b_offt <= offt && offt < 3 * EVEX_max_8b_offt) { + offt = offt - 2 * EVEX_max_8b_offt; + scale = 1; + } else if (3 * EVEX_max_8b_offt <= offt && offt < 5 * EVEX_max_8b_offt) { + offt = offt - 4 * EVEX_max_8b_offt; + scale = 2; + } + auto re = Xbyak::RegExp() + base + offt; + if (scale) { + re = re + reg_EVEX_max_8b_offt * scale; + } + if (bcast) { + return zword_b[re]; + } else { + return zword[re]; + } +} + +} // namespace gen +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_gen.h b/paddle/fluid/operators/math/jit_gen.h new file mode 100644 index 0000000000000000000000000000000000000000..6abf3434cc8d8f6ab2838ef822a4f6b948331802 --- /dev/null +++ b/paddle/fluid/operators/math/jit_gen.h @@ -0,0 +1,80 @@ +/* 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/platform/macros.h" + +#define XBYAK_USE_MMAP_ALLOCATOR +#include "xbyak/xbyak.h" +#include "xbyak/xbyak_util.h" + +DECLARE_bool(dump_jitcode); + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace gen { + +#define DECLARE_JIT_CODE(codename) \ + const char *name() const override { return #codename; } + +// Application Binary Interface +constexpr Xbyak::Operand::Code abi_param1(Xbyak::Operand::RDI), + abi_param2(Xbyak::Operand::RSI), abi_param3(Xbyak::Operand::RDX), + abi_param4(Xbyak::Operand::RCX), abi_not_param1(Xbyak::Operand::RCX); + +class JitCode : public Xbyak::CodeGenerator { + public: + explicit JitCode(size_t code_size = 256 * 1024, void *code_ptr = nullptr) + : Xbyak::CodeGenerator(code_size, code_ptr) {} + + virtual ~JitCode() {} + virtual const char *name() const = 0; + virtual void generate() = 0; + + template + const FUNC getCode() { + this->generate(); + const Xbyak::uint8 *code = CodeGenerator::getCode(); + if (FLAGS_dump_jitcode) { + this->dumpCode(code); + } + return reinterpret_cast(code); + } + DISABLE_COPY_AND_ASSIGN(JitCode); + + protected: + Xbyak::Reg64 param1{abi_param1}; + const int EVEX_max_8b_offt = 0x200; + const Xbyak::Reg64 reg_EVEX_max_8b_offt = rbp; + + void preCode(); + void postCode(); + void dumpCode(const Xbyak::uint8 *code) const; + void L(const char *label) { Xbyak::CodeGenerator::L(label); } + void L(const Xbyak::Label &label) { Xbyak::CodeGenerator::L(label); } + // Enhanced vector extension + Xbyak::Address EVEX_compress_addr(Xbyak::Reg64 base, int offt, + bool bcast = false); +}; + +} // namespace gen +} // 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 index 9088d0c7a6307c3fbd9707c719ec9e6f6c85fbdb..7b6027aa267803ff8ff830deabda536b1b27fec8 100644 --- a/paddle/fluid/operators/math/jit_kernel.h +++ b/paddle/fluid/operators/math/jit_kernel.h @@ -39,6 +39,7 @@ class Kernel { public: Kernel() = default; virtual ~Kernel() = default; + // TODO(TJ): below members should be deprecated. int num_{0}; int end_{0}; int rest_{0}; @@ -64,7 +65,7 @@ class KernelPool { template class VMulKernel : public Kernel { public: - virtual void Compute(const T *x, const T *y, T *z) const = 0; + void (*Compute)(const T *, const T *, T *, int); }; template @@ -151,6 +152,13 @@ class GRUKernel : public Kernel { virtual void ComputeHtPart2(T *gates, const T *ht_1, T *ht) const = 0; }; +template +class CRFDecodeKernel : public Kernel { + public: + virtual void Compute(const int seq_len, const T *x, const T *w, T *alpha, + int *track) const = 0; +}; + } // namespace jitkernel } // namespace math } // namespace operators diff --git a/paddle/fluid/operators/math/jit_kernel_blas.cc b/paddle/fluid/operators/math/jit_kernel_blas.cc index c88b17b012d1b9cd59220f6b37cb2ecb8a1551a4..8a988f8f482e4a4963f70c39bccd89387c1e0059 100644 --- a/paddle/fluid/operators/math/jit_kernel_blas.cc +++ b/paddle/fluid/operators/math/jit_kernel_blas.cc @@ -15,6 +15,12 @@ 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" + +#ifdef PADDLE_WITH_XBYAK +#include "paddle/fluid/operators/math/jit_code.h" +#endif + #ifdef PADDLE_WITH_MKLML #include "paddle/fluid/platform/dynload/mklml.h" #endif @@ -27,65 +33,86 @@ namespace paddle { namespace operators { namespace math { namespace jitkernel { - namespace jit = platform::jit; +template +void VMulRefer(const T* x, const T* y, T* z, int n) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] * y[i]; + } +} + +#ifdef PADDLE_WITH_MKLML +template +void VMulMKL(const T* x, const T* y, T* z, int n); + +template <> +void VMulMKL(const float* x, const float* y, float* z, int n) { + platform::dynload::vsMul(n, x, y, z); +} +template <> +void VMulMKL(const double* x, const double* y, double* z, int n) { + platform::dynload::vdMul(n, x, y, z); +} +#endif + /* VMUL JitKernel */ -template +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]; - } + static inline std::string name(int d) { + PADDLE_THROW("DType should be either float or double"); } -}; - + static inline bool useJIT(int d) { return false; } + static inline bool useMKL(int d) { return false; } + + explicit VMulKernelImpl(int d) : VMulKernel() { +#ifdef PADDLE_WITH_XBYAK + if (useJIT(d)) { + // roughly estimate the size of code + size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8; + jitcode_.reset(new gen::VMulJitCode(d, sz > 4096 ? sz : 4096)); + this->Compute = + jitcode_->getCode(); + return; + } +#endif #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); \ + if (useMKL(d)) { + this->Compute = VMulMKL; + return; + } +#endif + this->Compute = VMulRefer; } -#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); \ - } +#ifdef PADDLE_WITH_XBYAK -FOR_EACH_ISA(MKL_FLOAT, kGT16); -FOR_EACH_ISA_BLOCK(MKL_DOUBLE); + private: + std::unique_ptr jitcode_{nullptr}; #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); +#ifdef PADDLE_WITH_XBYAK +template <> +bool VMulKernelImpl::useJIT(int d) { + return gen::VMulJitCode::init(d); +} #endif -#ifdef __AVX2__ -INTRI8_FLOAT(jit::avx2); -#endif -#ifdef __AVX512F__ -INTRI8_FLOAT(jit::avx512f); + +#ifdef PADDLE_WITH_MKLML +template <> +bool VMulKernelImpl::useMKL(int d) { + return jit::MayIUse(jit::avx512f) && d > 512; +} + +template <> +bool VMulKernelImpl::useMKL(int d) { + return true; +} #endif -// TODO(TJ): eq16 test and complete avx512 -#undef INTRI8_FLOAT -#undef MKL_FLOAT -#undef MKL_DOUBLE + +REGISTER_JITKERNEL(vmul, VMulKernel); /* VADD JitKernel */ template @@ -465,13 +492,12 @@ INTRI_COMMON_FLOAT(jit::avx512f, kGT16); #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); +REGISTER_JITKERNEL_DEPRECATED(vadd, VAddKernel); +REGISTER_JITKERNEL_DEPRECATED(vscal, VScalKernel); +REGISTER_JITKERNEL_DEPRECATED(vaddb, VAddBiasKernel); +REGISTER_JITKERNEL_DEPRECATED(vrelu, VReluKernel); +REGISTER_JITKERNEL_DEPRECATED(vaddrelu, VAddReluKernel); +REGISTER_JITKERNEL_DEPRECATED(videntity, VIdentityKernel); } // namespace jitkernel } // namespace math diff --git a/paddle/fluid/operators/math/jit_kernel_crf_decode.cc b/paddle/fluid/operators/math/jit_kernel_crf_decode.cc new file mode 100644 index 0000000000000000000000000000000000000000..a4861c347e44ad86a066861d3375b556302a84bc --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_crf_decode.cc @@ -0,0 +1,296 @@ +/* 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 +#include "paddle/fluid/operators/math/jit_kernel_macro.h" +#ifdef __AVX__ +#include +#endif + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +namespace jit = platform::jit; + +/* CRF Decode JitKernel */ +template +class CRFDecodeKernelImpl : public CRFDecodeKernel { + public: + explicit CRFDecodeKernelImpl(int tag_num) : CRFDecodeKernel() { + this->num_ = tag_num; + } + void Compute(const int seq_len, const T* x, const T* w, T* alpha, + int* track) const override { + constexpr int state_trans_base_idx = 2; + for (int i = 0; i < this->num_; ++i) { + alpha[i] = w[i] + x[i]; + } + for (int k = 1; k < seq_len; ++k) { + for (int i = 0; i < this->num_; ++i) { + T max_score = -std::numeric_limits::max(); + int max_j = 0; + for (int j = 0; j < this->num_; ++j) { + T score = alpha[(k - 1) * this->num_ + j] + + w[(j + state_trans_base_idx) * this->num_ + i]; + if (score > max_score) { + max_score = score; + max_j = j; + } + } + alpha[k * this->num_ + i] = max_score + x[k * this->num_ + i]; + track[k * this->num_ + i] = max_j; + } + } + } +}; + +#define INIT_ALPHA(step_size) \ + /* Setup the alpha initial value.*/ \ + int i_offset = 0; \ + int last_offset = this->rest_ - step_size; \ + for (int i = 0; i <= this->end_; ++i) { \ + /* weights, input and alpha values. */ \ + __m256 w_content, x_content, alpha_content; \ + /* Load the relevant data into the variables from un-aligned address.*/ \ + w_content = _mm256_loadu_ps(w + i_offset); \ + x_content = _mm256_loadu_ps(x + i_offset); \ + alpha_content = _mm256_add_ps(w_content, x_content); \ + _mm256_storeu_ps(alpha + i_offset, alpha_content); \ + i_offset += step_size; \ + if (i == this->end_ - 1) { \ + if (this->rest_ > 0) { \ + i_offset += last_offset; \ + } else { \ + break; \ + } \ + } \ + } + +#define UPDATE_ALPHA(step_size) \ + /* Update the alpha and track values. */ \ + __m256 x_content = _mm256_loadu_ps(x + seq_offset + this->num_ + j_offset); \ + max_score = _mm256_add_ps(max_score, x_content); \ + _mm256_storeu_ps(alpha + seq_offset + this->num_ + j_offset, max_score); \ + _mm256_storeu_si256( \ + reinterpret_cast<__m256i*>(track + seq_offset + this->num_ + j_offset), \ + max_j); \ + /* Calculate the offset of next step*/ \ + j_offset += step_size; \ + if (j == this->end_ - 1) { \ + if (this->rest_ > 0) { \ + j_offset += last_offset; \ + } else { \ + break; \ + } \ + } + +#define INTRIAVX_FLOAT(block) \ + template <> \ + CRFDecodeKernelImpl::CRFDecodeKernelImpl( \ + int tag_num) \ + : CRFDecodeKernel() { \ + this->num_ = tag_num; \ + this->end_ = this->num_ / AVX_FLOAT_BLOCK; \ + this->rest_ = this->num_ % AVX_FLOAT_BLOCK; \ + } \ + template <> \ + void CRFDecodeKernelImpl::Compute( \ + const int seq_len, const float* x, const float* w, float* alpha, \ + int* track) const { \ + INIT_ALPHA(AVX_FLOAT_BLOCK) \ + /* Use the column-major strategy to get the location of maximum score.*/ \ + int seq_offset = 0; \ + constexpr int state_trans_base_idx = 2; \ + for (int k = 1; k < seq_len; ++k) { \ + int j_offset = 0; \ + for (int j = 0; j <= this->end_; ++j) { \ + /* Initialize the variables of maximum score and location.*/ \ + __m256 max_score = _mm256_set1_ps(-std::numeric_limits::max()); \ + __m256i max_j = _mm256_set1_epi32(0); \ + /* Calculate the offset of transition_weights.*/ \ + int trans_offset = state_trans_base_idx * this->num_ + j_offset; \ + for (int i = 0; i < this->num_; ++i) { \ + /* Initalize the content of alpha variable with related offset.*/ \ + __m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i); \ + /* Obtain the content of weights from un-aligned address.*/ \ + __m256 w_content = _mm256_loadu_ps(w + trans_offset); \ + __m256 score_v = _mm256_add_ps(alpha_content, w_content); \ + __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); \ + /* According to the mask value, update the index of the max_score.*/ \ + /* AVX instructions.*/ \ + __m128i lo_max_j = _mm256_extractf128_si256(max_j, 0); \ + __m128i hi_max_j = _mm256_extractf128_si256(max_j, 1); \ + __m128i lo_mask = _mm256_extractf128_si256((__m256i)mask, 0); \ + __m128i hi_mask = _mm256_extractf128_si256((__m256i)mask, 1); \ + lo_max_j = _mm_andnot_si128(lo_mask, lo_max_j); \ + hi_max_j = _mm_andnot_si128(hi_mask, hi_max_j); \ + lo_mask = _mm_and_si128(lo_mask, _mm_set1_epi32(i)); \ + hi_mask = _mm_and_si128(hi_mask, _mm_set1_epi32(i)); \ + lo_max_j = _mm_or_si128(lo_mask, lo_max_j); \ + hi_max_j = _mm_or_si128(hi_mask, hi_max_j); \ + max_j = _mm256_insertf128_si256(max_j, lo_max_j, 0); \ + max_j = _mm256_insertf128_si256(max_j, hi_max_j, 1); \ + /* AVX done*/ \ + /* Update the max_score value.*/ \ + max_score = _mm256_max_ps(max_score, score_v); \ + trans_offset += this->num_; \ + } \ + UPDATE_ALPHA(AVX_FLOAT_BLOCK) \ + } \ + seq_offset += this->num_; \ + } \ + } + +#define INTRIAVX2_FLOAT(isa, block) \ + template <> \ + CRFDecodeKernelImpl::CRFDecodeKernelImpl(int tag_num) \ + : CRFDecodeKernel() { \ + this->num_ = tag_num; \ + this->end_ = this->num_ / AVX2_FLOAT_BLOCK; \ + this->rest_ = this->num_ % AVX2_FLOAT_BLOCK; \ + } \ + template <> \ + void CRFDecodeKernelImpl::Compute( \ + const int seq_len, const float* x, const float* w, float* alpha, \ + int* track) const { \ + INIT_ALPHA(AVX2_FLOAT_BLOCK) \ + /* Use the column-major strategy to get the location of maximum score.*/ \ + int seq_offset = 0; \ + constexpr int state_trans_base_idx = 2; \ + for (int k = 1; k < seq_len; ++k) { \ + int j_offset = 0; \ + for (int j = 0; j <= this->end_; ++j) { \ + /* Initialize the variables of maximum score and location.*/ \ + __m256 max_score = _mm256_set1_ps(-std::numeric_limits::max()); \ + __m256i max_j = _mm256_set1_epi32(0); \ + /* Calculate the offset of transition_weights.*/ \ + int trans_offset = state_trans_base_idx * this->num_ + j_offset; \ + for (int i = 0; i < this->num_; ++i) { \ + /* Initalize the content of alpha variable with related offset.*/ \ + __m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i); \ + /* Obtain the content of weights from un-aligned address.*/ \ + __m256 w_content = _mm256_loadu_ps(w + trans_offset); \ + __m256 score_v = _mm256_add_ps(alpha_content, w_content); \ + __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); \ + /* According to the mask value, update the index of the max_score.*/ \ + /* AVX2 instructions.*/ \ + max_j = _mm256_or_si256( \ + _mm256_andnot_si256((__m256i)mask, max_j), \ + _mm256_and_si256((__m256i)mask, _mm256_set1_epi32(i))); \ + /* Update the max_score value.*/ \ + max_score = _mm256_max_ps(max_score, score_v); \ + trans_offset += this->num_; \ + } \ + UPDATE_ALPHA(AVX2_FLOAT_BLOCK) \ + } \ + seq_offset += this->num_; \ + } \ + } + +#define INTRIAVX512_FLOAT(block) \ + template <> \ + CRFDecodeKernelImpl::CRFDecodeKernelImpl( \ + int tag_num) \ + : CRFDecodeKernel() { \ + this->num_ = tag_num; \ + this->end_ = this->num_ / AVX512_FLOAT_BLOCK; \ + this->rest_ = this->num_ % AVX512_FLOAT_BLOCK; \ + } \ + template <> \ + void CRFDecodeKernelImpl::Compute( \ + const int seq_len, const float* x, const float* w, float* alpha, \ + int* track) const { \ + INIT_ALPHA(AVX512_FLOAT_BLOCK) \ + /* Use the column-major strategy to get the location of maximum score.*/ \ + int seq_offset = 0; \ + constexpr int state_trans_base_idx = 2; \ + for (int k = 1; k < seq_len; ++k) { \ + int j_offset = 0; \ + for (int j = 0; j <= this->end_; ++j) { \ + /* Initialize the variables of maximum score and location.*/ \ + __m512 max_score = _mm512_set1_ps(-std::numeric_limits::max()); \ + __m512i max_j = _mm512_setzero_si512(); \ + /* Calculate the offset of transition_weights.*/ \ + int trans_offset = state_trans_base_idx * this->num_ + j_offset; \ + for (int i = 0; i < this->num_; ++i) { \ + /* Initalize the content of alpha variable with related offset.*/ \ + __m512 alpha_content = _mm512_set1_ps(*(alpha + seq_offset + i)); \ + /* Obtain the content of weights from un-aligned address.*/ \ + __m512 w_content = _mm512_loadu_ps(w + trans_offset); \ + __m512 score_v = _mm512_add_ps(alpha_content, w_content); \ + __mmask16 mask = _mm512_cmp_ps_mask(score_v, max_score, _CMP_GT_OS); \ + /* AVX512 instructions.*/ \ + max_j = _mm512_mask_set1_epi32(max_j, mask, i); \ + /* Update the max_score value.*/ \ + max_score = _mm512_max_ps(max_score, score_v); \ + trans_offset += this->num_; \ + } \ + /* Update the alpha and track values.*/ \ + __m512 x_content = \ + _mm512_loadu_ps(x + seq_offset + this->num_ + j_offset); \ + max_score = _mm512_add_ps(max_score, x_content); \ + _mm512_storeu_ps(alpha + seq_offset + this->num_ + j_offset, \ + max_score); \ + _mm512_storeu_si512(reinterpret_cast<__m512i*>(track + seq_offset + \ + this->num_ + j_offset), \ + max_j); \ + /* Calculate the offset of next step*/ \ + j_offset += AVX512_FLOAT_BLOCK; \ + if (j == this->end_ - 1) { \ + if (this->rest_ > 0) { \ + j_offset += last_offset; \ + } else { \ + break; \ + } \ + } \ + } \ + seq_offset += this->num_; \ + } \ + } + +#ifdef __AVX__ +INTRIAVX_FLOAT(kEQ8); +INTRIAVX_FLOAT(kGT8LT16); +INTRIAVX_FLOAT(kEQ16); +INTRIAVX_FLOAT(kGT16); +#endif +#ifdef __AVX2__ +INTRIAVX2_FLOAT(jit::avx2, kEQ8); +INTRIAVX2_FLOAT(jit::avx2, kGT8LT16); +INTRIAVX2_FLOAT(jit::avx2, kEQ16); +INTRIAVX2_FLOAT(jit::avx2, kGT16); +#endif +#ifdef __AVX512F__ +INTRIAVX2_FLOAT(jit::avx512f, kEQ8); +INTRIAVX2_FLOAT(jit::avx512f, kGT8LT16); +INTRIAVX512_FLOAT(kEQ16); +INTRIAVX512_FLOAT(kGT16); +#endif + +#undef INTRIAVX512_FLOAT +#undef INTRIAVX2_FLOAT +#undef INTRIAVX_FLOAT +#undef INIT_ALPHA +#undef UPDATE_ALPHA + +REGISTER_JITKERNEL_DEPRECATED(crf_decode, CRFDecodeKernel); + +} // 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 index c4247580f491a7ca26259528ca74dd92e35785a9..d7c177e6782e19e199542e10e1d62587ee0df4cf 100644 --- a/paddle/fluid/operators/math/jit_kernel_exp.cc +++ b/paddle/fluid/operators/math/jit_kernel_exp.cc @@ -250,7 +250,7 @@ INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); #undef MKL_FLOAT #undef MKL_DOUBLE -REGISTER_JITKERNEL(vexp, VExpKernel); +REGISTER_JITKERNEL_DEPRECATED(vexp, VExpKernel); /* VSigmoid JitKernel */ template @@ -396,7 +396,7 @@ INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); #undef INTRI_GT16_FLOAT #undef INTRI_VSIGMOID -REGISTER_JITKERNEL(vsigmoid, VSigmoidKernel); +REGISTER_JITKERNEL_DEPRECATED(vsigmoid, VSigmoidKernel); /* VTanh JitKernel */ template @@ -531,7 +531,7 @@ INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); #undef INTRI_GT16_FLOAT #undef INTRI_VTANH -REGISTER_JITKERNEL(vtanh, VTanhKernel); +REGISTER_JITKERNEL_DEPRECATED(vtanh, VTanhKernel); #undef JITKERNEL_NEW_ACT_IMPL diff --git a/paddle/fluid/operators/math/jit_kernel_macro.h b/paddle/fluid/operators/math/jit_kernel_macro.h index d8e55f2673560ff6afa34376b73275b57a8ceea1..a8169ea48ae3eee5a8cba291be4496c4c6074221 100644 --- a/paddle/fluid/operators/math/jit_kernel_macro.h +++ b/paddle/fluid/operators/math/jit_kernel_macro.h @@ -21,8 +21,71 @@ namespace operators { namespace math { namespace jitkernel { -namespace jit = platform::jit; +#define JITKERNEL_DEFINE_NAME(ker_key, ker_class) \ + template <> \ + std::string ker_class##Impl::name(int d) { \ + std::string key(#ker_key "f"); \ + if (useJIT(d)) { \ + /* only jit code need record d*/ \ + return key + "jit" + std::to_string(d); \ + } else if (useMKL(d)) { \ + return key + "mkl"; \ + } else { \ + return key + "any"; \ + } \ + } \ + template <> \ + std::string ker_class##Impl::name(int d) { \ + std::string key(#ker_key "d"); \ + /* jit code do not support double yet*/ \ + if (useMKL(d)) { \ + return key + "mkl"; \ + } else { \ + return key + "any"; \ + } \ + } + +#define JITKERNEL_DECLARE(ker_class, ker_dtype) \ + template <> \ + std::shared_ptr> \ + KernelPool::Get, int>(int d) + +#define JITKERNEL_FIND_KEY(ker_class, ker_dtype) \ + std::string key = ker_class##Impl::name(d) + +#define JITKERNEL_IMPL(ker_class, ker_dtype) \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>(d)) + +#define REGISTER_JITKERNEL_WITH_DTYPE(ker_class, ker_dtype, marco_declare, \ + macro_find_key, macro_impl) \ + marco_declare(ker_class, ker_dtype) { \ + macro_find_key(ker_class, ker_dtype); \ + if (kers_.find(key) == kers_.end()) { \ + std::shared_ptr> p; \ + 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_ARGS(ker_key, ker_class, marco_define_name, \ + marco_declare, macro_find_key, macro_impl) \ + marco_define_name(ker_key, ker_class); \ + REGISTER_JITKERNEL_WITH_DTYPE(ker_class, float, JITKERNEL_DECLARE, \ + JITKERNEL_FIND_KEY, JITKERNEL_IMPL); \ + REGISTER_JITKERNEL_WITH_DTYPE(ker_class, double, JITKERNEL_DECLARE, \ + JITKERNEL_FIND_KEY, JITKERNEL_IMPL) + +#define REGISTER_JITKERNEL(ker_key, ker_class) \ + REGISTER_JITKERNEL_ARGS(ker_key, ker_class, JITKERNEL_DEFINE_NAME, \ + JITKERNEL_DECLARE, JITKERNEL_FIND_KEY, \ + JITKERNEL_IMPL) + +namespace jit = platform::jit; +// TODO(TJ): below defines are deprecated, would be remove recently #define SEARCH_BLOCK(macro_, ker, dtype, isa) \ if (d < AVX_FLOAT_BLOCK) { \ macro_(ker, dtype, isa, kLT8); \ @@ -47,44 +110,42 @@ namespace jit = platform::jit; 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>( \ +#define JITKERNEL_NEW_IMPL_DEPRECATED(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 JITKERNEL_WITH_DTYPE_DEPRECATED(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 REGISTER_JITKERNEL_DEPRECATED(ker_key, ker_class) \ + JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, float, f, \ + JITKERNEL_DECLARE, JITKERNEL_KEY, \ + JITKERNEL_NEW_IMPL_DEPRECATED); \ + JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, double, d, \ + JITKERNEL_DECLARE, JITKERNEL_KEY, \ + JITKERNEL_NEW_IMPL_DEPRECATED) + +#define REGISTER_JITKERNEL_ARGS_DEPRECATED(ker_key, ker_class, marco_declare, \ + macro_key, macro_impl) \ + JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, float, f, marco_declare, \ + macro_key, macro_impl); \ + JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, double, d, \ + marco_declare, macro_key, macro_impl) #define FOR_EACH_ISA(macro_, block) \ macro_(jit::avx512f, block); \ diff --git a/paddle/fluid/operators/math/jit_kernel_rnn.cc b/paddle/fluid/operators/math/jit_kernel_rnn.cc index fab293f7d03eb923995fa4cd99af955a34faa6a4..d0932a37bb85bbc41f662a106c8ef5693a72efeb 100644 --- a/paddle/fluid/operators/math/jit_kernel_rnn.cc +++ b/paddle/fluid/operators/math/jit_kernel_rnn.cc @@ -179,23 +179,23 @@ class LSTMKernelImpl : public LSTMKernel { /* 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_); + vmul_d_->Compute(gates, gates + d_, gates + d_, d_); + vmul_d_->Compute(ct_1, gates + d2_, gates + d2_, d_); 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); + vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_); } 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); + vmul_d_->Compute(gates, gates + d_, ct, d_); /* 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); + vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_); } private: @@ -289,36 +289,36 @@ class PeepholeKernelImpl : public LSTMKernel { 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_); + vmul_d_->Compute(wp_data, ct_1, checked, d_); + vmul_d_->Compute(wp_data + d_, ct_1, checked + d_, 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_); + vmul_d_->Compute(gates, gates + d_, gates + d_, d_); + vmul_d_->Compute(ct_1, gates + d2_, gates + d2_, d_); vadd_d_->Compute(gates + d_, gates + d2_, ct); /* get ogated*/ - vmul_d_->Compute(wp_data + d2_, ct, gates + d_); + vmul_d_->Compute(wp_data + d2_, ct, gates + d_, 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); + vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_); } 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); + vmul_d_->Compute(gates, gates + d_, ct, d_); /* get outgated, put W_oc * C_t on igated */ - vmul_d_->Compute(wp_data + d2_, ct, gates + d_); + vmul_d_->Compute(wp_data + d2_, ct, gates + d_, 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); + vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_); } private: @@ -352,8 +352,8 @@ class PeepholeKernelImpl : public LSTMKernel { act_cell, d)); \ } -REGISTER_JITKERNEL_ARGS(lstm, LSTMKernel, JITKERNEL_DECLARE_LSTM, - JITKERNEL_KEY_LSTM, JITKERNEL_NEW_LSTM_IMPL); +REGISTER_JITKERNEL_ARGS_DEPRECATED(lstm, LSTMKernel, JITKERNEL_DECLARE_LSTM, + JITKERNEL_KEY_LSTM, JITKERNEL_NEW_LSTM_IMPL); #undef INTRI8_FLOAT #undef JITKERNEL_DECLARE_LSTM @@ -378,13 +378,13 @@ class GRUKernelImpl : public GRUKernel { void ComputeH1(T* gates, T* ht) const override { act_gate_d_->Compute(gates, gates); act_state_d_->Compute(gates + d2_, gates + d2_); - vmul_d_->Compute(gates, gates + d2_, ht); + vmul_d_->Compute(gates, gates + d2_, ht, d_); } void ComputeHtPart1(T* gates, const T* ht_1, T* ht) const override { // W: {W_update, W_reset; W_state} act_gate_d2_->Compute(gates, gates); - vmul_d_->Compute(ht_1, gates + d_, ht); + vmul_d_->Compute(ht_1, gates + d_, ht, d_); } void ComputeHtPart2(T* gates, const T* ht_1, T* ht) const override { @@ -472,8 +472,8 @@ INTRI8_FLOAT(jit::avx512f); p = std::dynamic_pointer_cast>( \ std::make_shared>(act_gate, act_state, d)); -REGISTER_JITKERNEL_ARGS(gru, GRUKernel, JITKERNEL_DECLARE_GRU, - JITKERNEL_KEY_GRU, JITKERNEL_NEW_GRU_IMPL); +REGISTER_JITKERNEL_ARGS_DEPRECATED(gru, GRUKernel, JITKERNEL_DECLARE_GRU, + JITKERNEL_KEY_GRU, JITKERNEL_NEW_GRU_IMPL); #undef INTRI8_FLOAT #undef JITKERNEL_NEW_GRU_IMPL diff --git a/paddle/fluid/operators/math/jit_kernel_test.cc b/paddle/fluid/operators/math/jit_kernel_test.cc index c9e6ab740da4e39de3dc41f9df352b88e696c38d..667a95fe1a247cf9d3d63dae74f7e0fa9c2309ca 100644 --- a/paddle/fluid/operators/math/jit_kernel_test.cc +++ b/paddle/fluid/operators/math/jit_kernel_test.cc @@ -369,12 +369,12 @@ void lstm_ctht_better( 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); + vmul_d->Compute(gates, gates + d, gates + d, d); + vmul_d->Compute(ct_1, gates + d2, gates + d2, d); 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); + vmul_d->Compute(gates + d2, gates + d * 3, ht, d); } TEST(JitKernel, lstm) { @@ -578,7 +578,7 @@ void vmul_mkl(const int n, const float* x, const float* y, float* z) { TEST(JitKernel, vmul) { namespace jit = paddle::operators::math::jitkernel; - for (int d : {7, 8, 15, 16, 30, 256, 512}) { + for (int d : {7, 8, 15, 16, 20, 30, 256, 512, 1000, 1024}) { std::vector x(d), y(d); std::vector zref(d), ztgt(d); RandomVec(d, x.data()); @@ -616,7 +616,7 @@ TEST(JitKernel, vmul) { auto ttgts = GetCurrentUS(); for (int i = 0; i < repeat; ++i) { - ker->Compute(x_data, y_data, ztgt_data); + ker->Compute(x_data, y_data, ztgt_data, d); } auto ttgte = GetCurrentUS(); @@ -800,8 +800,8 @@ TEST(JitKernel, pool) { EXPECT_TRUE(std::dynamic_pointer_cast(pvmul_f) != std::dynamic_pointer_cast(pvmul_d)); - const auto& pvmul_from_key = jit::KernelPool::Instance().Get("vmulf4"); + const auto& pvmul_from_key = jit::KernelPool::Instance().Get("vmulfjit4"); EXPECT_EQ(pvmul_f, pvmul_from_key); - const auto& pvmul_from_key2 = jit::KernelPool::Instance().Get("vmulf5"); + const auto& pvmul_from_key2 = jit::KernelPool::Instance().Get("vmulfjit"); EXPECT_TRUE(pvmul_from_key2 == nullptr); } diff --git a/paddle/fluid/operators/math/pooling.cc b/paddle/fluid/operators/math/pooling.cc index b871851798e48e6b598cb4ab8e2e42db478a3820..8df43bb616179e2487534e0acabb71b09b87e1af 100644 --- a/paddle/fluid/operators/math/pooling.cc +++ b/paddle/fluid/operators/math/pooling.cc @@ -31,7 +31,7 @@ class Pool2dFunctor { const framework::Tensor& input, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* output) { + bool exclusive, framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_height = input.dims()[2]; const int input_width = input.dims()[3]; @@ -68,7 +68,8 @@ class Pool2dFunctor { pool_process.compute(input_data[h * input_width + w], &ele); } } - int pool_size = (hend - hstart) * (wend - wstart); + int pool_size = exclusive ? (hend - hstart) * (wend - wstart) + : ksize_height * ksize_width; pool_process.finalize(static_cast(pool_size), &ele); output_data[ph * output_width + pw] = ele; } @@ -93,7 +94,7 @@ class Pool2dGradFunctor { const framework::Tensor& output, const framework::Tensor& output_grad, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_grad_process, - framework::Tensor* input_grad) { + bool exclusive, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_height = input.dims()[2]; const int input_width = input.dims()[3]; @@ -124,7 +125,8 @@ class Pool2dGradFunctor { int wstart = pw * stride_width - padding_width; int wend = std::min(wstart + ksize_width, input_width); wstart = std::max(wstart, 0); - int pool_size = (hend - hstart) * (wend - wstart); + int pool_size = exclusive ? (hend - hstart) * (wend - wstart) + : ksize_height * ksize_width; float scale = 1.0 / pool_size; for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { @@ -249,7 +251,7 @@ class Pool3dFunctor { const framework::Tensor& input, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* output) { + bool exclusive, framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_depth = input.dims()[2]; const int input_height = input.dims()[3]; @@ -300,7 +302,9 @@ class Pool3dFunctor { } } int pool_size = - (dend - dstart) * (hend - hstart) * (wend - wstart); + exclusive + ? (dend - dstart) * (hend - hstart) * (wend - wstart) + : ksize_depth * ksize_height * ksize_width; pool_process.finalize(static_cast(pool_size), &ele); output_data[output_idx] = ele; } @@ -326,7 +330,7 @@ class Pool3dGradFunctor { const framework::Tensor& output, const framework::Tensor& output_grad, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_grad_process, - framework::Tensor* input_grad) { + bool exclusive, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_depth = input.dims()[2]; const int input_height = input.dims()[3]; @@ -369,7 +373,9 @@ class Pool3dGradFunctor { wstart = std::max(wstart, 0); int pool_size = - (dend - dstart) * (hend - hstart) * (wend - wstart); + exclusive + ? (dend - dstart) * (hend - hstart) * (wend - wstart) + : ksize_depth * ksize_height * ksize_width; float scale = 1.0 / pool_size; for (int d = dstart; d < dend; ++d) { for (int h = hstart; h < hend; ++h) { diff --git a/paddle/fluid/operators/math/pooling.cu b/paddle/fluid/operators/math/pooling.cu index b1c76350d1724629bae175abf47e6671a1532242..a689eb42242e551caa3470f34f7e8d7e80b6dfbe 100644 --- a/paddle/fluid/operators/math/pooling.cu +++ b/paddle/fluid/operators/math/pooling.cu @@ -29,7 +29,7 @@ __global__ void KernelPool2D(const int nthreads, const T* input_data, const int ksize_width, const int stride_height, const int stride_width, const int padding_height, const int padding_width, PoolProcess pool_process, - T* output_data) { + bool exclusive, T* output_data) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int pw = index % output_width; @@ -52,7 +52,8 @@ __global__ void KernelPool2D(const int nthreads, const T* input_data, pool_process.compute(input_data[h * input_width + w], &ele); } } - int pool_size = (hend - hstart) * (wend - wstart); + int pool_size = exclusive ? (hend - hstart) * (wend - wstart) + : ksize_height * ksize_width; pool_process.finalize(static_cast(pool_size), &ele); output_data[index] = ele; } @@ -65,7 +66,7 @@ __global__ void KernelPool2DGrad( const int input_width, const int output_height, const int output_width, const int ksize_height, const int ksize_width, const int stride_height, const int stride_width, const int padding_height, const int padding_width, - PoolProcess pool_process, T* input_grad) { + PoolProcess pool_process, bool exclusive, T* input_grad) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int offsetW = index % input_width + padding_width; @@ -95,7 +96,8 @@ __global__ void KernelPool2DGrad( int wend = min(wstart + ksize_width, input_width); hstart = max(hstart, 0); wstart = max(wstart, 0); - int pool_size = (hend - hstart) * (wend - wstart); + int pool_size = exclusive ? (hend - hstart) * (wend - wstart) + : ksize_height * ksize_width; int output_sub_idx = ph * output_width + pw; pool_process.compute(input, output_data[output_sub_idx], output_grad[output_sub_idx], @@ -163,7 +165,7 @@ class Pool2dFunctor { const framework::Tensor& input, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* output) { + bool exclusive, framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_height = input.dims()[2]; @@ -189,7 +191,8 @@ class Pool2dFunctor { KernelPool2D<<>>( nthreads, input_data, input_channels, input_height, input_width, output_height, output_width, ksize_height, ksize_width, stride_height, - stride_width, padding_height, padding_width, pool_process, output_data); + stride_width, padding_height, padding_width, pool_process, exclusive, + output_data); } }; @@ -208,7 +211,7 @@ class Pool2dGradFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* input_grad) { + bool exclusive, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_height = input.dims()[2]; @@ -236,7 +239,7 @@ class Pool2dGradFunctor { nthreads, input_data, output_data, output_grad_data, input_channels, input_height, input_width, output_height, output_width, ksize_height, ksize_width, stride_height, stride_width, padding_height, padding_width, - pool_process, input_grad_data); + pool_process, exclusive, input_grad_data); } }; @@ -313,16 +316,14 @@ template class Pool2dGradFunctor; template -__global__ void KernelPool3D(const int nthreads, const T* input_data, - const int channels, const int input_depth, - const int input_height, const int input_width, - const int output_depth, const int output_height, - const int output_width, const int ksize_depth, - const int ksize_height, const int ksize_width, - const int stride_depth, const int stride_height, - const int stride_width, const int padding_depth, - const int padding_height, const int padding_width, - PoolProcess pool_process, T* output_data) { +__global__ void KernelPool3D( + const int nthreads, const T* input_data, const int channels, + const int input_depth, const int input_height, const int input_width, + const int output_depth, const int output_height, const int output_width, + const int ksize_depth, const int ksize_height, const int ksize_width, + const int stride_depth, const int stride_height, const int stride_width, + const int padding_depth, const int padding_height, const int padding_width, + PoolProcess pool_process, bool exclusive, T* output_data) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int pw = index % output_width; @@ -351,7 +352,9 @@ __global__ void KernelPool3D(const int nthreads, const T* input_data, } } } - int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + int pool_size = exclusive + ? (dend - dstart) * (hend - hstart) * (wend - wstart) + : ksize_depth * ksize_height * ksize_width; pool_process.finalize(static_cast(pool_size), &ele); output_data[index] = ele; } @@ -366,7 +369,7 @@ __global__ void KernelPool3DGrad( const int ksize_height, const int ksize_width, const int stride_depth, const int stride_height, const int stride_width, const int padding_depth, const int padding_height, const int padding_width, PoolProcess pool_process, - T* input_grad) { + bool exclusive, T* input_grad) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int offsetW = index % input_width + padding_width; @@ -409,7 +412,9 @@ __global__ void KernelPool3DGrad( dstart = max(dstart, 0); hstart = max(hstart, 0); wstart = max(wstart, 0); - int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + int pool_size = + exclusive ? (dend - dstart) * (hend - hstart) * (wend - wstart) + : ksize_depth * ksize_height * ksize_width; int output_sub_idx = (pd * output_height + ph) * output_width + pw; pool_process.compute(input, output_data[output_sub_idx], output_grad[output_sub_idx], @@ -484,7 +489,7 @@ class Pool3dFunctor { const framework::Tensor& input, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* output) { + bool exclusive, framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_depth = input.dims()[2]; @@ -517,7 +522,7 @@ class Pool3dFunctor { nthreads, input_data, input_channels, input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, stride_width, - padding_depth, padding_height, padding_width, pool_process, + padding_depth, padding_height, padding_width, pool_process, exclusive, output_data); } }; @@ -537,7 +542,7 @@ class Pool3dGradFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* input_grad) { + bool exclusive, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_depth = input.dims()[2]; @@ -573,7 +578,7 @@ class Pool3dGradFunctor { input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, stride_width, padding_depth, padding_height, - padding_width, pool_process, input_grad_data); + padding_width, pool_process, exclusive, input_grad_data); } }; diff --git a/paddle/fluid/operators/math/pooling.h b/paddle/fluid/operators/math/pooling.h index 120f5919803806e0d3b7dc8eaf530ae89819b84d..0f64e321bf01eea69767af020ed8c1a75e31acb5 100644 --- a/paddle/fluid/operators/math/pooling.h +++ b/paddle/fluid/operators/math/pooling.h @@ -89,7 +89,7 @@ class Pool2dFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_compute, - framework::Tensor* output); + bool exclusive, framework::Tensor* output); }; template @@ -101,7 +101,7 @@ class Pool2dGradFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_compute, - framework::Tensor* input_grad); + bool exclusive, framework::Tensor* input_grad); }; template @@ -123,7 +123,7 @@ class Pool3dFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_compute, - framework::Tensor* output); + bool exclusive, framework::Tensor* output); }; template @@ -135,7 +135,7 @@ class Pool3dGradFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_compute, - framework::Tensor* input_grad); + bool exclusive, framework::Tensor* input_grad); }; template diff --git a/paddle/fluid/operators/math/selected_rows_functor.cc b/paddle/fluid/operators/math/selected_rows_functor.cc index 08f57dd45ad76946cbcafb98a3414003ed9d67a9..75946740375d74043960b68e94eb048b3bab4b79 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cc +++ b/paddle/fluid/operators/math/selected_rows_functor.cc @@ -12,9 +12,8 @@ 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 #include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" @@ -230,8 +229,24 @@ template struct SelectedRowsAddToTensor; // add or mul. namespace scatter { -size_t FindPos(const std::vector& rows, int64_t value) { - return std::find(rows.begin(), rows.end(), value) - rows.begin(); +template +typename std::enable_if< + std::is_floating_point::value && + std::is_same::value>::type +elementwise_add_to(const DeviceContext& ctx, BlasT* blas, + size_t data_len, const T* in, T* out) { + blas->AXPY(data_len, 1., in, out); +} + +template +typename std::enable_if< + !std::is_floating_point::value && + std::is_same::value>::type +elementwise_add_to(const DeviceContext& ctx, BlasT* blas, + size_t data_len, const T* in, T* out) { + for (int64_t i = 0; i < data_len; i++) { + out[i] += in[i]; + } } template @@ -246,48 +261,84 @@ struct MergeAdd { void operator()(const platform::CPUDeviceContext& context, const framework::SelectedRows& input, framework::SelectedRows* output) { - framework::SelectedRows& out = *output; - std::vector input_rows(input.rows()); + std::vector inputs; + inputs.push_back(&input); + (*this)(context, inputs, output); + } - std::map> merge_row_map; - for (size_t i = 0; i < input_rows.size(); ++i) { - merge_row_map[input_rows[i]].push_back(i); + void operator()(const platform::CPUDeviceContext& context, + const std::vector& inputs, + framework::SelectedRows* output) { + if (inputs.size() == 0) { + VLOG(3) << "no input! return"; + return; } - - 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()); - - T* out_data = out.mutable_value()->mutable_data( + const framework::SelectedRows* has_value_input = nullptr; + for (auto* in : inputs) { + if (in->rows().size() > 0) { + has_value_input = in; + break; + } + } + if (has_value_input == nullptr) { + VLOG(3) << "no input has value! just return" << std::endl; + return; + } + auto input_width = has_value_input->value().dims()[1]; + auto input_height = has_value_input->height(); + framework::SelectedRows& out = *output; + std::set merged_row_set; + for (auto* input : inputs) { + if (input->rows().size() == 0) { + continue; + } + PADDLE_ENFORCE_EQ(input_width, input->value().dims()[1], + "all input should have same " + "dimension except for the first one"); + PADDLE_ENFORCE_EQ(input_height, input->height(), + "all input should have same height"); + merged_row_set.insert(input->rows().begin(), input->rows().end()); + } + std::vector merge_rows(merged_row_set.begin(), + merged_row_set.end()); + std::unordered_map rows_to_id; + for (size_t i = 0; i < merge_rows.size(); ++i) { + rows_to_id[merge_rows[i]] = i; + } + out.set_rows(merge_rows); + out.set_height(input_height); + out.mutable_value()->mutable_data( framework::make_ddim( {static_cast(merge_rows.size()), input_width}), context.GetPlace()); - 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]; - } + + math::SetConstant constant_functor; + constant_functor(context, out.mutable_value(), 0.0); + + auto* out_data = out.mutable_value()->data(); + + auto blas = math::GetBlas(context); + for (auto* input : inputs) { + if (input->rows().size() == 0) { + continue; + } + auto* input_data = input->value().data(); + auto& input_rows = input->rows(); + + for (size_t i = 0; i < input_rows.size(); i++) { + size_t out_i = rows_to_id[input_rows[i]]; + elementwise_add_to( + context, &blas, static_cast(input_width), + &input_data[i * input_width], &out_data[out_i * input_width]); } } - - out.set_rows(merge_rows); } }; template struct MergeAdd; template struct MergeAdd; +template struct MergeAdd; +template struct MergeAdd; template struct UpdateToTensor { diff --git a/paddle/fluid/operators/math/selected_rows_functor.cu b/paddle/fluid/operators/math/selected_rows_functor.cu index ba8eccf82042b679f69a32f9d053f05ac8fb9a99..a4fa6f5c898c541120a874f962b0f6a817736510 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cu +++ b/paddle/fluid/operators/math/selected_rows_functor.cu @@ -18,6 +18,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { @@ -118,7 +119,7 @@ struct SelectedRowsAddTensor { auto* out_data = output->data(); SetConstant functor; - functor(context, output, 0.0); + functor(context, output, static_cast(0)); const int block_size = 256; dim3 threads(block_size, 1); @@ -136,6 +137,9 @@ struct SelectedRowsAddTensor { template struct SelectedRowsAddTensor; template struct SelectedRowsAddTensor; +template struct SelectedRowsAdd; +template struct SelectedRowsAddTensor; template struct SelectedRowsAddTo { @@ -175,6 +179,8 @@ template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; namespace { template @@ -227,6 +233,8 @@ template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; namespace scatter { @@ -267,10 +275,15 @@ struct MergeAdd { void operator()(const platform::CUDADeviceContext& context, const framework::SelectedRows& input, framework::SelectedRows* output) { - framework::SelectedRows& out = *output; framework::Vector input_rows(input.rows()); + if (input_rows.size() == 0) { + return; + } + + framework::SelectedRows& out = *output; std::set row_set(input_rows.begin(), input_rows.end()); - std::vector merge_rows(row_set.begin(), row_set.end()); + std::vector merge_rows_cpu(row_set.begin(), row_set.end()); + framework::Vector merge_rows(merge_rows_cpu); auto input_width = input.value().dims()[1]; @@ -282,7 +295,7 @@ struct MergeAdd { context.GetPlace()); math::SetConstant constant_functor; - constant_functor(context, out.mutable_value(), 0.0); + constant_functor(context, out.mutable_value(), static_cast(0)); auto* out_data = out.mutable_value()->data(); auto* input_data = input.value().data(); @@ -296,12 +309,80 @@ struct MergeAdd { out.mutable_rows()->CUDAMutableData(context.GetPlace()), out.rows().size(), input_width); } + + void operator()(const platform::CUDADeviceContext& context, + const std::vector& inputs, + framework::SelectedRows* output) { + if (inputs.size() == 0) { + VLOG(3) << "no input! return"; + return; + } + const framework::SelectedRows* has_value_input = nullptr; + for (auto* in : inputs) { + if (in->rows().size() > 0) { + has_value_input = in; + break; + } + } + if (has_value_input == nullptr) { + VLOG(3) << "no input has value! just return" << std::endl; + return; + } + auto input_width = has_value_input->value().dims()[1]; + auto input_height = has_value_input->height(); + framework::SelectedRows& out = *output; + std::set merged_row_set; + for (auto* input : inputs) { + if (input->rows().size() == 0) { + continue; + } + PADDLE_ENFORCE_EQ(input_width, input->value().dims()[1], + "all input should have same " + "dimension except for the first one"); + PADDLE_ENFORCE_EQ(input_height, input->height(), + "all input should have same height"); + merged_row_set.insert(input->rows().begin(), input->rows().end()); + } + std::vector merge_rows_cpu(merged_row_set.begin(), + merged_row_set.end()); + framework::Vector merge_rows(merge_rows_cpu); + + out.set_rows(merge_rows); + out.set_height(input_height); + 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(), static_cast(0)); + + auto* out_data = out.mutable_value()->data(); + + const int block_size = 256; + dim3 threads(block_size, 1); + + for (auto* input : inputs) { + if (input->rows().size() == 0) { + continue; + } + auto* input_data = input->value().data(); + auto& input_rows = input->rows(); + dim3 grid1(input_rows.size(), 1); + + MergeAddKernel<<>>( + input_data, input_rows.CUDAData(context.GetPlace()), out_data, + out.mutable_rows()->CUDAMutableData(context.GetPlace()), + out.rows().size(), input_width); + } + } }; template struct MergeAdd; template struct MergeAdd; template struct MergeAdd; template struct MergeAdd; +template struct MergeAdd; template __global__ void UpdateToTensorKernel(const T* selected_rows, diff --git a/paddle/fluid/operators/math/selected_rows_functor.h b/paddle/fluid/operators/math/selected_rows_functor.h index 900be86f91c6658a5265189a6745316c6471209e..521c53dd0d71707c13c4364c5ee59943a03d4a2d 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.h +++ b/paddle/fluid/operators/math/selected_rows_functor.h @@ -83,104 +83,9 @@ struct MergeAdd { void operator()(const DeviceContext& context, const framework::SelectedRows& input, 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); - } + void operator()(const DeviceContext& context, + const std::vector& inputs, + framework::SelectedRows* output); }; template diff --git a/paddle/fluid/operators/math/selected_rows_functor_test.cc b/paddle/fluid/operators/math/selected_rows_functor_test.cc index 835589356042b44c9fa5988aed726434fd66910a..f15b37a1e3f0ae9c7612c4f74470472393ff4ad6 100644 --- a/paddle/fluid/operators/math/selected_rows_functor_test.cc +++ b/paddle/fluid/operators/math/selected_rows_functor_test.cc @@ -302,6 +302,64 @@ TEST(selected_rows_functor, cpu_merge_add_int) { EXPECT_EQ(out_data[1 * row_numel], 2); EXPECT_EQ(out_data[2 * row_numel], 1); } + +TEST(selected_rows_functor, cpu_merge_add_multi) { + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + set_const; + + int64_t height = 10; + int64_t row_numel = 8; + + std::vector rows1{5, 2, 5, 3, 5}; + 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); + set_const(ctx, in1_value, 1.0); + + std::vector rows2{2, 5, 3, 5, 3}; + 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); + set_const(ctx, in2_value, 1.0); + + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + output->set_height(height); + paddle::operators::math::scatter::MergeAdd + merge_add_functor; + + std::vector inputs; + inputs.push_back(selected_rows1.get()); + inputs.push_back(selected_rows2.get()); + merge_add_functor(ctx, inputs, output.get()); + + EXPECT_EQ(output->height(), height); + EXPECT_EQ(output->value().dims(), + paddle::framework::make_ddim({3, row_numel})); + + std::vector ret_rows{2, 3, 5}; + EXPECT_EQ(output->rows(), ret_rows); + + auto* out_data = output->value().data(); + for (size_t i = 0; i < ret_rows.size(); ++i) { + for (size_t j = 0; j < row_numel; ++j) { + EXPECT_EQ(out_data[i * row_numel + j], ret_rows[i]); + } + } +} + TEST(selected_rows_functor, cpu_sum_to) { paddle::platform::CPUPlace cpu_place; paddle::platform::CPUDeviceContext ctx(cpu_place); @@ -318,6 +376,7 @@ TEST(selected_rows_functor, cpu_sum_to) { 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{ @@ -327,6 +386,7 @@ TEST(selected_rows_functor, cpu_sum_to) { 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()}; diff --git a/paddle/fluid/operators/math/selected_rows_functor_test.cu.cc b/paddle/fluid/operators/math/selected_rows_functor_test.cu.cc index cfb4055d09ad955076669520512a6ef025a4dd47..73d83fa2e43f14445c969648cd469b0e32d644c7 100644 --- a/paddle/fluid/operators/math/selected_rows_functor_test.cu.cc +++ b/paddle/fluid/operators/math/selected_rows_functor_test.cu.cc @@ -242,3 +242,67 @@ TEST(selected_rows_functor, gpu_add_to) { // row9: 2.0 + 3.0 EXPECT_EQ(tensor1_cpu_data[9 * row_numel + 6], 5.0); } + +TEST(selected_rows_functor, gpu_merge_add) { + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDADeviceContext& ctx = + *reinterpret_cast( + paddle::platform::DeviceContextPool::Instance().Get(gpu_place)); + paddle::operators::math::SetConstant + set_const; + + int64_t height = 10; + int64_t row_numel = 8; + + std::vector rows1{5, 2, 5, 3, 5}; + 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}), + gpu_place); + set_const(ctx, in1_value, 1.0); + + std::vector rows2{2, 5, 3, 5, 3}; + 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}), + gpu_place); + set_const(ctx, in2_value, 1.0); + + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + output->set_height(height); + paddle::operators::math::scatter::MergeAdd< + paddle::platform::CUDADeviceContext, float> + merge_add_functor; + + std::vector inputs; + inputs.push_back(selected_rows1.get()); + inputs.push_back(selected_rows2.get()); + merge_add_functor(ctx, inputs, output.get()); + + paddle::framework::Tensor output_cpu; + paddle::framework::TensorCopy(output->value(), cpu_place, ctx, &output_cpu); + ctx.Wait(); + + EXPECT_EQ(output->height(), height); + EXPECT_EQ(output->value().dims(), + paddle::framework::make_ddim({3, row_numel})); + + std::vector ret_rows{2, 3, 5}; + EXPECT_EQ(output->rows(), ret_rows); + + auto* out_data = output_cpu.data(); + for (size_t i = 0; i < ret_rows.size(); ++i) { + for (size_t j = 0; j < row_numel; ++j) { + EXPECT_EQ(out_data[i * row_numel + j], ret_rows[i]); + } + } +} diff --git a/paddle/fluid/operators/math/sequence_pooling.cc b/paddle/fluid/operators/math/sequence_pooling.cc index 7be8539a7b0f1890898fd386a3056601fda8a7c3..6d491dbf1ed162ef07fda4c07e95cc57108486fd 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cc +++ b/paddle/fluid/operators/math/sequence_pooling.cc @@ -31,7 +31,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class MaxSeqPoolFunctor { public: void operator()(const platform::CPUDeviceContext& context, @@ -70,7 +70,41 @@ class MaxSeqPoolFunctor { } } }; +// Instantisation of Max Sequence Pooling for test phase eg. no need to fill +// index buffer +template +class MaxSeqPoolFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::LoDTensor& input, framework::Tensor* output, + framework::Tensor* index) { + auto in_dims = input.dims(); + auto out_dims = output->dims(); + PADDLE_ENFORCE_GT(in_dims.size(), 1); + PADDLE_ENFORCE_GT(out_dims.size(), 1); + for (int64_t i = 1; i < in_dims.size(); ++i) { + PADDLE_ENFORCE_EQ(in_dims[i], out_dims[i]); + } + + auto starts = input.lod()[0]; + const T* in_data = input.data(); + T* out_data = output->data(); + int64_t num_seq = out_dims[0]; + int64_t dim = output->numel() / num_seq; + for (int64_t i = 0; i < num_seq; ++i) { + std::memcpy(&out_data[i * dim], &in_data[starts[i] * dim], + dim * sizeof(T)); + for (size_t j = starts[i] + 1; j < starts[i + 1]; ++j) { + for (int64_t k = 0; k < dim; ++k) { + if (in_data[j * dim + k] > out_data[i * dim + k]) { + out_data[i * dim + k] = in_data[j * dim + k]; + } + } + } + } + } +}; template class MaxSeqPoolGradFunctor { public: @@ -188,11 +222,16 @@ class SequencePoolFunctor { /* max pool has index output */ void operator()(const platform::CPUDeviceContext& context, const std::string pooltype, const framework::LoDTensor& input, - framework::Tensor* output, + framework::Tensor* output, bool is_test, framework::Tensor* index = nullptr) { if (pooltype == "MAX") { - math::MaxSeqPoolFunctor max_pool; - max_pool(context, input, output, index); + if (is_test) { + math::MaxSeqPoolFunctor max_pool; + max_pool(context, input, output, index); + } else { + math::MaxSeqPoolFunctor max_pool; + max_pool(context, input, output, index); + } return; } if (pooltype == "LAST") { @@ -200,6 +239,7 @@ class SequencePoolFunctor { last_pool(context, input, output); return; } + if (pooltype == "FIRST") { math::FirstSeqPoolFunctor first_pool; first_pool(context, input, output); diff --git a/paddle/fluid/operators/math/sequence_pooling.cu b/paddle/fluid/operators/math/sequence_pooling.cu index a92aef805a0434f2ebcbc62d4e5eaef0cfb21bfa..0015fafbc892912424dfa6dbd1778438d384ca19 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cu +++ b/paddle/fluid/operators/math/sequence_pooling.cu @@ -133,7 +133,7 @@ class SequencePoolFunctor { public: void operator()(const platform::CUDADeviceContext& context, const std::string pooltype, const framework::LoDTensor& input, - framework::Tensor* output, + framework::Tensor* output, bool is_test, framework::Tensor* index = nullptr) { auto& lod = input.lod()[0]; const size_t item_dim = output->numel() / output->dims()[0]; diff --git a/paddle/fluid/operators/math/sequence_pooling.h b/paddle/fluid/operators/math/sequence_pooling.h index 8dcbee65d0b63a137e5f422ec8667cc950641b4a..a1046ea2160d0ae9c2251612c97d3f2640b0aad1 100644 --- a/paddle/fluid/operators/math/sequence_pooling.h +++ b/paddle/fluid/operators/math/sequence_pooling.h @@ -28,7 +28,7 @@ class SequencePoolFunctor { /* max pool has index output */ void operator()(const DeviceContext& context, const std::string pooltype, const framework::LoDTensor& input, framework::Tensor* output, - framework::Tensor* index = nullptr); + bool is_test = false, framework::Tensor* index = nullptr); }; template diff --git a/paddle/fluid/operators/math/softmax.cu b/paddle/fluid/operators/math/softmax.cu index 3effe776258cb541dbba32f63eda457d917011f4..ce183ed3649055aab31eb6e3f44f2224475957e9 100644 --- a/paddle/fluid/operators/math/softmax.cu +++ b/paddle/fluid/operators/math/softmax.cu @@ -96,12 +96,15 @@ template class SoftmaxCUDNNFunctor; template class SoftmaxCUDNNFunctor; template class SoftmaxGradCUDNNFunctor; template class SoftmaxGradCUDNNFunctor; +template class SoftmaxGradCUDNNFunctor; template class SoftmaxFunctor; template class SoftmaxFunctor; template class SoftmaxFunctor; template class SoftmaxGradFunctor; template class SoftmaxGradFunctor; +template class SoftmaxGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/fluid/operators/mean_op.cu b/paddle/fluid/operators/mean_op.cu index 91e0ab28efc21d4376524c8ecf66b429d51d8847..413b8ace67bd0a36849373812950834523b62216 100644 --- a/paddle/fluid/operators/mean_op.cu +++ b/paddle/fluid/operators/mean_op.cu @@ -15,11 +15,15 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/mean_op.h" +#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( mean, ops::MeanKernel, - ops::MeanKernel); + ops::MeanKernel, + ops::MeanKernel); REGISTER_OP_CUDA_KERNEL( mean_grad, ops::MeanGradKernel, - ops::MeanGradKernel); + ops::MeanGradKernel, + ops::MeanGradKernel); diff --git a/paddle/fluid/operators/mean_op.h b/paddle/fluid/operators/mean_op.h index 362e9f9ae8b2f0f77198e3f3939211ae1117b27b..360b2f68a749f630d3c7ed009c16cb51ec150581 100644 --- a/paddle/fluid/operators/mean_op.h +++ b/paddle/fluid/operators/mean_op.h @@ -55,8 +55,7 @@ class MeanGradKernel : public framework::OpKernel { IG->mutable_data(context.GetPlace()); T ig_size = static_cast(IG->numel()); - Eigen::DSizes bcast(ig_size); - + Eigen::DSizes bcast(static_cast(ig_size)); EigenVector::Flatten(*IG).device( *context.template device_context().eigen_device()) = (EigenVector::From(*OG) / ig_size).broadcast(bcast); diff --git a/paddle/fluid/operators/merge_ids_op.cc b/paddle/fluid/operators/merge_ids_op.cc index c6ec4ab047d5e91625e646fd26108d2e477cdce5..6e0e13698097ade36449f2e8ff6ab981a1b24311 100644 --- a/paddle/fluid/operators/merge_ids_op.cc +++ b/paddle/fluid/operators/merge_ids_op.cc @@ -20,13 +20,16 @@ namespace operators { class MergeIdsOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); - AddInput( - "X", - "(LoDTensors) multi input tensor with shape{batch_num, N}, N is the " - "size of embedding table") + AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}") + .AsDuplicable(); + AddInput("Rows", "(LoDTensor) the input ids with shape{row_size, 1}, ") + .AsDuplicable(); + AddInput("X", + "(LoDTensors) multi input tensor with shape{Rows, N}, N is the " + "size of embedding table") + .AsDuplicable(); + AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors.") .AsDuplicable(); - AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors."); AddComment(R"DOC( Merge multi LoDTensor's into one according to Ids's shard num. @@ -79,15 +82,19 @@ class MergeIdsOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Ids"), "MergeIdsOp must has input Ids."); - PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has input X."); - PADDLE_ENFORCE(ctx->HasOutput("Out"), "MergeIdsOp must has output Out."); + PADDLE_ENFORCE(ctx->HasInputs("Ids"), + "MergeIdsOp must has multi input Ids."); + PADDLE_ENFORCE(ctx->HasInputs("Rows"), + "MergeIdsOp must has multi input Rows."); + PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has multi input X."); + PADDLE_ENFORCE(ctx->HasOutputs("Out"), + "MergeIdsOp must has multi output Out."); auto ids_var_type = ctx->GetInputsVarType("Ids").front(); - auto ids_dims = ctx->GetInputDim("Ids"); + auto ids_dims = ctx->GetInputsDim("Ids"); if (ids_var_type == framework::proto::VarType::LOD_TENSOR) { - PADDLE_ENFORCE_EQ(ids_dims.size(), 2); - PADDLE_ENFORCE_EQ(ids_dims[1], 1); + PADDLE_ENFORCE_EQ(ids_dims[0].size(), 2); + PADDLE_ENFORCE_EQ(ids_dims[0][1], 1); } auto x_var_type = ctx->GetInputsVarType("X"); for (auto &var_type : x_var_type) { diff --git a/paddle/fluid/operators/merge_ids_op.h b/paddle/fluid/operators/merge_ids_op.h index 83712a8519c6817151e1922c606c0fdd4682a2db..fef9e023d02f45e21ec409ad398ba7d9bdd36880 100644 --- a/paddle/fluid/operators/merge_ids_op.h +++ b/paddle/fluid/operators/merge_ids_op.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once +#include +#include #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/tensor_util.h" @@ -30,59 +32,70 @@ class MergeIdsOpKernel : public framework::OpKernel { if (!platform::is_cpu_place(place)) { PADDLE_THROW("MergeIds do not support GPU kernel"); } - VLOG(3) << "run in MergeIdsOpKernel"; - const auto *ids_var = ctx.InputVar("Ids"); - PADDLE_ENFORCE(ids_var->IsType(), - "only support to merge Ids of LoDTensor"); + const auto ids = ctx.MultiInput("Ids"); + const auto row_ids = ctx.MultiInput("Rows"); + const auto x_tensors = ctx.MultiInput("X"); + auto outs = ctx.MultiOutput("Out"); - const auto &ids_tensor = ids_var->Get(); - const auto &ids_dims = ids_tensor.dims(); - const int64_t *ids = ids_tensor.data(); + PADDLE_ENFORCE_EQ(row_ids.size(), x_tensors.size(), + "the number of Rows and X should be the same"); + PADDLE_ENFORCE_EQ(ids.size(), outs.size(), + "the number of Ids and Out should be the same"); - auto x_tensors = ctx.MultiInput("X"); + int row_ids_size = 0; + int row_size = 0; + int embedding_size = 0; - auto *out = ctx.Output("Out"); + for (int i = 0; i < x_tensors.size(); ++i) { + const auto *x_tensor = x_tensors[i]; + const auto *row_id = row_ids[i]; - int batch_size = 0; - int embedding_size = 0; - for (auto &input : x_tensors) { - if (framework::product(input->dims()) != 0) { - if (embedding_size == 0) { - embedding_size = input->dims()[1]; - } - PADDLE_ENFORCE_EQ(embedding_size, input->dims()[1], - "embedding size of all input should be the same"); - batch_size += input->dims()[0]; + if (embedding_size == 0) { + embedding_size = x_tensor->dims()[1]; } + PADDLE_ENFORCE_EQ(embedding_size, x_tensor->dims()[1], + "embedding size of all input should be the same"); + row_size += x_tensor->dims()[0]; + row_ids_size += row_id->dims()[0]; } + PADDLE_ENFORCE_EQ( - batch_size, ids_dims[0], - "the batch size of ids and merged embedding value should be the same"); + row_size, row_ids_size, + "the merged X dim[0] and merged Rows dim[0] should be the same"); + + std::unordered_map> + selected_rows_idx_map; + for (int i = 0; i < x_tensors.size(); ++i) { + const auto *row_id = row_ids[i]; + + for (int j = 0; j < row_id->numel(); ++j) { + int64_t key = row_id->data()[j]; + std::tuple val = std::make_tuple(i, j); + selected_rows_idx_map.insert(std::make_pair(key, val)); + } + } + PADDLE_ENFORCE_EQ(row_ids_size, selected_rows_idx_map.size(), + "the rows and tensor map size should be the same"); + + for (int i = 0; i < outs.size(); ++i) { + auto *out_ids = ids[i]; + auto *out = outs[i]; - const size_t shard_num = x_tensors.size(); + out->set_lod(out_ids->lod()); - if (shard_num == 1) { - VLOG(3) << "only one shard, we can copy the data directly"; - TensorCopy(*x_tensors[0], place, out); - } else { - std::vector in_indexs(shard_num, 0); + int nums = static_cast(out_ids->dims()[0]); auto *out_data = out->mutable_data( - framework::make_ddim({batch_size, embedding_size}), place); - // copy data from ins[shard_num] to out. - for (int i = 0; i < ids_dims[0]; ++i) { - int64_t id = ids[i]; - size_t shard_id = static_cast(id) % shard_num; - int index = in_indexs[shard_id]; - memcpy(out_data + embedding_size * i, - x_tensors[shard_id]->data() + index * embedding_size, + framework::make_ddim({nums, embedding_size}), place); + for (int j = 0; j < nums; ++j) { + int id = out_ids->data()[j]; + auto row_tuple = selected_rows_idx_map[id]; + int64_t row_idx = std::get<1>(row_tuple); + const auto *x_tensor = x_tensors[std::get<0>(row_tuple)]; + + memcpy(out_data + embedding_size * j, + x_tensor->data() + row_idx * embedding_size, sizeof(T) * embedding_size); - in_indexs[shard_id] += 1; - } - - for (size_t i = 0; i < shard_num; ++i) { - PADDLE_ENFORCE_EQ(in_indexs[i], x_tensors[i]->dims()[0], - "after merge, all data in x_tensor should be used"); } } } diff --git a/paddle/fluid/operators/mul_op.cu.cc b/paddle/fluid/operators/mul_op.cu.cc index 81f3e42bf412fa4d2cb48405f2f8ee49b6aa0b67..6c5a83c6a50c463502171f09bbf18e17e43917b5 100644 --- a/paddle/fluid/operators/mul_op.cu.cc +++ b/paddle/fluid/operators/mul_op.cu.cc @@ -20,6 +20,7 @@ namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel, ops::MulKernel, ops::MulKernel); -REGISTER_OP_CUDA_KERNEL(mul_grad, - ops::MulGradKernel, - ops::MulGradKernel); +REGISTER_OP_CUDA_KERNEL( + mul_grad, ops::MulGradKernel, + ops::MulGradKernel, + ops::MulGradKernel); diff --git a/paddle/fluid/operators/pool_cudnn_op.cu.cc b/paddle/fluid/operators/pool_cudnn_op.cu.cc index 31f083565fddee66aea1485ed71f41b6199f4502..4a332ce10b59b21d2518684237ce0bbf1bbfa75a 100644 --- a/paddle/fluid/operators/pool_cudnn_op.cu.cc +++ b/paddle/fluid/operators/pool_cudnn_op.cu.cc @@ -41,6 +41,7 @@ class PoolCUDNNOpKernel : public framework::OpKernel { T *output_data = output->mutable_data(ctx.GetPlace()); std::string pooling_type = ctx.Attr("pooling_type"); + bool exclusive = ctx.Attr("exclusive"); std::vector ksize = ctx.Attr>("ksize"); std::vector strides = ctx.Attr>("strides"); std::vector paddings = ctx.Attr>("paddings"); @@ -72,7 +73,8 @@ class PoolCUDNNOpKernel : public framework::OpKernel { if (pooling_type == "max") { pooling_mode = PoolingMode::kMaximum; } else { - pooling_mode = PoolingMode::kAverage; + pooling_mode = exclusive ? PoolingMode::kAverageExclusive + : PoolingMode::kAverageInclusive; } cudnnPoolingDescriptor_t cudnn_pool_desc = @@ -101,6 +103,7 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { Tensor *input_grad = ctx.Output(framework::GradVarName("X")); std::string pooling_type = ctx.Attr("pooling_type"); + bool exclusive = ctx.Attr("exclusive"); std::vector ksize = ctx.Attr>("ksize"); std::vector strides = ctx.Attr>("strides"); std::vector paddings = ctx.Attr>("paddings"); @@ -141,7 +144,8 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { pooling_mode = PoolingMode::kMaximum; } } else { - pooling_mode = PoolingMode::kAverage; + pooling_mode = exclusive ? PoolingMode::kAverageExclusive + : PoolingMode::kAverageInclusive; } cudnnPoolingDescriptor_t cudnn_pool_desc = @@ -174,7 +178,8 @@ REGISTER_OP_KERNEL(pool2d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel); REGISTER_OP_KERNEL(pool2d_grad, CUDNN, plat::CUDAPlace, ops::PoolCUDNNGradOpKernel, - ops::PoolCUDNNGradOpKernel); + ops::PoolCUDNNGradOpKernel, + ops::PoolCUDNNGradOpKernel); REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel, diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index 24a5346b031008531fcefff0e6f1c31da33d1c3b..484cb65746612343fafc49fe61b607f2e919cf4f 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -180,6 +180,12 @@ void Pool2dOpMaker::Make() { "operator." "If global_pooling = true, paddings and ksize will be ignored.") .SetDefault({0, 0}); + AddAttr( + "exclusive", + "(bool, default True) When true, will exclude the zero-padding in the " + "averaging calculating, otherwise, include the zero-padding. Note, it " + "is only used when pooling_type is avg. The defalut is True.") + .SetDefault(true); AddAttr( "use_cudnn", "(bool, default false) Only used in cudnn kernel, need install cudnn") @@ -236,6 +242,23 @@ Example: W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1 $$ + For exclusive = true: + $$ + hstart = i * strides[0] - paddings[0] + hend = hstart + ksize[0] + wstart = j * strides[1] - paddings[1] + wend = wstart + ksize[1] + Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{ksize[0] * ksize[1]} + $$ + For exclusive = false: + $$ + hstart = max(0, i * strides[0] - paddings[0]) + hend = min(H, hstart + ksize[0]) + wstart = max(0, j * strides[1] - paddings[1]) + wend = min(W, wstart + ksize[1]) + Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} + $$ + )DOC"); } @@ -283,6 +306,12 @@ void Pool3dOpMaker::Make() { "If global_pooling = true, ksize and paddings will be ignored.") .SetDefault({0, 0, 0}); // TODO(Chengduo): Add checker. (Currently, // TypedAttrChecker don't support vector type.) + AddAttr( + "exclusive", + "(bool, default True) When true, will exclude the zero-padding in the " + "averaging calculating, otherwise, include the zero-padding. Note, it " + "is only used when pooling_type is avg. The defalut is True.") + .SetDefault(true); AddAttr( "use_cudnn", diff --git a/paddle/fluid/operators/pool_op.h b/paddle/fluid/operators/pool_op.h index a63963ca926bb94ff99e5cfe6dbcb2b15075bcb8..c0594b7e3cc5602a44bb01951a22c2135ba5c7ce 100644 --- a/paddle/fluid/operators/pool_op.h +++ b/paddle/fluid/operators/pool_op.h @@ -69,6 +69,7 @@ class PoolKernel : public framework::OpKernel { std::vector ksize = context.Attr>("ksize"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); + bool exclusive = context.Attr("exclusive"); if (context.Attr("global_pooling")) { for (size_t i = 0; i < ksize.size(); ++i) { paddings[i] = 0; @@ -84,7 +85,7 @@ class PoolKernel : public framework::OpKernel { pool2d_forward; paddle::operators::math::MaxPool pool_process; pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, - out); + true, out); } else if (pooling_type == "avg") { paddle::operators::math::Pool2dFunctor< @@ -92,7 +93,7 @@ class PoolKernel : public framework::OpKernel { pool2d_forward; paddle::operators::math::AvgPool pool_process; pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, - out); + exclusive, out); } } break; case 3: { @@ -102,14 +103,14 @@ class PoolKernel : public framework::OpKernel { pool3d_forward; paddle::operators::math::MaxPool pool_process; pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, - out); + true, out); } else if (pooling_type == "avg") { paddle::operators::math::Pool3dFunctor< DeviceContext, paddle::operators::math::AvgPool, T> pool3d_forward; paddle::operators::math::AvgPool pool_process; pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, - out); + exclusive, out); } } break; default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } @@ -131,6 +132,7 @@ class PoolGradKernel : public framework::OpKernel { std::vector ksize = context.Attr>("ksize"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); + bool exclusive = context.Attr("exclusive"); if (context.Attr("global_pooling")) { for (size_t i = 0; i < ksize.size(); ++i) { @@ -157,7 +159,7 @@ class PoolGradKernel : public framework::OpKernel { pool2d_backward; paddle::operators::math::AvgPoolGrad pool_process; pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, - paddings, pool_process, in_x_grad); + paddings, pool_process, exclusive, in_x_grad); } } break; case 3: { @@ -172,7 +174,7 @@ class PoolGradKernel : public framework::OpKernel { pool3d_backward; paddle::operators::math::AvgPoolGrad pool_process; pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, - paddings, pool_process, in_x_grad); + paddings, pool_process, exclusive, in_x_grad); } } break; default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } diff --git a/paddle/fluid/operators/prefetch_op.cc b/paddle/fluid/operators/prefetch_op.cc index 0519c15e13aac99802ff0f95b975712b36b44246..490dfa41be2de987c51b7f06d988ce27980aa5f2 100644 --- a/paddle/fluid/operators/prefetch_op.cc +++ b/paddle/fluid/operators/prefetch_op.cc @@ -42,7 +42,8 @@ class PrefetchOp : public framework::OperatorBase { auto& ctx = *pool.Get(place); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); std::vector rets; for (size_t i = 0; i < ins.size(); i++) { @@ -69,6 +70,7 @@ class PrefetchOpMaker : public framework::OpProtoAndCheckerMaker { "(LoDTensor) result " "to be fetched from parameter server") .AsDuplicable(); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr>( "epmap", "(string vector, default 127.0.0.1:6164)" diff --git a/paddle/fluid/operators/read_op.cc b/paddle/fluid/operators/read_op.cc index a0d640b2020958af53a4405ae886eadb2a1e117e..a0b70938d354cbb3bf10a9c8c589ba5153624f45 100644 --- a/paddle/fluid/operators/read_op.cc +++ b/paddle/fluid/operators/read_op.cc @@ -33,6 +33,19 @@ class ReadInferShape : public framework::InferShapeBase { reader_dims.size(), out_names.size(), "The reader's dim number doesn't match the output number."); ctx->SetOutputsDim("Out", reader_dims); + if (!ctx->IsRuntime()) { + auto in_desc = + boost::get(ctx->GetInputVarPtrs("Reader")[0]); + auto in_lod_levels = in_desc->GetLoDLevels(); + auto out_var_ptrs = ctx->GetOutputVarPtrs("Out"); + PADDLE_ENFORCE_EQ(in_lod_levels.size(), out_var_ptrs.size(), + "LoDLevels of Input(Reader) must be the same as the " + "number of Outputs(Out)."); + for (size_t i = 0; i < out_var_ptrs.size(); ++i) { + auto* out_desc = boost::get(out_var_ptrs[i]); + out_desc->SetLoDLevel(in_lod_levels[i]); + } + } } }; diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc index 4d34b8a1686efb1fc30020f0d27e9a3c3a6c0866..0399ff41007fbe10da8d53a05671eb0cfb475a5f 100644 --- a/paddle/fluid/operators/recv_op.cc +++ b/paddle/fluid/operators/recv_op.cc @@ -42,7 +42,8 @@ class RecvOp : public framework::OperatorBase { auto& ctx = *pool.Get(place); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); std::vector rets; for (size_t i = 0; i < outs.size(); i++) { @@ -73,6 +74,7 @@ This operator can get variables from server side. "Server endpoints in the order of input " "variables for mapping") .SetDefault({}); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr("sync_mode", "(int, default 0)" "sync recv or async recv.") diff --git a/paddle/fluid/operators/ref_by_trainer_id_op.cc b/paddle/fluid/operators/ref_by_trainer_id_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..6cb651af6dc3d8e301365968787c199acc4c60ee --- /dev/null +++ b/paddle/fluid/operators/ref_by_trainer_id_op.cc @@ -0,0 +1,79 @@ +/* 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/ref_by_trainer_id_op.h" +#include + +namespace paddle { +namespace operators { + +class RefByTrainerIdOp : public framework::OperatorWithKernel { + public: + RefByTrainerIdOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInputs("X"), + "Input(X) of RefByTrainerIdOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("TrainerId"), + "Input(TrainerId) of RefByTrainerIdOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of RefByTrainerIdOp should not be null."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("TrainerId").size(), 1, + "TrainerId should be a scalar."); + // Out's shape is determined at runtime. + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType( + ctx.MultiInput("X")[0]->type()), + ctx.GetPlace()); + } +}; + +class RefByTrainerIdOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor) Input tensor list.").AsDuplicable(); + AddInput("TrainerId", "(Tensor) Scalar int, the trainer id runtime value."); + AddOutput("Out", "(Tensor) Return one tensor reference of X[trainer_id]"); + AddComment(R"DOC( +**RefByTrainerId operator** + +Return a reference of a tensor, using trainer_id as the index to find from the input. + +$$Out = X[TrainerId]$$ +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_WITHOUT_GRADIENT(ref_by_trainer_id, ops::RefByTrainerIdOp, + ops::RefByTrainerIdOpMaker); +REGISTER_OP_CPU_KERNEL( + ref_by_trainer_id, + ops::RefByTrainerIdKernel, + ops::RefByTrainerIdKernel, + ops::RefByTrainerIdKernel, + ops::RefByTrainerIdKernel); diff --git a/paddle/fluid/operators/ref_by_trainer_id_op.cu.cc b/paddle/fluid/operators/ref_by_trainer_id_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..b98e2b5c9c7341f2a424fb4b32ff1e8bc45a056c --- /dev/null +++ b/paddle/fluid/operators/ref_by_trainer_id_op.cu.cc @@ -0,0 +1,26 @@ +/* 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/ref_by_trainer_id_op.h" + +REGISTER_OP_CUDA_KERNEL( + ref_by_trainer_id, + paddle::operators::RefByTrainerIdKernel, + paddle::operators::RefByTrainerIdKernel, + paddle::operators::RefByTrainerIdKernel, + paddle::operators::RefByTrainerIdKernel); diff --git a/paddle/fluid/operators/ref_by_trainer_id_op.h b/paddle/fluid/operators/ref_by_trainer_id_op.h new file mode 100644 index 0000000000000000000000000000000000000000..2ce577544ae2437b9297da2190fd09b435d5173c --- /dev/null +++ b/paddle/fluid/operators/ref_by_trainer_id_op.h @@ -0,0 +1,48 @@ +/* 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. */ + +#pragma once + +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +template +class RefByTrainerIdKernel : public framework::OpKernel { + public: + virtual void Compute(const framework::ExecutionContext& context) const { + auto* out = context.Output("Out"); + auto in_list = context.MultiInput("X"); + auto* trainer_id_t = context.Input("TrainerId"); + int64_t trainer_id = 0; + auto* trainer_id_data = trainer_id_t->data(); + if (platform::is_gpu_place(context.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + auto stream = context.cuda_device_context().stream(); + memory::Copy<>(platform::CPUPlace(), &trainer_id, + boost::get(context.GetPlace()), + trainer_id_data, sizeof(int64_t), stream); +#endif + } else { + trainer_id = *trainer_id_data; + } + PADDLE_ENFORCE_LT(trainer_id, in_list.size()); + out->mutable_data(context.GetPlace()); + out->ShareDataWith(*(in_list[trainer_id])); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/rmsprop_op.h b/paddle/fluid/operators/rmsprop_op.h index 797cd45fdcdbd5c3567d1676f37e148304ee6e2d..389c84d2464090ff9bd9e8b471cd0103c86a347a 100644 --- a/paddle/fluid/operators/rmsprop_op.h +++ b/paddle/fluid/operators/rmsprop_op.h @@ -179,8 +179,8 @@ class RmspropOpKernel : public framework::OpKernel { 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"); + PADDLE_ENFORCE_EQ(&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; @@ -198,8 +198,8 @@ class RmspropOpKernel : public framework::OpKernel { 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"); + PADDLE_ENFORCE_EQ(&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()), @@ -243,8 +243,8 @@ class RmspropOpKernel : public framework::OpKernel { 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"); + PADDLE_ENFORCE_EQ(&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()), diff --git a/paddle/fluid/operators/scale_op.cu b/paddle/fluid/operators/scale_op.cu index 04c802da12958a53626f533833c2709110531136..349f39360b8e3100a7f844d3e2d3768053c37c58 100644 --- a/paddle/fluid/operators/scale_op.cu +++ b/paddle/fluid/operators/scale_op.cu @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/scale_op.h" +#include "paddle/fluid/platform/float16.h" +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( scale, @@ -20,4 +22,6 @@ REGISTER_OP_CUDA_KERNEL( paddle::operators::ScaleKernel, paddle::operators::ScaleKernel, paddle::operators::ScaleKernel); + int64_t>, + paddle::operators::ScaleKernel); diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index 40404295266899c6ac2f7b1e08fdf7db40958794..8ca2877d8adad643089587fcee0917affa537f7d 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -39,7 +39,8 @@ class SendBarrierOp : public framework::OperatorBase { std::vector eps = Attr>("endpoints"); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); VLOG(3) << "SendBarrierOp sync"; @@ -67,6 +68,7 @@ This operator will send a send barrier signal to list_and_serv op, so that the Parameter Server would knew all variables have been sent. )DOC"); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr>("endpoints", "(string vector, default 127.0.0.1:6164)" "Server endpoints to send variables to.") diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index 48322ac7fd54a2e4cc3405a2c4dcddfc273f5a66..be1dc4bf14c3394963822b065ca088afbfacd858 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -44,7 +44,8 @@ class SendOp : public framework::OperatorBase { auto& ctx = *pool.Get(place); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); std::vector rets; for (size_t i = 0; i < ins.size(); i++) { @@ -79,6 +80,7 @@ This operator will send variables to listen_and_serve op at the parameter server "(int, default 0)" "sync send or async send.") .SetDefault(0); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr>("epmap", "(string vector, default 127.0.0.1:6164)" "Server endpoints in the order of input " diff --git a/paddle/fluid/operators/sequence_pool_op.cc b/paddle/fluid/operators/sequence_pool_op.cc index 15d3f064eb7b025dc9a85b2aabad24186061cbd4..217bb1610fd3f02f0f72d3b7750ebcdfad243f48 100644 --- a/paddle/fluid/operators/sequence_pool_op.cc +++ b/paddle/fluid/operators/sequence_pool_op.cc @@ -47,6 +47,7 @@ class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker { "(Tensor) This tensor is used for the sequence max-pooling " "to record the max indexes.") .AsIntermediate(); + AddAttr("is_test", "").SetDefault(false); AddAttr( "pooltype", "(string, default 'AVERAGE') the pooling pooltype of SequencePoolOp.") diff --git a/paddle/fluid/operators/sequence_pool_op.h b/paddle/fluid/operators/sequence_pool_op.h index 2aa20792f24305a106c500a3d7a6e3d363bc31d8..f2e4a55dee49664b2fc09813f6dba5f68aaf11d5 100644 --- a/paddle/fluid/operators/sequence_pool_op.h +++ b/paddle/fluid/operators/sequence_pool_op.h @@ -32,10 +32,6 @@ class SequencePoolKernel : public framework::OpKernel { auto* in = context.Input("X"); auto* out = context.Output("Out"); std::string pooltype = context.Attr("pooltype"); - Tensor* index = nullptr; - if (pooltype == "MAX") { - index = context.Output("MaxIndex"); - } auto dims = in->dims(); auto lod = in->lod(); @@ -48,13 +44,22 @@ class SequencePoolKernel : public framework::OpKernel { dims[0] = lod[0].size() - 1; out->Resize({dims}); out->mutable_data(context.GetPlace()); - if (pooltype == "MAX") { + Tensor* index = nullptr; + + const bool is_test = context.Attr("is_test"); + + // Do not create index buffer for inference (is_test) mode + // TODO(jczaja): Skip index buffer creation for other devices eg. GPU + if (pooltype == "MAX" && + (is_test == false || + platform::is_cpu_place(context.GetPlace()) == false)) { + index = context.Output("MaxIndex"); index->Resize({dims}); index->mutable_data(context.GetPlace()); } math::SequencePoolFunctor pool; pool(context.template device_context(), pooltype, *in, out, - index); + is_test, index); } }; diff --git a/paddle/fluid/operators/sign_op.cc b/paddle/fluid/operators/sign_op.cc index f3985dcc027f974e0213a73ea9a21e268d77615f..6837856a6da804e27af2cd6c83052c04f17140d8 100644 --- a/paddle/fluid/operators/sign_op.cc +++ b/paddle/fluid/operators/sign_op.cc @@ -67,4 +67,5 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(sign, ops::SignOp, ops::SignOpMaker, ops::SignGradMaker); REGISTER_OP_CPU_KERNEL( - sign, ops::SignKernel); + sign, ops::SignKernel, + ops::SignKernel); diff --git a/paddle/fluid/operators/sign_op.cu b/paddle/fluid/operators/sign_op.cu index e0d7a87e6485a74dd1cdee1a05abc42eef460990..817e0fbbd511462f161633242d28e63062676eb9 100644 --- a/paddle/fluid/operators/sign_op.cu +++ b/paddle/fluid/operators/sign_op.cu @@ -13,7 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/sign_op.h" +#include "paddle/fluid/platform/float16.h" REGISTER_OP_CUDA_KERNEL( sign, - paddle::operators::SignKernel); + paddle::operators::SignKernel, + paddle::operators::SignKernel, + paddle::operators::SignKernel); diff --git a/paddle/fluid/operators/softmax_cudnn_op.cu.cc b/paddle/fluid/operators/softmax_cudnn_op.cu.cc index f6e241af0634650f4a32be6a4547617f8ec3ee60..ad3e5543f10ae05865565110ba2231c897c205b8 100644 --- a/paddle/fluid/operators/softmax_cudnn_op.cu.cc +++ b/paddle/fluid/operators/softmax_cudnn_op.cu.cc @@ -80,4 +80,5 @@ REGISTER_OP_KERNEL(softmax, CUDNN, plat::CUDAPlace, ops::SoftmaxCUDNNKernel); REGISTER_OP_KERNEL(softmax_grad, CUDNN, plat::CUDAPlace, ops::SoftmaxGradCUDNNKernel, - ops::SoftmaxGradCUDNNKernel); + ops::SoftmaxGradCUDNNKernel, + ops::SoftmaxGradCUDNNKernel); diff --git a/paddle/fluid/operators/softmax_op.cu.cc b/paddle/fluid/operators/softmax_op.cu.cc index 5fb4f011d9b47cebc4a23bcce47eada825263343..19359b7eef5126d84f0707d39095a74ae4561186 100644 --- a/paddle/fluid/operators/softmax_op.cu.cc +++ b/paddle/fluid/operators/softmax_op.cu.cc @@ -23,4 +23,5 @@ REGISTER_OP_CUDA_KERNEL( ops::SoftmaxKernel); REGISTER_OP_CUDA_KERNEL( softmax_grad, ops::SoftmaxGradKernel, - ops::SoftmaxGradKernel); + ops::SoftmaxGradKernel, + ops::SoftmaxGradKernel); diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cc b/paddle/fluid/operators/softmax_with_cross_entropy_op.cc index 1a9324ec862fc3dd7ce669c5fed94527cac22b8f..2900221485e6ec097796ac38936ce31f8382c86a 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cc +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cc @@ -44,6 +44,12 @@ class SoftmaxWithCrossEntropyOpMaker "(bool, default: false), A flag to indicate whether to interpretate " "the given labels as soft labels.") .SetDefault(false); + AddAttr( + "numeric_stable_mode", + "(bool, default: false), A flag to indicate whether to use more " + "numerically stable algorithm. This flag is only valid when " + "soft_label is false and GPU is used.") + .SetDefault(false); AddAttr( "ignore_index", "(int, default -100), Specifies a target value that is ignored and" diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu index a07c17348ebb3f768d1c8be65c2d31e3c130bd23..6d48796191dd13a45f0c7267bfaf05489f528a9d 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu @@ -17,6 +17,7 @@ limitations under the License. */ #include #include "paddle/fluid/operators/math/cross_entropy.h" #include "paddle/fluid/operators/softmax_with_cross_entropy_op.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { @@ -117,8 +118,8 @@ using BlockReduceTempStorage = typename BlockReduce::TempStorage; // Make sure that BlockDim <= feature_size // This kernel is used to calculate the max element of each row template -__global__ void RowReductionForMax(const T* logits_data, T* max_data, - int feature_size) { +static __global__ void RowReductionForMax(const T* logits_data, T* max_data, + int feature_size) { __shared__ BlockReduceTempStorage temp_storage; auto beg_idx = feature_size * blockIdx.x + threadIdx.x; @@ -141,9 +142,10 @@ __global__ void RowReductionForMax(const T* logits_data, T* max_data, } // Make sure that BlockDim <= feature_size -template -__global__ void RowReductionForDiffMaxSum(const T* logits_data, T* max_data, - T* softmax, int feature_size) { +template +static __global__ void RowReductionForDiffMaxSum(const T* logits_data, + T* max_data, T* softmax, + int feature_size) { __shared__ BlockReduceTempStorage temp_storage; auto beg_idx = feature_size * blockIdx.x + threadIdx.x; @@ -153,24 +155,34 @@ __global__ void RowReductionForDiffMaxSum(const T* logits_data, T* max_data, softmax[beg_idx] = logits_data[beg_idx] - block_max; T diff_max_sum = real_exp(softmax[beg_idx]); - beg_idx += BlockDim; - while (beg_idx < end_idx) { - softmax[beg_idx] = logits_data[beg_idx] - block_max; - diff_max_sum += real_exp(softmax[beg_idx]); - beg_idx += BlockDim; + auto idx = beg_idx + BlockDim; + while (idx < end_idx) { + softmax[idx] = logits_data[idx] - block_max; + diff_max_sum += real_exp(softmax[idx]); + idx += BlockDim; } diff_max_sum = BlockReduce(temp_storage).Reduce(diff_max_sum, cub::Sum()); if (threadIdx.x == 0) max_data[blockIdx.x] = real_log(diff_max_sum); + + if (!CalculateLogSoftmax) return; + __syncthreads(); + diff_max_sum = max_data[blockIdx.x]; + softmax[beg_idx] -= diff_max_sum; + beg_idx += BlockDim; + while (beg_idx < end_idx) { + softmax[beg_idx] -= diff_max_sum; + beg_idx += BlockDim; + } + if (threadIdx.x == 0) max_data[blockIdx.x] = 0; } // Make sure that BlockDim <= feature_size template -__global__ void RowReductionForSoftmaxAndCrossEntropy(const T* logits_data, - const T* labels_data, - T* loss_data, T* softmax, - int feature_size) { +static __global__ void RowReductionForSoftmaxAndCrossEntropy( + const T* logits_data, const T* labels_data, T* loss_data, T* softmax, + int feature_size) { __shared__ BlockReduceTempStorage temp_storage; auto beg_idx = feature_size * blockIdx.x + threadIdx.x; @@ -194,11 +206,134 @@ __global__ void RowReductionForSoftmaxAndCrossEntropy(const T* logits_data, } template -__global__ void SetSoftmaxToOneWhenFeatureSizeIsOne(T* out, int batch_size) { +struct HardLabelSoftmaxWithCrossEntropyFunctor { + public: + HardLabelSoftmaxWithCrossEntropyFunctor(const T* logits, + const int64_t* labels, T* loss, + T* log_softmax, int feature_size) + : logits_(logits), + labels_(labels), + loss_(loss), + log_softmax_(log_softmax), + feature_size_(feature_size) {} + + __device__ void operator()(int idx) const { + auto row_idx = idx / feature_size_; + auto col_idx = idx % feature_size_; + if (col_idx != labels_[row_idx]) { + log_softmax_[idx] = real_exp(log_softmax_[idx]); + } else { + auto softmax = log_softmax_[idx]; + log_softmax_[idx] = real_exp(softmax); + loss_[row_idx] = -softmax; + } + } + + private: + const T* logits_; + const int64_t* labels_; + T* loss_; + T* log_softmax_; + int feature_size_; +}; + +template +struct HardLabelSoftmaxWithCrossEntropyFunctorWithIgnoreIdx { + public: + HardLabelSoftmaxWithCrossEntropyFunctorWithIgnoreIdx(const T* logits, + const int64_t* labels, + T* loss, T* log_softmax, + int feature_size, + int ignore_idx) + : logits_(logits), + labels_(labels), + loss_(loss), + log_softmax_(log_softmax), + feature_size_(feature_size), + ignore_idx_(ignore_idx) {} + + __device__ void operator()(int idx) const { + auto row_idx = idx / feature_size_; + auto col_idx = idx % feature_size_; + if (col_idx != labels_[row_idx] || col_idx == ignore_idx_) { + log_softmax_[idx] = real_exp(log_softmax_[idx]); + } else { + auto softmax = log_softmax_[idx]; + log_softmax_[idx] = real_exp(softmax); + loss_[row_idx] = -softmax; + } + } + + private: + const T* logits_; + const int64_t* labels_; + T* loss_; + T* log_softmax_; + int feature_size_; + int ignore_idx_; +}; + +template +static __global__ void SetSoftmaxToOneWhenFeatureSizeIsOne(T* out, + int batch_size) { auto idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < batch_size) out[idx] = static_cast(1); } +template +static void HardLabelSoftmaxWithCrossEntropy( + const platform::CUDADeviceContext& ctx, const T* logits_data, + const int64_t* labels_data, T* loss_data, T* softmax_data, int batch_size, + int feature_size, int ignore_idx) { + constexpr int kMaxBlockDim = 512; + int block_dim = feature_size >= kMaxBlockDim + ? kMaxBlockDim + : (1 << static_cast(std::log2(feature_size))); + auto stream = ctx.stream(); + +#define CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(BlockDim) \ + case BlockDim: { \ + RowReductionForMax<<>>( \ + logits_data, loss_data, feature_size); \ + RowReductionForDiffMaxSum<<>>( \ + logits_data, loss_data, softmax_data, feature_size); \ + platform::ForRange for_range( \ + ctx, batch_size* feature_size); \ + if (ignore_idx >= 0 && ignore_idx < feature_size) { \ + for_range(HardLabelSoftmaxWithCrossEntropyFunctorWithIgnoreIdx( \ + logits_data, labels_data, loss_data, softmax_data, feature_size, \ + ignore_idx)); \ + } else { \ + for_range(HardLabelSoftmaxWithCrossEntropyFunctor( \ + logits_data, labels_data, loss_data, softmax_data, feature_size)); \ + } \ + } break + + switch (block_dim) { + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(512); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(256); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(128); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(64); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(32); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(16); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(8); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(4); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(2); + case 1: + SetSoftmaxToOneWhenFeatureSizeIsOne<<<(batch_size + kMaxBlockDim - 1) / + kMaxBlockDim, + kMaxBlockDim, 0, stream>>>( + softmax_data, batch_size); + cudaMemsetAsync(loss_data, 0, batch_size * sizeof(T), stream); + break; + default: + PADDLE_THROW("BlockDim must be 2^n in softmax_with_cross_entropy_op"); + break; + } +#undef CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL +} + template static void SoftmaxWithCrossEntropyFusedKernel(const T* logits_data, const T* labels_data, @@ -237,7 +372,7 @@ static void SoftmaxWithCrossEntropyFusedKernel(const T* logits_data, kMaxBlockDim, kMaxBlockDim, 0, stream>>>( softmax_data, batch_size); - cudaMemsetAsync(loss_data, 0, batch_size, stream); + cudaMemsetAsync(loss_data, 0, batch_size * sizeof(T), stream); break; default: PADDLE_THROW("BlockDim must be 2^n in softmax_with_cross_entropy_op"); @@ -272,11 +407,21 @@ class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel { logits_data, labels_data, softmax_data, loss_data, batch_size, feature_size, context.cuda_device_context().stream()); } else { - math::SoftmaxCUDNNFunctor()(context.cuda_device_context(), logits, - softmax); - math::CrossEntropyFunctor()( - context.cuda_device_context(), loss, softmax, labels, false, - ignore_index); + if (!context.Attr("numeric_stable_mode")) { + math::SoftmaxCUDNNFunctor()(context.cuda_device_context(), logits, + softmax); + math::CrossEntropyFunctor()( + context.cuda_device_context(), loss, softmax, labels, false, + ignore_index); + } else { + int batch_size = logits->dims()[0]; + int feature_size = logits->dims()[1]; + auto* logits_data = logits->data(); + auto* labels_data = labels->data(); + HardLabelSoftmaxWithCrossEntropy( + context.cuda_device_context(), logits_data, labels_data, loss_data, + softmax_data, batch_size, feature_size, ignore_index); + } } } }; diff --git a/paddle/fluid/operators/split_ids_op.cc b/paddle/fluid/operators/split_ids_op.cc index c867c46873ae7ddbdbda280351e4ab28235bcc08..243f81e296fb95a2c7e9f717950b8a958ad98852 100644 --- a/paddle/fluid/operators/split_ids_op.cc +++ b/paddle/fluid/operators/split_ids_op.cc @@ -20,20 +20,27 @@ namespace operators { class SplitIdsOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); - AddOutput("Out", "(LoDTensor) The outputs of the input Ids.") + AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}") + .AsDuplicable(); + + AddOutput("Out", "(LoDTensors) The outputs of the input Ids.") .AsDuplicable(); AddComment(R"DOC( Split a LoDTensor of Ids into multi LoDTensors, the number is pserver's number Example: Input: - X = [1,2,3,4,5,6] + X = [[1,2,3,4,5,6],[2,3]] Out(3 output): - out0 = [3, 6] - out1 = [1, 4] - out2 = [2, 5] + if compress is True: + out0 = [3, 3, 6] + out1 = [1, 4] + out2 = [2, 2, 5] + else: + out0 = [3, 6] + out1 = [1, 4] + out2 = [2, 5] )DOC"); } }; @@ -43,16 +50,24 @@ class SplitIdsOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Ids"), "SplitIdsOp must has input Ids."); + PADDLE_ENFORCE(ctx->HasInputs("Ids"), "SplitIdsOp must has input Ids."); PADDLE_ENFORCE(ctx->HasOutputs("Out"), "SplitIdsOp must has output Out."); auto ids_var_type = ctx->GetInputsVarType("Ids").front(); - auto ids_dims = ctx->GetInputDim("Ids"); + auto ids_dims = ctx->GetInputsDim("Ids"); if (ids_var_type == framework::proto::VarType::LOD_TENSOR) { - PADDLE_ENFORCE_EQ(ids_dims.size(), 2); - PADDLE_ENFORCE_EQ(ids_dims[1], 1); + PADDLE_ENFORCE_EQ(ids_dims[0].size(), 2); } } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType( + ctx.MultiInput("Ids").front()->type()), + ctx.GetPlace()); + } }; class SplitIdsOpInferVarType : public framework::VarTypeInference { @@ -66,12 +81,28 @@ class SplitIdsOpInferVarType : public framework::VarTypeInference { } }; +class SplitIdsOpGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto grad = new framework::OpDesc(); + grad->SetType("concat"); + grad->SetInput("X", OutputGrad("Out")); + grad->SetOutput("Out", InputGrad("Ids")); + grad->SetAttr("axis", 0); + return std::unique_ptr(grad); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(split_ids, ops::SplitIdsOp, ops::SplitIdsOpMaker, - ops::SplitIdsOpInferVarType); + ops::SplitIdsOpGradMaker, ops::SplitIdsOpInferVarType); + REGISTER_OP_CPU_KERNEL( split_ids, ops::SplitIdsOpKernel, ops::SplitIdsOpKernel); diff --git a/paddle/fluid/operators/split_ids_op.h b/paddle/fluid/operators/split_ids_op.h index c4af5a65fc5f81c1af7c1fdcca637ca37c940637..69ac6c5a6b9a8b318520eb9a3ff89a3a6be48339 100644 --- a/paddle/fluid/operators/split_ids_op.h +++ b/paddle/fluid/operators/split_ids_op.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once +#include +#include #include #include #include "paddle/fluid/framework/op_registry.h" @@ -31,19 +33,39 @@ class SplitIdsOpKernel : public framework::OpKernel { PADDLE_THROW("SplitIds do not support GPU kernel"); } - const auto *ids_var = ctx.InputVar("Ids"); + const auto ids_vars = ctx.MultiInputVar("Ids"); + + PADDLE_ENFORCE_GT(ids_vars.size(), 0, "The number of Ids should > 0"); + auto *ids_var = ids_vars[0]; + if (ids_var->IsType()) { - const auto &ids_dims = ctx.Input("Ids")->dims(); - const T *ids = ctx.Input("Ids")->data(); + int batch_size = 0; + const auto ids_tensors = ctx.MultiInput("Ids"); + for (size_t i = 0; i < ids_tensors.size(); ++i) { + batch_size += ids_tensors[i]->dims()[0]; + } + VLOG(4) << "Get Total BatchSize is: " << batch_size; + + std::vector all_ids(batch_size); + int offset = 0; + for (size_t i = 0; i < ids_tensors.size(); ++i) { + const auto *ids = ids_tensors[i]; + std::memcpy(all_ids.data() + offset, ids->data(), + ids->numel() * sizeof(T)); + offset += ids->numel(); + } + + std::set st(all_ids.begin(), all_ids.end()); + all_ids.assign(st.begin(), st.end()); + auto outs = ctx.MultiOutput("Out"); const size_t shard_num = outs.size(); - std::vector> out_ids; out_ids.resize(outs.size()); // split id by their shard_num. - for (int i = 0; i < ids_dims[0]; ++i) { - T id = ids[i]; + for (int i = 0; i < all_ids.size(); ++i) { + T id = all_ids[i]; size_t shard_id = static_cast(id) % shard_num; out_ids[shard_id].push_back(id); } @@ -64,7 +86,7 @@ class SplitIdsOpKernel : public framework::OpKernel { PADDLE_ENFORCE_EQ(ids_dims[0], static_cast(ids_selected_rows->rows().size()), ""); - const T *ids = ids_selected_rows->value().data(); + const T *ids_data = ids_selected_rows->value().data(); const auto &ids_rows = ids_selected_rows->rows(); auto outs = ctx.MultiOutput("Out"); const size_t shard_num = outs.size(); @@ -87,7 +109,7 @@ class SplitIdsOpKernel : public framework::OpKernel { T *output = out->mutable_value()->mutable_data(ddim, place); for (int64_t i = 0; i < ddim[0]; ++i) { memcpy(output + i * row_width, - ids + id_to_index[out->rows()[i]] * row_width, + ids_data + id_to_index[out->rows()[i]] * row_width, row_width * sizeof(T)); } } diff --git a/paddle/fluid/operators/split_selected_rows_op.cc b/paddle/fluid/operators/split_selected_rows_op.cc index 76615a9405d7a8e3fa9dba8d01a956209e02ae8f..0e7b1463d1ba81aed53e0e3f3a90d2a1fbf0ffbc 100644 --- a/paddle/fluid/operators/split_selected_rows_op.cc +++ b/paddle/fluid/operators/split_selected_rows_op.cc @@ -22,9 +22,9 @@ class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "The input SelectedRows."); AddOutput("Out", "The outputs of the input SelectedRows.").AsDuplicable(); - AddAttr>("height_sections", - "Height for each output SelectedRows.") - .SetDefault(std::vector({})); + AddAttr>("height_sections", + "Height for each output SelectedRows.") + .SetDefault(std::vector({})); AddComment(R"DOC( Split a SelectedRows with a specified rows section. diff --git a/paddle/fluid/operators/split_selected_rows_op.h b/paddle/fluid/operators/split_selected_rows_op.h index 0e9ce165b98845f4745ee70b028513ea31cc6657..af64607fafc6544047714e731846a2440be219b8 100644 --- a/paddle/fluid/operators/split_selected_rows_op.h +++ b/paddle/fluid/operators/split_selected_rows_op.h @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { namespace operators { -static int FindOutIdx(int row, const std::vector& abs_sections) { +static int FindOutIdx(int row, const std::vector& abs_sections) { for (size_t i = 1; i < abs_sections.size(); ++i) { if (row < abs_sections[i]) { return i - 1; @@ -30,9 +30,9 @@ static int FindOutIdx(int row, const std::vector& abs_sections) { return abs_sections.size() - 1; } -static std::vector ToAbsoluteSection( - const std::vector& height_sections) { - std::vector abs_sections; +static std::vector ToAbsoluteSection( + const std::vector& height_sections) { + std::vector abs_sections; abs_sections.resize(height_sections.size()); abs_sections[0] = 0; for (size_t i = 1; i < height_sections.size(); ++i) { @@ -47,7 +47,7 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* x = ctx.Input("X"); auto outs = ctx.MultiOutput("Out"); - auto height_sections = ctx.Attr>("height_sections"); + auto height_sections = ctx.Attr>("height_sections"); auto abs_sections = ToAbsoluteSection(height_sections); diff --git a/paddle/fluid/operators/spp_op.h b/paddle/fluid/operators/spp_op.h index 08cb7849d20443862b66ea6096c095b294c7242c..35d9737ee01fe1505cbe30e8ed735e6b92cb8df2 100644 --- a/paddle/fluid/operators/spp_op.h +++ b/paddle/fluid/operators/spp_op.h @@ -56,12 +56,14 @@ class SppKernel : public framework::OpKernel { math::Pool2dFunctor, T> pool_forward; math::MaxPool max_process; pool_forward(context.template device_context(), *in_x, - kernel_size, strides, paddings, max_process, &out_level); + kernel_size, strides, paddings, max_process, true, + &out_level); } else if (pooling_type == "avg") { math::Pool2dFunctor, T> pool_forward; math::AvgPool avg_process; pool_forward(context.template device_context(), *in_x, - kernel_size, strides, paddings, avg_process, &out_level); + kernel_size, strides, paddings, avg_process, true, + &out_level); } // flatten pooling output shape int output_flatten_w = in_x->dims()[1] * bins * bins; @@ -154,7 +156,7 @@ class SppGradKernel : public framework::OpKernel { math::AvgPoolGrad avg_process; pool_backward(context.template device_context(), *in_x, *&out_level, *&outgrad_level, kernel_size, strides, - paddings, avg_process, in_x_grad); + paddings, avg_process, true, in_x_grad); } } } diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc index 34dbac2ab8dcc9bd2b91e2daa2f42806057f5f56..d19ac9839c90a116265b761e3b1b3f855e2d95e8 100644 --- a/paddle/fluid/operators/sum_op.cc +++ b/paddle/fluid/operators/sum_op.cc @@ -67,6 +67,7 @@ class SumOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { auto x_vars = ctx.MultiInputVar("X"); + auto x_vars_name = ctx.Inputs("X"); framework::LibraryType library{framework::LibraryType::kPlain}; framework::DataLayout layout{framework::DataLayout::kAnyLayout}; @@ -81,15 +82,18 @@ class SumOp : public framework::OperatorWithKernel { if (x_vars[0]->IsType()) { int dtype = -1; - for (auto& x_var : x_vars) { - auto& lod_tensor = x_var->Get(); - if (lod_tensor.numel() == 0) { + for (size_t idx = 0; idx < x_vars.size(); ++idx) { + PADDLE_ENFORCE(x_vars[idx] != nullptr, + "Input var[%s] should not be nullptr", x_vars_name[idx]); + // FIXME(zcd): The input x_var may be SelectedRows or LoDTensor. + auto tensor = framework::GetTensorFromVar(*x_vars[idx]); + if (tensor->numel() == 0) { continue; } if (dtype == -1) { - dtype = framework::ToDataType(lod_tensor.type()); + dtype = framework::ToDataType(tensor->type()); } else { - PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(lod_tensor.type())); + PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(tensor->type())); } } PADDLE_ENFORCE_NE(dtype, -1, diff --git a/paddle/fluid/operators/sum_op.cu b/paddle/fluid/operators/sum_op.cu index 89bcd1bbc86dc29cb7b98cbef3057a8f98c74555..db4c2d6c115f04b436db00854ca4b02fea09866b 100644 --- a/paddle/fluid/operators/sum_op.cu +++ b/paddle/fluid/operators/sum_op.cu @@ -11,10 +11,13 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/sum_op.h" +#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( sum, ops::SumKernel, ops::SumKernel, ops::SumKernel, - ops::SumKernel); + ops::SumKernel, + ops::SumKernel); diff --git a/paddle/fluid/operators/sum_op.h b/paddle/fluid/operators/sum_op.h index 11987c61aebaad00f8a71f1b909c83c44ddc8b0e..19b2c68c823adbed82319f7b04992baedd5d41f9 100644 --- a/paddle/fluid/operators/sum_op.h +++ b/paddle/fluid/operators/sum_op.h @@ -61,7 +61,7 @@ class SumKernel : public framework::OpKernel { if (start != 2) { math::SetConstant constant_functor; constant_functor(context.template device_context(), - out, 0.0); + out, static_cast(0)); } } @@ -83,79 +83,54 @@ class SumKernel : public framework::OpKernel { } } } else if (out_var->IsType()) { - std::unique_ptr in0; - if (in_place) { - // If is in_place, we store the input[0] to in0 - auto &in_sel0 = in_vars[0]->Get(); - auto &rows = in_sel0.rows(); -#ifdef PADDLE_WITH_CUDA - std::vector rows_in_cpu; - rows_in_cpu.reserve(rows.size()); - for (auto item : rows) { - rows_in_cpu.push_back(item); - } - in0.reset(new framework::SelectedRows(rows_in_cpu, in_sel0.height())); -#else - in0.reset(new framework::SelectedRows(rows, in_sel0.height())); -#endif - in0->mutable_value()->ShareDataWith(in_sel0.value()); + if (in_place && in_vars.size() < 2) { + return; } - auto get_selected_row = [&](size_t i) -> const SelectedRows & { - if (i == 0 && in0) { - return *in0.get(); - } else { - return in_vars[i]->Get(); + std::vector inputs; + SelectedRows temp_in0; + + if (in_place) { + auto &in0 = in_vars[0]->Get(); + temp_in0.set_height(in0.height()); + temp_in0.set_rows(in0.rows()); + framework::TensorCopy(in0.value(), in0.place(), + context.device_context(), + temp_in0.mutable_value()); + inputs.push_back(&temp_in0); + for (size_t i = 1; i < in_vars.size(); ++i) { + auto &in = in_vars[i]->Get(); + if (in.rows().size() > 0) { + inputs.push_back(&in); + } + } + } else { + for (auto &in_var : in_vars) { + auto &in = in_var->Get(); + if (in.rows().size() > 0) { + inputs.push_back(&in_var->Get()); + } } - }; + } auto *out = context.Output("Out"); out->mutable_rows()->clear(); - auto *out_value = out->mutable_value(); - - // Runtime InferShape - size_t first_dim = 0; - for (size_t i = 0; i < in_num; i++) { - auto &sel_row = get_selected_row(i); - first_dim += sel_row.rows().size(); - } - std::vector in_dim; - for (size_t i = 0; i < in_num; i++) { - auto &sel_row = get_selected_row(i); - if (sel_row.rows().size() > 0) { - in_dim = framework::vectorize(sel_row.value().dims()); + bool has_data = false; + for (auto &in : inputs) { + if (in->rows().size() > 0) { + has_data = true; break; } } - if (in_dim.empty()) { - VLOG(3) << "WARNING: all the inputs are empty"; - in_dim = - framework::vectorize(get_selected_row(in_num - 1).value().dims()); + if (has_data) { + math::scatter::MergeAdd merge_add; + merge_add(context.template device_context(), inputs, + out); } else { - in_dim[0] = static_cast(first_dim); - } - - out_value->Resize(framework::make_ddim(in_dim)); - out_value->mutable_data(context.GetPlace()); - // if all the input sparse vars are empty, no need to - // merge these vars. - if (first_dim == 0UL) { - return; - } - - math::SelectedRowsAddTo functor; - - int64_t offset = 0; - for (size_t i = 0; i < in_num; i++) { - auto &sel_row = get_selected_row(i); - if (sel_row.rows().size() == 0) { - continue; - } - PADDLE_ENFORCE_EQ(out->height(), sel_row.height()); - functor(context.template device_context(), sel_row, - offset, out); - offset += sel_row.value().numel(); + // no data, just set a empty out tensor. + out->mutable_value()->mutable_data(framework::make_ddim({0}), + context.GetPlace()); } } else if (out_var->IsType()) { auto &out_array = *out_var->GetMutable(); diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index d4ba0f9c33c91811647f9d19a332f139c16b0eb2..673f86da76ee0712b4d941f5b33594f89926b973 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -223,7 +223,9 @@ class TensorRTEngineKernel : public framework::OpKernel { // Add outputs for (auto& output : output_maps) { - engine->DeclareOutput(output); + if (!engine->HasDeclared(output)) { + engine->DeclareOutput(output); + } } engine->FreezeNetwork(); diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc index e2b7b6b8e447381229e4ad594b7974bc0aa159d5..b5426e17aac19dc07ee56545fac8472d9ef0d93c 100644 --- a/paddle/fluid/operators/test_send_nccl_id.cc +++ b/paddle/fluid/operators/test_send_nccl_id.cc @@ -92,7 +92,7 @@ TEST(SendNcclId, RPCServer) { std::string ep = string::Sprintf("127.0.0.1:%d", port); distributed::RPCClient* client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(0); LOG(INFO) << "connect to server" << ep; client->AsyncSendVar(ep, dev_ctx, scope, NCCL_ID_VARNAME); diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index aa907595cb7cf165974caa69fe8eb0370471732d..e3132ae76f624f3338d749e4fcebbd0ecd7ffe79 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -29,7 +29,7 @@ class CPUUniformRandomKernel : public framework::OpKernel { if (out_var->IsType()) { tensor = out_var->GetMutable(); } else if (out_var->IsType()) { - auto shape = ctx.Attr>("shape"); + auto shape = ctx.Attr>("shape"); auto *selected_rows = out_var->GetMutable(); tensor = selected_rows->mutable_value(); tensor->Resize(framework::make_ddim(shape)); @@ -67,7 +67,7 @@ class UniformRandomOp : public framework::OperatorWithKernel { 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) { @@ -94,7 +94,7 @@ This operator initializes a tensor with random values sampled from a uniform distribution. The random result is in set [min, max]. )DOC"); - AddAttr>("shape", "The shape of the output tensor"); + AddAttr>("shape", "The shape of the output tensor"); AddAttr("min", "Minimum value of uniform random. [default -1.0].") .SetDefault(-1.0f); AddAttr("max", "Maximun value of uniform random. [default 1.0].") diff --git a/paddle/fluid/operators/uniform_random_op.cu b/paddle/fluid/operators/uniform_random_op.cu index bbb692b0ddfc18e8a62c0d2a6bac88f9932f6704..2bb0ecc139f7096d1b61150e0a2d4fb095338749 100644 --- a/paddle/fluid/operators/uniform_random_op.cu +++ b/paddle/fluid/operators/uniform_random_op.cu @@ -48,7 +48,7 @@ class GPUUniformRandomKernel : public framework::OpKernel { if (out_var->IsType()) { tensor = out_var->GetMutable(); } else if (out_var->IsType()) { - auto shape = context.Attr>("shape"); + auto shape = context.Attr>("shape"); tensor = out_var->GetMutable()->mutable_value(); tensor->Resize(framework::make_ddim(shape)); } else { diff --git a/paddle/fluid/platform/cudnn_helper.h b/paddle/fluid/platform/cudnn_helper.h index bb8b14bb9fa41942c3aa653ca224c0842fbf9a00..07bb02be1962f758e50cab1f27de43e89f3953c3 100644 --- a/paddle/fluid/platform/cudnn_helper.h +++ b/paddle/fluid/platform/cudnn_helper.h @@ -76,8 +76,9 @@ enum class DataLayout { // Not use enum class PoolingMode { kMaximum, - kAverage, kMaximumDeterministic, + kAverageExclusive, + kAverageInclusive, }; #if CUDNN_VERSION < 6000 @@ -91,8 +92,10 @@ inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) { switch (mode) { case PoolingMode::kMaximumDeterministic: return CUDNN_POOLING_MAX; - case PoolingMode::kAverage: + case PoolingMode::kAverageExclusive: return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + case PoolingMode::kAverageInclusive: + return CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING; case PoolingMode::kMaximum: return CUDNN_POOLING_MAX; default: @@ -105,8 +108,10 @@ inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) { switch (mode) { case PoolingMode::kMaximumDeterministic: return CUDNN_POOLING_MAX_DETERMINISTIC; - case PoolingMode::kAverage: + case PoolingMode::kAverageExclusive: return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + case PoolingMode::kAverageInclusive: + return CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING; case PoolingMode::kMaximum: return CUDNN_POOLING_MAX; default: @@ -341,6 +346,28 @@ class ScopedPoolingDescriptor { DISABLE_COPY_AND_ASSIGN(ScopedPoolingDescriptor); }; +class ScopedSpatialTransformerDescriptor { + public: + ScopedSpatialTransformerDescriptor() { + PADDLE_ENFORCE(dynload::cudnnCreateSpatialTransformerDescriptor(&desc_)); + } + ~ScopedSpatialTransformerDescriptor() { + PADDLE_ENFORCE(dynload::cudnnDestroySpatialTransformerDescriptor(desc_)); + } + + template + inline cudnnSpatialTransformerDescriptor_t descriptor(const int nbDims, + const int dimA[]) { + PADDLE_ENFORCE(dynload::cudnnSetSpatialTransformerNdDescriptor( + desc_, CUDNN_SAMPLER_BILINEAR, CudnnDataType::type, nbDims, dimA)); + return desc_; + } + + private: + cudnnSpatialTransformerDescriptor_t desc_; + DISABLE_COPY_AND_ASSIGN(ScopedSpatialTransformerDescriptor); +}; + inline bool CanCUDNNBeUsed(const framework::ExecutionContext& ctx) { bool use_cudnn = ctx.Attr("use_cudnn"); use_cudnn &= paddle::platform::is_gpu_place(ctx.GetPlace()); diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index 38cfe6cf62724f4ac9fa23ae9211055a8504b3f0..36e7f293482899168b00815c73aea9cea461bed5 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -32,23 +32,25 @@ platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) { "'Place' is not supported, Please re-compile with WITH_GPU " "option"); } - return it->second.get(); + return it->second.get().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; +template +inline void EmplaceDeviceContext( + std::map>>* + map_ptr, + platform::Place p) { + using PtrType = std::unique_ptr; + map_ptr->emplace(p, std::async(std::launch::deferred, [=] { + // lazy evaluation. i.e., only create device context at + // first `Get` + return PtrType(new DevCtx(boost::get(p))); + })); } DeviceContextPool::DeviceContextPool( const std::vector& places) { PADDLE_ENFORCE_GT(places.size(), 0); - using PtrType = std::unique_ptr; std::set set; for (auto& p : places) { set.insert(p); @@ -57,16 +59,13 @@ DeviceContextPool::DeviceContextPool( for (auto& p : set) { if (platform::is_cpu_place(p)) { #ifdef PADDLE_WITH_MKLDNN - device_contexts_.emplace( - p, PtrType(new MKLDNNDeviceContext(boost::get(p)))); + EmplaceDeviceContext(&device_contexts_, p); #else - device_contexts_.emplace( - p, PtrType(new CPUDeviceContext(boost::get(p)))); + EmplaceDeviceContext(&device_contexts_, p); #endif } else if (platform::is_gpu_place(p)) { #ifdef PADDLE_WITH_CUDA - device_contexts_.emplace( - p, PtrType(new CUDADeviceContext(boost::get(p)))); + EmplaceDeviceContext(&device_contexts_, p); #else PADDLE_THROW( "'CUDAPlace' is not supported, Please re-compile with WITH_GPU " @@ -74,9 +73,8 @@ DeviceContextPool::DeviceContextPool( #endif } else if (platform::is_cuda_pinned_place(p)) { #ifdef PADDLE_WITH_CUDA - device_contexts_.emplace( - p, - PtrType(new CUDAPinnedDeviceContext(boost::get(p)))); + EmplaceDeviceContext( + &device_contexts_, p); #else PADDLE_THROW( "'CUDAPlace' is not supported, Please re-compile with WITH_GPU " @@ -161,65 +159,31 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface { allocations_; }; -class CudnnHolder { - public: - CudnnHolder(const cudaStream_t* stream, const CUDAPlace& place) - : workspace_(nullptr), stream_(stream), place_(place) { - PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_)); - PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, *stream_)); - } - - cudnnHandle_t cudnn_handle() const { return cudnn_handle_; } +CudnnHolder::CudnnHolder(const cudaStream_t* stream, const CUDAPlace& place) + : workspace_(nullptr), workspace_len_(0), stream_(stream), place_(place) { + PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_)); + PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, *stream_)); +} - void RunFunc(const std::function& cudnn_func, - size_t required_workspace_len) { - std::lock_guard lock(mtx_); - if (required_workspace_len > WorkspaceSize()) { - ReallocateWorkspace(required_workspace_len); - } - cudnn_func(WorkspacePtr()); - } - - ~CudnnHolder() { PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_)); } - - private: - size_t WorkspaceSize() const { - if (workspace_ == nullptr) { - return 0; - } else { - return workspace_->size(); - } +CudnnHolder::~CudnnHolder() { + PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_)); + if (workspace_ != nullptr) { + paddle::memory::Free(place_, workspace_); } +} - void* WorkspacePtr() const { - if (workspace_ == nullptr) { - return nullptr; - } else { - return workspace_->ptr(); - } +void CudnnHolder::ReallocateWorkspace(size_t required_workspace_len) { + if (required_workspace_len <= workspace_len_) { + return; } - - void ReallocateWorkspace(size_t required_workspace_len) { - if (required_workspace_len <= WorkspaceSize()) { - return; - } - if (workspace_ != nullptr) { - // Maybe someone is using the current workspace - PADDLE_ENFORCE(cudaStreamSynchronize(*stream_)); - workspace_.reset(); - } - workspace_ = paddle::memory::Alloc(place_, required_workspace_len, - memory::Allocator::kFluxHuge); + if (workspace_ != nullptr) { + // Maybe someone is using the current workspace + PADDLE_ENFORCE(cudaStreamSynchronize(*stream_)); + paddle::memory::Free(place_, workspace_); } - - cudnnHandle_t cudnn_handle_; - std::unique_ptr workspace_; - - const cudaStream_t* stream_; // not owned; - const CUDAPlace place_; - - std::mutex mtx_; -}; + workspace_ = paddle::memory::Alloc(place_, required_workspace_len); + workspace_len_ = required_workspace_len; +} CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place), cudnn_holder_(nullptr) { @@ -240,12 +204,12 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace 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; + LOG_FIRST_N(WARNING, 1) << "Please NOTE: 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_)); } @@ -287,9 +251,8 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_holder_->cudnn_handle(); } -void CUDADeviceContext::RunCudnnFuncWithWorkspace( - const std::function& cudnn_func, size_t workspace_len) const { - cudnn_holder_->RunFunc(cudnn_func, workspace_len); +CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const { + return CudnnWorkspaceHandle(cudnn_holder_.get()); } cudaStream_t CUDADeviceContext::stream() const { return stream_; } diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h index 942e13a724339dc85ed1fc72c11e208ddce36dbb..df248f9bb15591d5015ad01278797ec7e31ef9d1 100644 --- a/paddle/fluid/platform/device_context.h +++ b/paddle/fluid/platform/device_context.h @@ -10,6 +10,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include // NOLINT #include #include // NOLINT #include @@ -72,7 +73,60 @@ struct DefaultDeviceContextType { #ifdef PADDLE_WITH_CUDA class EigenCudaStreamDevice; -class CudnnHolder; +class CudnnHolder { + public: + CudnnHolder(const cudaStream_t* stream, const CUDAPlace& place); + ~CudnnHolder(); + cudnnHandle_t cudnn_handle() const { return cudnn_handle_; } + + private: + friend class CudnnWorkspaceHandle; + void ReallocateWorkspace(size_t required_workspace_len); + + template + void RunFuncImpl(Callback&& cudnn_func, size_t required_workspace_len) { + if (required_workspace_len > workspace_len_) { + ReallocateWorkspace(required_workspace_len); + } + cudnn_func(workspace_); + } + + std::mutex& Mutex() { return mtx_; } + + cudnnHandle_t cudnn_handle_; + void* workspace_; + size_t workspace_len_; + + const cudaStream_t* stream_; // not owned; + const CUDAPlace place_; + + std::mutex mtx_; +}; + +class CudnnWorkspaceHandle { + public: + /*! \brief The lock would not be acquired when constructor calls. + * The lock would be acquired when RunFunc() is called first time. */ + inline explicit CudnnWorkspaceHandle(CudnnHolder* holder) : holder_(holder) {} + + /*! \brief Thread which call RunFunc() would acquire the lock first + * before invoking cudnn functions. */ + template + inline void RunFunc(Callback&& cudnn_func, size_t required_workspace_len) { + if (!guard_) { + guard_.reset(new std::lock_guard(holder_->Mutex())); + } + holder_->RunFuncImpl(std::forward(cudnn_func), + required_workspace_len); + } + + CudnnWorkspaceHandle(CudnnWorkspaceHandle&&) = default; + CudnnWorkspaceHandle& operator=(CudnnWorkspaceHandle&&) = delete; + + private: + CudnnHolder* holder_; // not own + std::unique_ptr> guard_; +}; class CUDADeviceContext : public DeviceContext { public: @@ -100,10 +154,14 @@ class CUDADeviceContext : public DeviceContext { /*! \brief Return cudnn handle in the device context. */ cudnnHandle_t cudnn_handle() const; - /*! \brief Run a cudnn function with the workspace provided by - * CUDADeviceContext */ - void RunCudnnFuncWithWorkspace(const std::function& cudnn_func, - size_t workspace_len) const; + /*! \brief Return a cudnn workspace handle to call multiple cudnn + * functions without interrupting by other threads. + * Once the first cudnn function is called by the handle, a lock + * would be acquired to prevent other threads from accessing the + * workspace. Once the handle is destructed, the lock would be released. + * CudnnWorkspaceHandle is an RAII object to implement thread-safe + * sequential cudnn function calls. */ + CudnnWorkspaceHandle cudnn_workspace_handle() const; /*! \brief Return cuda stream in the device context. */ cudaStream_t stream() const; @@ -223,9 +281,6 @@ 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) { @@ -237,7 +292,8 @@ class DeviceContextPool { private: static DeviceContextPool* pool; - std::map> device_contexts_; + std::map>> + device_contexts_; DISABLE_COPY_AND_ASSIGN(DeviceContextPool); }; diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h index e6353f67ef118072a2d8e49111e8ecc486589998..d3d754b6f58d25a9dfacafaf55d50b353a71ee6d 100644 --- a/paddle/fluid/platform/dynload/cudnn.h +++ b/paddle/fluid/platform/dynload/cudnn.h @@ -65,44 +65,51 @@ extern void EnforceCUDNNLoaded(const char* fn_name); * include all needed cudnn functions in HPPL * different cudnn version has different interfaces **/ -#define CUDNN_DNN_ROUTINE_EACH(__macro) \ - __macro(cudnnSetTensor4dDescriptor); \ - __macro(cudnnSetTensor4dDescriptorEx); \ - __macro(cudnnSetTensorNdDescriptor); \ - __macro(cudnnGetTensorNdDescriptor); \ - __macro(cudnnGetConvolutionNdForwardOutputDim); \ - __macro(cudnnGetConvolutionForwardAlgorithm); \ - __macro(cudnnCreateTensorDescriptor); \ - __macro(cudnnDestroyTensorDescriptor); \ - __macro(cudnnCreateFilterDescriptor); \ - __macro(cudnnSetFilter4dDescriptor); \ - __macro(cudnnSetFilterNdDescriptor); \ - __macro(cudnnGetFilterNdDescriptor); \ - __macro(cudnnSetPooling2dDescriptor); \ - __macro(cudnnSetPoolingNdDescriptor); \ - __macro(cudnnGetPoolingNdDescriptor); \ - __macro(cudnnDestroyFilterDescriptor); \ - __macro(cudnnCreateConvolutionDescriptor); \ - __macro(cudnnCreatePoolingDescriptor); \ - __macro(cudnnDestroyPoolingDescriptor); \ - __macro(cudnnSetConvolution2dDescriptor); \ - __macro(cudnnDestroyConvolutionDescriptor); \ - __macro(cudnnSetConvolutionNdDescriptor); \ - __macro(cudnnGetConvolutionNdDescriptor); \ - __macro(cudnnDeriveBNTensorDescriptor); \ - __macro(cudnnCreate); \ - __macro(cudnnDestroy); \ - __macro(cudnnSetStream); \ - __macro(cudnnActivationForward); \ - __macro(cudnnConvolutionForward); \ - __macro(cudnnConvolutionBackwardBias); \ - __macro(cudnnGetConvolutionForwardWorkspaceSize); \ - __macro(cudnnTransformTensor); \ - __macro(cudnnPoolingForward); \ - __macro(cudnnPoolingBackward); \ - __macro(cudnnSoftmaxBackward); \ - __macro(cudnnSoftmaxForward); \ - __macro(cudnnGetVersion); \ +#define CUDNN_DNN_ROUTINE_EACH(__macro) \ + __macro(cudnnSetTensor4dDescriptor); \ + __macro(cudnnSetTensor4dDescriptorEx); \ + __macro(cudnnSetTensorNdDescriptor); \ + __macro(cudnnGetTensorNdDescriptor); \ + __macro(cudnnGetConvolutionNdForwardOutputDim); \ + __macro(cudnnGetConvolutionForwardAlgorithm); \ + __macro(cudnnCreateTensorDescriptor); \ + __macro(cudnnDestroyTensorDescriptor); \ + __macro(cudnnCreateFilterDescriptor); \ + __macro(cudnnSetFilter4dDescriptor); \ + __macro(cudnnSetFilterNdDescriptor); \ + __macro(cudnnGetFilterNdDescriptor); \ + __macro(cudnnSetPooling2dDescriptor); \ + __macro(cudnnSetPoolingNdDescriptor); \ + __macro(cudnnGetPoolingNdDescriptor); \ + __macro(cudnnDestroyFilterDescriptor); \ + __macro(cudnnCreateConvolutionDescriptor); \ + __macro(cudnnCreatePoolingDescriptor); \ + __macro(cudnnDestroyPoolingDescriptor); \ + __macro(cudnnSetConvolution2dDescriptor); \ + __macro(cudnnDestroyConvolutionDescriptor); \ + __macro(cudnnSetConvolutionNdDescriptor); \ + __macro(cudnnGetConvolutionNdDescriptor); \ + __macro(cudnnDeriveBNTensorDescriptor); \ + __macro(cudnnCreateSpatialTransformerDescriptor); \ + __macro(cudnnSetSpatialTransformerNdDescriptor); \ + __macro(cudnnDestroySpatialTransformerDescriptor); \ + __macro(cudnnSpatialTfGridGeneratorForward); \ + __macro(cudnnSpatialTfGridGeneratorBackward); \ + __macro(cudnnSpatialTfSamplerForward); \ + __macro(cudnnSpatialTfSamplerBackward); \ + __macro(cudnnCreate); \ + __macro(cudnnDestroy); \ + __macro(cudnnSetStream); \ + __macro(cudnnActivationForward); \ + __macro(cudnnConvolutionForward); \ + __macro(cudnnConvolutionBackwardBias); \ + __macro(cudnnGetConvolutionForwardWorkspaceSize); \ + __macro(cudnnTransformTensor); \ + __macro(cudnnPoolingForward); \ + __macro(cudnnPoolingBackward); \ + __macro(cudnnSoftmaxBackward); \ + __macro(cudnnSoftmaxForward); \ + __macro(cudnnGetVersion); \ __macro(cudnnGetErrorString); CUDNN_DNN_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP) diff --git a/paddle/fluid/platform/init.cc b/paddle/fluid/platform/init.cc index 8612b88ef7a6c8c7a282eb8a6830687833d0aa9d..53a1cdad0cca72fa0835dae3bdd915d6ce2a87e8 100644 --- a/paddle/fluid/platform/init.cc +++ b/paddle/fluid/platform/init.cc @@ -119,21 +119,49 @@ void InitDevices(bool init_p2p, const std::vector devices) { platform::SetNumThreads(FLAGS_paddle_num_threads); #endif - if (platform::jit::MayIUse(platform::jit::avx512f)) { -#ifndef __AVX512F__ - LOG(WARNING) << "AVX512F is available, Please re-compile on local machine"; + if (platform::jit::MayIUse(platform::jit::avx)) { +#ifndef __AVX__ + LOG(WARNING) << "AVX is available, Please re-compile on local machine"; #endif } - if (platform::jit::MayIUse(platform::jit::avx2)) { -#ifndef __AVX2__ - LOG(WARNING) << "AVX2 is available, Please re-compile on local machine"; + +// Throw some informations when CPU instructions mismatch. +#define AVX_GUIDE(compiletime, runtime) \ + LOG(FATAL) \ + << "This version is compiled on higher instruction(" #compiletime \ + ") system, you may encounter illegal instruction error running on" \ + " your local CPU machine. Please reinstall the " #runtime \ + " version or compile from source code." + +#ifdef __AVX512F__ + if (!platform::jit::MayIUse(platform::jit::avx512f)) { + if (platform::jit::MayIUse(platform::jit::avx2)) { + AVX_GUIDE(AVX512, AVX2); + } else if (platform::jit::MayIUse(platform::jit::avx)) { + AVX_GUIDE(AVX512, AVX); + } else { + AVX_GUIDE(AVX512, NonAVX); + } + } #endif + +#ifdef __AVX2__ + if (!platform::jit::MayIUse(platform::jit::avx2)) { + if (platform::jit::MayIUse(platform::jit::avx)) { + AVX_GUIDE(AVX2, AVX); + } else { + AVX_GUIDE(AVX2, NonAVX); + } } - if (platform::jit::MayIUse(platform::jit::avx)) { -#ifndef __AVX__ - LOG(WARNING) << "AVX is available, Please re-compile on local machine"; #endif + +#ifdef __AVX__ + if (!platform::jit::MayIUse(platform::jit::avx)) { + AVX_GUIDE(AVX, NonAVX); } +#endif + +#undef AVX_GUIDE } void InitGLOG(const std::string &prog_name) { diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index c0a2543ba5d8ff8f34cb6231c51cb5053a6a9481..814012e6c1fad414d10f5a64af283bed57e11fe3 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -187,6 +187,29 @@ class MKLDNNHandler { return mem_p; } + std::shared_ptr AcquireMemory( + const std::shared_ptr& user_memory_p, + const std::shared_ptr& target_memory_p, + const std::string& suffix, + std::vector& pipeline) { // NOLINT + auto local_key = key_ + suffix; + auto key_reorder_p = key_ + suffix + "reorder_p"; + + auto stored_reorder_p = std::static_pointer_cast( + dev_ctx_.GetBlob(key_reorder_p)); + + if (stored_reorder_p) { + pipeline.push_back(*stored_reorder_p); + } else { + auto reorder_p = + std::make_shared(*user_memory_p, *target_memory_p); + dev_ctx_.SetBlob(key_reorder_p, reorder_p); + pipeline.push_back(*reorder_p); + } + + return target_memory_p; + } + std::shared_ptr AcquireMemory( mkldnn::memory::primitive_desc& mpd, // NOLINT mkldnn::memory::primitive_desc& user_mpd, // NOLINT diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index da46a1abe12258b47b2fd4afb5f146daf15e026d..56bf9e31a35fdec5b7f04849068ff96ac9776c0e 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -226,7 +226,7 @@ RecordBlock::~RecordBlock() { void EnableProfiler(ProfilerState state) { PADDLE_ENFORCE(state != ProfilerState::kDisabled, - "Can't enbale profling, since the input state is ", + "Can't enable profiling, since the input state is ", "ProfilerState::kDisabled"); std::lock_guard l(profiler_mu); diff --git a/paddle/fluid/platform/stream_callback_manager.h b/paddle/fluid/platform/stream_callback_manager.h index 6c984065aa5fa1a8875aebe84051ab396bc417ec..0e88a439cf6ca83e3d98725f58875adeeea86be0 100644 --- a/paddle/fluid/platform/stream_callback_manager.h +++ b/paddle/fluid/platform/stream_callback_manager.h @@ -24,8 +24,6 @@ namespace paddle { namespace platform { -using StreamCallback = std::function; - class StreamCallbackManager; struct StreamCallbackContext { @@ -35,7 +33,7 @@ struct StreamCallbackContext { : manager_(manager), callback_(callback) {} const StreamCallbackManager *manager_; // do not own - StreamCallback callback_; + std::function callback_; }; class StreamCallbackManager { @@ -45,16 +43,18 @@ class StreamCallbackManager { template inline void AddCallback(Callback &&callback) const { - AddCallbackWithStreamAndErrorInfo( - [=](cudaStream_t, cudaError_t) { callback(); }); - } - - template - inline void AddCallbackWithStreamAndErrorInfo(Callback &&callback) const { - auto *stream_callback_context = new StreamCallbackContext(this, callback); - PADDLE_ENFORCE(cudaStreamAddCallback( - stream_, StreamCallbackManager::StreamCallbackFunc, - stream_callback_context, 0)); + auto *stream_callback_context = + new StreamCallbackContext(this, std::forward(callback)); + PADDLE_ENFORCE( +#if CUDA_VERSION >= 10000 + cudaLaunchHostFunc(stream_, StreamCallbackManager::StreamCallbackFunc, + stream_callback_context) +#else + cudaStreamAddCallback(stream_, + StreamCallbackManager::StreamCallbackFunc, + stream_callback_context, 0) +#endif + ); // NOLINT } void Wait() const { thread_pool_.reset(new ThreadPool(1)); } @@ -63,17 +63,21 @@ class StreamCallbackManager { const cudaStream_t stream_; mutable std::unique_ptr thread_pool_; - // cudaStreamCallback cannot call CUDA API inside, so we have to use - // thread_pool here +// cudaStreamCallback cannot call CUDA API inside, so we have to use +// thread_pool here +#if CUDA_VERSION >= 10000 + static void CUDART_CB StreamCallbackFunc(void *user_data) +#else static void CUDART_CB StreamCallbackFunc(cudaStream_t stream, - cudaError_t status, - void *user_data) { + cudaError_t status, void *user_data) +#endif + { auto *callback_context_ptr = reinterpret_cast(user_data); callback_context_ptr->manager_->thread_pool_->enqueue([=]() { std::unique_ptr callback_context( callback_context_ptr); - callback_context->callback_(stream, status); + callback_context->callback_(); }); } }; diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index 3b22718a8c6f994dbc2dc3e7aaa19a7163f716ba..d3b0d4a22954c1d67dc9551b997dcffa0625cbeb 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -57,6 +57,18 @@ struct variant_caster> { auto caster = make_caster(); if (!load_success_ && caster.load(src, convert)) { load_success_ = true; + + if (std::is_same>::value) { + auto caster_ints = make_caster>(); + if (caster_ints.load(src, convert)) { + VLOG(4) << "This value are floats and int64_ts satisfy " + "simultaneously, will set it's type to " + "std::vector"; + value = cast_op>(caster_ints); + return true; + } + } + value = cast_op(caster); return true; } @@ -259,6 +271,8 @@ void BindOpDesc(pybind11::module *m) { pybind11::enum_(*m, "AttrType", "") .value("INT", pd::proto::AttrType::INT) .value("INTS", pd::proto::AttrType::INTS) + .value("LONG", pd::proto::AttrType::LONG) + .value("LONGS", pd::proto::AttrType::LONGS) .value("FLOAT", pd::proto::AttrType::FLOAT) .value("FLOATS", pd::proto::AttrType::FLOATS) .value("STRING", pd::proto::AttrType::STRING) diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 5f15a29f4c3e9b1412912fe4723642d1ede60346..fc821e04a0baf9278295da18ee5a69afcf2c4605 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -821,6 +821,24 @@ All parameter, weight, gradient are variables in Paddle. [](BuildStrategy &self, bool b) { self.enable_data_balance_ = b; }) // FIXME(chengudo): enable_data_balance seems not important + .def_property( + "enable_sequential_execution", + [](const BuildStrategy &self) { + return self.enable_sequential_execution_; + }, + [](BuildStrategy &self, bool b) { + self.enable_sequential_execution_ = b; + }, + R"DOC(The type is BOOL. If set True, the execution order of ops would be the same as what is in the program. Default False.)DOC") + .def_property( + "remove_unnecessary_lock", + [](const BuildStrategy &self) { + return self.remove_unnecessary_lock_; + }, + [](BuildStrategy &self, bool b) { + self.remove_unnecessary_lock_ = b; + }, + R"DOC(The type is BOOL. If set True, some locks in GPU ops would be released and ParallelExecutor would run faster. Default False.)DOC") .def_property( "fuse_elewise_add_act_ops", [](const BuildStrategy &self) { diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index 5a71382fb14b64989502c34d8ac0aa13c62bc7d0..d7676f89ab5e781f910f98d03e72d5f7c1023a9a 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -147,13 +147,11 @@ function cmake_gen() { -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} -DCUDNN_ROOT=/usr/ -DWITH_TESTING=${WITH_TESTING:-ON} - -DWITH_FAST_BUNDLE_TEST=ON -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DWITH_CONTRIB=${WITH_CONTRIB:-ON} - -DWITH_INFERENCE=${WITH_INFERENCE:-ON} -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} -DINFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR} -DWITH_ANAKIN=${WITH_ANAKIN:-OFF} @@ -181,12 +179,10 @@ EOF -DWITH_PYTHON=${WITH_PYTHON:-ON} \ -DCUDNN_ROOT=/usr/ \ -DWITH_TESTING=${WITH_TESTING:-ON} \ - -DWITH_FAST_BUNDLE_TEST=ON \ -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \ -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \ -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \ - -DWITH_INFERENCE=${WITH_INFERENCE:-ON} \ -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} \ -DINFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR} \ -DWITH_ANAKIN=${WITH_ANAKIN:-OFF} \ @@ -653,7 +649,7 @@ function gen_capi_package() { function gen_fluid_lib() { mkdir -p ${PADDLE_ROOT}/build cd ${PADDLE_ROOT}/build - if [[ ${WITH_C_API:-OFF} == "OFF" && ${WITH_INFERENCE:-ON} == "ON" ]] ; then + if [[ ${WITH_C_API:-OFF} == "OFF" ]] ; then cat <