diff --git a/.travis.yml b/.travis.yml index fe4eb2d1570fb3dbb38128400e6f438b99236c0d..3391e2c3cab9938c9dc5705b51367c707d3bbe9d 100644 --- a/.travis.yml +++ b/.travis.yml @@ -16,34 +16,14 @@ env: - JOB=check_style - JOB=build_android addons: - apt: - packages: - - gcc-4.8 - - g++-4.8 - - git - - build-essential - - python - - python-pip - - python2.7-dev - - python-wheel - - libboost-dev - - curl - - swig - - graphviz - - clang-format-3.8 - - automake - - libtool - - ccache ssh_known_hosts: 13.229.163.131 before_install: - - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt - - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit - | function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; } script: - | # 43min timeout - if [[ "$JOB" != "doc" ]]; then timeout 2580 paddle/scripts/paddle_docker_build.sh ${JOB}; else paddle/scripts/paddle_build.sh ${JOB}; fi; + paddle/scripts/paddle_docker_build.sh ${JOB} if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi; - | if [[ "$JOB" != "doc" ]]; then exit 0; fi; diff --git a/README.md b/README.md index d06375a444dd65675bdd75baccf8445c1638a87c..a3b13fe79cc33927e5cc3e091926b111688a941b 100644 --- a/README.md +++ b/README.md @@ -75,19 +75,19 @@ We provide [English](http://www.paddlepaddle.org/docs/develop/documentation/en/g You might want to start from this online interactive book that can run in a Jupyter Notebook. -- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/cluster_train_en.html) +- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/cluster/index_en.html) You can run distributed training jobs on MPI clusters. -- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/k8s_en.html) +- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/cluster/multi_cluster/k8s_en.html) You can also run distributed training jobs on Kubernetes clusters. -- [Python API](http://www.paddlepaddle.org/docs/develop/documentation/en/api/index_en.html) +- [Python API](http://www.paddlepaddle.org/docs/develop/api/en/overview.html) Our new API enables much shorter programs. -- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/dev/contribute_to_paddle_en.html) +- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/dev/contribute_to_paddle_en.html) We appreciate your contributions! diff --git a/benchmark/cluster/vgg16/vgg16_fluid.py b/benchmark/cluster/vgg16/vgg16_fluid.py index 6c47f6535c3cfc31ba165ccec07ac95a9bfd0a7f..6c7d2c10363d3e311dfae455f3dd8fcfc51077a0 100644 --- a/benchmark/cluster/vgg16/vgg16_fluid.py +++ b/benchmark/cluster/vgg16/vgg16_fluid.py @@ -240,8 +240,6 @@ def main(): t = fluid.DistributeTranspiler() t.transpile( - optimize_ops, - params_grads, trainer_id=args.task_index, pservers=args.ps_hosts, trainers=trainers) diff --git a/contrib/float16/.gitignore b/contrib/float16/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..dd28d354f4160b4be68b46a7bebcdf2097d5811a --- /dev/null +++ b/contrib/float16/.gitignore @@ -0,0 +1 @@ +*.inference.model diff --git a/contrib/float16/float16_benchmark.md b/contrib/float16/float16_benchmark.md new file mode 100644 index 0000000000000000000000000000000000000000..b51d6bde92fa04d2268afa36b9c4bd18bc28fe73 --- /dev/null +++ b/contrib/float16/float16_benchmark.md @@ -0,0 +1,97 @@ +# float16 benchmark + +## Description +We want to compare the inference benchmark of float16 vs float32 on the "image_classification" example on Nvidia Tesla V100 GPU, where we can enable the tensor core computation for float16 mode. We test Vgg16 and Resnet50 on the imagenet data set, and Vgg16 and Resnet32 on the cifar10 data set. For completeness, we also add the inference benchmark of Vgg16 and Resnet50 on imagenet data set tested on Nvidia GeForce GTX 1080 Ti GPU. + +For more details about tensor core, please refer to https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/ + +## Test environment +- GPU: single Nvidia Tesla V100 or single Nvidia GeForce GTX 1080 Ti +- CUDNN: 7.1.1 +- CUDA: 9.0 +- Code: https://github.com/PaddlePaddle/Paddle/pull/10331 (Tensor core is enabled in float16 mode) + +## Benchmark on V100 +All times are in ms (millisecond) averaged over 1000 iterations tested on a single Nvidia V100 GPU with respective to different mini-batch(mb) sizes. + +### Vgg16 on imagenet (flowers data set: image.shape = [3, 224, 224]): + +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:| +|float32| 14.01 | 9.70 | 22.99 | 28.26 | 53.87 | 84.42 | 178.95 | +|float16| 3.32 | 4.11 | 5.88 | 9.41 | 16.54 | 30.47 | 60.23 | +|Speedup| 4.22 | 2.36  | 3.91 | 3.00 | 3.26  | 2.77 | 2.97 | + +Total time spent on conv op for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:| +|float32| 11.95 | 6.96 | 18.65 | 21.42 | 41.35 | 60.58 | 130.11 | +|float16| 1.78 | 2.10 | 2.93 | 4.55 | 7.99 | 14.63 | 28.67 | +|Speedup| 6.71 | 3.31  | 6.37 | 4.71 | 5.18  | 4.14 | 4.54 | + + +### Resnet50 on imagenet (flowers data set: image.shape = [3, 224, 224]): + +Total inference time for one batch: + +|       | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| +|float32| 7.03 | 7.41 | 9.16 | 12.55 | 21.13 | 38.27 | 67.93 | 127.02 | +|float16| 6.13 | 6.32 | 6.24 | 7.40 | 10.90 | 18.18 | 33.20 | 64.52 | +|Speedup| 1.15 | 1.17  | 1.47  | 1.70 | 1.94  | 2.11 | 2.05 | 1.97 | + +Total time spent on conv op for one batch: + +|       | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| +|float32| 5.43 | 5.46 | 6.50 | 8.36 | 13.80 | 24.45 | 41.21 | 73.44 | +|float16| 4.19 | 4.30 | 3.96 | 4.21 | 5.63 | 8.77 | 15.24 | 28.40 | +|Speedup| 1.30 | 1.27  | 1.64  | 1.99 | 2.45  | 2.79 | 2.70 | 2.59 | + + +### Vgg16 on cifar10 (image.shape = [3, 32, 32]): + +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | mb=256 | mb=512 | +|-------|-----:|-----:|-----:|-----:|------:|------:|------:|-------:|-------:|-------:| +|float32| 3.13 | 3.17 | 3.19 | 3.58 | 3.98 | 6.23 | 8.42 | 13.44 | 24.19 | 44.97 | +|float16| 2.72 | 2.77 | 2.76 | 2,88 | 2.96 | 3.24 | 4.01 | 5.78 | 9.65 | 17.37 | +|Speedup| 1.15 | 1.14 | 1.16 | 1.24 | 1.34 | 1.92  | 2.10 | 2.33  | 2.51 | 2.59 | + + +### Resnet32 on cifar10 (image.shape = [3, 32, 32]): + +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | mb=256 | mb=512 | +|-------|-----:|-----:|-----:|-----:|------:|------:|------:|-------:|-------:|-------:| +|float32| 3.11 | 3.14 | 2.99 | 3.04 | 3.10 | 3.28 | 4.47 | 6.86 | 11.63 | 21.16 | +|float16| 3.70 | 3.81 | 3.75 | 3.83 | 3.77 | 3.97 | 3.92 | 4.15 | 6.41 | 11.02 | +|Speedup|     |     |     |     |       | | 1.14  | 1.65 | 1.81 | 1.92 | + + +## Benchmark on 1080 Ti +All times are in ms (millisecond) averaged over 1000 iterations tested on a single Nvidia GeForce GTX 1080 Ti GPU with respective to different mini-batch(mb) sizes. + +### Vgg16 on imagenet (flowers data set: image.shape = [3, 224, 224]): +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | +|-------|-----: |-----: |-----: |-----: |------: |-------:| +|float32| 5.60 | 9.38 | 15.86 | 29.79 | 57.60 | 117.73 | +|float16| 4.99 | 7.79 | 13.47 | 26.02 | 52.30 | 102.34 | +|Speedup| 1.12 | 1.20  | 1.18 | 1.15 | 1.10  | 1.15 | + + +### Resnet50 on imagenet (flowers data set: image.shape = [3, 224, 224]): +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |-------:|-------:| +|float32| 5.63 | 6.23 | 8.85 | 14.71 | 26.07 | 52.86 | 108.95 | +|float16| 5.89 | 6.44 | 7.94 | 12.57 | 22.03 | 45.06 | 92.68 | +|Speedup| |  | 1.12  | 1.17 | 1.18  | 1.17 | 1.18 | diff --git a/contrib/float16/float16_inference_demo.py b/contrib/float16/float16_inference_demo.py new file mode 100644 index 0000000000000000000000000000000000000000..063227d5d2586d66ad4091133a8edf014da839f8 --- /dev/null +++ b/contrib/float16/float16_inference_demo.py @@ -0,0 +1,362 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function +from float16_transpiler import Float16Transpiler + +import argparse +import paddle +import paddle.fluid as fluid +import contextlib +import math +import sys +import numpy as np +import os + +parser = argparse.ArgumentParser( + 'Float16 inference accuracy test and benchmark.') +parser.add_argument( + '--train_batch_size', type=int, default=16, help="Batch size for training.") +parser.add_argument( + '--inf_batch_size', type=int, default=32, help="Batch size for inference.") +parser.add_argument( + '--repeat', type=int, default=1, help="How many times to run the test.") +parser.add_argument( + '--data_set', + type=str, + default='cifar10', + choices=['cifar10', 'imagenet'], + help="Optional dataset for benchmark.") +parser.add_argument( + '--model', + type=str, + default='vgg', + choices=['vgg', 'resnet'], + help="Optional model for benchmark.") +parser.add_argument( + '--threshold', + type=float, + default=0.005, + help='Save inference model when test accuracy reach this threshold.') +parser.add_argument('--learning_rate', type=float, default=0.001) +args = parser.parse_args() + + +def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'): + conv1 = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=False) + return fluid.layers.batch_norm(input=conv1, act=act) + + +def shortcut(input, ch_out, stride): + ch_in = input.shape[1] + if ch_in != ch_out: + return conv_bn_layer(input, ch_out, 1, stride, 0, None) + else: + return input + + +def basicblock(input, ch_out, stride): + short = shortcut(input, ch_out, stride) + conv1 = conv_bn_layer(input, ch_out, 3, stride, 1) + conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1, act=None) + return fluid.layers.elementwise_add(x=short, y=conv2, act='relu') + + +def bottleneck(input, ch_out, stride): + short = shortcut(input, ch_out * 4, stride) + conv1 = conv_bn_layer(input, ch_out, 1, stride, 0) + conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1) + conv3 = conv_bn_layer(conv2, ch_out * 4, 1, 1, 0, act=None) + return fluid.layers.elementwise_add(x=short, y=conv3, act='relu') + + +def layer_warp(block_func, input, ch_out, count, stride): + res_out = block_func(input, ch_out, stride) + for i in range(1, count): + res_out = block_func(res_out, ch_out, 1) + return res_out + + +def resnet_imagenet(input, depth=50): + cfg = { + 18: ([2, 2, 2, 1], basicblock), + 34: ([3, 4, 6, 3], basicblock), + 50: ([3, 4, 6, 3], bottleneck), + 101: ([3, 4, 23, 3], bottleneck), + 152: ([3, 8, 36, 3], bottleneck) + } + stages, block_func = cfg[depth] + conv1 = conv_bn_layer(input, ch_out=64, filter_size=7, stride=2, padding=3) + pool1 = fluid.layers.pool2d( + input=conv1, pool_type='avg', pool_size=3, pool_stride=2) + res1 = layer_warp(block_func, pool1, 64, stages[0], 1) + res2 = layer_warp(block_func, res1, 128, stages[1], 2) + res3 = layer_warp(block_func, res2, 256, stages[2], 2) + res4 = layer_warp(block_func, res3, 512, stages[3], 2) + pool2 = fluid.layers.pool2d( + input=res4, + pool_size=7, + pool_type='avg', + pool_stride=1, + global_pooling=True) + return pool2 + + +def resnet_cifar10(input, depth=32): + assert (depth - 2) % 6 == 0 + + n = (depth - 2) // 6 + + conv1 = conv_bn_layer( + input=input, ch_out=16, filter_size=3, stride=1, padding=1) + res1 = layer_warp(basicblock, conv1, 16, n, 1) + res2 = layer_warp(basicblock, res1, 32, n, 2) + res3 = layer_warp(basicblock, res2, 64, n, 2) + pool = fluid.layers.pool2d( + input=res3, pool_size=8, pool_type='avg', pool_stride=1) + return pool + + +def vgg16(input): + def conv_block(input, num_filter, groups, dropouts): + return fluid.nets.img_conv_group( + input=input, + pool_size=2, + pool_stride=2, + conv_num_filter=[num_filter] * groups, + conv_filter_size=3, + conv_act='relu', + conv_with_batchnorm=True, + conv_batchnorm_drop_rate=dropouts, + pool_type='max') + + conv1 = conv_block(input, 64, 2, [0.3, 0]) + conv2 = conv_block(conv1, 128, 2, [0.4, 0]) + conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) + conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) + conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) + + drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) + fc1 = fluid.layers.fc(input=drop, size=4096, act=None) + bn = fluid.layers.batch_norm(input=fc1, act='relu') + drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) + fc2 = fluid.layers.fc(input=drop2, size=4096, act=None) + return fc2 + + +def train(place, save_dirname): + if args.data_set == "cifar10": + class_dim = 10 + data_shape = [3, 32, 32] + elif args.data_set == "imagenet": + class_dim = 102 + data_shape = [3, 224, 224] + else: + raise ValueError("%s dataset is not supported" % data_set) + + images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + if args.model == "vgg": + print("train vgg") + net = vgg16(images) + elif args.model == "resnet": + print("train resnet") + if args.data_set == "cifar10": + net = resnet_cifar10(images) + elif args.data_set == "imagenet": + net = resnet_imagenet(images) + else: + raise ValueError("%s dataset is not supported" % args.data_set) + else: + raise ValueError("%s network is not supported" % args.model) + + predict = fluid.layers.fc(input=net, size=class_dim, act='softmax') + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + acc = fluid.layers.accuracy(input=predict, label=label) + + #Test program + test_program = fluid.default_main_program().clone(for_test=True) + optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate) + optimizer.minimize(avg_cost) + + BATCH_SIZE = args.train_batch_size + PASS_NUM = 100 + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.flowers.train() + if args.data_set == 'imagenet' else paddle.dataset.cifar.train10(), + buf_size=128 * 10), + batch_size=args.train_batch_size) + + test_reader = paddle.batch( + paddle.dataset.flowers.test() + if args.data_set == 'imagenet' else paddle.dataset.cifar.test10(), + batch_size=args.inf_batch_size) + + exe = fluid.Executor(place) + feeder = fluid.DataFeeder(place=place, feed_list=[images, label]) + + exe.run(fluid.default_startup_program()) + main_program = fluid.default_main_program() + + for pass_id in range(PASS_NUM): + for batch_id, data in enumerate(train_reader()): + train_image = np.array( + map(lambda x: x[0].reshape(data_shape), data)).astype("float32") + train_label = np.array(map(lambda x: x[1], data)).astype("int64") + train_label = train_label.reshape([-1, 1]) + + exe.run(main_program, + feed={'pixel': train_image, + 'label': train_label}) + + if (batch_id % 100) == 0: + acc_list = [] + avg_loss_list = [] + for tid, test_data in enumerate(test_reader()): + test_image = np.array( + map(lambda x: x[0].reshape(data_shape), + test_data)).astype("float32") + test_label = np.array(map(lambda x: x[1], + test_data)).astype("int64") + test_label = test_label.reshape([-1, 1]) + + loss_t, acc_t = exe.run( + program=test_program, + feed={"pixel": test_image, + "label": test_label}, + fetch_list=[avg_cost, acc]) + if math.isnan(float(loss_t)): + sys.exit("got NaN loss, training failed.") + acc_list.append(float(acc_t)) + avg_loss_list.append(float(loss_t)) + + acc_value = np.array(acc_list).mean() + avg_loss_value = np.array(avg_loss_list).mean() + + print( + 'PassID {0:1}, BatchID {1:04}, Test Loss {2:2.2}, Accuracy {3:2.2}'. + format(pass_id, batch_id + 1, + float(avg_loss_value), float(acc_value))) + + if acc_value > args.threshold: + print( + 'Save inference model with test accuracy of {0} at {1}'. + format(float(acc_value), save_dirname)) + fluid.io.save_inference_model(save_dirname, ["pixel"], + [predict], exe) + return + + +def test_accuracy(executor, inference_program, feed_target_names, + fetch_targets): + if args.data_set == "cifar10": + data_shape = [3, 32, 32] + elif args.data_set == "imagenet": + data_shape = [3, 224, 224] + else: + raise ValueError("%s dataset is not supported" % data_set) + + test_reader = paddle.batch( + paddle.dataset.cifar.test10() + if args.data_set == "cifar10" else paddle.dataset.flowers.test(), + batch_size=args.inf_batch_size) + + test_num = 0 + correct_num = 0 + + for test_data in test_reader(): + test_image = np.array( + map(lambda x: x[0].reshape(data_shape), test_data)).astype( + "float32") + test_label = np.array(map(lambda x: x[1], test_data)).astype("int64") + test_label = test_label.reshape([-1, 1]) + + results = executor.run(program=inference_program, + feed={feed_target_names[0]: test_image}, + fetch_list=fetch_targets) + + prediction = np.argmax(results[0], axis=1).reshape([-1, 1]) + correct_num += np.sum(prediction == test_label) + test_num += test_label.size + + print("{0} out of {1} predictions are correct.".format(correct_num, + test_num)) + print("Test accuray is {0}.".format(float(correct_num) / float(test_num))) + + +def infer(place, save_dirname): + exe = fluid.Executor(place) + inference_scope = fluid.core.Scope() + + with fluid.scope_guard(inference_scope): + # Use fluid.io.load_inference_model to obtain the inference program desc, + # the feed_target_names (the names of variables that will be feeded + # data using feed operators), and the fetch_targets (variables that + # we want to obtain data from using fetch operators). + print("Load inference model from {0}".format(save_dirname)) + [inference_program, feed_target_names, + fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) + + print("The test set accuracy of inference in float mode is:") + test_accuracy(exe, inference_program, feed_target_names, fetch_targets) + + float16_inference_program = inference_program.clone() + t = Float16Transpiler() + t.transpile(float16_inference_program, place) + + print("The test set accuracy of inference in float16 mode is:") + test_accuracy(exe, float16_inference_program, feed_target_names, + fetch_targets) + + fp16_save_dirname = "float16_" + save_dirname + fluid.io.save_inference_model(fp16_save_dirname, feed_target_names, + fetch_targets, exe, + float16_inference_program) + + +@contextlib.contextmanager +def scope_prog_guard(): + prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.core.Scope() + with fluid.scope_guard(scope): + with fluid.program_guard(prog, startup_prog): + yield + + +if __name__ == "__main__": + if not fluid.core.is_compiled_with_cuda(): + raise Exception("This test requires CUDA GPUs!") + + place = fluid.CUDAPlace(0) + if not fluid.core.is_float16_supported(place): + raise Exception( + "This test requires compute capability of CUDA GPU >= 5.3!") + + for i in range(args.repeat): + with scope_prog_guard(): + save_dirname = "image_classification_" + args.data_set + "_" + args.model + ".inference.model" + train(place, save_dirname) + infer(place, save_dirname) diff --git a/contrib/float16/float16_inference_report.md b/contrib/float16/float16_inference_report.md new file mode 100644 index 0000000000000000000000000000000000000000..67623a4d8d58b2ab84031ac4b290931ab540973b --- /dev/null +++ b/contrib/float16/float16_inference_report.md @@ -0,0 +1,163 @@ +## Introduction +Working with deep neural networks (DNN) is a two-stage process. First we train DNN using labeled examples of inputs and desired outputs to obtain the model parameters (weights), then we deploy DNN along with the trained weights to run inference on unknown inputs. Typically, these weights are in float data type and hence we run inference in float mode using these weights. This post focuses on the discussion of how to use low precision float16 data type to represent these trained weights and run inference in float16 mode as well as the advantages of float16 inference over its float counterpart by showing some experiment results. + +## What is float16? +float16 (or FP16) is a half-precision floating-point format that uses 16 bits in memory to represent a value. The advantage over 32-bit single-precision floating-point format (commonly known as float data type) is that it requires half the storage and bandwidth at the expense of precision and range. Fortunately, DNN inference has high tolerance against the loss of precision and range when using float16 to represent the weights and the inference accuracy will only be minimally affected in most cases. This gives us the opportunity to use float16 data type to speedup the inference. + +Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type. + +## Why float16? +The trend in today's deep learning community is to use bigger and deeper model. This translates to larger memory footprint, higher computation demands, and as a result higher energy consumption on computing devices. The advantages of float16 over float are correspondingly three-fold: + +1. We only need half the memory size to load the same model using float16 representations. Moreover, most of the intermediate results generated during float16 inference are also of float16 data type. This makes the whole memory footprint of float16 inference roughly about half of its float counterpart. This is especially useful when deploying inference on mobile devices with limited available memory. Also given the same available memory, the maximum batch size for float16 inference is about twice that for float inference. + +2. Because float16 occupies less memory than float, in theory hardware devices can achieve much higher floating point operators per second (FLOPS) for float16 data than float data. Right now, an outstanding example of hardware devices that actually deliver such advantages is Nvidia's latest Volta architecture GPUs, including Tesla V100 and Titan V. Moreover float16 takes less time to read from or write to memory and hence float16 can make inference more efficient especially in memory-bound applications where the performance is largely affected by how fast it is to read and write data. + +3. From the energy efficiency perspective, the energy needed to read, write, and compute float16 data is much less that its float counterpart, which can significantly reduce the battery power consumption on mobile devices or the total cost of ownership (TCO) of data centers. + +## Fluid implementation of float16 inference +### Overview +Fluid use [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block. + +### Basic requirement +When an operator is run by an executor, it uses a kernel to perform computations on tensors contained in the input variables, and then write the results to the tensors in the output variables. Each operator has multiple kernels for different combinations of data types, devices, and library types, respectively. The operator will select the appropriate kernel to run based on, among other things, the data type of the input tensors. By default, every Fluid operator has a kernel for float data type that takes float inputs and generates float outputs. + +This means that if we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will makes the program run in float mode and gives us a final output of float data type. + +The same principle applies if we want a program to run in float16 mode. We provide input variable of float16 data type to the first operator and every subsequent operator will invoke the float16 kernel until we get the final output in float16 data type. So the preliminary requirements for float16 inference is to add float16 kernels to operators that are needed in a specific kind of neural networks. Our current focus is on Convolutional Neural Networks (CNN) and hence we have added float16 kernels to the following operators: convolution, pooling, GEMM, elementwise addition, batch norm, dropout, various activations including relu and tanh, and softmax. + +### float16 transpiler +Furthermore, we need a float16 transpiler to achieve the following usage code: + +```python +# Get the float32 inference program and load the associated float32 weights +[inference_program, feed_target_names, + fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) + +# Prepare the float input data +batch_size = 1 +tensor_img = numpy.random.rand(batch_size, 3, 32, 32).astype(numpy.float32) + +# Running inference_program in float mode +float_results = exe.run(inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + +# Use float16 transpiler to speedup +float16_inference_program = float_inference_program.clone() +t = Float16Transpiler() +t.transpile(float16_inference_program, GPUPlace) + +# Running float16_inference_program in float16 mode using the same input data +float16_results = exe.run(float16_inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + +# Do some tests to verify the correctness of float16 inference +... +np.testing.assert_almost_equal(float_results, float16_results, ...) +... + +# Save the float16 inference program and float16 weights for future deployment +fluid.io.save_inference_model(fp16_save_dirname, feed_target_names, + fetch_targets, exe, + float16_inference_program) +``` + +In this scenario, we already have a float32 inference program and some associated float32 weights that can do float32 inference. We can easily use the `transpile` method of the `Float16Transpiler` class to do certain modifications to the existing program and weights so that we have a new float16 program and the associated float16 weights. + +We can then run various inference experiments in float16 mode and save the float16 program and weights on disk for future deployment. To enhance the code usability, we maintain a consistent API so that user can use the same float32 input data to run inference program in either float32 and float16 mode and obtain output data both of float32 data type. This requires us to add some cast operators in the program to convert between float16 tensor and float32 tensor. + +The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference). + +### Experiment results +We provide demo codes that can be used to reproduce the experiment results by doing: +```bash +git clone https://github.com/PaddlePaddle/Paddle.git +cd Paddle +# This line will generate a paddle development docker image with cuda 8 and cudnn 7 +# If you want test on cuda 9 instead, change the line 5 in Paddle/Dockerfile +# from `FROM nvidia/cuda:8.0-cudnn7-devel-ubuntu16.04` +# to `FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04` and similarly for other configurations +nvidia-docker build -t paddle:float16 . +# After running this, different results will be written to different log files in Paddle/contrib/float16/ +nvidia-docker run -it -v $PWD:/paddle paddle:float16 /paddle/contrib/float16/run_float16_demo.sh +``` + +#### Correctness +As is mentioned before, DNN inference has been found to be tolerant against the loss of precision and range incured by float16 and we want to see how good this tolerance is. + +We train a resnet32 model using cifar10 data set, save it when test set accuracy is above 60%, and then test the inference accuracy on the 10000 examples of the cifar10 test set in float16 and float32 mode, respectively. + +We repeat the test ten times and get the following results: + +| | float16 | float32 | +|--------|--------:|--------: | +| # 1 | 62.75% | 62.72% | +| # 2 | 61.27% | 61.28% | +| # 3 | 62.24% | 62.23% | +| # 4 | 64.16% | 64.17% | +| # 5 | 60.75% | 60.77% | +| # 6 | 63.25% | 63.24% | +| # 7 | 62.15% | 62.13% | +| # 8 | 62.05% | 62.02% | +| # 9 | 65.19% | 65.20% | +| #10 | 62.53% | 62.48% | +| average| 62.63% | 62.62% | + +We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over 10 tests. + +#### Performance benchmark +Currently, Fluid inference in float16 mode is only supported on Nvidia GPU device. There is no motivation to support float16 inference on non-ARM CPUs because float16 is not natively supported there and float16 calculation will only be slower than its float counterpart. + +Nvidia started to support its native float16 data type (which has the same internal memory representation as Fluid float16 class) on CUDA 7.5. Moreover, float16 speedups on common computational intensive tasks including GEMM (general matrix-matrix multiplication) and convolution are supported since cublas 7.5 and cuDNN 5.0. + +Recently, the introduction of [tensor core](https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/) in volta architecture GPUs and the support of tensor core calculation in CUDA 9.0 and cuDNN 7 make float16 truly superior to float in certain deep learning applications. + +We thus benchmark the float16 inference performance on a single Nvidia Tesla V100 GPU (volta architecture and with tensor cores) and compare it with its float32 counterpart. All the following results are in ms (millisecond) averaged over 1000 mini-batches with respective to different mini-batch(mb) sizes. + +Average inference time for one mini-batch on Vgg16 model tested on imagenet data set: + +| total | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:| +|float32| 14.01 | 9.70 | 22.99 | 28.26 | 53.87 | 84.42 | 178.95 | +|float16| 3.32 | 4.11 | 5.88 | 9.41 | 16.54 | 30.47 | 60.23 | +|Speedup| 4.22 | 2.36  | 3.91 | 3.00 | 3.26  | 2.77 | 2.97 | + +We can see that float16 inference provides 2x ~ 4x speedup on different batch sizes. + +Convolution operation is ususally the computational bottleneck of CNN, so we also check the average time spent on the Fluid convolution operators for one mini-batch as follows: + +|conv op| mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:| +|float32| 11.95 | 6.96 | 18.65 | 21.42 | 41.35 | 60.58 | 130.11 | +|float16| 1.78 | 2.10 | 2.93 | 4.55 | 7.99 | 14.63 | 28.67 | +|Speedup| 6.71 | 3.31  | 6.37 | 4.71 | 5.18  | 4.14 | 4.54 | + +Fluid convolution operator uses cuDNN 7 to implement the kernel and we can see that with the help of tensor core, float16 convolution is significantly faster than its float32 counterpart, which makes the overall float16 inference performance much better. + +Similarly, we also list the benchmark results of Resnet50 model tested on imagenet data set: + +| total | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| +|float32| 7.03 | 7.41 | 9.16 | 12.55 | 21.13 | 38.27 | 67.93 | 127.02 | +|float16| 6.13 | 6.32 | 6.24 | 7.40 | 10.90 | 18.18 | 33.20 | 64.52 | +|Speedup| 1.15 | 1.17  | 1.47  | 1.70 | 1.94  | 2.11 | 2.05 | 1.97 | + +|conv op| mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| +|float32| 5.43 | 5.46 | 6.50 | 8.36 | 13.80 | 24.45 | 41.21 | 73.44 | +|float16| 4.19 | 4.30 | 3.96 | 4.21 | 5.63 | 8.77 | 15.24 | 28.40 | +|Speedup| 1.30 | 1.27  | 1.64  | 1.99 | 2.45  | 2.79 | 2.70 | 2.59 | + +We find that the speedup provided by float16 inference starts relatively small at 1.15x for batch size 1 and gradually increase to about 2x for larger batch sizes. Similar trend can be found for the time spent on the convolution operator. Note that right now the tensor core will only be utilized in the convolution operation when certain dimentional requirements are met for the input data and filter. The speedup by float16 inference for Resnet50 is smaller than the Vgg16 counterpart partially because the convolution operation in Resnet is much simpler than the Vgg counterpart and this makes the tensor core less utilized in Resnet than in Vgg. + +We also did the same benchmark on a Nvidia GeForce GTX 1080 Ti GPU that does not support tensor core. The results show that for Vgg16, float16 inference provides consistent small speedup (around 1.15x) for all mini-batch sizes, while for Resnet50, float16 inference is slower than its float32 counterpart in small batch sizes (mb = 1 and 2) and then deliver around 1.15x speedup for all larger batch sizes. By comparing the benchmarks on 1080 Ti and V100, we find that tensor core, which is specialized for float16 computations, is a critical component for high performance float16 inference. + +Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/contrib/float16/float16_benchmark.md) for comprehensive benchmark results. + +### Summary +1. Fluid is now able to run inference in float16 mode via a float16 transpiler. We currently support CNN programs, including Vgg and Resnet, to run in float16 inference mode. +2. The accuracy of float16 inference is verified to be almost identical to the float32 counterpart at least on CNNs. +3. float16 inference provides significant speedup on large and computationally intensive Vgg16 network on image net data set. For the much smaller and simpler Resnet50, the speedup provided by float16 inference is less significant than on Vgg16 but still favorable especially for large batch size. +4. We cannot achieve the superior float16 inference performance without the help of the newly introduced tensor cores on the Nvidia Volta architecture GPUs. diff --git a/contrib/float16/float16_transpiler.py b/contrib/float16/float16_transpiler.py new file mode 100644 index 0000000000000000000000000000000000000000..91ba101edb65cd45bd5e37a0c6ad25e515593a81 --- /dev/null +++ b/contrib/float16/float16_transpiler.py @@ -0,0 +1,256 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import paddle.fluid.core as core +from paddle.fluid.framework import Program +from paddle.fluid.executor import global_scope + + +class Float16Transpiler: + def transpile(self, program, place, scope=None): + ''' + Transpile the program desc and cast the weights to float16 data type to + enable float16 inference. + + Since the operator in a program desc will automatically choose the + right compute kernel to run based on the data type of the input tensor. + We actually don't need to change the program desc to run in float16 mode. + + However, in this way, users who are used to feeding and fetching tensors + of float32 data type when running typical inference may find it confusing + and difficult to run inference in float16 mode as they need to convert + input data to float16 dtype and then convert the results back to float32 + dtype to match the rest of code. + + So this function appends cast ops to the program desc where necessary so + that users are able to run inference in float16 mode while providing input + tensor (feed_holder) of float data type and obtaining output tensor + (fetch_holder) of float data type. + + Moreover, it is desired that when we have the scope and program desc to run + inference in float32 mode, we can use a single API to do the necessary + modification and then user can run float16 inference on the fly. To make + this happen, this function also create new parameters in the scope to have the + converted float16 weights and change the operators in program desc to use + these new parameters. + + :param program: program to transpile + :type program: Program + :param place: inference place + :type place: Place + :param scope: inference scope + :type scope: Scope + ''' + if not isinstance(program, Program): + raise TypeError("program should be as Program type") + if not isinstance(place, core.CPUPlace) and not isinstance( + place, core.CUDAPlace): + raise TypeError("place should be as CPUPlace/CUDAPlace type") + if scope is None: + scope = global_scope() + if not isinstance(scope, core.Scope): + raise TypeError("scope should be as Scope type or None") + + self.scope = scope + self.place = place + self.block = program.block(0) + self.input_map = {} # store the input names should be adjusted + + self._modify_feed_fetch() + self._convert_param_to_float16() + self._adjust_input(skip=True) + self._remove_unused_var() + + # TODO(luotao): use clone() method to flush the program.desc in force, + # since some large program.desc will not be flushed immediately. + # And a better solution will be considered later. + program = program.clone() + + # ====================== private transpiler functions ===================== + def _adjust_input(self, skip=False): + ''' + Change the input variable name in operators. + + When we are in the process of modifying a program desc, we usually + replace some variables with some other variables, where we create + a dictionary input_map to record the one-to-one correspondence + between each old variable and the new one. + + After that, this function will search all the operators that use the + old variables and change the info in op to use the new variables. There + maybe some exceptions to this rule when we are using the float16 transpiler + and insert cast ops to cast float32 variable to float16 one. After we + insert the cast op to cast var_1 to var_1_fp16, we don't want to change + the input of cast op to var_1_fp16 after using this function. + ''' + skip_ops = {"cast"} + for i in range(len(self.block.ops)): + current_op = self.block.ops[i] + if skip and current_op.type in skip_ops: + continue + for input_arg in current_op.input_arg_names: + if input_arg in self.input_map: + current_op.rename_input(input_arg, + self.input_map[input_arg]) + + def _remove_unused_var(self): + ''' + remove unused varibles in program + ''' + args = [] + for i in range(len(self.block.ops)): + current_op = self.block.ops[i] + args += current_op.input_arg_names + args += current_op.output_arg_names + args = list(set(args)) # unique the input and output arguments + + for var in self.block.vars.keys(): + if var not in args: + self.block.remove_var(var) + + def _modify_feed_fetch(self): + ''' + Modify feed fetch op/vars for float16 inference. + + For each feed op: + feed_op->feed_target_var + + Change it to: + feed_op->feed_target_var->cast_op(from other dtype to float16)->tmp_var + + For each fetch op: + fetch_target_var->fetch_op + + Change it to: + tmp_var->cast_op(from float16 to other dtype)->fetch_target_var->fetch_op + + :return: None + ''' + + def find_op(var): + # It is possible that var.op is not up to date after some + # modifications to program desc. Here we force to make it up to date. + var.op = None + for op in self.block.ops: + if var.name in op.output_arg_names: + var.op = op + break + + if var.op is None: + raise ValueError("The target variable must have an " + "associated operator that generates it.") + + i = 0 + while i < len(self.block.ops): + cur_op = self.block.ops[i] + if cur_op.type == "feed": + var_name = cur_op.output("Out")[0] + tmp_var_name = var_name + ".fp16" + var = self.block.vars[var_name] + tmp_var = self.block.create_var( + name=tmp_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape, + persistable=var.persistable) + self.block.insert_op( + i + 1, + type="cast", + inputs={"X": var}, + outputs={"Out": tmp_var}, + attrs={ + 'in_dtype': int(var.dtype), + 'out_dtype': int(tmp_var.dtype) + }) + self.input_map[var_name] = tmp_var_name + i = i + 1 + elif cur_op.type == "fetch": + var_name = cur_op.input("X")[0] + tmp_var_name = var_name + ".fp16" + var = self.block.vars[var_name] + tmp_var = self.block.create_var( + name=tmp_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape, + persistable=var.persistable) + find_op(var) + var.op.rename_output(var_name, tmp_var_name) + self.block.insert_op( + i, + type="cast", + inputs={"X": tmp_var}, + outputs={"Out": var}, + attrs={ + 'in_dtype': int(tmp_var.dtype), + 'out_dtype': int(var.dtype) + }) + i = i + 1 + i = i + 1 + + def _convert_param_to_float16(self): + def _get_no_fp16_conversion_var_names(): + ''' + Get the set of input variable names that shouldn't be converted to float16. + + When we want to run inference in float16 mode, most parameters need to be + firstly converted to float16. However, there are some parameters that + shouldn't be converted to float16 because the corresponding operator + requires float32 parameters even in float16 mode (when the input data is + of float16 data type). Currently, the only operator that has this exclusion + is the batch norm op. + + :return: set of input variable names + :type var_names: set + ''' + op_names = {'batch_norm'} + var_names = [] + for op in self.block.ops: + if op.type in op_names: + var_names += op.input_arg_names + return set(var_names) + + def _should_be_converted(var): + return var.persistable and \ + var.name not in self.no_conversion_vars and \ + var.type != core.VarDesc.VarType.FEED_MINIBATCH and \ + var.type != core.VarDesc.VarType.FETCH_LIST + + self.no_conversion_vars = _get_no_fp16_conversion_var_names() + conversion_var_list = filter(_should_be_converted, + self.block.vars.values()) + for var in conversion_var_list: + fp16_var_name = var.name + ".fp16" + fp16_var = self.block.create_parameter( + name=fp16_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape) + + # cast the data in the tensor of the original var to float16 + # data type and store it in the tensor of the new float16 var + self.scope.var(fp16_var_name) + fp16_tensor = self.scope.find_var(fp16_var_name).get_tensor() + tensor = np.array(self.scope.find_var(var.name).get_tensor()) + # After the old tensor data is converted to np.float16, view(np.uint16) + # is used so that the internal memory of the numpy array will be + # reinterpreted to be of np.uint16 data type, which is binded to fluid + # float16 data type via the help of pybind in tensor_py.h. + fp16_tensor.set( + tensor.astype(np.float16).view(np.uint16), self.place) + + # old var will be replaced by the fp16 var in program desc + self.input_map[var.name] = fp16_var_name + self.block.remove_var(var.name) diff --git a/contrib/float16/run_float16_demo.sh b/contrib/float16/run_float16_demo.sh new file mode 100755 index 0000000000000000000000000000000000000000..d8a34ee67b8fab214fa6e96104304689211f84da --- /dev/null +++ b/contrib/float16/run_float16_demo.sh @@ -0,0 +1,117 @@ +#!/bin/bash + +BUILD_PATH=/paddle/fp16_build +WHEEL_PATH=$BUILD_PATH/python/dist +INFER_PATH=$BUILD_PATH/paddle/fluid/inference/tests/book +DEMO_PATH=/paddle/contrib/float16 + +# Use the single most powerful CUDA GPU on your machine +export CUDA_VISIBLE_DEVICES=0 + +# Build the PaddlePaddle Fluid wheel package and install it. +mkdir -p $BUILD_PATH && cd $BUILD_PATH +cmake .. -DWITH_AVX=OFF \ + -DWITH_MKL=OFF \ + -DWITH_GPU=ON \ + -DWITH_TESTING=ON \ + -DWITH_TIMER=ON \ + -DWITH_PROFILER=ON \ + -DWITH_FLUID_ONLY=ON +make -j `nproc` +pip install -U "$WHEEL_PATH/$(ls $WHEEL_PATH)" + +cd $DEMO_PATH +# Clear previous log results +rm -f *.log + +# Test the float16 inference accuracy of resnet32 on cifar10 data set +stdbuf -oL python float16_inference_demo.py \ + --data_set=cifar10 \ + --model=resnet \ + --threshold=0.6 \ + --repeat=10 \ + 2>&1 | tee -a float16_inference_accuracy.log + +# Sleep to cool down the GPU for consistent benchmarking +sleep 2m + +# benchmarking parameters +REPEAT=1000 +MAXIMUM_BATCH_SIZE=512 + +for ((batch_size = 1; batch_size <= MAXIMUM_BATCH_SIZE; batch_size *= 2)); +do + + # Test inference benchmark of vgg16 on imagenet + stdbuf -oL python float16_inference_demo.py \ + --data_set=imagenet \ + --model=vgg \ + --threshold=0.001 \ + --repeat=1 \ + + $INFER_PATH/test_inference_image_classification_vgg \ + --data_set=imagenet \ + --dirname=$DEMO_PATH/image_classification_imagenet_vgg.inference.model \ + --fp16_dirname=$DEMO_PATH/float16_image_classification_imagenet_vgg.inference.model \ + --repeat=$REPEAT \ + --batch_size=$batch_size \ + --skip_cpu=true \ + 2>&1 | tee -a imagenet_vgg16_benchmark.log + + sleep 2m + + # Test inference benchmark of resnet50 on imagenet + stdbuf -oL python float16_inference_demo.py \ + --data_set=imagenet \ + --model=resnet \ + --threshold=0.001 \ + --repeat=1 \ + + $INFER_PATH/test_inference_image_classification_resnet \ + --data_set=imagenet \ + --dirname=$DEMO_PATH/image_classification_imagenet_resnet.inference.model \ + --fp16_dirname=$DEMO_PATH/float16_image_classification_imagenet_resnet.inference.model \ + --repeat=$REPEAT \ + --batch_size=$batch_size \ + --skip_cpu=true \ + 2>&1 | tee -a imagenet_resnet50_benchmark.log + + sleep 2m + + # Test inference benchmark of vgg16 on cifar10 + stdbuf -oL python float16_inference_demo.py \ + --data_set=cifar10 \ + --model=vgg \ + --threshold=0.001 \ + --repeat=1 \ + + $INFER_PATH/test_inference_image_classification_vgg \ + --data_set=cifar10 \ + --dirname=$DEMO_PATH/image_classification_cifar10_vgg.inference.model \ + --fp16_dirname=$DEMO_PATH/float16_image_classification_cifar10_vgg.inference.model \ + --repeat=$REPEAT \ + --batch_size=$batch_size \ + --skip_cpu=true \ + 2>&1 | tee -a cifar10_vgg16_benchmark.log + + sleep 1m + + # Test inference benchmark of resnet32 on cifar10 + stdbuf -oL python float16_inference_demo.py \ + --data_set=cifar10 \ + --model=resnet \ + --threshold=0.001 \ + --repeat=1 \ + + $INFER_PATH/test_inference_image_classification_vgg \ + --data_set=cifar10 \ + --dirname=$DEMO_PATH/image_classification_cifar10_resnet.inference.model \ + --fp16_dirname=$DEMO_PATH/float16_image_classification_cifar10_resnet.inference.model \ + --repeat=$REPEAT \ + --batch_size=$batch_size \ + --skip_cpu=true \ + 2>&1 | tee -a cifar10_resnet32_benchmark.log + + sleep 1m + +done diff --git a/doc/fluid/design/dist_train/distributed_traing_review.md b/doc/fluid/design/dist_train/distributed_traing_review.md new file mode 100644 index 0000000000000000000000000000000000000000..74066a3c2bf7b5e43b18c271de3e18271ee80925 --- /dev/null +++ b/doc/fluid/design/dist_train/distributed_traing_review.md @@ -0,0 +1,48 @@ +# Parallelism, Asynchronous, Synchronous, Codistillation + + +For valuable models, it’s worth using more hardware resources to reduce the training time and improve the final model quality. This doc discuss various solutions, their empirical results and some latest researches. + +# Model Parallelism +In some situations, larger and more complex models can improve the model quality. Sometimes, such models cannot fit in one device. Sometimes, parts of the model can be executed in parallel to improve speed. Model Parallelism address the issues by partitioning a single model and place the shards on several devices for execution. + +A common way of model parallelism is partition the logic of “gradient application” to parameter servers, while leaving the forward and backward computation at training servers. + +More flexible model parallelism is challenging. For example, multi-level-single-direction LSTM can be partitioned by layers, while such solution is not helpful for bi-directional LSTM. Different models can have quite different ways of partitioning and the benefits also depend on the underlying hardware. Framework needs to provide flexible APIs for user to define the customized partition scheme. For example, in TensorFlow, user can use tf.device() to specify the device placement. In MxNet, mx.AttrScope(ctx_group='dev1') does similar things. Recent research proposes to automatically find the optimal partition scheme with Reinforcement Learning, which is essentially solution space search algorithm that could cost a lot of extra hardware sources. + +# Data Parallelism +Data Parallelism runs the same model on multiple devices, each taking in a partition of the input batch. It’s more commonly used for a few reasons. It generally applies to common SGD mini-batch training. Compared with model parallelism, which requires users to carefully partition their model and tune for good performance, data parallelism usually involves no more than calling an extra API and speed up is more predictable. + +# Asynchronous Training +In asynchronous training, it usually involves a set of trainers and a set of parameter servers. The parameter servers collectively hold a single copy of shared parameters. While the trainers each holds a unique copy of model and trains the model independently. Each trainer pulls parameters from parameter servers and sends gradients to the parameter servers independently. Similarly the parameter servers applies the gradients to parameters as soon as the gradients are received and sends parameters whenever they are requested. + +In theory, asynchronous training is not safe and unstable. Each trainer is very likely using stale copy of parameters and parameters are also likely to apply stale gradients. However, in practice, especially for large-scale nonconvex optimization, it is effective [1]. Compared with synchronous solution, which will be discussed later, asynchronous distributed training is easier to implement and scales to a few dozen workers without losing much performance due to network communication or other overhead. Besides, asynchronous training can make progress even in case of random trainer failure in the cluster. + +Many production models, such as [3], are trained with distributed asynchronous solutions due to its scalability and effectiveness in practice. However, asynchronous training has its limitations. Usually, it’s not as stable as synchronous training. A warm-up phase is sometimes needed. Learning rate is usually smaller compared with synchronous training and decay is also often needed. Normally, asynchronous training doesn’t scale beyond 100 trainers. In other words, when putting more trainers beyond that, the model cannot converge faster. + +# Synchronous Training +Unlike asynchronous training, synchronous training requires step barriers. Parameter servers needs to wait for gradients from all trainers before they are applied to parameters and trainers will always pull the latest parameters. + +An obvious advantage of synchronous training is that the behavior is more clearly defined. Usually, it's more stable than asynchronous training. Learning rate can be set larger and for some vision tasks, the final accuracy can be slightly higher. (In my practical experience, for some models, it can actually be worse). + +Synchronous training usually faces scalability and performance issues, if not carefully implemented or deployed. In [2], native synchronous training can be 20%~40% slower than asynchronous training. A common trick to avoid slowness, discussed in [1] and [2], is to have backups. N+M replicas are scheduled while only the first N is needed for the training step the proceed. + +Similar to asynchronous training, the benefit of synchronous training diminishes quickly. Depending on the models, increasing the number of trainers (effectively batch size) beyond a point won’t delivers faster converge time or better final model quality. + +# Codistillation +Codistillation is a technique that tries to scale the training further. A few training instance (each training instance can be distributed) are performed during the same period. Each training instance has extra losses that comes from the prediction of other training instances. (likey teacher and student) The training process converges faster and usually converge to a better model quality. [4] + + +# Reference + +[1] Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. + +[2] Jianmin Chen, Rajat Monga, Samy Bengio, and Rafal Jozefowicz. Revisiting distributed synchronous SGD. + +[3] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. + +[4] LARGE SCALE DISTRIBUTED NEURAL NETWORK TRAINING THROUGH ONLINE DISTILLATION + + + + diff --git a/doc/v2/howto/cluster/multi_cluster/index_en.rst b/doc/v2/howto/cluster/multi_cluster/index_en.rst index b69bd5b2dbf1967d65558da06812d76f431c1d5a..9bc1eb2e3796d95dd69b165e916e263ea34b87f6 100644 --- a/doc/v2/howto/cluster/multi_cluster/index_en.rst +++ b/doc/v2/howto/cluster/multi_cluster/index_en.rst @@ -8,28 +8,28 @@ The user's cluster environment is not the same. To facilitate everyone's deploym .. toctree:: :maxdepth: 1 - k8s_cn.md - k8s_distributed_cn.md + k8s_en.md + k8s_distributed_en.md `OpenMPI `_ is a mature high-performance parallel computing framework, which is widely used in the field of HPC. The following guide describes how to use OpenMPI to build PaddlePaddle's cluster training task: .. toctree:: :maxdepth: 1 - openmpi_cn.md + openmpi_en.md `Fabric `_ is a convenient tool for program deployment and management. We provide a way to deploy and manage with Fabric. If you want to know more about it, please read the following guidelines: .. toctree:: :maxdepth: 1 - fabric_cn.md + fabric_en.md We also support the deployment of PaddlePaddle on AWS. Learn more about: .. toctree:: :maxdepth: 1 - k8s_aws_cn.md + k8s_aws_en.md -The examples can be found under `cluster_train_v2 `_ . \ No newline at end of file +The examples can be found under `cluster_train_v2 `_ . diff --git a/paddle/capi/Matrix.cpp b/paddle/capi/Matrix.cpp index 24b0020636c0141b87dc80f5079f7342ec28157c..733d49cacfda17ad19b7bd7918be73c1fd14a64f 100644 --- a/paddle/capi/Matrix.cpp +++ b/paddle/capi/Matrix.cpp @@ -108,7 +108,7 @@ paddle_error paddle_matrix_get_row(paddle_matrix mat, paddle_error paddle_matrix_get_shape(paddle_matrix mat, uint64_t* height, uint64_t* width) { - if (mat == nullptr) return kPD_NULLPTR; + if (mat == nullptr || cast(mat)->mat == nullptr) return kPD_NULLPTR; if (height != nullptr) { *height = cast(mat)->mat->getHeight(); } diff --git a/paddle/cuda/include/hl_base.h b/paddle/cuda/include/hl_base.h index 402302a5bfa877be33bf3bbd6ba05c8f3e48c3ba..77f5d82dbe2cad183491033736bac85961b6d320 100644 --- a/paddle/cuda/include/hl_base.h +++ b/paddle/cuda/include/hl_base.h @@ -229,6 +229,11 @@ extern __thread cudaStream_t default_stream; // __shfl has been deprecated as of CUDA 9.0. #if CUDA_VERSION < 9000 +template +__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) { + return __shfl_down(val, delta); +} + template __forceinline__ __device__ T __shfl_sync(unsigned, T val, int src_line, int width) { diff --git a/paddle/cuda/src/hl_top_k.cu b/paddle/cuda/src/hl_top_k.cu index 59ba552f560dab904d4983e0778ff57be9477c3e..b17290557c4f635a963d88525409b9373b057a4b 100644 --- a/paddle/cuda/src/hl_top_k.cu +++ b/paddle/cuda/src/hl_top_k.cu @@ -12,9 +12,9 @@ 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 "hl_base.h" -#include "hl_sparse.ph" -#include "hl_top_k.h" +#include "paddle/cuda/include/hl_base.h" +#include "paddle/cuda/include/hl_sparse.ph" +#include "paddle/cuda/include/hl_top_k.h" #include "paddle/utils/Logging.h" // using namespace hppl; @@ -244,8 +244,9 @@ __device__ __forceinline__ void blockReduce(Pair* shTopK, if (--beamSize == 0) break; __syncthreads(); + // NOTE(zcd): temporary solution unsigned mask = 0u; - // CREATE_SHFL_MASK(mask, tid < len); + CREATE_SHFL_MASK(mask, true); if (tid == maxId[0]) { if (beam < maxLength) { diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index 756fd2afd62717da55ba696950abbb279b82d4b9..327409914eb50212decb3eaba68831a7a682a095 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -53,7 +53,7 @@ void BroadcastOpHandle::RunImpl() { Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var); - // NOTE(zcd): the Place of input can be get from in_tensor and in_var_handle , + // NOTE(zcd): the Place of input can get from in_tensor and in_var_handle , // maybe they are different, because the Place that getting from in_tensor is // determined at runtime, the other is determined at building SSA graph stage. // If they are different, DataTransform should be applied. Currently, it has diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index 32576423a62a1a12f085d565e7ff267145bf979c..d70f26026c28867e592a9f8e37cc53e6c1d6d85e 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -93,6 +93,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { RunImpl(scope, place); } +bool OperatorBase::HasInputs(const std::string& name) const { + if (inputs_.find(name) != inputs_.end()) { + return true; + } else { + return false; + } +} + std::string OperatorBase::Input(const std::string& name) const { auto& ins = Inputs(name); PADDLE_ENFORCE_LE(ins.size(), 1UL, @@ -109,6 +117,14 @@ const std::vector& OperatorBase::Inputs( return it->second; } +bool OperatorBase::HasOutputs(const std::string& name) const { + if (outputs_.find(name) != outputs_.end()) { + return true; + } else { + return false; + } +} + std::string OperatorBase::Output(const std::string& name) const { auto& outs = Outputs(name); PADDLE_ENFORCE_LE(outs.size(), 1UL, @@ -220,13 +236,18 @@ void OperatorBase::CheckAllInputOutputSet() const { if (op_info == nullptr || op_info->proto_ == nullptr) return; for (auto& in : op_info->Proto().inputs()) { - PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(), - "Type %s's input %s is not set", Type(), in.name()); + if (!in.dispensable()) { + PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(), + "Operator %s's input, %s, is not set", Type(), in.name()); + } } for (auto& out : op_info->Proto().outputs()) { - PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(), - "Type %s's output %s is not set", Type(), out.name()); + if (!out.dispensable()) { + PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(), + "Operator %s's output, %s, is not set", Type(), + out.name()); + } } } @@ -332,6 +353,9 @@ class RuntimeInferShapeContext : public InferShapeContext { : op_(op), scope_(scope) {} bool HasInput(const std::string& name) const override { + if (!op_.HasInputs(name)) { + return false; + } auto& ins = Inputs(name); size_t length = ins.size(); if (length == 0) { @@ -345,6 +369,9 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasOutput(const std::string& name) const override { + if (!op_.HasOutputs(name)) { + return false; + } auto& outs = Outputs(name); size_t length = outs.size(); if (length == 0) { @@ -358,6 +385,9 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasInputs(const std::string& name) const override { + if (!op_.HasInputs(name)) { + return false; + } auto inputs = op_.Inputs(name); if (inputs.empty()) { return false; @@ -371,6 +401,9 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasOutputs(const std::string& name) const override { + if (!op_.HasOutputs(name)) { + return false; + } auto outputs = op_.Outputs(name); if (outputs.empty()) { return false; diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 826cc57b725ab4b52e5d67ab82e939cbd62a8460..d373c48b1a75c5f75c7520b56f230bc2c146b174 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -105,6 +105,7 @@ class OperatorBase { const VariableNameMap& Inputs() const { return inputs_; } const VariableNameMap& Outputs() const { return outputs_; } + bool HasInputs(const std::string& name) const; //! Get a input with argument's name described in `op_proto` std::string Input(const std::string& name) const; //! Get a input which has multiple variables. @@ -112,6 +113,7 @@ class OperatorBase { //! Get all inputs variable names std::vector InputVars() const; + bool HasOutputs(const std::string& name) const; //! Get a output with argument's name described in `op_proto` std::string Output(const std::string& name) const; //! Get an output which has multiple variables. diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index 4b5866ad5dd5c2e33830b61e575ab92954cecb39..288789d6e484100820c937e6081701f1e9245706 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1,4 +1,4 @@ -if(WITH_TESTING) - nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) - nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda) -endif() +nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) +nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda) +set(ENGINE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/engine.cc) +add_subdirectory(convert) diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..572e29515f89685c528023cf6569ef8e559d6874 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -0,0 +1,3 @@ +nv_test(test_tensorrt_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES}) +nv_test(test_tensorrt_activation_op SRCS test_activation_op.cc ${ENGINE_FILE} activation_op.cc + DEPS ${FLUID_CORE_MODULES} activation_op) diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..543784289cfc51048057e467d36fdd1f334eb903 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -0,0 +1,40 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class ReluOpConverter : public OpConverter { + public: + ReluOpConverter() {} + void operator()(const framework::OpDesc& op) override { + LOG(INFO) << "convert a fluid relu op to tensorrt activation layer whose " + "type is Relu"; + const nvinfer1::ITensor* input_tensor = + engine_->GetITensor(op.Input("X")[0]); + nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER( + engine_, Activation, *const_cast(input_tensor), + nvinfer1::ActivationType::kRELU); + engine_->SetITensor(op.Output("Out")[0], layer->getOutput(0)); + } +}; + +REGISTER_TRT_OP_CONVERTER(relu, ReluOpConverter); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..431500b90e144e2a30fe705b72e93452f806ca65 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -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. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class Conv2dOpConverter : public OpConverter { + public: + Conv2dOpConverter() {} + void operator()(const framework::OpDesc& op) override { + LOG(INFO) + << "convert a fluid conv2d op to tensorrt conv layer without bias"; + } +}; + +REGISTER_TRT_OP_CONVERTER(conv2d, Conv2dOpConverter); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/mul_op.cc b/paddle/fluid/inference/tensorrt/convert/mul_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f9834ab156c9dcc11f4e89075b7bf5457cf00268 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/mul_op.cc @@ -0,0 +1,33 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class MulOpConverter : public OpConverter { + public: + MulOpConverter() {} + void operator()(const framework::OpDesc& op) override { + LOG(INFO) << "convert a fluid mul op to tensorrt fc layer without bias"; + } +}; + +REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h new file mode 100644 index 0000000000000000000000000000000000000000..f8ca219bb83c6db3ba640464233c6cbd337926a6 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -0,0 +1,91 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/inference/tensorrt/engine.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * Convert Op from Fluid to TensorRT Engine. + */ +class OpConverter { + public: + OpConverter() {} + virtual void operator()(const framework::OpDesc& op) {} + + void Execute(const framework::OpDesc& op, TensorRTEngine* engine) { + std::string type = op.Type(); + auto it = converters_.find(type); + PADDLE_ENFORCE(it != converters_.end(), "no OpConverter for optype [%s]", + type); + it->second->SetEngine(engine); + (*it->second)(op); + } + + static OpConverter& Global() { + static auto* x = new OpConverter; + return *x; + } + + template + void Register(const std::string& key) { + converters_[key] = new T; + } + + // convert fluid op to tensorrt layer + void ConvertOp(const framework::OpDesc& op, TensorRTEngine* engine) { + OpConverter::Global().Execute(op, engine); + } + + // convert fluid block to tensorrt network + void ConvertBlock(const framework::BlockDesc& block, TensorRTEngine* engine) { + for (auto op : block.AllOps()) { + OpConverter::Global().Execute(*op, engine); + } + } + + void SetEngine(TensorRTEngine* engine) { engine_ = engine; } + + virtual ~OpConverter() {} + + // TensorRT engine + TensorRTEngine* engine_{nullptr}; + + private: + // registered op converter map, whose key is the fluid op type, and value is + // the pointer position of corresponding OpConverter class. + std::unordered_map converters_; + // fluid inference scope + framework::Scope* scope_{nullptr}; +}; + +#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ + struct trt_##op_type__##_converter { \ + trt_##op_type__##_converter() { \ + OpConverter::Global().Register(#op_type__); \ + } \ + }; \ + trt_##op_type__##_converter trt_##op_type__##_converter__; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..0f390bee1f5967547ea21e6d42a1b4300de58d04 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc @@ -0,0 +1,94 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/place.h" + +USE_OP(relu); + +namespace paddle { +namespace inference { +namespace tensorrt { + +void compare(float input, float expect) { + framework::Scope scope; + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); + + // init fluid op and variable + auto x_var = scope.Var("X"); + auto x_tensor = x_var->GetMutable(); + x_tensor->Resize({1, 1}); + std::vector init; + init.push_back(input); + framework::TensorFromVector(init, ctx, x_tensor); + + auto out_var = scope.Var("Out"); + auto out_tensor = out_var->GetMutable(); + out_tensor->Resize({1, 1}); + out_tensor->mutable_data(place); + + framework::OpDesc op_desc; + op_desc.SetType("relu"); + op_desc.SetInput("X", {"X"}); + op_desc.SetOutput("Out", {"Out"}); + + auto relu_op = framework::OpRegistry::CreateOp(op_desc); + + // run fluid op + relu_op->Run(scope, place); + std::vector out1; + framework::TensorToVector(*out_tensor, ctx, &out1); + + // init tensorrt op + cudaStream_t stream; + ASSERT_EQ(0, cudaStreamCreate(&stream)); + TensorRTEngine* engine = new TensorRTEngine(1, 1 << 10, &stream); + engine->InitNetwork(); + engine->DeclareInput("X", nvinfer1::DataType::kFLOAT, + nvinfer1::DimsCHW{1, 1, 1}); + + OpConverter op_converter; + op_converter.ConvertOp(op_desc, engine); + + engine->DeclareOutput("Out"); + engine->FreezeNetwork(); + engine->SetInputFromCPU("X", &input, 1 * sizeof(float)); + + // run tensorrt op + engine->Execute(1); + + float out2; + engine->GetOutputInCPU("Out", &out2, 1 * sizeof(float)); + + ASSERT_EQ(out1[0], out2); + ASSERT_EQ(out1[0], expect); + + delete engine; + cudaStreamDestroy(stream); +} + +TEST(OpConverter, ConvertRelu) { + compare(1, 1); // relu(1) = 1 + compare(-5, 0); // relu(-5) = 0 +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc new file mode 100644 index 0000000000000000000000000000000000000000..5c5ac10394d3cb76ac4b8409c7fa79eb5e9a90b2 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -0,0 +1,37 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(BlockConverter, ConvertBlock) { + framework::ProgramDesc prog; + auto* block = prog.MutableBlock(0); + auto* mul_op = block->AppendOp(); + mul_op->SetType("mul"); + auto* conv2d_op = block->AppendOp(); + conv2d_op->SetType("conv2d"); + + OpConverter converter; + converter.ConvertBlock(*block, nullptr /*TensorRTEngine*/); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index 03a25f8e8b52b452cb621aeddf3563cc21a033df..df123a59079acc5f549e733b412ab302aa397a92 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -80,8 +80,8 @@ nvinfer1::ITensor* TensorRTEngine::DeclareInput(const std::string& name, PADDLE_ENFORCE(infer_network_ != nullptr, "should initnetwork first"); auto* input = infer_network_->addInput(name.c_str(), dtype, dim); PADDLE_ENFORCE(input, "infer network add input %s failed", name); - buffer_sizes_[name] = kDataTypeSize[static_cast(dtype)] * AccumDims(dim); + TensorRTEngine::SetITensor(name, input); return input; } @@ -99,6 +99,19 @@ void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer* layer, int offset, buffer_sizes_[name] = 0; } +void TensorRTEngine::DeclareOutput(const std::string& name) { + PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate output name %s", + name); + + auto* output = TensorRTEngine::GetITensor(name); + PADDLE_ENFORCE(output != nullptr); + output->setName(name.c_str()); + infer_network_->markOutput(*output); + // output buffers' size can only be decided latter, set zero here to mark this + // and will reset latter. + buffer_sizes_[name] = 0; +} + void* TensorRTEngine::GetOutputInGPU(const std::string& name) { return buffer(name); } @@ -110,7 +123,6 @@ void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, PADDLE_ENFORCE(it != buffer_sizes_.end()); PADDLE_ENFORCE_GT(it->second, 0); PADDLE_ENFORCE_GE(max_size, it->second); - PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buffer(name), it->second, cudaMemcpyDeviceToHost, *stream_)); } @@ -126,10 +138,24 @@ void*& TensorRTEngine::buffer(const std::string& name) { void TensorRTEngine::SetInputFromCPU(const std::string& name, void* data, size_t size) { void* buf = buffer(name); + cudaMemcpyAsync(buf, data, size, cudaMemcpyHostToDevice, *stream_); PADDLE_ENFORCE_EQ( 0, cudaMemcpyAsync(buf, data, size, cudaMemcpyHostToDevice, *stream_)); } +void TensorRTEngine::SetITensor(const std::string& name, + nvinfer1::ITensor* tensor) { + PADDLE_ENFORCE(tensor != nullptr); + PADDLE_ENFORCE_EQ(0, itensor_map_.count(name), "duplicate itensor name %s", + name); + itensor_map_[name] = tensor; +} + +nvinfer1::ITensor* TensorRTEngine::GetITensor(const std::string& name) { + PADDLE_ENFORCE(itensor_map_.count(name), "no itensor %s", name); + return itensor_map_[name]; +} + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index d6d4c2f8a2ced08ed8481e92e131f6e2bed9ec05..ec919b943d3281dd675b15e2f14adb7b3487f46f 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -80,6 +80,8 @@ class TensorRTEngine : public EngineBase { // name. void DeclareOutput(const nvinfer1::ILayer* layer, int offset, const std::string& name); + // Set the itensor_map_[name] as the network's output, and set its name. + void DeclareOutput(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 @@ -98,6 +100,10 @@ class TensorRTEngine : public EngineBase { // LOW EFFICENCY! Get output to CPU, this will trigger a memory copy from GPU // to CPU. void GetOutputInCPU(const std::string& name, void* dst, size_t max_size); + // Fill an ITensor into map itensor_map_. + void SetITensor(const std::string& name, nvinfer1::ITensor* tensor); + // Get an ITensor called name. + nvinfer1::ITensor* GetITensor(const std::string& name); nvinfer1::ICudaEngine* engine() { return infer_engine_.get(); } nvinfer1::INetworkDefinition* network() { return infer_network_.get(); } @@ -113,6 +119,8 @@ class TensorRTEngine : public EngineBase { std::vector buffers_; // max data size for the buffers. std::unordered_map buffer_sizes_; + std::unordered_map + itensor_map_; // TensorRT related internal members template diff --git a/paddle/fluid/inference/tensorrt/test_engine.cc b/paddle/fluid/inference/tensorrt/test_engine.cc index c6e1c71cdc8eea48ab6197057d1d389f03c9cf54..a08b78f930d30d674247a713fadd3e42e5ada350 100644 --- a/paddle/fluid/inference/tensorrt/test_engine.cc +++ b/paddle/fluid/inference/tensorrt/test_engine.cc @@ -70,7 +70,6 @@ TEST_F(TensorRTEngineTest, add_layer) { engine_->Execute(1); LOG(INFO) << "to get output"; - // void* y_v = float y_cpu; engine_->GetOutputInCPU("y", &y_cpu, sizeof(float)); diff --git a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc index 1a685b9e2ebcd7d4b5a057b506bccf6adcd9952c..c4fd1e298b0daea85db2a407d04ad2d7bcdee0f0 100644 --- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc +++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc @@ -16,9 +16,12 @@ limitations under the License. */ #include "gtest/gtest.h" #include "paddle/fluid/inference/tests/test_helper.h" +DEFINE_string(data_set, "cifar10", "Data set to test"); DEFINE_string(dirname, "", "Directory of the inference model."); +DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model."); DEFINE_int32(batch_size, 1, "Batch size of input data"); DEFINE_int32(repeat, 1, "Running the inference program repeat times"); +DEFINE_bool(skip_cpu, false, "Skip the cpu test"); TEST(inference, image_classification) { if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) { @@ -35,20 +38,31 @@ TEST(inference, image_classification) { paddle::framework::LoDTensor input; // Use normilized image pixels as input data, // which should be in the range [0.0, 1.0]. - SetupTensor(&input, {FLAGS_batch_size, 3, 32, 32}, - static_cast(0), static_cast(1)); + if (FLAGS_data_set == "cifar10") { + SetupTensor(&input, {FLAGS_batch_size, 3, 32, 32}, + static_cast(0), static_cast(1)); + } else if (FLAGS_data_set == "imagenet") { + SetupTensor(&input, {FLAGS_batch_size, 3, 224, 224}, + static_cast(0), static_cast(1)); + } else { + LOG(FATAL) << "Only cifar10 or imagenet is supported."; + } + std::vector cpu_feeds; cpu_feeds.push_back(&input); paddle::framework::LoDTensor output1; - std::vector cpu_fetchs1; - cpu_fetchs1.push_back(&output1); - - // Run inference on CPU - LOG(INFO) << "--- CPU Runs: ---"; - TestInference( - dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat); - LOG(INFO) << output1.dims(); + if (!FLAGS_skip_cpu) { + std::vector cpu_fetchs1; + cpu_fetchs1.push_back(&output1); + + // Run inference on CPU + LOG(INFO) << "--- CPU Runs: ---"; + LOG(INFO) << "Batch size is " << FLAGS_batch_size; + TestInference( + dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat); + LOG(INFO) << output1.dims(); + } #ifdef PADDLE_WITH_CUDA paddle::framework::LoDTensor output2; @@ -57,24 +71,27 @@ TEST(inference, image_classification) { // Run inference on CUDA GPU LOG(INFO) << "--- GPU Runs: ---"; + LOG(INFO) << "Batch size is " << FLAGS_batch_size; TestInference( dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat); LOG(INFO) << output2.dims(); - CheckError(output1, output2); + if (!FLAGS_skip_cpu) { + CheckError(output1, output2); + } // float16 inference requires cuda GPUs with >= 5.3 compute capability - if (paddle::platform::GetCUDAComputeCapability(0) >= 53) { + if (!FLAGS_fp16_dirname.empty() && + paddle::platform::GetCUDAComputeCapability(0) >= 53) { paddle::framework::LoDTensor output3; std::vector cpu_fetchs3; cpu_fetchs3.push_back(&output3); LOG(INFO) << "--- GPU Runs in float16 mode: ---"; - std::string fp16_dirname = dirname; - fp16_dirname.replace(fp16_dirname.find("book/"), - std::string("book/").size(), "book/float16_"); + LOG(INFO) << "Batch size is " << FLAGS_batch_size; + TestInference( - fp16_dirname, cpu_feeds, cpu_fetchs3, FLAGS_repeat); + FLAGS_fp16_dirname, cpu_feeds, cpu_fetchs3, FLAGS_repeat); CheckError(output2, output3); } diff --git a/paddle/fluid/operators/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/batch_norm_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..0e4a56d4a45a732cfcf43b09228bc0c44df5924c --- /dev/null +++ b/paddle/fluid/operators/batch_norm_mkldnn_op.cc @@ -0,0 +1,325 @@ +/* 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 "mkldnn.hpp" +#include "paddle/fluid/operators/batch_norm_op.h" +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using paddle::platform::MKLDNNDeviceContext; +using paddle::platform::MKLDNNMemDesc; +using mkldnn::memory; + +template +using EigenArrayMap = + Eigen::Map>; +template +using ConstEigenArrayMap = + Eigen::Map>; +template +using EigenVectorArrayMap = Eigen::Map>; +template +using ConstEigenVectorArrayMap = + Eigen::Map>; + +namespace { +template +struct bn_type_traits { + using op_type = T; + using op_desc = typename op_type::desc; + using op_prim = typename op_type::primitive_desc; +}; + +template +void copy_to_weights(T scale_begin, T scale_end, T shift_begin, T shift_end, + Container *c) { + auto it = std::begin(*c); + + std::copy(scale_begin, scale_end, std::inserter(*c, it)); + std::copy( + shift_begin, shift_end, + std::inserter(*c, std::next(it, std::distance(scale_begin, scale_end)))); +} + +template +void run_batch_norm_op(Args &&... args) { + Op batch_norm_op{args...}; + + std::vector pipeline; + pipeline.push_back(batch_norm_op); + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); +} + +template +inline void *cast_const_to_void(const T *t) { + return static_cast(const_cast(t)); +} +} // namespace + +template +class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto data_layout_str = ctx.Attr("data_layout"); + auto data_layout = framework::StringToDataLayout(data_layout_str); + PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, + "MKLDNN batch normalization handles only NCHW data layout"); + + const float epsilon = ctx.Attr("epsilon"); + const float momentum = ctx.Attr("momentum"); + const bool is_test = ctx.Attr("is_test"); + + const auto *x = ctx.Input("X"); + const auto *mean = ctx.Input("Mean"); + const auto *variance = ctx.Input("Variance"); + + auto &dev_ctx = ctx.template device_context(); + auto mkldnn_engine = dev_ctx.GetEngine(); + + auto *y = ctx.Output("Y"); + auto *mean_out = ctx.Output("MeanOut"); + auto *variance_out = ctx.Output("VarianceOut"); + auto *batch_mean = ctx.Output("SavedMean"); + auto *batch_variance = ctx.Output("SavedVariance"); + + const auto *scale = ctx.Input("Scale"); + const auto *shift = ctx.Input("Bias"); + + y->mutable_data(ctx.GetPlace()); + mean_out->mutable_data(ctx.GetPlace()); + variance_out->mutable_data(ctx.GetPlace()); + + if (!is_test) { + batch_mean->mutable_data(ctx.GetPlace()); + batch_variance->mutable_data(ctx.GetPlace()); + } + + auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring + : mkldnn::prop_kind::forward_training; + + auto dims = paddle::framework::vectorize2int(x->dims()); + + auto src_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + auto dst_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + + auto src_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine}; + auto dst_pd = mkldnn::memory::primitive_desc{dst_md, mkldnn_engine}; + + auto src = mkldnn::memory{src_pd, cast_const_to_void(x->data())}; + auto dst = mkldnn::memory{dst_pd, y->data()}; + + unsigned flags = mkldnn::use_scale_shift; + if (is_test) flags |= mkldnn::use_global_stats; + + using bn_fwd_types = bn_type_traits; + auto batch_norm_fwd_desc = + bn_fwd_types::op_desc{propagation, src_md, epsilon, flags}; + auto batch_norm_fwd_pd = + bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + + const unsigned int ic = dims[1]; + + // MKLDNN requires a single piece of memory for scale and shift/bias data + const size_t scaleshift_size = 2 * ic; + std::vector scaleshift_data; + scaleshift_data.reserve(scaleshift_size); + + copy_to_weights(scale->data(), scale->data() + ic, shift->data(), + shift->data() + ic, &scaleshift_data); + + auto scaleshift_memory = mkldnn::memory{ + batch_norm_fwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + + if (is_test) { + auto mean_memory = mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), + cast_const_to_void(mean->data())}; + + auto variance_memory = + mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), + cast_const_to_void(variance->data())}; + + run_batch_norm_op( + batch_norm_fwd_pd, src, (const mkldnn::primitive::at &)mean_memory, + (const mkldnn::primitive::at &)variance_memory, scaleshift_memory, + dst); + } else { + auto mean_memory = + mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), + cast_const_to_void(batch_mean->data())}; + + auto variance_memory = + mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), + cast_const_to_void(batch_variance->data())}; + + run_batch_norm_op(batch_norm_fwd_pd, src, + scaleshift_memory, dst, + mean_memory, variance_memory); + } + + if (!is_test) { + const unsigned int in = dims[0]; + const unsigned int sample_size = x->numel() / in / ic; + + // saved_xx is use just in this batch of data + EigenVectorArrayMap saved_mean_e( + batch_mean->mutable_data(ctx.GetPlace()), ic); + EigenVectorArrayMap saved_variance_e( + batch_variance->mutable_data(ctx.GetPlace()), ic); + saved_mean_e.setZero(); + saved_variance_e.setZero(); + + const unsigned int x_arr_size = in * ic; + ConstEigenArrayMap x_arr(x->data(), sample_size, x_arr_size); + for (unsigned int nc = 0; nc < x_arr_size; ++nc) { + saved_mean_e(nc % ic) += x_arr.col(nc).sum(); + } + saved_mean_e /= in * sample_size; + for (unsigned int nc = 0; nc < x_arr_size; ++nc) { + saved_variance_e(nc % ic) += + (x_arr.col(nc) - saved_mean_e(nc % ic)).matrix().squaredNorm(); + } + saved_variance_e /= in * sample_size; + + ConstEigenVectorArrayMap mean_arr{mean->data(), ic}; + ConstEigenVectorArrayMap variance_arr{variance->data(), ic}; + + EigenVectorArrayMap running_mean_arr( + mean_out->mutable_data(ctx.GetPlace()), ic); + EigenVectorArrayMap running_var_arr( + variance_out->mutable_data(ctx.GetPlace()), ic); + + auto one_minus_momentum = 1. - momentum; + running_mean_arr = + mean_arr * momentum + saved_mean_e * one_minus_momentum; + running_var_arr = + variance_arr * momentum + saved_variance_e * one_minus_momentum; + } + } +}; + +template +class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext &ctx) const override { + auto data_layout_str = ctx.Attr("data_layout"); + auto data_layout = framework::StringToDataLayout(data_layout_str); + PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, + "MKLDNN batch normalization handles only NCHW data layout"); + + auto &dev_ctx = ctx.template device_context(); + auto mkldnn_engine = dev_ctx.GetEngine(); + + const float epsilon = ctx.Attr("epsilon"); + + const auto *x = ctx.Input("X"); + const auto *scale = ctx.Input("Scale"); + const auto *shift = ctx.Input("Bias"); + const auto *batch_mean = ctx.Input("SavedMean"); + const auto *batch_variance = ctx.Input("SavedVariance"); + + const auto *diff_y = ctx.Input(framework::GradVarName("Y")); + auto *diff_x = ctx.Output(framework::GradVarName("X")); + auto *diff_scale = ctx.Output(framework::GradVarName("Scale")); + auto *diff_shift = ctx.Output(framework::GradVarName("Bias")); + + diff_x->mutable_data(ctx.GetPlace()); + diff_scale->mutable_data(ctx.GetPlace()); + diff_shift->mutable_data(ctx.GetPlace()); + + auto dims = paddle::framework::vectorize2int(x->dims()); + unsigned flags = mkldnn::use_scale_shift | !mkldnn::use_global_stats; + + auto src_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + auto dst_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + auto diff_src_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + auto diff_dst_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + + using bn_bwd_types = bn_type_traits; + using bn_fwd_types = bn_type_traits; + + auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ + mkldnn::prop_kind::forward_training, src_md, epsilon, flags}; + auto batch_norm_fwd_pd = + bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + + auto batch_norm_bwd_desc = bn_bwd_types::op_desc{ + mkldnn::prop_kind::backward, diff_dst_md, dst_md, epsilon, flags}; + auto batch_norm_bwd_pd = bn_bwd_types::op_prim{ + batch_norm_bwd_desc, mkldnn_engine, batch_norm_fwd_pd}; + + auto src = mkldnn::memory{{src_md, mkldnn_engine}, + cast_const_to_void(x->data())}; + + auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(), + cast_const_to_void(batch_mean->data())}; + + auto variance = + mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(), + cast_const_to_void(batch_variance->data())}; + + auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine}, + cast_const_to_void(diff_y->data())}; + + const unsigned int ic = dims[1]; + + const size_t scaleshift_size = 2 * ic; + + std::vector scaleshift_data; + scaleshift_data.reserve(scaleshift_size); + copy_to_weights(scale->data(), scale->data() + ic, shift->data(), + shift->data() + ic, &scaleshift_data); + + auto scaleshift_memory = mkldnn::memory{ + batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + + std::vector diff_scaleshift_data; + diff_scaleshift_data.reserve(scaleshift_size); + copy_to_weights(diff_scale->data(), diff_scale->data() + ic, + diff_shift->data(), diff_shift->data() + ic, + &diff_scaleshift_data); + + auto diff_scaleshift_memory = + mkldnn::memory{batch_norm_bwd_pd.diff_weights_primitive_desc(), + diff_scaleshift_data.data()}; + + auto diff_src = mkldnn::memory{{diff_src_md, mkldnn_engine}, + static_cast(diff_x->data())}; + + run_batch_norm_op( + batch_norm_bwd_pd, src, mean, variance, diff_dst, scaleshift_memory, + diff_src, diff_scaleshift_memory); + + auto it = std::begin(diff_scaleshift_data); + std::copy(it, std::next(it, ic), diff_scale->data()); + std::copy(std::next(it, ic), std::end(diff_scaleshift_data), + diff_shift->data()); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace, + ops::BatchNormMKLDNNOpKernel); +REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace, + ops::BatchNormMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index f8b2505ccfb143f9f74cf0b16d92e8e1ca059709..b4bd40d0311bf10ec1fddabab2ee131fe02baf52 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -15,6 +15,9 @@ limitations under the License. */ #include "paddle/fluid/operators/batch_norm_op.h" #include #include "paddle/fluid/framework/data_layout.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif namespace paddle { namespace operators { @@ -106,7 +109,18 @@ class BatchNormOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ(bn_param_type, framework::ToDataType( ctx.Input("Variance")->type()), "Variance input should be of float type"); - return framework::OpKernelType(input_data_type, ctx.GetPlace()); + + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; + } +#endif + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, + library_); } }; @@ -151,6 +165,9 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { "Variance of the current mini batch, " "will apply to output when training") .AsIntermediate(); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Batch Normalization. @@ -349,8 +366,19 @@ class BatchNormGradOp : public framework::OperatorWithKernel { if (t == nullptr) { PADDLE_THROW("can't find Y@GRAD"); } - return framework::OpKernelType(framework::ToDataType(t->type()), - ctx.GetPlace()); + + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; + } +#endif + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + layout, library_); } }; @@ -474,6 +502,7 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker { op->SetInput(framework::GradVarName("Y"), OutputGrad("Y")); op->SetInput("Scale", Input("Scale")); + op->SetInput("Bias", Input("Bias")); op->SetInput("SavedMean", Output("SavedMean")); op->SetInput("SavedVariance", Output("SavedVariance")); diff --git a/paddle/fluid/operators/beam_search_decode_op.cc b/paddle/fluid/operators/beam_search_decode_op.cc index 4a8dfd4b54227070c2143b180f8ab92753885550..68fb988afd8af4e9ac3acb4506c1c31fcf85e5a3 100644 --- a/paddle/fluid/operators/beam_search_decode_op.cc +++ b/paddle/fluid/operators/beam_search_decode_op.cc @@ -23,16 +23,54 @@ struct BeamSearchDecodeFunctor { BeamSearchDecodeFunctor(const LoDTensorArray& step_ids, const LoDTensorArray& step_scores, LoDTensor* id_tensor, LoDTensor* score_tensor) - : step_ids_(step_ids), - step_scores_(step_scores), + : step_ids_origin_(step_ids), + step_scores_origin_(step_scores), id_tensor_(id_tensor), - score_tensor_(score_tensor) {} + score_tensor_(score_tensor) { + tensor_on_gpu_ = false; + // First make a copy of GPU data on CPU + if (platform::is_gpu_place(step_ids_origin_[0].place())) { + tensor_on_gpu_ = true; + platform::DeviceContextPool& pool = + platform::DeviceContextPool::Instance(); + auto* dev_ctx = pool.Get(step_ids_origin_[0].place()); + // Copy all tensors in the input tensor array + for (auto& step_id : step_ids_origin_) { + framework::LoDTensor out; + dev_ctx->Wait(); + framework::TensorCopy(step_id, platform::CPUPlace(), *dev_ctx, &out); + dev_ctx->Wait(); + + out.set_lod(step_id.lod()); + step_ids_.push_back(out); + } + } + if (platform::is_gpu_place(step_scores_origin_[0].place())) { + tensor_on_gpu_ = true; + platform::DeviceContextPool& pool = + platform::DeviceContextPool::Instance(); + auto* dev_ctx = pool.Get(step_scores_origin_[0].place()); + // Copy all tensors in the input tensor array + for (auto& step_score : step_scores_origin_) { + framework::LoDTensor out; + dev_ctx->Wait(); + framework::TensorCopy(step_score, platform::CPUPlace(), *dev_ctx, &out); + dev_ctx->Wait(); + + out.set_lod(step_score.lod()); + step_scores_.push_back(out); + } + } + } template void operator()() const; - const LoDTensorArray& step_ids_; - const LoDTensorArray& step_scores_; + bool tensor_on_gpu_; + const LoDTensorArray& step_ids_origin_; + const LoDTensorArray& step_scores_origin_; + LoDTensorArray step_ids_ = LoDTensorArray(); + LoDTensorArray step_scores_ = LoDTensorArray(); LoDTensor* id_tensor_; LoDTensor* score_tensor_; }; @@ -40,8 +78,14 @@ struct BeamSearchDecodeFunctor { template void BeamSearchDecodeFunctor::operator()() const { BeamSearchDecoder beam_search_decoder; - beam_search_decoder.PackAllSteps(step_ids_, step_scores_, id_tensor_, - score_tensor_); + // Check if the tensor is on GPU. If so, use the CPU copy instead + if (tensor_on_gpu_) { + beam_search_decoder.PackAllSteps(step_ids_, step_scores_, id_tensor_, + score_tensor_); + } else { + beam_search_decoder.PackAllSteps(step_ids_origin_, step_scores_origin_, + id_tensor_, score_tensor_); + } } template <> diff --git a/paddle/fluid/operators/bilinear_tensor_product_op.h b/paddle/fluid/operators/bilinear_tensor_product_op.h index ca80e6085c4f8b242cd26803202475cd50474bcd..7191711a731676298219d2f4bd95fdece27250da 100644 --- a/paddle/fluid/operators/bilinear_tensor_product_op.h +++ b/paddle/fluid/operators/bilinear_tensor_product_op.h @@ -61,9 +61,9 @@ class BilinearTensorProductKernel : public framework::OpKernel { auto output_col_vec = output_mat.chip(i, 1); Tensor weight_mat = weight->Slice(i, i + 1).Resize(framework::make_ddim({x_dim, y_dim})); - math::gemm(dev_ctx, CblasNoTrans, CblasNoTrans, - batch_size, y_dim, x_dim, 1, x->data(), - weight_mat.data(), 0, left_mul.data()); + math::GetBlas(dev_ctx).GEMM( + CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, x->data(), + weight_mat.data(), 0, left_mul.data()); output_col_vec.device(place) = (left_mul_mat * y_mat).sum(Eigen::DSizes(1)); } @@ -125,6 +125,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { set_zero(dev_ctx, d_y, static_cast(0)); } + auto blas = math::GetBlas(ctx); + // Caculate the Output(X@Grad) and Output(Y@Grad). if (d_x || d_y) { Eigen::DSizes bcast_for_x(1, y_dim); @@ -138,18 +140,16 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { output_vec.reshape(Eigen::DSizes(batch_size, 1)) .broadcast(bcast_for_x) * y_mat; - math::gemm( - dev_ctx, CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1, - y_scale.data(), weight_i.data(), 1, d_x->data()); + blas.GEMM(CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1, + y_scale.data(), weight_i.data(), 1, d_x->data()); } if (d_y) { x_scale_mat.device(place) = output_vec.reshape(Eigen::DSizes(batch_size, 1)) .broadcast(bcast_for_y) * x_mat; - math::gemm( - dev_ctx, CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, - x_scale.data(), weight_i.data(), 1, d_y->data()); + blas.GEMM(CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, + x_scale.data(), weight_i.data(), 1, d_y->data()); } } } @@ -166,9 +166,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { output_vec.reshape(Eigen::DSizes(batch_size, 1)) .broadcast(bcast_for_weight) * x_mat; - math::gemm(dev_ctx, CblasTrans, CblasNoTrans, x_dim, - y_dim, batch_size, 1, x_scale.data(), - y->data(), 0, d_weight_i.data()); + blas.GEMM(CblasTrans, CblasNoTrans, x_dim, y_dim, batch_size, 1, + x_scale.data(), y->data(), 0, d_weight_i.data()); } } diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc index 0e0622e290f42811c83c354d749ef32a2d9dcadb..2b2a9dc8319f964875371214168ce04cb67fc818 100644 --- a/paddle/fluid/operators/cross_entropy_op.cc +++ b/paddle/fluid/operators/cross_entropy_op.cc @@ -164,11 +164,13 @@ or not. But the output only shares the LoD information with input X. } // namespace paddle namespace ops = paddle::operators; +using CPUCtx = paddle::platform::CPUDeviceContext; + REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(cross_entropy_grad, ops::CrossEntropyGradientOp); -REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, - ops::CrossEntropyOpKernel); +REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, + ops::CrossEntropyOpKernel); REGISTER_OP_CPU_KERNEL(cross_entropy_grad, - ops::CrossEntropyGradientOpKernel, - ops::CrossEntropyGradientOpKernel); + ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel); diff --git a/paddle/fluid/operators/cross_entropy_op.cu b/paddle/fluid/operators/cross_entropy_op.cu index 6449149d4b55962e84baafffc0c2c03f8108866f..30dbd5bd3d39dd2992c3dd91364003bb7715a2eb 100644 --- a/paddle/fluid/operators/cross_entropy_op.cu +++ b/paddle/fluid/operators/cross_entropy_op.cu @@ -14,98 +14,11 @@ limitations under the License. */ #include "paddle/fluid/operators/cross_entropy_op.h" -namespace paddle { -namespace operators { - -namespace { - -template -__global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X, - const int64_t* label, const int N, - const int D) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - int idx = i * D + label[i]; - dX[idx] = -dY[i] / X[idx]; - } -} - -template -__global__ void SoftCrossEntropyGradientKernel(T* dX, const T* dY, const T* X, - const T* label, const int N, - const int D) { - int ids = blockIdx.x * blockDim.x + threadIdx.x; - if (ids < N * D) { - int row_ids = ids / D; - dX[ids] = -label[ids] * dY[row_ids] / X[ids]; - } -} -} // namespace - -template -class CrossEntropyOpCUDAKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "This kernel only runs on GPU device."); - const Tensor* x = ctx.Input("X"); - const Tensor* label = ctx.Input("Label"); - Tensor* y = ctx.Output("Y"); - y->mutable_data(ctx.GetPlace()); - - math::CrossEntropyFunctor()( - ctx.template device_context(), y, x, label, - ctx.Attr("soft_label")); - } -}; - -template -class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "This kernel only runs on GPU device."); - - const Tensor* x = ctx.Input("X"); - const Tensor* label = ctx.Input("Label"); - Tensor* dx = ctx.Output(framework::GradVarName("X")); - dx->mutable_data(ctx.GetPlace()); - - const T* dy_data = - ctx.Input(framework::GradVarName("Y"))->data(); - T* dx_data = dx->mutable_data(ctx.GetPlace()); - const T* x_data = x->data(); - - int64_t batch_size = x->dims()[0]; - int64_t class_num = x->dims()[1]; - - int block = 512; - int grid = (batch_size * class_num + block - 1) / block; - - auto& dev_ctx = ctx.template device_context(); - auto stream = dev_ctx.stream(); - - if (ctx.Attr("soft_label")) { - auto* label_data = label->data(); - SoftCrossEntropyGradientKernel<<>>( - dx_data, dy_data, x_data, label_data, batch_size, class_num); - } else { - math::SetConstant functor; - functor(dev_ctx, dx, 0); - auto* label_data = label->data(); - grid = (batch_size + block - 1) / block; - CrossEntropyGradientKernel<<>>( - dx_data, dy_data, x_data, label_data, batch_size, class_num); - } - } -}; - -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel, - ops::CrossEntropyOpCUDAKernel); +using CUDACtx = paddle::platform::CUDADeviceContext; +REGISTER_OP_CUDA_KERNEL(cross_entropy, + ops::CrossEntropyOpKernel, + ops::CrossEntropyOpKernel); REGISTER_OP_CUDA_KERNEL(cross_entropy_grad, - ops::CrossEntropyGradientOpCUDAKernel, - ops::CrossEntropyGradientOpCUDAKernel); + ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel); diff --git a/paddle/fluid/operators/cross_entropy_op.h b/paddle/fluid/operators/cross_entropy_op.h index 6da3a24dc89a85fe432b6350d3af7b0e84337c9d..19a2aec92b267ece94685ce34604b7d1cfa5d209 100644 --- a/paddle/fluid/operators/cross_entropy_op.h +++ b/paddle/fluid/operators/cross_entropy_op.h @@ -17,69 +17,106 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/math/cross_entropy.h" #include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; -template -using EigenMatrix = framework::EigenMatrix; -template +template class CrossEntropyOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), - "This kernel only runs on CPU."); - const Tensor* x = ctx.Input("X"); - const Tensor* labels = ctx.Input("Label"); - Tensor* y = ctx.Output("Y"); + auto* x = ctx.Input("X"); + auto* labels = ctx.Input("Label"); + auto* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); - math::CrossEntropyFunctor()( - ctx.template device_context(), y, x, labels, + math::CrossEntropyFunctor()( + ctx.template device_context(), y, x, labels, ctx.Attr("soft_label")); } }; template +class XeSoftlabelGradFunctor { + public: + XeSoftlabelGradFunctor(T* dx, + const T* dy, // NOLINT + const T* x, // NOLINT + const T* label, // NOLINT + size_t num_classes) + : dx_(dx), dy_(dy), x_(x), label_(label), num_classes_(num_classes) {} + + HOSTDEVICE void operator()(size_t i) { + auto row_ids = i / num_classes_; + dx_[i] = -label_[i] * dy_[row_ids] / x_[i]; + } + + private: + T* dx_; + const T* dy_; + const T* x_; + const T* label_; + size_t num_classes_; +}; + +template +class XeGradFunctor { + public: + XeGradFunctor(T* dx, + const T* dy, // NOLINT + const T* x, // NOLINT + const int64_t* label, // NOLINT + size_t num_classes) + : dx_(dx), dy_(dy), x_(x), label_(label), num_classes_(num_classes) {} + + HOSTDEVICE void operator()(size_t sample_id) { + auto x_is_true_offset = sample_id * num_classes_ + label_[sample_id]; + for (size_t x_offset = sample_id * num_classes_; + x_offset < (sample_id + 1) * num_classes_; ++x_offset) { + dx_[x_offset] = x_offset != x_is_true_offset + ? static_cast(0) + : -dy_[sample_id] / x_[x_offset]; + } + } + + private: + T* dx_; + const T* dy_; + const T* x_; + const int64_t* label_; + size_t num_classes_; +}; + +template class CrossEntropyGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), - "This kernel only runs on CPU."); - const Tensor* x = ctx.Input("X"); - const Tensor* dy = ctx.Input(framework::GradVarName("Y")); - const Tensor* label = ctx.Input("Label"); - Tensor* dx = ctx.Output(framework::GradVarName("X")); - T* dx_data = dx->mutable_data(ctx.GetPlace()); + auto* x = ctx.Input("X"); + auto* dy = ctx.Input(framework::GradVarName("Y")); + auto* label = ctx.Input("Label"); + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dx_data = dx->mutable_data(ctx.GetPlace()); int64_t class_num = x->dims()[1]; if (ctx.Attr("soft_label")) { - auto x_mat = EigenMatrix::From(*x); - auto dy_mat = EigenMatrix::From(*dy); - auto lbl_mat = EigenMatrix::From(*label); - auto dx_mat = EigenMatrix::From(*dx); - - dx_mat.device(*ctx.template device_context() - .eigen_device()) = - -(lbl_mat * - dy_mat.broadcast(Eigen::DSizes(1, class_num)) / x_mat); + XeSoftlabelGradFunctor functor(dx_data, dy->data(), x->data(), + label->data(), + static_cast(class_num)); + platform::ForRange for_range( + ctx.template device_context(), + static_cast(dx->numel())); + for_range(functor); } else { - int64_t batch_size = x->dims()[0]; - const T* dy_data = dy->data(); - const T* x_data = x->data(); - const int64_t* label_data = label->data(); - - math::SetConstant functor; - functor(ctx.template device_context(), dx, 0); - - for (int64_t i = 0; i < batch_size; ++i) { - PADDLE_ASSERT(label_data[i] >= 0 || label_data[i] < class_num); - int64_t index = i * class_num + label_data[i]; - dx_data[index] = math::TolerableValue()(-dy_data[i] / x_data[index]); - } + XeGradFunctor functor(dx_data, dy->data(), x->data(), + label->data(), + static_cast(class_num)); + platform::ForRange for_range( + ctx.template device_context(), + static_cast(dy->numel())); + for_range(functor); } } }; diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc index 7ca694886e9209a49e214352f5babc473a1f275a..e6ee28ea8d920ef80fead258a9bd0d5f6762c879 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/detail/grpc_server.cc @@ -211,6 +211,11 @@ void AsyncGRPCServer::WaitClientGet(int count) { } } +void AsyncGRPCServer::WaitServerReady() { + std::unique_lock lock(this->mutex_ready_); + condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); +} + void AsyncGRPCServer::RunSyncUpdate() { ::grpc::ServerBuilder builder; builder.AddListeningPort(address_, ::grpc::InsecureServerCredentials(), @@ -244,6 +249,12 @@ void AsyncGRPCServer::RunSyncUpdate() { t_prefetch_.reset(new std::thread( std::bind(&AsyncGRPCServer::HandleRequest, this, cq_prefetch_.get(), "cq_prefetch", prefetch_register))); + + { + std::lock_guard lock(this->mutex_ready_); + ready_ = 1; + } + condition_ready_.notify_all(); // wait server server_->Wait(); t_send_->join(); diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h index 99b87b8c6cb3e597778b88c395e4abf400d82c39..7f9cae21ccca8dd51f9fbe98148d01a51ac6eb84 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/detail/grpc_server.h @@ -45,8 +45,9 @@ class RequestBase; class AsyncGRPCServer final { public: explicit AsyncGRPCServer(const std::string &address, bool sync_mode) - : address_(address), sync_mode_(sync_mode) {} + : address_(address), sync_mode_(sync_mode), ready_(0) {} + void WaitServerReady(); void RunSyncUpdate(); // functions to sync server barrier status. @@ -118,6 +119,10 @@ class AsyncGRPCServer final { framework::ProgramDesc *program_; framework::Executor *executor_; int selected_port_; + + std::mutex mutex_ready_; + std::condition_variable condition_ready_; + int ready_; }; }; // namespace detail diff --git a/paddle/fluid/operators/elementwise_op_function.h b/paddle/fluid/operators/elementwise_op_function.h index 953aedc85064ee803ab02afd427a5a6f22096f94..8b052611f80ddf874ca48c1c58e13346528a834e 100644 --- a/paddle/fluid/operators/elementwise_op_function.h +++ b/paddle/fluid/operators/elementwise_op_function.h @@ -22,6 +22,7 @@ limitations under the License. */ #ifdef __NVCC__ #include #include +#include "paddle/fluid/platform/cuda_device_function.h" #include "paddle/fluid/platform/cuda_primitives.h" constexpr int ELEMWISE_MAX_BLOCK_DIM = 1024; #endif @@ -336,43 +337,6 @@ static void ElemwiseGradBroadcast1CPU(const T* x, const T* y, const T* out, } #ifdef __NVCC__ - -template -__device__ T reduceSum(T val, int tid, int len) { - // NOTE(zcd): The warp size should be taken from the - // parameters of the GPU but not specified as 32 simply. - // To make the reduceSum more efficiently, - // I use Warp-Level Parallelism and assume the Warp size - // is 32 which may be different for different GPU, - // but most card's warp size is 32. - const int warpSize = 32; - __shared__ T shm[warpSize]; - unsigned mask = 0u; - CREATE_SHFL_MASK(mask, tid < len); - - for (int offset = warpSize / 2; offset > 0; offset /= 2) - val += platform::__shfl_down_sync(mask, val, offset); - - if (tid < warpSize) shm[tid] = 0; - - __syncthreads(); - - if (tid % warpSize == 0) { - shm[tid / warpSize] = val; - } - __syncthreads(); - - CREATE_SHFL_MASK(mask, tid < warpSize); - - if (tid < warpSize) { - val = shm[tid]; - for (int offset = warpSize / 2; offset > 0; offset /= 2) - val += platform::__shfl_down_sync(mask, val, offset); - } - - return val; -} - template static __global__ void ElemwiseGradBroadcast1CUDAKernel( const T* x, const T* y, const T* out, const T* dout, int h, int w, @@ -395,7 +359,7 @@ static __global__ void ElemwiseGradBroadcast1CUDAKernel( if (dy) { h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h; - val = reduceSum(val, tid, h); + val = paddle::platform::reduceSum(val, tid, h); if (threadIdx.x == 0) { dy[j] = val; } @@ -472,7 +436,7 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel( if (dy) { int h = pre * post; h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h; - val = reduceSum(val, tid, h); + val = paddle::platform::reduceSum(val, tid, h); if (threadIdx.x == 0) { dy[j] = val; } diff --git a/paddle/fluid/operators/gru_unit_op.h b/paddle/fluid/operators/gru_unit_op.h index 15d91ca30593871e2b343eb0e5c0b76aa8055968..49e657a272cdcf919f4ae88c159d7ef6bdae9c93 100644 --- a/paddle/fluid/operators/gru_unit_op.h +++ b/paddle/fluid/operators/gru_unit_op.h @@ -87,10 +87,10 @@ class GRUUnitKernel : public framework::OpKernel { const T* weight_data = weight->data(); T* gate_data = gate->data(); T* reset_hidden_prev_data = reset_hidden_prev->data(); - math::gemm( - context.template device_context(), false, false, - batch_size, 2 * frame_size, frame_size, 1, hidden_prev_data, frame_size, - weight_data, frame_size * 2, 1, gate_data, frame_size * 3); + auto blas = math::GetBlas(context); + blas.GEMM(false, false, batch_size, 2 * frame_size, frame_size, 1, + hidden_prev_data, frame_size, weight_data, frame_size * 2, 1, + gate_data, frame_size * 3); // calculate activited gate Eigen::array extents({{batch_size, frame_size}}); @@ -103,11 +103,10 @@ class GRUUnitKernel : public framework::OpKernel { g.slice(r_offsets, extents), g.slice(r_offsets, extents)); auto r = g.slice(r_offsets, extents); // reset gate r_h_p.device(place) = r * h_p; // reset previous hidden state - math::gemm( - context.template device_context(), false, false, - batch_size, frame_size, frame_size, 1, reset_hidden_prev_data, - frame_size, weight_data + frame_size * frame_size * 2, frame_size, 1, - gate_data + frame_size * 2, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size, frame_size, 1, + reset_hidden_prev_data, frame_size, + weight_data + frame_size * frame_size * 2, frame_size, 1, + gate_data + frame_size * 2, frame_size * 3); Eigen::array c_offsets({{0, frame_size * 2}}); ActCompute(context.Attr("activation"), place, @@ -188,11 +187,11 @@ class GRUUnitGradKernel : public framework::OpKernel { ActGradCompute(context.Attr("activation"), place, c, c, d_g.slice(c_offsets, extents), d_h * u); // backward for reset_hidden_prev - math::gemm( - context.template device_context(), false, true, - batch_size, frame_size, frame_size, 1, gate_grad_data + frame_size * 2, - frame_size * 3, weight_data + frame_size * frame_size * 2, frame_size, - 0, reset_hidden_prev_grad_data, frame_size); + auto blas = math::GetBlas(context); + blas.GEMM(false, true, batch_size, frame_size, frame_size, 1, + gate_grad_data + frame_size * 2, frame_size * 3, + weight_data + frame_size * frame_size * 2, frame_size, 0, + reset_hidden_prev_grad_data, frame_size); // backward for unactivated reset gate ActGradCompute(context.Attr("gate_activation"), place, r, r, d_g.slice(r_offsets, extents), d_r_h_p * h_p); @@ -200,18 +199,15 @@ class GRUUnitGradKernel : public framework::OpKernel { if (weight_grad) { T* weight_grad_data = weight_grad->mutable_data(context.GetPlace()); // backward for state_weight - math::gemm( - context.template device_context(), true, false, - frame_size, frame_size, batch_size, 1, reset_hidden_prev_data, - frame_size, gate_grad_data + frame_size * 2, frame_size * 3, 0, - weight_grad_data + frame_size * frame_size * 2, frame_size); + blas.GEMM(true, false, frame_size, frame_size, batch_size, 1, + reset_hidden_prev_data, frame_size, + gate_grad_data + frame_size * 2, frame_size * 3, 0, + weight_grad_data + frame_size * frame_size * 2, frame_size); // backward for update_gate_weight and reset_gate_weight - math::gemm( - context.template device_context(), true, false, - frame_size, frame_size * 2, batch_size, 1, hidden_prev_data, - frame_size, gate_grad_data, frame_size * 3, 0, weight_grad_data, - frame_size * 2); + blas.GEMM(true, false, frame_size, frame_size * 2, batch_size, 1, + hidden_prev_data, frame_size, gate_grad_data, frame_size * 3, 0, + weight_grad_data, frame_size * 2); } // backward for hidden_prev if (hidden_prev_grad) { @@ -219,11 +215,9 @@ class GRUUnitGradKernel : public framework::OpKernel { hidden_prev_grad->mutable_data(context.GetPlace()); auto d_h_p = EigenMatrix::From(*hidden_prev_grad); d_h_p.device(place) = d_r_h_p * r + d_h * (u.constant(T(1)) - u); - math::gemm( - context.template device_context(), false, true, - batch_size, frame_size, frame_size * 2, 1, gate_grad_data, - frame_size * 3, weight_data, frame_size * 2, 1, hidden_prev_grad_data, - frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size * 2, 1, + gate_grad_data, frame_size * 3, weight_data, frame_size * 2, 1, + hidden_prev_grad_data, frame_size); } // backward for input if (input_grad) { diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index f22f8b261030c0c536e2118351ec2aa1a9be6cd0..59b94511552874e1557d8a9d7a687af14f96c31c 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -66,12 +66,7 @@ static void ParallelExecuteBlocks( for (size_t i = 0; i < fs.size(); ++i) fs[i].wait(); } -static void SavePort(std::shared_ptr rpc_service) { - std::ofstream port_file; - port_file.open("/tmp/paddle.selected_port"); - port_file << rpc_service->GetSelectedPort(); - port_file.close(); -} +std::atomic_int ListenAndServOp::selected_port_{0}; ListenAndServOp::ListenAndServOp(const std::string &type, const framework::VariableNameMap &inputs, @@ -79,15 +74,27 @@ ListenAndServOp::ListenAndServOp(const std::string &type, const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} -int ListenAndServOp::GetSelectedPort() const { - return rpc_service_->GetSelectedPort(); -} - void ListenAndServOp::Stop() { rpc_service_->Push(LISTEN_TERMINATE_MESSAGE); server_thread_->join(); } +void ListenAndServOp::SavePort(const std::string &file_path) const { + // NOTE: default write file to /tmp/paddle.selected_port + selected_port_ = rpc_service_->GetSelectedPort(); + + std::ofstream port_file; + port_file.open(file_path); + port_file << selected_port_.load(); + port_file.close(); + VLOG(4) << "selected port written to " << file_path; +} + +void ListenAndServOp::WaitServerReady() { + while (selected_port_.load() == 0) { + } +} + void ListenAndServOp::RunSyncLoop(framework::Executor *executor, framework::ProgramDesc *program, framework::Scope *recv_scope, @@ -318,9 +325,13 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, // start the server listening after all member initialized. server_thread_.reset(new std::thread(RunServer, rpc_service_)); VLOG(3) << "wait server thread to become ready..."; - sleep(5); + rpc_service_->WaitServerReady(); + // Write to a file of server selected port for python use. - SavePort(rpc_service_); + std::string file_path = + string::Sprintf("/tmp/paddle.%d.selected_port", + static_cast(::getpid())); + SavePort(file_path); if (sync_mode) { RunSyncLoop(&executor, program, &recv_scope, prefetch_block); } else { diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index 5c8fc31c9774a0f2e8233824459b29b42469bd1a..f52a55c5c2d6902df6cb7e0a0d7242c6e86dc786 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include +#include #include #include @@ -39,8 +40,6 @@ class ListenAndServOp : public framework::OperatorBase { const framework::VariableNameMap& outputs, const framework::AttributeMap& attrs); - int GetSelectedPort() const; - void RunSyncLoop(framework::Executor* executor, framework::ProgramDesc* program, framework::Scope* recv_scope, @@ -49,14 +48,25 @@ class ListenAndServOp : public framework::OperatorBase { void RunAsyncLoop(framework::Executor* executor, framework::ProgramDesc* program) const; + void SavePort( + const std::string& file_path = "/tmp/paddle.selected_port") const; + + void WaitServerReady(); + + int GetSelectedPort() { return selected_port_; } + void Stop() override; void RunImpl(const framework::Scope& scope, const platform::Place& dev_place) const override; + static void ResetPort() { selected_port_ = 0; } + protected: mutable std::shared_ptr rpc_service_; mutable std::shared_ptr server_thread_; + // FIXME(wuyi): it's static so that the operator can be cloned. + static std::atomic_int selected_port_; }; } // namespace operators diff --git a/paddle/fluid/operators/math/blas_impl.cu.h b/paddle/fluid/operators/math/blas_impl.cu.h new file mode 100644 index 0000000000000000000000000000000000000000..89935829ab35a52dd85bcaf906b53e41d576cf3f --- /dev/null +++ b/paddle/fluid/operators/math/blas_impl.cu.h @@ -0,0 +1,151 @@ +// 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/math_function.h" +#include "paddle/fluid/platform/dynload/cublas.h" + +namespace paddle { +namespace operators { +namespace math { + +template +struct CUBlas; + +template <> +struct CUBlas { + template + static void GEMM(ARGS... args) { + PADDLE_ENFORCE(platform::dynload::cublasSgemm(args...)); + } +}; + +template <> +struct CUBlas { + template + static void GEMM(ARGS... args) { + PADDLE_ENFORCE(platform::dynload::cublasDgemm(args...)); + } +}; + +template <> +struct CUBlas { + using float16 = platform::float16; + + static void GEMM(cublasHandle_t handle, cublasOperation_t transa, + cublasOperation_t transb, int m, int n, int k, + const float16 *alpha, const float16 *A, int lda, + const float16 *B, int ldb, const float16 *beta, float16 *C, + int ldc) { + PADDLE_ENFORCE( + platform::dynload::cublasHgemm(handle, transa, transb, m, n, k, + reinterpret_cast(alpha), + reinterpret_cast(A), lda, + reinterpret_cast(B), ldb, + reinterpret_cast(beta), + reinterpret_cast<__half *>(C), ldc)); + } +}; + +template <> +template +void Blas::GEMM(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, + const int M, const int N, + const int K, const T alpha, + const T *A, const T *B, + const T beta, T *C) const { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + + CUBlas::GEMM(context_.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, + B, ldb, A, lda, &beta, C, N); +} + +template <> +template <> +inline void Blas::GEMM( + const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB, const int M, + const int N, const int K, const platform::float16 alpha, + const platform::float16 *A, const platform::float16 *B, + const platform::float16 beta, platform::float16 *C) const { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + + // TODO(kexinzhao): add processing code for compute capability < 53 case + PADDLE_ENFORCE_GE(context_.GetComputeCapability(), 53, + "cublas fp16 gemm requires GPU compute capability >= 53"); + +#if CUDA_VERSION >= 8000 + float h_alpha = static_cast(alpha); + float h_beta = static_cast(beta); + + cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT; +#if CUDA_VERSION >= 9000 + if (context_.GetComputeCapability() >= 70) { + PADDLE_ENFORCE(platform::dynload::cublasSetMathMode( + context_.cublas_handle(), CUBLAS_TENSOR_OP_MATH)); + algo = CUBLAS_GEMM_DFALT_TENSOR_OP; + } else { + PADDLE_ENFORCE(platform::dynload::cublasSetMathMode( + context_.cublas_handle(), CUBLAS_DEFAULT_MATH)); + } +#endif // CUDA_VERSION >= 9000 + + // cublasHgemm does true FP16 computation which is slow for non-Volta + // GPUs. So use cublasGemmEx instead which does pesudo FP16 computation: + // input/output in fp16, computation in fp32, which can also be accelerated + // using tensor cores in volta GPUs. + PADDLE_ENFORCE(platform::dynload::cublasGemmEx( + context_.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, B, + CUDA_R_16F, ldb, A, CUDA_R_16F, lda, &h_beta, C, CUDA_R_16F, N, + CUDA_R_32F, algo)); +#else + // CUDA 7.5 does not support cublasGemmEx, hence we fall back to use hgemm + CUBlas::GEMM(context_.cublas_handle(), cuTransB, cuTransA, + N, M, K, &h_alpha, h_B, ldb, h_A, lda, + &h_beta, h_C, N); +#endif // CUDA_VERSION >= 8000 +} + +template <> +template +void Blas::GEMM( + const bool transA, const bool transB, const int M, const int N, const int K, + const T alpha, const T *A, const int lda, const T *B, const int ldb, + const T beta, T *C, const int ldc) const { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; + CUBlas::GEMM(context_.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, + B, ldb, A, lda, &beta, C, ldc); +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/blas_impl.h b/paddle/fluid/operators/math/blas_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..f6d6669765865386116532c5c65c689aa170eaa6 --- /dev/null +++ b/paddle/fluid/operators/math/blas_impl.h @@ -0,0 +1,74 @@ +// 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/math_function.h" + +namespace paddle { +namespace operators { +namespace math { + +template +struct CBlas; + +template <> +struct CBlas { + template + static void GEMM(ARGS... args) { + cblas_sgemm(args...); + } +}; + +template <> +struct CBlas { + template + static void GEMM(ARGS... args) { + cblas_dgemm(args...); + } +}; + +template <> +struct CBlas { + static void GEMM(...) { PADDLE_THROW("float16 GEMM not supported on CPU"); } +}; + +template <> +template +void Blas::GEMM(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, + const int M, const int N, + const int K, const T alpha, + const T *A, const T *B, + const T beta, T *C) const { + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + CBlas::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, + beta, C, ldc); +} + +template <> +template +void Blas::GEMM( + const bool transA, const bool transB, const int M, const int N, const int K, + const T alpha, const T *A, const int lda, const T *B, const int ldb, + const T beta, T *C, const int ldc) const { + CBlas::GEMM(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, + transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, + lda, B, ldb, beta, C, ldc); +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/cross_entropy.cu b/paddle/fluid/operators/math/cross_entropy.cu index 6d2ba2bd0d653ecf83f9e2abc1413ae551dc8bb7..0de58d5fddd84d33f708c4c73e5a19dc2fe8a86b 100644 --- a/paddle/fluid/operators/math/cross_entropy.cu +++ b/paddle/fluid/operators/math/cross_entropy.cu @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/cross_entropy.h" +#include "paddle/fluid/platform/cuda_device_function.h" #include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { @@ -30,66 +31,22 @@ __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label, } } -template -__device__ __forceinline__ T sum_single_warp(T val) { - val += platform::__shfl_down_sync(0, val, 16); - val += platform::__shfl_down_sync(0, val, 8); - val += platform::__shfl_down_sync(0, val, 4); - val += platform::__shfl_down_sync(0, val, 2); - val += platform::__shfl_down_sync(0, val, 1); - return val; -} - -// CUDA do not support dynamic arrary in template -// https://stackoverflow.com/questions/20497209 -template -struct SharedMemory { - // Ensure that we won't compile any un-specialized types - __device__ T* GetPointer() { return NULL; } -}; - -template <> -struct SharedMemory { - __device__ float* GetPointer() { - extern __shared__ float s_float[]; - return s_float; - } -}; - -template <> -struct SharedMemory { - __device__ double* GetPointer() { - extern __shared__ double s_double[]; - return s_double; - } -}; - template __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, const int class_num) { int tid = threadIdx.x; - SharedMemory d_sum_shared; - T* d_sum = d_sum_shared.GetPointer(); - d_sum[tid] = 0; + T val = 0; - int cur_idx = tid; - int next_idx = blockIdx.x * class_num + tid; - while (cur_idx < class_num) { - d_sum[tid] += - math::TolerableValue()(std::log(X[next_idx])) * label[next_idx]; - next_idx += blockDim.x; - cur_idx += blockDim.x; + 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]; } - __syncthreads(); - for (unsigned int stride = blockDim.x >> 1; stride >= 32; stride >>= 1) { - if (tid < stride) d_sum[tid] += d_sum[tid + stride]; - __syncthreads(); + val = paddle::platform::reduceSum(val, tid, blockDim.x); + if (threadIdx.x == 0) { + Y[blockIdx.x] = -val; } - - T val = d_sum[tid]; - val = sum_single_warp(val); - if (tid == 0) Y[blockIdx.x] = -val; } } // namespace @@ -113,9 +70,7 @@ class CrossEntropyFunctor { ? 512 : pow(2, static_cast(std::log2(class_num))); - SoftCrossEntropyKernel<<< - batch_size, block, block * sizeof(T), - reinterpret_cast(ctx).stream()>>>( + SoftCrossEntropyKernel<<>>( loss_data, prob_data, label_data, class_num); } else { const int64_t* label_data = labels->data(); diff --git a/paddle/fluid/operators/math/detail/gru_cpu_kernel.h b/paddle/fluid/operators/math/detail/gru_cpu_kernel.h index 26e6adafdfcd28dcc51109b9dec812f1c82c8a0e..b6f4ab93777f2b5eb13a4a3172028d87c4546017 100644 --- a/paddle/fluid/operators/math/detail/gru_cpu_kernel.h +++ b/paddle/fluid/operators/math/detail/gru_cpu_kernel.h @@ -43,8 +43,8 @@ void hl_naive_gru_forward_reset_output(OpResetOutput op_reset_output, r_prev_out = prev_output_value[i]; } - op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, - r_value_reset_output, active_gate); + op_reset_output(&r_value_update_gate, &r_value_reset_gate, &r_prev_out, + &r_value_reset_output, active_gate); update_gate[i] = r_value_update_gate; reset_gate[i] = r_value_reset_gate; @@ -71,8 +71,8 @@ void hl_naive_gru_forward_final_output(OpFinalOutput op_final_output, r_prev_out = prev_output_value[i]; } - op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, - r_output, active_node); + op_final_output(&r_value_update_gate, &r_value_frame_state, &r_prev_out, + &r_output, active_node); frame_state[i] = r_value_frame_state; output_value[i] = r_output; @@ -99,8 +99,8 @@ void hl_avx_gru_forward_reset_output(OpResetOutput op_reset_output, r_prev_out = (reinterpret_cast<__m256 *>(prev_output_value))[i]; } - op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, - r_value_reset_output, active_gate); + op_reset_output(&r_value_update_gate, &r_value_reset_gate, &r_prev_out, + &r_value_reset_output, active_gate); update_gate[i] = r_value_update_gate; reset_gate[i] = r_value_reset_gate; @@ -129,8 +129,8 @@ void hl_avx_gru_forward_final_output(OpFinalOutput op_final_output, r_prev_out = (reinterpret_cast<__m256 *>(prev_output_value))[i]; } - op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, - r_output, active_node); + op_final_output(&r_value_update_gate, &r_value_frame_state, &r_prev_out, + &r_output, active_node); frame_state[i] = r_value_frame_state; (reinterpret_cast<__m256 *>(output_value))[i] = r_output; @@ -213,9 +213,9 @@ void hl_naive_gru_backward_state_grad(OpStateGrad op_state_grad, T *gate_value, r_prev_out_grad = prev_out_grad[i]; } - op_state_grad(r_update_gate_value, r_update_gate_grad, r_frame_state_value, - r_frame_state_grad, r_prev_out_value, r_prev_out_grad, - r_out_grad, active_node); + op_state_grad(&r_update_gate_value, &r_update_gate_grad, + &r_frame_state_value, &r_frame_state_grad, &r_prev_out_value, + &r_prev_out_grad, &r_out_grad, active_node); update_gate_grad[i] = r_update_gate_grad; frame_state_grad[i] = r_frame_state_grad; @@ -258,9 +258,9 @@ void hl_naive_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value, r_prev_out_grad = prev_out_grad[i]; } - op_reset_grad(r_update_gate_value, r_update_gate_grad, r_reset_gate_value, - r_reset_gate_grad, r_prev_out_value, r_prev_out_grad, - r_reset_output_grad, active_gate); + op_reset_grad(&r_update_gate_value, &r_update_gate_grad, + &r_reset_gate_value, &r_reset_gate_grad, &r_prev_out_value, + &r_prev_out_grad, &r_reset_output_grad, active_gate); update_gate_grad[i] = r_update_gate_grad; reset_gate_grad[i] = r_reset_gate_grad; @@ -302,9 +302,9 @@ void hl_avx_gru_backward_state_grad(OpStateGrad op_state_grad, T *gate_value, r_prev_out_grad = (reinterpret_cast<__m256 *>(prev_out_grad))[i]; } - op_state_grad(r_update_gate_value, r_update_gate_grad, r_frame_state_value, - r_frame_state_grad, r_prev_out_value, r_prev_out_grad, - r_out_grad, active_node); + op_state_grad(&r_update_gate_value, &r_update_gate_grad, + &r_frame_state_value, &r_frame_state_grad, &r_prev_out_value, + &r_prev_out_grad, &r_out_grad, active_node); update_gate_grad[i] = r_update_gate_grad; frame_state_grad[i] = r_frame_state_grad; @@ -350,9 +350,9 @@ void hl_avx_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value, r_prev_out_grad = (reinterpret_cast<__m256 *>(prev_out_grad))[i]; } - op_reset_grad(r_update_gate_value, r_update_gate_grad, r_reset_gate_value, - r_reset_gate_grad, r_prev_out_value, r_prev_out_grad, - r_reset_output_grad, active_gate); + op_reset_grad(&r_update_gate_value, &r_update_gate_grad, + &r_reset_gate_value, &r_reset_gate_grad, &r_prev_out_value, + &r_prev_out_grad, &r_reset_output_grad, active_gate); update_gate_grad[i] = r_update_gate_grad; reset_gate_grad[i] = r_reset_gate_grad; diff --git a/paddle/fluid/operators/math/detail/gru_gpu_kernel.h b/paddle/fluid/operators/math/detail/gru_gpu_kernel.h index da25a7d2137cfe5160e28c4e590dd5c43cd99ccf..813d69f6aba722609a0523a5be71d32f91f76d59 100644 --- a/paddle/fluid/operators/math/detail/gru_gpu_kernel.h +++ b/paddle/fluid/operators/math/detail/gru_gpu_kernel.h @@ -55,8 +55,8 @@ __global__ void KeGruForwardResetOutput(OpResetOutput op_reset_output, r_prev_out = prev_output_value[frame_idx]; } - op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, - r_value_reset_output, active_gate); + op_reset_output(&r_value_update_gate, &r_value_reset_gate, &r_prev_out, + &r_value_reset_output, active_gate); gate_value[frame_idx + frame_size * 0] = r_value_update_gate; gate_value[frame_idx + frame_size * 1] = r_value_reset_gate; @@ -93,8 +93,8 @@ __global__ void KeGruForwardFinalOutput(OpFinalOutput op_final_output, r_prev_out = prev_output_value[frame_idx]; } - op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, - r_output, active_node); + op_final_output(&r_value_update_gate, &r_value_frame_state, &r_prev_out, + &r_output, active_node); gate_value[frame_idx + frame_size * 2] = r_value_frame_state; output_value[frame_idx] = r_output; @@ -137,9 +137,9 @@ __global__ void KeGruBackwardStateGrad(OpStateGrad op_state_grad, T *gate_value, r_prev_out_grad = prev_out_grad[frame_idx]; } - op_state_grad(r_update_gate_value, r_update_gate_grad, r_frame_state_value, - r_frame_state_grad, r_prev_out_value, r_prev_out_grad, - r_out_grad, active_node); + op_state_grad(&r_update_gate_value, &r_update_gate_grad, &r_frame_state_value, + &r_frame_state_grad, &r_prev_out_value, &r_prev_out_grad, + &r_out_grad, active_node); gate_grad[frame_idx + frame_size * 0] = r_update_gate_grad; gate_grad[frame_idx + frame_size * 2] = r_frame_state_grad; @@ -185,9 +185,9 @@ __global__ void KeGruBackwardResetGrad(OpResetGrad op_reset_grad, T *gate_value, r_reset_output_grad = reset_output_grad[frame_idx]; } - op_reset_grad(r_update_gate_value, r_update_gate_grad, r_reset_gate_value, - r_reset_gate_grad, r_prev_out_value, r_prev_out_grad, - r_reset_output_grad, active_gate); + op_reset_grad(&r_update_gate_value, &r_update_gate_grad, &r_reset_gate_value, + &r_reset_gate_grad, &r_prev_out_value, &r_prev_out_grad, + &r_reset_output_grad, active_gate); gate_grad[frame_idx + frame_size * 0] = r_update_gate_grad; gate_grad[frame_idx + frame_size * 1] = r_reset_gate_grad; diff --git a/paddle/fluid/operators/math/detail/gru_kernel.h b/paddle/fluid/operators/math/detail/gru_kernel.h index 991f2e758c2c3b4acc93c1eb84dbd196d623b5cc..f6d192358bd84eb56a2e01eb36f28d8832ef271f 100644 --- a/paddle/fluid/operators/math/detail/gru_kernel.h +++ b/paddle/fluid/operators/math/detail/gru_kernel.h @@ -12,11 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#pragma once +#include #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/platform/hostdevice.h" -#include - // TODO(guosheng): refine code style in gru_kernel namespace paddle { namespace operators { @@ -28,25 +28,25 @@ namespace forward { template class gru_resetOutput { public: - HOSTDEVICE void operator()(T &value_update_gate, T &value_reset_gate, - T &prev_out, T &value_reset_output, + HOSTDEVICE void operator()(T *value_update_gate, T *value_reset_gate, + T *prev_out, T *value_reset_output, ActivationType act_gate) { - value_update_gate = activation(value_update_gate, act_gate); - value_reset_gate = activation(value_reset_gate, act_gate); - value_reset_output = prev_out * value_reset_gate; + *value_update_gate = activation(*value_update_gate, act_gate); + *value_reset_gate = activation(*value_reset_gate, act_gate); + *value_reset_output = (*prev_out) * (*value_reset_gate); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &value_update_gate, - __m256 &value_reset_gate, __m256 &prev_out, - __m256 &value_reset_output, + HOSTDEVICE void operator()(__m256 *value_update_gate, + __m256 *value_reset_gate, __m256 *prev_out, + __m256 *value_reset_output, ActivationType act_gate) { - value_update_gate = activation(value_update_gate, act_gate); - value_reset_gate = activation(value_reset_gate, act_gate); - value_reset_output = _mm256_mul_ps(prev_out, value_reset_gate); + *value_update_gate = activation(*value_update_gate, act_gate); + *value_reset_gate = activation(*value_reset_gate, act_gate); + *value_reset_output = _mm256_mul_ps(*prev_out, *value_reset_gate); } #endif #endif @@ -55,25 +55,25 @@ class gru_resetOutput { template class gru_finalOutput { public: - HOSTDEVICE void operator()(T &value_update_gate, T &value_frame_state, - T &prev_out, T &value_output, + HOSTDEVICE void operator()(T *value_update_gate, T *value_frame_state, + T *prev_out, T *value_output, ActivationType act_input) { - value_frame_state = activation(value_frame_state, act_input); - value_output = prev_out - (value_update_gate * prev_out) + - (value_update_gate * value_frame_state); + *value_frame_state = activation(*value_frame_state, act_input); + *value_output = *prev_out - ((*value_update_gate) * (*prev_out)) + + ((*value_update_gate) * (*value_frame_state)); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &value_update_gate, - __m256 &value_frame_state, __m256 &prev_out, - __m256 &value_output, ActivationType act_input) { - value_frame_state = activation(value_frame_state, act_input); - value_output = _mm256_add_ps( - _mm256_sub_ps(prev_out, _mm256_mul_ps(value_update_gate, prev_out)), - _mm256_mul_ps(value_update_gate, value_frame_state)); + HOSTDEVICE void operator()(__m256 *value_update_gate, + __m256 *value_frame_state, __m256 *prev_out, + __m256 *value_output, ActivationType act_input) { + *value_frame_state = activation(*value_frame_state, act_input); + *value_output = _mm256_add_ps( + _mm256_sub_ps(*prev_out, _mm256_mul_ps(*value_update_gate, *prev_out)), + _mm256_mul_ps(*value_update_gate, *value_frame_state)); } #endif #endif @@ -85,37 +85,38 @@ namespace backward { template class gru_stateGrad { public: - HOSTDEVICE void operator()(T &value_update_gate, T &grad_update_gate, - T &value_frame_state, T &grad_frame_state, - T &value_prev_out, T &grad_prev_out, - T &grad_output, ActivationType act_input) { - grad_update_gate = (grad_output * value_frame_state); - grad_update_gate -= (grad_output * value_prev_out); - grad_prev_out -= (grad_output * value_update_gate); - grad_prev_out += grad_output; - grad_frame_state = activation(grad_output * value_update_gate, - value_frame_state, act_input); + HOSTDEVICE void operator()(T *value_update_gate, T *grad_update_gate, + T *value_frame_state, T *grad_frame_state, + T *value_prev_out, T *grad_prev_out, + T *grad_output, ActivationType act_input) { + *grad_update_gate = (*grad_output * (*value_frame_state)); + *grad_update_gate -= (*grad_output * (*value_prev_out)); + *grad_prev_out -= (*grad_output * (*value_update_gate)); + *grad_prev_out += *grad_output; + *grad_frame_state = activation(*grad_output * (*value_update_gate), + *value_frame_state, act_input); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &value_update_gate, - __m256 &grad_update_gate, - __m256 &value_frame_state, - __m256 &grad_frame_state, __m256 &value_prev_out, - __m256 &grad_prev_out, __m256 &grad_output, + HOSTDEVICE void operator()(__m256 *value_update_gate, + __m256 *grad_update_gate, + __m256 *value_frame_state, + __m256 *grad_frame_state, __m256 *value_prev_out, + __m256 *grad_prev_out, __m256 *grad_output, ActivationType act_input) { - grad_update_gate = _mm256_mul_ps(grad_output, value_frame_state); - grad_update_gate = _mm256_sub_ps( - grad_update_gate, _mm256_mul_ps(grad_output, value_prev_out)); - grad_prev_out = _mm256_add_ps( - _mm256_sub_ps(grad_prev_out, - _mm256_mul_ps(grad_output, value_update_gate)), - grad_output); - grad_frame_state = activation(_mm256_mul_ps(grad_output, value_update_gate), - value_frame_state, act_input); + *grad_update_gate = _mm256_mul_ps(*grad_output, *value_frame_state); + *grad_update_gate = _mm256_sub_ps( + *grad_update_gate, _mm256_mul_ps(*grad_output, *value_prev_out)); + *grad_prev_out = _mm256_add_ps( + _mm256_sub_ps(*grad_prev_out, + _mm256_mul_ps(*grad_output, *value_update_gate)), + *grad_output); + *grad_frame_state = + activation(_mm256_mul_ps(*grad_output, *value_update_gate), + *value_frame_state, act_input); } #endif #endif @@ -124,32 +125,34 @@ class gru_stateGrad { template class gru_resetGrad { public: - HOSTDEVICE void operator()(T &value_update_gate, T &grad_update_gate, - T &value_reset_gate, T &grad_reset_gate, - T &value_prev_out, T &grad_prev_out, - T &grad_reset_output, ActivationType act_gate) { - grad_reset_gate = (grad_reset_output * value_prev_out); - grad_prev_out += (grad_reset_output * value_reset_gate); - grad_update_gate = - activation(grad_update_gate, value_update_gate, act_gate); - grad_reset_gate = activation(grad_reset_gate, value_reset_gate, act_gate); + HOSTDEVICE void operator()(T *value_update_gate, T *grad_update_gate, + T *value_reset_gate, T *grad_reset_gate, + T *value_prev_out, T *grad_prev_out, + T *grad_reset_output, ActivationType act_gate) { + *grad_reset_gate = (*grad_reset_output * (*value_prev_out)); + *grad_prev_out += (*grad_reset_output * (*value_reset_gate)); + *grad_update_gate = + activation(*grad_update_gate, *value_update_gate, act_gate); + *grad_reset_gate = + activation(*grad_reset_gate, *value_reset_gate, act_gate); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &value_update_gate, - __m256 &grad_update_gate, __m256 &value_reset_gate, - __m256 &grad_reset_gate, __m256 &value_prev_out, - __m256 &grad_prev_out, __m256 &grad_reset_output, + HOSTDEVICE void operator()(__m256 *value_update_gate, + __m256 *grad_update_gate, __m256 *value_reset_gate, + __m256 *grad_reset_gate, __m256 *value_prev_out, + __m256 *grad_prev_out, __m256 *grad_reset_output, ActivationType act_gate) { - grad_reset_gate = _mm256_mul_ps(grad_reset_output, value_prev_out); - grad_prev_out = _mm256_add_ps( - grad_prev_out, _mm256_mul_ps(grad_reset_output, value_reset_gate)); - grad_update_gate = - activation(grad_update_gate, value_update_gate, act_gate); - grad_reset_gate = activation(grad_reset_gate, value_reset_gate, act_gate); + *grad_reset_gate = _mm256_mul_ps(*grad_reset_output, *value_prev_out); + *grad_prev_out = _mm256_add_ps( + *grad_prev_out, _mm256_mul_ps(*grad_reset_output, *value_reset_gate)); + *grad_update_gate = + activation(*grad_update_gate, *value_update_gate, act_gate); + *grad_reset_gate = + activation(*grad_reset_gate, *value_reset_gate, act_gate); } #endif #endif diff --git a/paddle/fluid/operators/math/gru_compute.cc b/paddle/fluid/operators/math/gru_compute.cc index 3f044b775138c495052bec3d19121bf26c37cb37..d786250271231179b46ae704c9bd013efe26d910 100644 --- a/paddle/fluid/operators/math/gru_compute.cc +++ b/paddle/fluid/operators/math/gru_compute.cc @@ -25,21 +25,21 @@ struct GRUUnitFunctor { const detail::ActivationType active_node, const detail::ActivationType active_gate) { #ifndef __NVCC__ + auto blas = math::GetBlas(context); if (value.prev_out_value) { - math::gemm( - context, false, false, batch_size, frame_size * 2, frame_size, 1, - value.prev_out_value, frame_size, value.gate_weight, frame_size * 2, - 1, value.gate_value, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size * 2, frame_size, 1, + value.prev_out_value, frame_size, value.gate_weight, + frame_size * 2, 1, value.gate_value, frame_size * 3); } detail::forward_reset_output(detail::forward::gru_resetOutput(), value, frame_size, batch_size, active_gate); if (value.prev_out_value) { - math::gemm( - context, false, false, batch_size, frame_size, frame_size, 1, - value.reset_output_value, frame_size, value.state_weight, frame_size, - 1, value.gate_value + frame_size * 2, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size, frame_size, 1, + value.reset_output_value, frame_size, value.state_weight, + frame_size, 1, value.gate_value + frame_size * 2, + frame_size * 3); } detail::forward_final_output(detail::forward::gru_finalOutput(), value, @@ -58,36 +58,32 @@ struct GRUUnitGradFunctor { #ifndef __NVCC__ detail::backward_state_grad(detail::backward::gru_stateGrad(), value, grad, frame_size, batch_size, active_node); - + auto blas = math::GetBlas(context); if (value.prev_out_value && grad.prev_out_grad) { - math::gemm( - context, false, true, batch_size, frame_size, frame_size, 1, - grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight, - frame_size, 0, grad.reset_output_grad, frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size, 1, + grad.gate_grad + frame_size * 2, frame_size * 3, + value.state_weight, frame_size, 0, grad.reset_output_grad, + frame_size); if (grad.state_weight_grad) { - math::gemm( - context, true, false, frame_size, frame_size, batch_size, 1, - value.reset_output_value, frame_size, - grad.gate_grad + frame_size * 2, frame_size * 3, 1, - grad.state_weight_grad, frame_size); + blas.GEMM(true, false, frame_size, frame_size, batch_size, 1, + value.reset_output_value, frame_size, + grad.gate_grad + frame_size * 2, frame_size * 3, 1, + grad.state_weight_grad, frame_size); } } detail::backward_reset_grad(detail::backward::gru_resetGrad(), value, grad, frame_size, batch_size, active_gate); - if (grad.prev_out_grad && value.prev_out_value) { - math::gemm( - context, false, true, batch_size, frame_size, frame_size * 2, 1, - grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1, - grad.prev_out_grad, frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size * 2, 1, + grad.gate_grad, frame_size * 3, value.gate_weight, + frame_size * 2, 1, grad.prev_out_grad, frame_size); if (grad.gate_weight_grad) { - math::gemm( - context, true, false, frame_size, frame_size * 2, batch_size, 1, - value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1, - grad.gate_weight_grad, frame_size * 2); + blas.GEMM(true, false, frame_size, frame_size * 2, batch_size, 1, + value.prev_out_value, frame_size, grad.gate_grad, + frame_size * 3, 1, grad.gate_weight_grad, frame_size * 2); } } #endif diff --git a/paddle/fluid/operators/math/gru_compute.cu b/paddle/fluid/operators/math/gru_compute.cu index 27caf3383dd6cd94779391b722bba1d6b74772c0..f26bec41095789c197841f4d8362a229b07a2af0 100644 --- a/paddle/fluid/operators/math/gru_compute.cu +++ b/paddle/fluid/operators/math/gru_compute.cu @@ -9,6 +9,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include #include "paddle/fluid/operators/math/detail/gru_gpu_kernel.h" #include "paddle/fluid/operators/math/detail/gru_kernel.h" #include "paddle/fluid/operators/math/gru_compute.h" @@ -36,12 +37,11 @@ struct GRUUnitFunctor { threads = dim3(32, 32); grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32); } - + auto blas = math::GetBlas(context); if (value.prev_out_value) { - math::gemm( - context, false, false, batch_size, frame_size * 2, frame_size, 1, - value.prev_out_value, frame_size, value.gate_weight, frame_size * 2, - 1, value.gate_value, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size * 2, frame_size, 1, + value.prev_out_value, frame_size, value.gate_weight, + frame_size * 2, 1, value.gate_value, frame_size * 3); } if (batch_size == 1) { @@ -61,10 +61,10 @@ struct GRUUnitFunctor { } if (value.prev_out_value) { - math::gemm( - context, false, false, batch_size, frame_size, frame_size, 1, - value.reset_output_value, frame_size, value.state_weight, frame_size, - 1, value.gate_value + frame_size * 2, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size, frame_size, 1, + value.reset_output_value, frame_size, value.state_weight, + frame_size, 1, value.gate_value + frame_size * 2, + frame_size * 3); } if (batch_size == 1) { @@ -121,18 +121,19 @@ struct GRUUnitGradFunctor { grad.output_grad, frame_size, batch_size, active_node); } + auto blas = math::GetBlas(context); + if (value.prev_out_value && grad.prev_out_grad) { - math::gemm( - context, false, true, batch_size, frame_size, frame_size, 1, - grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight, - frame_size, 0, grad.reset_output_grad, frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size, 1, + grad.gate_grad + frame_size * 2, frame_size * 3, + value.state_weight, frame_size, 0, grad.reset_output_grad, + frame_size); if (grad.state_weight_grad) { - math::gemm( - context, true, false, frame_size, frame_size, batch_size, 1, - value.reset_output_value, frame_size, - grad.gate_grad + frame_size * 2, frame_size * 3, 1, - grad.state_weight_grad, frame_size); + blas.GEMM(true, false, frame_size, frame_size, batch_size, 1, + value.reset_output_value, frame_size, + grad.gate_grad + frame_size * 2, frame_size * 3, 1, + grad.state_weight_grad, frame_size); } } @@ -153,16 +154,14 @@ struct GRUUnitGradFunctor { } if (grad.prev_out_grad && value.prev_out_value) { - math::gemm( - context, false, true, batch_size, frame_size, frame_size * 2, 1, - grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1, - grad.prev_out_grad, frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size * 2, 1, + grad.gate_grad, frame_size * 3, value.gate_weight, + frame_size * 2, 1, grad.prev_out_grad, frame_size); if (grad.gate_weight_grad) { - math::gemm( - context, true, false, frame_size, frame_size * 2, batch_size, 1, - value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1, - grad.gate_weight_grad, frame_size * 2); + blas.GEMM(true, false, frame_size, frame_size * 2, batch_size, 1, + value.prev_out_value, frame_size, grad.gate_grad, + frame_size * 3, 1, grad.gate_weight_grad, frame_size * 2); } } } diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc index b5ae41c8f9d7aeb8e410b795fb9fbbd57ec69d4b..b63676f961bcd488797aca887c281a7d351cfca0 100644 --- a/paddle/fluid/operators/math/math_function.cc +++ b/paddle/fluid/operators/math/math_function.cc @@ -24,72 +24,6 @@ namespace math { using float16 = paddle::platform::float16; -template <> -void gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const float16* B, const float16 beta, - float16* C) { - PADDLE_THROW("float16 GEMM not supported on CPU"); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float alpha, const float* A, const float* B, const float beta, - float* C) { - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - cblas_sgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, - beta, C, ldc); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const double alpha, const double* A, const double* B, const double beta, - double* C) { - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - cblas_dgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, - beta, C, ldc); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const int lda, const float16* B, - const int ldb, const float16 beta, float16* C, const int ldc) { - PADDLE_THROW("float16 GEMM not supported on CPU"); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, const float alpha, - const float* A, const int lda, const float* B, const int ldb, - const float beta, float* C, const int ldc) { - cblas_sgemm(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, - transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, - lda, B, ldb, beta, C, ldc); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, - const double alpha, const double* A, const int lda, const double* B, - const int ldb, const double beta, double* C, const int ldc) { - cblas_dgemm(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, - transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, - lda, B, ldb, beta, C, ldc); -} - template <> void matmul( const platform::CPUDeviceContext& context, @@ -123,8 +57,8 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), + Blas(context).GEMM( + transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } @@ -152,8 +86,8 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), + Blas(context).GEMM( + transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } @@ -230,8 +164,8 @@ void batched_gemm( const float* Ak = &A[k * strideA]; const float* Bk = &B[k * strideB]; float* Ck = &C[k * M * N]; - gemm(context, transA, transB, M, N, K, - alpha, Ak, Bk, beta, Ck); + Blas(context).GEMM(transA, transB, M, N, K, + alpha, Ak, Bk, beta, Ck); } } @@ -246,8 +180,8 @@ void batched_gemm( const double* Ak = &A[k * strideA]; const double* Bk = &B[k * strideB]; double* Ck = &C[k * M * N]; - gemm(context, transA, transB, M, N, K, - alpha, Ak, Bk, beta, Ck); + Blas(context).GEMM(transA, transB, M, N, K, + alpha, Ak, Bk, beta, Ck); } } #endif diff --git a/paddle/fluid/operators/math/math_function.cu b/paddle/fluid/operators/math/math_function.cu index 2aa819625e0f5213a6001908e715bcc73d4747c3..7bf816ac190a9b848b12ea07e655449802a26bc3 100644 --- a/paddle/fluid/operators/math/math_function.cu +++ b/paddle/fluid/operators/math/math_function.cu @@ -25,157 +25,6 @@ namespace math { using float16 = paddle::platform::float16; -template <> -void gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const float16* B, const float16 beta, - float16* C) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - - // TODO(kexinzhao): add processing code for compute capability < 53 case - PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53, - "cublas fp16 gemm requires GPU compute capability >= 53"); - -#if CUDA_VERSION >= 8000 - float h_alpha = static_cast(alpha); - float h_beta = static_cast(beta); - - cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT; -#if CUDA_VERSION >= 9000 - if (context.GetComputeCapability() >= 70) { - PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(), - CUBLAS_TENSOR_OP_MATH)); - algo = CUBLAS_GEMM_DFALT_TENSOR_OP; - } else { - PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(), - CUBLAS_DEFAULT_MATH)); - } -#endif // CUDA_VERSION >= 9000 - - // cublasHgemm does true FP16 computation which is slow for non-Volta - // GPUs. So use cublasGemmEx instead which does pesudo FP16 computation: - // input/output in fp16, computation in fp32, which can also be accelerated - // using tensor cores in volta GPUs. - PADDLE_ENFORCE(platform::dynload::cublasGemmEx( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, B, - CUDA_R_16F, ldb, A, CUDA_R_16F, lda, &h_beta, C, CUDA_R_16F, N, - CUDA_R_32F, algo)); -#else - // CUDA 7.5 does not support cublasGemmEx, hence we fall back to use hgemm - const half h_alpha = static_cast(alpha); - const half h_beta = static_cast(beta); - const half* h_A = reinterpret_cast(A); - const half* h_B = reinterpret_cast(B); - half* h_C = reinterpret_cast(C); - - PADDLE_ENFORCE(platform::dynload::cublasHgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb, - h_A, lda, &h_beta, h_C, N)); -#endif // CUDA_VERSION >= 8000 -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float alpha, const float* A, const float* B, const float beta, - float* C) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - - PADDLE_ENFORCE(platform::dynload::cublasSgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, - lda, &beta, C, N)); -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const double alpha, const double* A, const double* B, const double beta, - double* C) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - PADDLE_ENFORCE(platform::dynload::cublasDgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, - lda, &beta, C, N)); -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const int lda, const float16* B, - const int ldb, const float16 beta, float16* C, const int ldc) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; - - const half h_alpha = static_cast(alpha); - const half h_beta = static_cast(beta); - const half* h_A = reinterpret_cast(A); - const half* h_B = reinterpret_cast(B); - half* h_C = reinterpret_cast(C); - - // TODO(kexinzhao): add processing code for compute capability < 53 case - PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53, - "cublas Hgemm requires GPU compute capability >= 53"); - PADDLE_ENFORCE(platform::dynload::cublasHgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb, - h_A, lda, &h_beta, h_C, ldc)); -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, const float alpha, - const float* A, const int lda, const float* B, const int ldb, - const float beta, float* C, const int ldc) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; - PADDLE_ENFORCE(platform::dynload::cublasSgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, - lda, &beta, C, ldc)); -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, - const double alpha, const double* A, const int lda, const double* B, - const int ldb, const double beta, double* C, const int ldc) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; - PADDLE_ENFORCE(platform::dynload::cublasDgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, - lda, &beta, C, ldc)); -} - template <> void matmul( const platform::CUDADeviceContext& context, @@ -200,8 +49,8 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), + Blas(context).GEMM( + transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } @@ -229,8 +78,8 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), + Blas(context).GEMM( + transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } @@ -258,8 +107,8 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), + Blas(context).GEMM( + transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } diff --git a/paddle/fluid/operators/math/math_function.h b/paddle/fluid/operators/math/math_function.h index cdd02974722045457aacdfa517c147751185f332..9950c09ea618d6c4250d66beb480d6f707813b54 100644 --- a/paddle/fluid/operators/math/math_function.h +++ b/paddle/fluid/operators/math/math_function.h @@ -42,6 +42,7 @@ int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, #include #include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/platform/device_context.h" @@ -56,17 +57,48 @@ namespace math { // Then matrixA: M * K, matrixB: K * N, matrixC : M * N // For more detailed info, please refer to // http://www.netlib.org/lapack/explore-html/d4/de2/sgemm_8f.html + +template +class Blas { + public: + explicit Blas(const DeviceContext& context) : context_(context) {} + + template + void GEMM(const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB, + const int M, const int N, const int K, const T alpha, const T* A, + const T* B, const T beta, T* C) const; + + template + void GEMM(const bool transA, const bool transB, const int M, const int N, + const int K, const T alpha, const T* A, const int lda, const T* B, + const int ldb, const T beta, T* C, const int ldc) const; + + private: + const DeviceContext& context_; +}; + template -void gemm(const DeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const T alpha, const T* A, const T* B, const T beta, T* C); +class BlasT : private Blas { + public: + using Blas::Blas; + + template + void GEMM(ARGS... args) const { + static_cast*>(this)->template GEMM(args...); + } +}; -// gemm wrapper with stride args for matrix uncontinuous in memory template -void gemm(const DeviceContext& context, const bool transA, const bool transB, - const int M, const int N, const int K, const T alpha, const T* A, - const int lda, const T* B, const int ldb, const T beta, T* C, - const int ldc); +inline BlasT GetBlas( + const framework::ExecutionContext& exe_ctx) { + return BlasT( + exe_ctx.template device_context()); +} + +template +inline BlasT GetBlas(const DeviceContext& dev_ctx) { + return BlasT(dev_ctx); +} // matrix multiply with continuous memory template @@ -137,3 +169,8 @@ struct RowwiseMean { } // namespace math } // namespace operators } // namespace paddle + +#include "paddle/fluid/operators/math/blas_impl.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/operators/math/blas_impl.cu.h" +#endif diff --git a/paddle/fluid/operators/math/math_function_test.cc b/paddle/fluid/operators/math/math_function_test.cc index 25a9d0111eee45b28adff012b705cbfa2407d2b6..6d11dc8c76799a72bd144e4103a6c65d5c94a649 100644 --- a/paddle/fluid/operators/math/math_function_test.cc +++ b/paddle/fluid/operators/math/math_function_test.cc @@ -14,6 +14,13 @@ #include "paddle/fluid/operators/math/math_function.h" #include "gtest/gtest.h" +template +inline paddle::operators::math::BlasT +GetBlas(const paddle::platform::CPUDeviceContext& context) { + return paddle::operators::math::GetBlas(context); +} + TEST(math_function, gemm_notrans_cblas) { paddle::framework::Tensor input1; paddle::framework::Tensor input2; @@ -34,9 +41,8 @@ TEST(math_function, gemm_notrans_cblas) { memcpy(input3_ptr, arr3, 8 * sizeof(float)); paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::gemm( - context, false, false, m, n, k, 1, input1_ptr, 3, input2_ptr + 1, 4, 1, - input3_ptr + 1, 4); + GetBlas(context).GEMM(false, false, m, n, k, 1, input1_ptr, 3, + input2_ptr + 1, 4, 1, input3_ptr + 1, 4); EXPECT_EQ(input3_ptr[0], 0); EXPECT_EQ(input3_ptr[1], 24); @@ -68,9 +74,8 @@ TEST(math_function, gemm_trans_clbas) { memcpy(input3_ptr, arr3, 8 * sizeof(float)); paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::gemm( - context, false, true, m, n, k, 1, input1_ptr, 3, input2_ptr + 3, 3, 1, - input3_ptr + 1, 4); + GetBlas(context).GEMM(false, true, m, n, k, 1, input1_ptr, 3, + input2_ptr + 3, 3, 1, input3_ptr + 1, 4); EXPECT_EQ(input3_ptr[0], 0); EXPECT_EQ(input3_ptr[1], 24); diff --git a/paddle/fluid/operators/math/math_function_test.cu b/paddle/fluid/operators/math/math_function_test.cu index b84bb9974930a1619edeb85de4281b19da22fbbd..3d03981b9f8a5ee5c302acce3d31157a16d8b67b 100644 --- a/paddle/fluid/operators/math/math_function_test.cu +++ b/paddle/fluid/operators/math/math_function_test.cu @@ -13,6 +13,7 @@ // limitations under the License. #include "gtest/gtest.h" #include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/device_context.h" void fill_fp16_data(paddle::platform::float16* in_ptr, size_t size, const std::vector& data) { @@ -178,6 +179,13 @@ TEST(math_function, trans_mul_notrans_fp16) { EXPECT_EQ(static_cast(out_ptr[8]), 29); } +template +inline paddle::operators::math::BlasT +GetBlas(const paddle::platform::CUDADeviceContext& context) { + return paddle::operators::math::GetBlas(context); +} + TEST(math_function, gemm_notrans_cublas_fp32) { paddle::framework::Tensor input1; paddle::framework::Tensor input2; @@ -210,8 +218,8 @@ TEST(math_function, gemm_notrans_cublas_fp32) { float* b = input2_gpu.data(); float* c = input3_gpu.mutable_data(gpu_place); - paddle::operators::math::gemm( - context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4); + GetBlas(context).GEMM(false, false, m, n, k, 1, a, 3, b + 1, 4, 1, + c + 1, 4); paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3); @@ -271,10 +279,9 @@ TEST(math_function, gemm_notrans_cublas_fp16) { paddle::platform::float16* c = input3_gpu.mutable_data(gpu_place); - paddle::operators::math::gemm( - context, false, false, m, n, k, paddle::platform::float16(1), a, 3, b + 1, - 4, paddle::platform::float16(1), c + 1, 4); + GetBlas(context).GEMM( + false, false, m, n, k, static_cast(1), a, 3, + b + 1, 4, static_cast(1), c + 1, 4); paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3); @@ -327,8 +334,8 @@ TEST(math_function, gemm_trans_cublas_fp32) { float* b = input2_gpu.data(); float* c = input3_gpu.mutable_data(gpu_place); - paddle::operators::math::gemm( - context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4); + GetBlas(context).GEMM(false, true, m, n, k, 1, a, 3, b + 3, 3, 1, + c + 1, 4); paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3); @@ -382,10 +389,9 @@ TEST(math_function, gemm_trans_cublas_fp16) { paddle::platform::float16* c = input3_gpu.mutable_data(gpu_place); - paddle::operators::math::gemm( - context, false, true, m, n, k, paddle::platform::float16(1), a, 3, b + 3, - 3, paddle::platform::float16(1), c + 1, 4); + GetBlas(context).GEMM( + false, true, m, n, k, static_cast(1), a, 3, + b + 3, 3, static_cast(1), c + 1, 4); paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3); diff --git a/paddle/fluid/operators/math/matmul.h b/paddle/fluid/operators/math/matmul.h index 0006c5062f3639da589eea44d47917d879933615..67efd1be5322b633e5dbc804e6b0a3db6519f497 100644 --- a/paddle/fluid/operators/math/matmul.h +++ b/paddle/fluid/operators/math/matmul.h @@ -131,8 +131,9 @@ class MatMulFunctor { if (!batchCount) { // regular matrix multiplication - gemm(context, transA, transB, M, N, kA, alpha, - a.data(), b.data(), beta, out->data()); + Blas(context).GEMM(transA, transB, M, N, kA, alpha, + a.data(), b.data(), beta, + out->data()); } else { // batched matrix multiplication batched_gemm( diff --git a/paddle/fluid/operators/momentum_op.cc b/paddle/fluid/operators/momentum_op.cc index 6c70970e15f0d63ebe2134c6bc8163339ba30e75..f13ec53905aa3d5b55b865c3514f36211c06a549 100644 --- a/paddle/fluid/operators/momentum_op.cc +++ b/paddle/fluid/operators/momentum_op.cc @@ -17,6 +17,8 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; + class MomentumOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -50,6 +52,12 @@ class MomentumOp : public framework::OperatorWithKernel { ctx->SetOutputDim("ParamOut", param_dim); ctx->SetOutputDim("VelocityOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class MomentumOpMaker : public framework::OpProtoAndCheckerMaker { diff --git a/paddle/fluid/operators/row_conv_op.cu b/paddle/fluid/operators/row_conv_op.cu index dd8e62aca47a3b34a3788a43cc0043a887af818f..082f761d37ebdecccfa69981856dea2daef818b8 100644 --- a/paddle/fluid/operators/row_conv_op.cu +++ b/paddle/fluid/operators/row_conv_op.cu @@ -14,7 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/row_conv_op.h" -#include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/cuda_device_function.h" namespace paddle { namespace operators { @@ -189,6 +189,10 @@ __global__ void RowConvGradFilterImproved(const T *in, const T *dout, } __syncthreads(); + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int i = 0; i < num_sequence; i++) { int start = static_cast(batch_indices[i]); int end = static_cast(batch_indices[i + 1]); @@ -220,7 +224,7 @@ __global__ void RowConvGradFilterImproved(const T *in, const T *dout, for (int offset = 16; offset > 0; offset = offset / 2) { // blockDim.x is 32. - val += platform::__shfl_down_sync(0, val, offset); + val += platform::__shfl_down_sync(mask, val, offset); } __syncthreads(); @@ -251,6 +255,10 @@ __global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence, T *sh_in = mem; T *sh_dout = &mem[block_x * block_y]; + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int i = 0; i < num_sequence; i++) { int start = static_cast(batch_indices[i]); int end = static_cast(batch_indices[i + 1]); @@ -276,7 +284,7 @@ __global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence, for (int offset = 16; offset > 0; offset = offset / 2) { // blockDim.x is 32. - val += platform::__shfl_down_sync(0, val, offset); + val += platform::__shfl_down_sync(mask, val, offset); } __syncthreads(); diff --git a/paddle/fluid/operators/save_op.cc b/paddle/fluid/operators/save_op.cc index f45d07ed90d52d204e9a3a5c2efe6df6b27ebfe6..dcc1b9ec204e9e273b8fd2b12f2423fc989ba502 100644 --- a/paddle/fluid/operators/save_op.cc +++ b/paddle/fluid/operators/save_op.cc @@ -106,6 +106,8 @@ class SaveOp : public framework::OperatorBase { auto out_kernel_type = framework::OpKernelType(out_dtype, place); framework::LoDTensor out; framework::TransDataType(in_kernel_type, out_kernel_type, tensor, &out); + // copy LoD info to the new tensor + out.set_lod(tensor.lod()); framework::SerializeToStream(fout, out, dev_ctx); } else { framework::SerializeToStream(fout, tensor, dev_ctx); diff --git a/paddle/fluid/operators/scale_op.cc b/paddle/fluid/operators/scale_op.cc index 1e938638c9182972a2ae2436166ff0aa49efd4be..7dcf33c989c3bcd905da8017ee36ec8ce8032911 100644 --- a/paddle/fluid/operators/scale_op.cc +++ b/paddle/fluid/operators/scale_op.cc @@ -35,7 +35,6 @@ class ScaleOp : public framework::OperatorWithKernel { } }; -template class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker) @@ -47,9 +46,9 @@ Scale operator $$Out = scale*X$$ )DOC"); - AddAttr("scale", - "(float, default 1.0)" - "The scaling factor of the scale operator.") + AddAttr("scale", + "(float, default 1.0)" + "The scaling factor of the scale operator.") .SetDefault(1.0); } }; @@ -73,8 +72,7 @@ class ScaleGradMaker : public framework::SingleGradOpDescMaker { namespace ops = paddle::operators; -REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, - ops::ScaleGradMaker); +REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, ops::ScaleGradMaker); REGISTER_OP_CPU_KERNEL( scale, ops::ScaleKernel, ops::ScaleKernel, diff --git a/paddle/fluid/operators/send_recv_op_test.cc b/paddle/fluid/operators/send_recv_op_test.cc index 93e55d410388a672eb749302162ea81de4c6cba1..eb51f301bfe2a97c65dd1fec23ff5a44f3843b05 100644 --- a/paddle/fluid/operators/send_recv_op_test.cc +++ b/paddle/fluid/operators/send_recv_op_test.cc @@ -116,6 +116,7 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs, void StartServerNet(bool is_sparse, std::atomic *initialized) { f::Scope scope; p::CPUPlace place; + VLOG(4) << "before init tensor"; if (is_sparse) { InitSelectedRowsInScope(place, &scope); } else { @@ -137,6 +138,7 @@ void StartServerNet(bool is_sparse, std::atomic *initialized) { attrs.insert({"PrefetchBlock", prefetch_block}); attrs.insert({"grad_to_block_id", std::vector({""})}); attrs.insert({"sync_mode", true}); + VLOG(4) << "before init op"; listen_and_serv_op = f::OpRegistry::CreateOp("listen_and_serv", {{"X", {"x1"}}}, {}, attrs); *initialized = true; @@ -149,7 +151,9 @@ TEST(SendRecvOp, CPUDense) { std::thread server_thread(StartServerNet, false, &initialized); while (!initialized) { } - sleep(5); // wait server to start + static_cast(listen_and_serv_op.get()) + ->WaitServerReady(); + // local net f::Scope scope; p::CPUPlace place; @@ -185,6 +189,7 @@ TEST(SendRecvOp, CPUDense) { listen_and_serv_op->Stop(); server_thread.join(); listen_and_serv_op.reset(nullptr); + paddle::operators::ListenAndServOp::ResetPort(); } TEST(SendRecvOp, CPUSparse) { @@ -193,7 +198,12 @@ TEST(SendRecvOp, CPUSparse) { std::thread server_thread(StartServerNet, true, &initialized); while (!initialized) { } - sleep(5); // wait server to start + auto *listen_and_serv_op_ptr = + static_cast( + listen_and_serv_op.get()); + ASSERT_TRUE(listen_and_serv_op_ptr != nullptr); + listen_and_serv_op_ptr->WaitServerReady(); + // local net f::Scope scope; p::CPUPlace place; @@ -201,10 +211,6 @@ TEST(SendRecvOp, CPUSparse) { InitSelectedRowsInScope(place, &scope); scope.Var("RPC_CLIENT_VAR"); f::AttributeMap attrs; - auto *listen_and_serv_op_ptr = - static_cast( - listen_and_serv_op.get()); - ASSERT_TRUE(listen_and_serv_op_ptr != nullptr); selected_port = listen_and_serv_op_ptr->GetSelectedPort(); std::string endpoint = paddle::string::Sprintf("127.0.0.1:%d", selected_port); attrs.insert({"endpoints", std::vector({endpoint})}); @@ -236,4 +242,5 @@ TEST(SendRecvOp, CPUSparse) { listen_and_serv_op->Stop(); server_thread.join(); listen_and_serv_op.reset(); + paddle::operators::ListenAndServOp::ResetPort(); } diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index 2741ba95bcfc1db3d74e0fb8c3f6fddf7d5a2caa..aa7b192e327704c02a26c86cc208ebe8a5cd7ba5 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -164,7 +164,9 @@ REGISTER_OPERATOR(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(softmax_grad, ops::SoftmaxOpGrad); REGISTER_OP_CPU_KERNEL( - softmax, ops::SoftmaxKernel); + softmax, ops::SoftmaxKernel, + ops::SoftmaxKernel); REGISTER_OP_CPU_KERNEL( softmax_grad, - ops::SoftmaxGradKernel); + ops::SoftmaxGradKernel, + ops::SoftmaxGradKernel); diff --git a/paddle/fluid/operators/softmax_op.cu.cc b/paddle/fluid/operators/softmax_op.cu.cc index 0c1f7cef7ab7b66358d80f6f0670e0d07536128c..5fb4f011d9b47cebc4a23bcce47eada825263343 100644 --- a/paddle/fluid/operators/softmax_op.cu.cc +++ b/paddle/fluid/operators/softmax_op.cu.cc @@ -19,6 +19,8 @@ namespace ops = paddle::operators; namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( softmax, ops::SoftmaxKernel, + ops::SoftmaxKernel, ops::SoftmaxKernel); -REGISTER_OP_CUDA_KERNEL(softmax_grad, - ops::SoftmaxGradKernel); +REGISTER_OP_CUDA_KERNEL( + softmax_grad, ops::SoftmaxGradKernel, + ops::SoftmaxGradKernel); diff --git a/paddle/fluid/operators/top_k_op.cc b/paddle/fluid/operators/top_k_op.cc index 2e4e8caed5327f4ca9038c376de2ec831354917e..942a5de3f90f20eabe691924a570b61509eccf76 100644 --- a/paddle/fluid/operators/top_k_op.cc +++ b/paddle/fluid/operators/top_k_op.cc @@ -75,4 +75,5 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(top_k, ops::TopkOp, ops::TopkOpMaker, paddle::framework::EmptyGradOpMaker); REGISTER_OP_CPU_KERNEL(top_k, - ops::TopkKernel); + ops::TopkKernel, + ops::TopkKernel); diff --git a/paddle/fluid/operators/top_k_op.cu b/paddle/fluid/operators/top_k_op.cu index d7f4d383ce0d9e1ff42fc12c96aaf0ceb532e5db..faaae1f9b6ae1e2c577017c34116dab95e261ab2 100644 --- a/paddle/fluid/operators/top_k_op.cu +++ b/paddle/fluid/operators/top_k_op.cu @@ -15,6 +15,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/top_k_op.h" #include "paddle/fluid/platform/assert.h" +#include "paddle/fluid/platform/cuda_device_function.h" namespace paddle { namespace operators { @@ -235,8 +236,13 @@ __device__ __forceinline__ void BlockReduce(Pair* sh_topk, int* maxid, sh_topk[tid] = topk[*beam]; } } + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + if (maxid[0] / 32 == warp) { - if (__shfl(*beam, (maxid[0]) % 32, 32) == MaxLength) break; + if (platform::__shfl_sync(mask, *beam, (maxid[0]) % 32, 32) == MaxLength) + break; } } } @@ -318,4 +324,5 @@ class TopkOpCUDAKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel, + paddle::operators::TopkOpCUDAKernel); diff --git a/paddle/fluid/platform/cuda_device_function.h b/paddle/fluid/platform/cuda_device_function.h new file mode 100644 index 0000000000000000000000000000000000000000..2405f33d4f0ad83611e57d07a47e787eab439285 --- /dev/null +++ b/paddle/fluid/platform/cuda_device_function.h @@ -0,0 +1,84 @@ +/* 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 + +namespace paddle { +namespace platform { + +// __shfl_down and __shfl have been deprecated as of CUDA 9.0. +#if CUDA_VERSION < 9000 +template +__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) { + return __shfl_down(val, delta); +} + +template +__forceinline__ __device__ T __shfl_sync(unsigned, T val, int src_line, + int width) { + return __shfl(val, src_line, width); +} +#define CREATE_SHFL_MASK(mask, predicate) mask = 0u; +#else +#define FULL_WARP_MASK 0xFFFFFFFF +#define CREATE_SHFL_MASK(mask, predicate) \ + mask = __ballot_sync(FULL_WARP_MASK, (predicate)) +template +__forceinline__ __device__ T __shfl_down_sync(unsigned mask, T val, int delta) { + return __shfl_down_sync(mask, val, delta); +} + +template +__forceinline__ __device__ T __shfl_sync(unsigned mask, T val, int src_line, + int width) { + return __shfl_sync(mask, val, src_line, width); +} +#endif + +template +__device__ T reduceSum(T val, int tid, int len) { + // NOTE(zcd): The warp size should be taken from the + // parameters of the GPU but not specified as 32 simply. + // To make the reduceSum more efficiently, + // I use Warp-Level Parallelism and assume the Warp size + // is 32 which may be different for different GPU, + // but most card's warp size is 32. + const int warpSize = 32; + __shared__ T shm[warpSize]; + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, tid < len); + + for (int offset = warpSize / 2; offset > 0; offset /= 2) + val += platform::__shfl_down_sync(mask, val, offset); + + if (tid < warpSize) shm[tid] = 0; + + if (tid % warpSize == 0) { + shm[tid / warpSize] = val; + } + __syncthreads(); + + CREATE_SHFL_MASK(mask, tid < warpSize); + + if (tid < warpSize) { + val = shm[tid]; + for (int offset = warpSize / 2; offset > 0; offset /= 2) + val += platform::__shfl_down_sync(mask, val, offset); + } + return val; +} + +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/cuda_primitives.h b/paddle/fluid/platform/cuda_primitives.h index 866ff30a8be7be124a72a8dc7e70ef4140ee716a..d535ed2f89df6a0b311ec068ecd92c8e3183cee7 100644 --- a/paddle/fluid/platform/cuda_primitives.h +++ b/paddle/fluid/platform/cuda_primitives.h @@ -65,19 +65,5 @@ CUDA_ATOMIC_WRAPPER(Add, double) { return __longlong_as_double(old); } #endif - -// __shfl_down has been deprecated as of CUDA 9.0. -#if CUDA_VERSION < 9000 -template -__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) { - return __shfl_down(val, delta); -} -#define CREATE_SHFL_MASK(mask, predicate) mask = 0u; -#else -#define FULL_WARP_MASK 0xFFFFFFFF -#define CREATE_SHFL_MASK(mask, predicate) \ - mask = __ballot_sync(FULL_WARP_MASK, (predicate)) -#endif - } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index b07427c8f6903e0100ca9a478881444d86501bcc..428d9ebceaabd987261c1dcd6e66faf044b702c0 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -18,7 +18,6 @@ limitations under the License. */ #include #include #include "paddle/fluid/platform/device_context.h" -#include "paddle/fluid/platform/profiler.pb.h" namespace paddle { namespace platform { diff --git a/paddle/function/RowConvOpGpu.cu b/paddle/function/RowConvOpGpu.cu index 9d8a6d80bb2c95a9aa8eb3666fd637402c932559..f820ee9a9713ce17547aa03945dc3c291ef50a59 100644 --- a/paddle/function/RowConvOpGpu.cu +++ b/paddle/function/RowConvOpGpu.cu @@ -12,8 +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 "RowConvOp.h" -#include "hl_base.h" +#include "paddle/cuda/include/hl_base.h" +#include "paddle/function/RowConvOp.h" namespace paddle { @@ -94,7 +94,7 @@ __global__ void KeRowConv2(real* y, } template <> -void RowConv(GpuMatrix& out, +void RowConv(GpuMatrix& out, // NOLINT const GpuMatrix& in, const GpuMatrix& filter, const GpuIVector& seq) { @@ -144,6 +144,10 @@ __global__ void KeRowConvBwWeight(real* dw, } __syncthreads(); + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int i = 0; i < numSeq; ++i) { const int start = starts[i]; const int end = starts[i + 1]; @@ -170,11 +174,10 @@ __global__ void KeRowConvBwWeight(real* dw, real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx + context - 1 - t]; __syncthreads(); // warp size and blockDim.x is 32. - val += __shfl_down(val, 16); - val += __shfl_down(val, 8); - val += __shfl_down(val, 4); - val += __shfl_down(val, 2); - val += __shfl_down(val, 1); + + for (int offset = 16; offset > 0; offset /= 2) + val += __shfl_down_sync(mask, val, offset); + __syncthreads(); if (tidx == 0) { sh_dw[t][tidy] += val; @@ -205,6 +208,10 @@ __global__ void KeRowConvBwWeight2(real* dw, __shared__ real sh_x[BLOCK_H][BLOCK_W]; __shared__ real sh_dy[BLOCK_H][BLOCK_W]; + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int i = 0; i < numSeq; ++i) { const int start = starts[i]; const int end = starts[i + 1]; @@ -230,11 +237,9 @@ __global__ void KeRowConvBwWeight2(real* dw, real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx]; __syncthreads(); // warp size and blockDim.x is 32. - val += __shfl_down(val, 16); - val += __shfl_down(val, 8); - val += __shfl_down(val, 4); - val += __shfl_down(val, 2); - val += __shfl_down(val, 1); + for (int offset = 16; offset > 0; offset /= 2) + val += __shfl_down_sync(mask, val, offset); + __syncthreads(); if (tidx == 0 && (gidx + tidy) < width) { @@ -323,8 +328,8 @@ template <> void RowConvGrad(const GpuMatrix& outG, const GpuMatrix& in, const GpuMatrix& filter, - GpuMatrix& inG, - GpuMatrix& filterG, + GpuMatrix& inG, // NOLINT + GpuMatrix& filterG, // NOLINT const GpuIVector& seq) { const size_t numSeq = seq.getSize() - 1; const size_t contextLength = filter.getHeight(); diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 0e84cb37392839d112448b0b3c12b042e7df838e..bcd6dfe1fda6b1243007b0c26a6e0087eedcc10c 100644 --- a/paddle/math/Matrix.cpp +++ b/paddle/math/Matrix.cpp @@ -2157,26 +2157,20 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat, int wend = wstart + sizeX; wstart = wstart < 0 ? 0 : wstart; wend = wend < (int)imgSizeW ? wend : (int)imgSizeW; - if (maskData == NULL) { - real tmp = -(real)FLT_MAX; - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - tmp = tmp < inputData[h * imgSizeW + w] - ? inputData[h * imgSizeW + w] - : tmp; - } - } - outData[ph * outputW + pw] = tmp; - } else { - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - if (outData[ph * outputW + pw] < inputData[h * imgSizeW + w]) { - outData[ph * outputW + pw] = inputData[h * imgSizeW + w]; - maskData[ph * outputW + pw] = h * imgSizeW + w; - } + + real maxval = -(real)FLT_MAX; + int max_index = -1; + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + if (maxval < inputData[h * imgSizeW + w]) { + maxval = inputData[h * imgSizeW + w]; + max_index = h * imgSizeW + w; } } } + + outData[ph * outputW + pw] = maxval; + if (maskData != NULL) maskData[ph * outputW + pw] = max_index; } } // compute offset diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index 53455fd86041f87d4050eba89f37c3435fed1ae1..1595cc9e8aad4d143ca62f84f812dbc791dc1d26 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -40,6 +40,7 @@ function print_usage() { ${BLUE}capi${NONE}: generate paddle CAPI package ${BLUE}fluid_inference_lib${NONE}: deploy fluid inference library ${BLUE}check_style${NONE}: run code style check + ${BLUE}cicheck${NONE}: run CI tasks " } @@ -453,6 +454,8 @@ function gen_capi_package() { } function gen_fluid_inference_lib() { + mkdir -p ${PADDLE_ROOT}/build + cd ${PADDLE_ROOT}/build if [ ${WITH_C_API:-OFF} == "OFF" ] ; then cat <feed_target_var - - Change it to: - feed_op->feed_target_var->cast_op(from other dtype to float16)->tmp_var - - For each fetch op: - fetch_target_var->fetch_op - - Change it to: - tmp_var->cast_op(from float16 to other dtype)->fetch_target_var->fetch_op - - :return: None - ''' - - def find_op(var): - # It is possible that var.op is not up to date after some - # modifications to program desc. Here we force to make it up to date. - var.op = None - for op in self.block.ops: - if var.name in op.output_arg_names: - var.op = op - break - - if var.op is None: - raise ValueError("The target variable must have an " - "associated operator that generates it.") - - i = 0 - while i < len(self.block.ops): - cur_op = self.block.ops[i] - if cur_op.type == "feed": - var_name = cur_op.output("Out")[0] - tmp_var_name = var_name + ".fp16" - var = self.block.vars[var_name] - tmp_var = self.block.create_var( - name=tmp_var_name.encode('ascii'), - type=var.type, - dtype=core.VarDesc.VarType.FP16, - shape=var.shape, - persistable=var.persistable) - self.block.insert_op( - i + 1, - type="cast", - inputs={"X": var}, - outputs={"Out": tmp_var}, - attrs={ - 'in_dtype': int(var.dtype), - 'out_dtype': int(tmp_var.dtype) - }) - self.input_map[var_name] = tmp_var_name - i = i + 1 - elif cur_op.type == "fetch": - var_name = cur_op.input("X")[0] - tmp_var_name = var_name + ".fp16" - var = self.block.vars[var_name] - tmp_var = self.block.create_var( - name=tmp_var_name.encode('ascii'), - type=var.type, - dtype=core.VarDesc.VarType.FP16, - shape=var.shape, - persistable=var.persistable) - find_op(var) - var.op.rename_output(var_name, tmp_var_name) - self.block.insert_op( - i, - type="cast", - inputs={"X": tmp_var}, - outputs={"Out": var}, - attrs={ - 'in_dtype': int(tmp_var.dtype), - 'out_dtype': int(var.dtype) - }) - i = i + 1 - i = i + 1 - - def _convert_param_to_float16(self): - def _get_no_fp16_conversion_var_names(): - ''' - Get the set of input variable names that shouldn't be converted to float16. - - When we want to run inference in float16 mode, most parameters need to be - firstly converted to float16. However, there are some parameters that - shouldn't be converted to float16 because the corresponding operator - requires float32 parameters even in float16 mode (when the input data is - of float16 data type). Currently, the only operator that has this exclusion - is the batch norm op. - - :return: set of input variable names - :type var_names: set - ''' - op_names = {'batch_norm'} - var_names = [] - for op in self.block.ops: - if op.type in op_names: - var_names += op.input_arg_names - return set(var_names) - - def _should_be_converted(var): - return var.persistable and \ - var.name not in self.no_conversion_vars and \ - var.type != core.VarDesc.VarType.FEED_MINIBATCH and \ - var.type != core.VarDesc.VarType.FETCH_LIST - - self.no_conversion_vars = _get_no_fp16_conversion_var_names() - conversion_var_list = filter(_should_be_converted, - self.block.vars.values()) - for var in conversion_var_list: - fp16_var_name = var.name + ".fp16" - fp16_var = self.block.create_parameter( - name=fp16_var_name.encode('ascii'), - type=var.type, - dtype=core.VarDesc.VarType.FP16, - shape=var.shape) - - # cast the data in the tensor of the original var to float16 - # data type and store it in the tensor of the new float16 var - self.scope.var(fp16_var_name) - fp16_tensor = self.scope.find_var(fp16_var_name).get_tensor() - tensor = np.array(self.scope.find_var(var.name).get_tensor()) - # After the old tensor data is converted to np.float16, view(np.uint16) - # is used so that the internal memory of the numpy array will be - # reinterpreted to be of np.uint16 data type, which is binded to fluid - # float16 data type via the help of pybind in tensor_py.h. - fp16_tensor.set( - tensor.astype(np.float16).view(np.uint16), self.place) - - # old var will be replaced by the fp16 var in program desc - self.input_map[var.name] = fp16_var_name - self.block.remove_var(var.name) diff --git a/python/paddle/fluid/inferencer.py b/python/paddle/fluid/inferencer.py index 3ea50bf196d00152e6579623c981ecbfb57b8e3b..58e027695a7100245dd424583e2cedeed3d165e6 100644 --- a/python/paddle/fluid/inferencer.py +++ b/python/paddle/fluid/inferencer.py @@ -12,18 +12,22 @@ # See the License for the specific language governing permissions and # limitations under the License. +import core + __all__ = ['Inferencer', ] class Inferencer(object): - def __init__(self, network_func, params, place=None): + def __init__(self, network_func, param_path=None, place=None): # 1. we need to generate a framework.Program by calling # network_func. Reference: fluid.program_guard in test_word2vec.py # 2. move the default_main_program to self.program. # 3. run the default_startup program. - self.params = params + + # 4. load params from param_path into scope + self.scope = core.Scope() self.place = place def infer(self, inputs): diff --git a/python/paddle/fluid/layer_helper.py b/python/paddle/fluid/layer_helper.py index 62933b512529bd04fab8c9ded12e636ecfae685c..86efd1ff51cf29485ee28b4d60ffb1439af1aad9 100644 --- a/python/paddle/fluid/layer_helper.py +++ b/python/paddle/fluid/layer_helper.py @@ -400,11 +400,11 @@ class LayerHelper(object): if isinstance(act, basestring): act = {'type': act} + if 'use_cudnn' in self.kwargs and self.kwargs.get('use_cudnn'): + act['use_cudnn'] = self.kwargs.get('use_cudnn') if 'use_mkldnn' in self.kwargs: act['use_mkldnn'] = self.kwargs.get('use_mkldnn') act_type = act.pop('type') - if 'use_mkldnn' in self.kwargs: - act['use_mkldnn'] = self.kwargs.get('use_mkldnn') tmp = input_var # NOTE(dzhwinter): some activation support inplace compution. if not core.IsInplace(act_type): diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index acfad45704d4ea9e28711c019db3563489aab3ff..0a6befd1485a1f79d63873c47a9fd74ab4214f57 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -50,8 +50,6 @@ def data(name, dtype(int|float): The type of data : float32, float_16, int etc type(VarType): The output type. By default it is LOD_TENSOR. lod_level(int): The LoD Level. 0 means the input data is not a sequence. - main_program(Program): Name of the main program that calls this - startup_program(Program): Name of the startup program stop_gradient(bool): A boolean that mentions whether gradient should flow. Returns: @@ -74,13 +72,15 @@ def data(name, if append_batch_size: shape = [-1] + shape # append batch size as -1 - return helper.create_global_variable( + data_var = helper.create_global_variable( name=name, shape=shape, dtype=dtype, type=type, stop_gradient=stop_gradient, lod_level=lod_level) + data_var.is_data = True + return data_var class BlockGuardServ(BlockGuard): diff --git a/python/paddle/fluid/layers/math_op_patch.py b/python/paddle/fluid/layers/math_op_patch.py index 08a0184c2c2ad5f3c3792fd0a12f0ab0c746849b..1754061c4ba6f5b97bced3548bc412dfb1b7932c 100644 --- a/python/paddle/fluid/layers/math_op_patch.py +++ b/python/paddle/fluid/layers/math_op_patch.py @@ -169,7 +169,9 @@ def monkey_patch_variable(): # a*b == b*a. Do not need to reverse explicitly ("__rmul__", "elementwise_mul", False), ("__div__", "elementwise_div", False), + ("__truediv__", "elementwise_div", False), ("__rdiv__", "elementwise_div", True), + ("__rtruediv__", "elementwise_div", True), ("__pow__", "elementwise_pow", False), ("__rpow__", "elementwise_pow", True), # for logical compare diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 7f16bf2a0c430213b2f52dafe8fa948b9e350f96..1786be22fdcd0d074b45bc94b3b0c4e8c41b4e8a 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -88,6 +88,7 @@ def fc(input, num_flatten_dims=1, param_attr=None, bias_attr=None, + use_cudnn=False, use_mkldnn=False, act=None, is_test=False, @@ -1496,6 +1497,7 @@ def batch_norm(input, bias_attr=None, data_layout='NCHW', in_place=False, + use_mkldnn=False, name=None, moving_mean_name=None, moving_variance_name=None, @@ -1574,9 +1576,12 @@ def batch_norm(input, "SavedMean": saved_mean, "SavedVariance": saved_variance }, - attrs={"momentum": momentum, - "epsilon": epsilon, - "is_test": is_test}) + attrs={ + "momentum": momentum, + "epsilon": epsilon, + "is_test": is_test, + "use_mkldnn": use_mkldnn + }) return helper.append_activation(batch_norm_out) diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 9ae43b3e93e4b7d337097a25379720c18dfd331c..0a314ddfd7c607a3bc7f7c746c4c4990fc4a52e2 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -28,7 +28,8 @@ from contextlib import contextmanager __all__ = [ 'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer', - 'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'Adadelta', 'ModelAverage' + 'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'Adadelta', 'ModelAverage', + 'Optimizer' ] diff --git a/python/paddle/fluid/params.py b/python/paddle/fluid/params.py deleted file mode 100644 index a5d257e53a2958acd1b8f6ef29d0f9f531b36678..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/params.py +++ /dev/null @@ -1,39 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from . import core - -__all__ = ['Params', ] - - -class Params(object): - def __init__(self, path=None): - self.scope = core.Scope() - - if path: - self._load(path) - - def _load(self, path): - # reference: load_persistables in io.py - pass - - def save(self, path): - # reference: save_persistables in io.py - pass - - def add_params(self, scope): - # take the keys from the scope, - # if not already exists in self.scope, - # add the key and value into self.scope. - pass diff --git a/python/paddle/fluid/tests/book/image_classification/notest_image_classification_resnet.py b/python/paddle/fluid/tests/book/image_classification/notest_image_classification_resnet.py index 5cbfdef91a64ae7c58d060edfb7b9f3bc8160f2b..17db38797cf19ae387f69f66daa42fc78cfcb7d5 100644 --- a/python/paddle/fluid/tests/book/image_classification/notest_image_classification_resnet.py +++ b/python/paddle/fluid/tests/book/image_classification/notest_image_classification_resnet.py @@ -64,15 +64,14 @@ def resnet_cifar10(input, depth=32): res3 = layer_warp(basicblock, res2, 32, 64, n, 2) pool = fluid.layers.pool2d( input=res3, pool_size=8, pool_type='avg', pool_stride=1) - return pool + predict = fluid.layers.fc(input=pool, size=10, act='softmax') + return predict def inference_network(): - classdim = 10 data_shape = [3, 32, 32] images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') - net = resnet_cifar10(images, 32) - predict = fluid.layers.fc(input=net, size=classdim, act='softmax') + predict = resnet_cifar10(images, 32) return predict diff --git a/python/paddle/fluid/tests/book/image_classification/notest_image_classification_vgg.py b/python/paddle/fluid/tests/book/image_classification/notest_image_classification_vgg.py index 8a6a5ff61a913ad6cbc609f8376afcbc621d60e2..e83afeed2f72635a40aa2ac21dc0c8611c309de4 100644 --- a/python/paddle/fluid/tests/book/image_classification/notest_image_classification_vgg.py +++ b/python/paddle/fluid/tests/book/image_classification/notest_image_classification_vgg.py @@ -43,15 +43,14 @@ def vgg16_bn_drop(input): bn = fluid.layers.batch_norm(input=fc1, act='relu') drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) fc2 = fluid.layers.fc(input=drop2, size=4096, act=None) - return fc2 + predict = fluid.layers.fc(input=fc2, size=10, act='softmax') + return predict def inference_network(): - classdim = 10 data_shape = [3, 32, 32] images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') - net = vgg16_bn_drop(images) - predict = fluid.layers.fc(input=net, size=classdim, act='softmax') + predict = vgg16_bn_drop(images) return predict diff --git a/python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py b/python/paddle/fluid/tests/book/notest_rnn_encoder_decoder.py similarity index 100% rename from python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py rename to python/paddle/fluid/tests/book/notest_rnn_encoder_decoder.py diff --git a/python/paddle/fluid/tests/book/test_fit_a_line.py b/python/paddle/fluid/tests/book/test_fit_a_line.py index 6dfc2997ae0328a41fe22d13dfa8fc51d4d021a6..ecb34699af0dc14782601702ab8afedbca7e1bfd 100644 --- a/python/paddle/fluid/tests/book/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/test_fit_a_line.py @@ -80,12 +80,7 @@ def train(use_cuda, save_dirname, is_local): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_image_classification.py b/python/paddle/fluid/tests/book/test_image_classification.py index 09f994c37020c4612223e4168be4cf535157f60b..dbcdb5766e7d20efdb12da0ea4c6f005d903849b 100644 --- a/python/paddle/fluid/tests/book/test_image_classification.py +++ b/python/paddle/fluid/tests/book/test_image_classification.py @@ -189,12 +189,7 @@ def train(net_type, use_cuda, save_dirname, is_local): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, @@ -252,26 +247,6 @@ def infer(use_cuda, save_dirname=None): fetch_targets, exe, inference_transpiler_program) - if use_cuda and fluid.core.is_float16_supported(place): - # Use float16_transpiler to speedup - fp16_transpiler_program = inference_transpiler_program.clone() - t.float16_transpile(fp16_transpiler_program, place) - - fp16_results = exe.run(fp16_transpiler_program, - feed={feed_target_names[0]: tensor_img}, - fetch_list=fetch_targets) - - assert len(results[0]) == len(fp16_results[0]) - for i in range(len(results[0])): - np.testing.assert_almost_equal( - results[0][i], fp16_results[0][i], decimal=2) - - print("float16 infer results: ", fp16_results[0]) - - fluid.io.save_inference_model("float16_" + save_dirname, - feed_target_names, fetch_targets, exe, - fp16_transpiler_program) - def main(net_type, use_cuda, is_local=True): if use_cuda and not fluid.core.is_compiled_with_cuda(): diff --git a/python/paddle/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/fluid/tests/book/test_label_semantic_roles.py index d9cd76952e31f8185512ab45f9f3ab2ce7d9da48..50ef29c4572f1b12fe9793bbf037cd7fe71a9e53 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -259,12 +259,7 @@ def train(use_cuda, save_dirname=None, is_local=True): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index 830d78df8b9e56b45f7e928562ef4b89e88f696d..46c6b9c29a265741a99655d5ac29244798f6fec2 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -231,12 +231,7 @@ def train_main(use_cuda, is_sparse, is_local=True): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_recognize_digits.py b/python/paddle/fluid/tests/book/test_recognize_digits.py index 5ec6890c1b0dabd2804a92071b63c9610299e67c..c115aa4d7d6b514f9207543730e5e76cb0d2040c 100644 --- a/python/paddle/fluid/tests/book/test_recognize_digits.py +++ b/python/paddle/fluid/tests/book/test_recognize_digits.py @@ -162,12 +162,7 @@ def train(nn_type, trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_recommender_system.py b/python/paddle/fluid/tests/book/test_recommender_system.py index 2172c275b8082689a6ff5f2c3c27a2ff4e92275a..d022dedbff805d597b68b5a47f7931f2dd946615 100644 --- a/python/paddle/fluid/tests/book/test_recommender_system.py +++ b/python/paddle/fluid/tests/book/test_recommender_system.py @@ -261,12 +261,7 @@ def train(use_cuda, save_dirname, is_local=True): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_understand_sentiment.py b/python/paddle/fluid/tests/book/test_understand_sentiment.py index dedd153778d7ad9caeb5fa7090a980bc7f177dea..241778e303036d068dc0a40e4574a02eb97ad134 100644 --- a/python/paddle/fluid/tests/book/test_understand_sentiment.py +++ b/python/paddle/fluid/tests/book/test_understand_sentiment.py @@ -213,12 +213,7 @@ def train(word_dict, trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py index 8929779de9448d036e1528b64330b37463ab3988..6dec0f6857e86b4b9c1c67af934aa9bfdb1c3df7 100644 --- a/python/paddle/fluid/tests/book/test_word2vec.py +++ b/python/paddle/fluid/tests/book/test_word2vec.py @@ -145,12 +145,7 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/understand_sentiment/notest_understand_sentiment_stacked_lstm.py b/python/paddle/fluid/tests/book/understand_sentiment/notest_understand_sentiment_stacked_lstm.py new file mode 100644 index 0000000000000000000000000000000000000000..9948e5c0234ed78237c94f9a25d6401619267d0d --- /dev/null +++ b/python/paddle/fluid/tests/book/understand_sentiment/notest_understand_sentiment_stacked_lstm.py @@ -0,0 +1,140 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import paddle +import paddle.fluid as fluid +from functools import partial + +CLASS_DIM = 2 +EMB_DIM = 128 +HID_DIM = 512 +STACKED_NUM = 3 + + +def stacked_lstm_net(data, input_dim, class_dim, emb_dim, hid_dim, stacked_num): + assert stacked_num % 2 == 1 + + emb = fluid.layers.embedding( + input=data, size=[input_dim, emb_dim], is_sparse=True) + + fc1 = fluid.layers.fc(input=emb, size=hid_dim) + lstm1, cell1 = fluid.layers.dynamic_lstm(input=fc1, size=hid_dim) + + inputs = [fc1, lstm1] + + for i in range(2, stacked_num + 1): + fc = fluid.layers.fc(input=inputs, size=hid_dim) + lstm, cell = fluid.layers.dynamic_lstm( + input=fc, size=hid_dim, is_reverse=(i % 2) == 0) + inputs = [fc, lstm] + + fc_last = fluid.layers.sequence_pool(input=inputs[0], pool_type='max') + lstm_last = fluid.layers.sequence_pool(input=inputs[1], pool_type='max') + + prediction = fluid.layers.fc(input=[fc_last, lstm_last], + size=class_dim, + act='softmax') + return prediction + + +def inference_network(word_dict): + data = fluid.layers.data( + name="words", shape=[1], dtype="int64", lod_level=1) + + dict_dim = len(word_dict) + net = stacked_lstm_net(data, dict_dim, CLASS_DIM, EMB_DIM, HID_DIM, + STACKED_NUM) + return net + + +def train_network(word_dict): + prediction = inference_network(word_dict) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + cost = fluid.layers.cross_entropy(input=prediction, label=label) + avg_cost = fluid.layers.mean(cost) + accuracy = fluid.layers.accuracy(input=prediction, label=label) + return avg_cost, accuracy + + +def train(use_cuda, save_path): + BATCH_SIZE = 128 + EPOCH_NUM = 5 + + word_dict = paddle.dataset.imdb.word_dict() + + train_data = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.imdb.train(word_dict), buf_size=1000), + batch_size=BATCH_SIZE) + + test_data = paddle.batch( + paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) + + def event_handler(event): + if isinstance(event, fluid.EndIteration): + if (event.batch_id % 10) == 0: + avg_cost, accuracy = trainer.test(reader=test_data) + + print('BatchID {1:04}, Loss {2:2.2}, Acc {3:2.2}'.format( + event.batch_id + 1, avg_cost, accuracy)) + + if accuracy > 0.01: # Low threshold for speeding up CI + trainer.params.save(save_path) + return + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + trainer = fluid.Trainer( + partial(train_network, word_dict), + optimizer=fluid.optimizer.Adagrad(learning_rate=0.002), + place=place, + event_handler=event_handler) + + trainer.train(train_data, EPOCH_NUM, event_handler=event_handler) + + +def infer(use_cuda, save_path): + params = fluid.Params(save_path) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + word_dict = paddle.dataset.imdb.word_dict() + inferencer = fluid.Inferencer( + partial(inference_network, word_dict), params, place=place) + + def create_random_lodtensor(lod, place, low, high): + data = np.random.random_integers(low, high, + [lod[-1], 1]).astype("int64") + res = fluid.LoDTensor() + res.set(data, place) + res.set_lod([lod]) + return res + + lod = [0, 4, 10] + tensor_words = create_random_lodtensor( + lod, place, low=0, high=len(word_dict) - 1) + results = inferencer.infer({'words': tensor_words}) + print("infer results: ", results) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + save_path = "understand_sentiment_stacked_lstm.inference.model" + train(use_cuda, save_path) + infer(use_cuda, save_path) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/word2vec/no_test_word2vec_new_api.py b/python/paddle/fluid/tests/book/word2vec/no_test_word2vec_new_api.py new file mode 100644 index 0000000000000000000000000000000000000000..35e163dc9df5a35ee5774b6b157366c4eabcb0f7 --- /dev/null +++ b/python/paddle/fluid/tests/book/word2vec/no_test_word2vec_new_api.py @@ -0,0 +1,149 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle +import paddle.fluid as fluid +import numpy as np +import math +import sys +from functools import partial + +PASS_NUM = 100 +EMBED_SIZE = 32 +HIDDEN_SIZE = 256 +N = 5 +BATCH_SIZE = 32 + + +def create_random_lodtensor(lod, place, low, high): + # The range of data elements is [low, high] + data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64") + res = fluid.LoDTensor() + res.set(data, place) + res.set_lod([lod]) + return res + + +word_dict = paddle.dataset.imikolov.build_dict() +dict_size = len(word_dict) + + +def inference_program(is_sparse): + first_word = fluid.layers.data(name='firstw', shape=[1], dtype='int64') + second_word = fluid.layers.data(name='secondw', shape=[1], dtype='int64') + third_word = fluid.layers.data(name='thirdw', shape=[1], dtype='int64') + forth_word = fluid.layers.data(name='forthw', shape=[1], dtype='int64') + + embed_first = fluid.layers.embedding( + input=first_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_second = fluid.layers.embedding( + input=second_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_third = fluid.layers.embedding( + input=third_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_forth = fluid.layers.embedding( + input=forth_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + + concat_embed = fluid.layers.concat( + input=[embed_first, embed_second, embed_third, embed_forth], axis=1) + hidden1 = fluid.layers.fc(input=concat_embed, + size=HIDDEN_SIZE, + act='sigmoid') + predict_word = fluid.layers.fc(input=hidden1, size=dict_size, act='softmax') + return predict_word + + +def train_program(is_sparse): + next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') + predict_word = inference_program(is_sparse) + cost = fluid.layers.cross_entropy(input=predict_word, label=next_word) + avg_cost = fluid.layers.mean(cost) + return avg_cost + + +def train(use_cuda, is_sparse, save_path): + train_reader = paddle.batch( + paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + def event_handler(event): + print type(event) + if isinstance(event, fluid.EndEpochEvent): + avg_cost = trainer.test(reader=paddle.dataset.imikolov.test( + word_dict, N)) + + if avg_cost < 5.0: + trainer.save_params(save_path) + return + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + + trainer = fluid.Trainer( + partial(train_program, is_sparse), + fluid.optimizer.SGD(learning_rate=0.001), + place=place) + trainer.train( + reader=train_reader, num_epochs=100, event_handler=event_handler) + + +def infer(use_cuda, is_sparse, save_path): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + partial(inference_program, is_sparse), + param_path=save_path, + place=place) + + lod = [0, 1] + first_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1) + second_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1) + third_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1) + fourth_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1) + result = inferencer.infer({ + 'firstw': first_word, + 'secondw': second_word, + 'thirdw': third_word, + 'forthw': fourth_word + }) + print(result) + + +def main(use_cuda, is_sparse): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + + save_path = "word2vec.inference.model" + train(use_cuda, is_sparse, save_path) + infer(use_cuda, is_sparse, save_path) + + +if __name__ == '__main__': + for use_cuda in (False, True): + for is_sparse in (False, True): + main(use_cuda=use_cuda, is_sparse=is_sparse) diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py new file mode 100644 index 0000000000000000000000000000000000000000..f6097d4b846e8da1c4ee3cc49b31f9873660056d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py @@ -0,0 +1,56 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +import paddle.fluid.core as core +from paddle.fluid.op import Operator +import paddle.fluid as fluid +from op_test import OpTest +from paddle.fluid.framework import grad_var_name +from test_batch_norm_op import TestBatchNormOpInference, TestBatchNormOpTraining, _reference_training, _reference_grad + + +class TestMKLDNNBatchNormOpTraining(TestBatchNormOpTraining): + def init_kernel_type(self): + self.use_mkldnn = True + self.data_formats = ["NCHW"] + + def ref_forward_backward(self, x, y_grad, scale, bias, mean, variance, + epsilon, momentum, shape, data_layout): + # run forward + y, saved_mean, saved_variance = _reference_training( + x, scale, bias, epsilon, data_layout) + mean_out = saved_mean * (1. - momentum) + momentum * mean + variance_out = saved_variance * (1. - momentum) + momentum * variance + # run backward + x_grad, scale_grad, bias_grad = _reference_grad( + x, y_grad, scale, saved_mean, saved_variance, epsilon, data_layout) + + return y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad + + +class TestMKLDNNBatchNormOpInference(TestBatchNormOpInference): + def init_kernel_type(self): + self.use_mkldnn = True + + def test_check_output(self): + place = core.CPUPlace() + data_format = "NCHW" + + self.check_with_place(place, data_format, self.dtype, [2, 3, 4, 5]) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_op.py b/python/paddle/fluid/tests/unittests/test_batch_norm_op.py index a0e78a460703778b46191b50c75e92bfbcaec411..4216d83653b27ec7f18034e576fbedbecc3f1cfe 100644 --- a/python/paddle/fluid/tests/unittests/test_batch_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_batch_norm_op.py @@ -158,6 +158,8 @@ def set_output_grad(scope, outputs, place, feed_dict=None): class TestBatchNormOpInference(unittest.TestCase): def setUp(self): self.dtype = np.float32 + self.use_mkldnn = False + self.init_kernel_type() def __assert_close(self, tensor, np_array, msg, atol=1e-4): self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg) @@ -230,6 +232,7 @@ class TestBatchNormOpInference(unittest.TestCase): # attrs is_test=True, data_layout=data_layout, + use_mkldnn=self.use_mkldnn, epsilon=epsilon) batch_norm_op.run(scope, place) @@ -254,10 +257,15 @@ class TestBatchNormOpInference(unittest.TestCase): [2, 3, 4, 5]) self.check_with_place(place, data_format, self.dtype, [2, 3]) + def init_kernel_type(self): + pass + class TestFP16BatchNormOpInference(TestBatchNormOpInference): def setUp(self): self.dtype = np.float16 + self.use_mkldnn = False + self.init_kernel_type() def test_check_output(self): places = [] @@ -274,9 +282,28 @@ class TestFP16BatchNormOpInference(TestBatchNormOpInference): class TestBatchNormOpTraining(unittest.TestCase): + def setUp(self): + self.use_mkldnn = False + self.data_formats = ["NCHW", "NHWC"] + self.init_kernel_type() + def __assert_close(self, tensor, np_array, msg, atol=1e-4): np.allclose(np.array(tensor), np_array, atol=atol) + def ref_forward_backward(self, x, y_grad, scale, bias, mean, variance, + epsilon, momentum, shape, data_layout): + # run forward + y, saved_mean, var_ref = _reference_training(x, scale, bias, epsilon, + data_layout) + mean_out = saved_mean * (1. - momentum) + momentum * mean + variance_out = var_ref * (1. - momentum) + momentum * variance + saved_variance = 1. / np.sqrt(var_ref + epsilon) + # run backward + x_grad, scale_grad, bias_grad = _reference_grad( + x, y_grad, scale, saved_mean, var_ref, epsilon, data_layout) + + return y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad + def test_forward_backward(self): def test_with_place(place, data_layout, shape): # attr @@ -295,16 +322,11 @@ class TestBatchNormOpTraining(unittest.TestCase): mean = np.zeros(scale_shape).astype(np.float32) variance = np.ones(scale_shape).astype(np.float32) - # run forward - y, saved_mean, var_ref = _reference_training(x, scale, bias, - epsilon, data_layout) - mean_out = saved_mean * (1. - momentum) + momentum * mean - variance_out = var_ref * (1. - momentum) + momentum * variance - saved_variance = 1. / np.sqrt(var_ref + epsilon) - # run backward y_grad = np.random.random_sample(shape).astype(np.float32) - x_grad, scale_grad, bias_grad = _reference_grad( - x, y_grad, scale, saved_mean, var_ref, epsilon, data_layout) + + y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad = self.ref_forward_backward( + x, y_grad, scale, bias, mean, variance, epsilon, momentum, + shape, data_layout) var_dict = locals() var_dict['y@GRAD'] = y_grad @@ -344,7 +366,8 @@ class TestBatchNormOpTraining(unittest.TestCase): "momentum": momentum, "epsilon": epsilon, "is_test": False, - "data_layout": data_layout + "data_layout": data_layout, + "use_mkldnn": self.use_mkldnn }) block.create_var(name='y@GRAD', dtype='float32', shape=y.shape) @@ -387,13 +410,17 @@ class TestBatchNormOpTraining(unittest.TestCase): print "op test forward passed: ", str(place), data_layout places = [core.CPUPlace()] + if core.is_compiled_with_cuda() and core.op_support_gpu("batch_norm"): places.append(core.CUDAPlace(0)) for place in places: - for data_format in ["NCHW", "NHWC"]: + for data_format in self.data_formats: test_with_place(place, data_format, [2, 3, 4, 5]) + def init_kernel_type(self): + pass + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py b/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py index 4ee00605e22ba45d9e46a8bba27712c3fd97872a..7976dd7c3f14390fb00bc8ab39121b6a686e3039 100644 --- a/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py +++ b/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py @@ -22,12 +22,12 @@ from paddle.fluid.op import Operator class TestBeamSearchDecodeOp(unittest.TestCase): def setUp(self): self.scope = core.Scope() - self.cpu_place = core.CPUPlace() + self.place = core.CPUPlace() def append_lod_tensor(self, tensor_array, lod, data): lod_tensor = core.LoDTensor() lod_tensor.set_lod(lod) - lod_tensor.set(data, self.cpu_place) + lod_tensor.set(data, self.place) tensor_array.append(lod_tensor) def test_get_set(self): @@ -71,7 +71,7 @@ class TestBeamSearchDecodeOp(unittest.TestCase): SentenceIds="sentence_ids", SentenceScores="sentence_scores") - beam_search_decode_op.run(self.scope, self.cpu_place) + beam_search_decode_op.run(self.scope, self.place) expected_lod = [[0, 4, 8], [0, 1, 3, 6, 9, 10, 13, 16, 19]] self.assertEqual(sentence_ids.lod(), expected_lod) @@ -84,5 +84,11 @@ class TestBeamSearchDecodeOp(unittest.TestCase): np.array_equal(np.array(sentence_scores), expected_data)) +class TestBeamSearchDecodeOpGPU(TestBeamSearchDecodeOp): + def setUp(self): + self.scope = core.Scope() + self.place = core.CUDAPlace(0) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_train.py b/python/paddle/fluid/tests/unittests/test_dist_train.py index c7fdd06f105e3b5fd906d3524d41df8f84160e63..77e9a8f7e72a9e0790ce1d1f48356abcca8eaccf 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_train.py +++ b/python/paddle/fluid/tests/unittests/test_dist_train.py @@ -34,7 +34,7 @@ class TestSendOp(unittest.TestCase): p.start() time.sleep(10) - with open("/tmp/paddle.selected_port", "r") as fn: + with open("/tmp/paddle.%d.selected_port" % p.pid, "r") as fn: selected_port = int(fn.readlines()[0]) self.init_client(place, selected_port) diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py index aeda67650205a473486b05fdd5b4364753fb18ca..0aada3deb0f047d21701b64af022ebad372d505b 100644 --- a/python/paddle/fluid/trainer.py +++ b/python/paddle/fluid/trainer.py @@ -12,44 +12,188 @@ # See the License for the specific language governing permissions and # limitations under the License. +import core +import framework +import executor +import data_feeder +import contextlib + +# optimizer is same as the parameter of Trainer.__init__. Rename it to opt_module +import optimizer as opt_module + __all__ = [ - 'Event', 'Trainer', + 'BeginEpochEvent', + 'EndEpochEvent', + 'BeginStepEvent', + 'EndStepEvent', ] -class Event(object): - BEGIN_EPOCH = 0 - END_EPOCH = 1 - BEGIN_STEP = 2 - END_STEP = 3 +class BeginEpochEvent(object): + def __init__(self, epoch_id): + self.epoch = epoch_id + + +class EndEpochEvent(object): + def __init__(self, epoch_id): + self.epoch = epoch_id - def __init__(self): - self.step = 0 - self.epoch = 0 - self.type = Event.BEGIN_EPOCH + +class BeginStepEvent(object): + def __init__(self, epoch_id, step_id): + self.epoch = epoch_id + self.step = step_id + + +class EndStepEvent(object): + def __init__(self, epoch_id, step_id): + self.epoch = epoch_id + self.step = step_id class Trainer(object): - def __init__(self, network_func, optimizer, params=None, place=None): + """ + + Args: + program_func(callable): A function which will return loss. The loss must be a scaler. + optimizer(optimizer.Optimizer): The optimizer should be an instance of Optimizer + place: The device place of this trainer. + """ + + def __init__(self, program_func, optimizer, param_path=None, place=None): # 1. we need to generate a framework.Program by calling - # network_func. Reference: fluid.program_guard in + # program_func. Reference: fluid.program_guard in # test_word2vec.py + self.scope = core.Scope() + + self.startup_program = framework.Program() + self.train_program = framework.Program() + + with framework.program_guard(self.train_program, self.startup_program): + loss = program_func() + if not isinstance(optimizer, opt_module.Optimizer): + raise TypeError( + "The optimizer should be an instance of Optimizer") + + optimizer.minimize(loss) + + self.place = Trainer._check_and_get_place(place) # 2. move the default_main_program to self.program and run the # default_startup program on an empty core.Scope() + # Run startup program + exe = executor.Executor(place) + exe.run(self.startup_program, scope=self.scope) + + if param_path: + # load params from param_path into scope + # TODO(yuyang): This depends on parameters implementation. + pass - # 3. call self.params.add_vars with the initialized scope, it - # will add the new vars of the initialized scope into - # self.params. - self.network_func = network_func - self.optimizer = optimizer - self.params = params - self.place = place # TODO(helin): support distributed training - def train(self, reader, num_epochs, event_handler): - pass + def train(self, + num_epochs, + event_handler, + reader=None, + parallel=False, + feed_order=None): + """ + Train the model. + + Args: + num_epochs: The number of epoch. An epoch will process all data in reader + event_handler: The event handler. A function with type (ev:Event)->void + reader: + parallel: True if use multi-CPUs or multi-GPUs + feed_order: Feeding order of reader. None will following the defining + order in program + + Returns: + + """ + if parallel: + raise NotImplementedError( + "Parallel Executor version of trainer is not implemented") + + self._train_by_executor(num_epochs, event_handler, reader, feed_order) def test(self, reader): pass + + def save_params(self, param_path): + # reference: save_persistables in io.py + pass + + @staticmethod + def _check_and_get_place(place): + """ + Check the type of place or get the default place + Args: + place(None|core.CUDAPlace|core.CPUPlace): the place that trainer will be executed on. + + Raises: + TypeError if the type mismatched. + + Returns: + the original place if it is not None. + if fluid is compiled with CUDA, returns CUDAPlace(0) by default. + Otherwise returns CPUPlace by default. + """ + if place is None: + if core.is_compiled_with_cuda(): + return core.CUDAPlace(0) + else: + return core.CPUPlace() + else: + if not isinstance(place, core.CUDAPlace) and not isinstance( + place, core.CPUPlace): + raise TypeError("Place should be either CUDAPlace or CPUPlace") + return place + + @contextlib.contextmanager + def _prog_and_scope_guard(self): + with framework.program_guard( + main_program=self.train_program, + startup_program=self.startup_program): + with executor.scope_guard(self.scope): + yield + + def _train_by_executor(self, num_epochs, event_handler, reader, feed_order): + """ + Train by Executor and single device. + + Args: + num_epochs: + event_handler: + reader: + feed_order: + + Returns: + + """ + with self._prog_and_scope_guard(): + exe = executor.Executor(self.place) + if feed_order is None: + feed_var_list = [ + var + for var in self.train_program.global_block( + ).vars.itervalues() + if hasattr(var, 'is_data') and var.is_data + ] + else: + feed_var_list = [ + self.train_program.global_block().var(var_name) + for var_name in feed_order + ] + + feeder = data_feeder.DataFeeder( + feed_list=feed_var_list, place=self.place) + for epoch_id in range(num_epochs): + event_handler(BeginEpochEvent(epoch_id)) + for step_id, data in enumerate(reader()): + event_handler(BeginStepEvent(epoch_id, step_id)) + exe.run(feed=feeder.feed(data), fetch_list=[]) + event_handler(EndStepEvent(epoch_id, step_id)) + event_handler(EndEpochEvent(epoch_id))