Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into gen_nccl_id_op

.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;

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!
 

benchmark/cluster/vgg16/run_vgg_dist.sh

+#!/bin/bash
+
+# Update to point to the source file.
+VGG_SRC="vgg16_fluid.py"
+
+export TRAINING_ROLE=PSERVER
+export TRAINERS=2
+export POD_IP=127.0.0.1
+export PADDLE_INIT_PORT=6174
+MKL_NUM_THREADS=1 python -u ${VGG_SRC} --local 0 --ps_host=127.0.0.1:6174 --trainer_hosts=127.0.0.1:6174 &
+
+# Need to wait for the ps to start first.
+sleep 10
+echo "done start ps"
+
+export TRAINING_ROLE=TRAINER
+export TRAINERS=2
+export POD_IP=127.0.0.1
+export PADDLE_INIT_PORT=6174
+CUDA_VISIBLE_DEVICES=4 MKL_NUM_THREADS=1 python -u ${VGG_SRC} --local 0 --ps_host=127.0.0.1:6174 --trainer_hosts=127.0.0.1:6174 --device=GPU --task_index=0 &
+CUDA_VISIBLE_DEVICES=5 MKL_NUM_THREADS=1 python -u ${VGG_SRC} --local 0 --ps_host=127.0.0.1:6174 --trainer_hosts=127.0.0.1:6174 --device=GPU --task_index=1 &

benchmark/cluster/vgg16/vgg16_fluid.py

@@ -200,18 +200,19 @@ def main():
                 num_samples += len(data)
                 train_pass_acc.add(value=acc, weight=b_size)
                 print(
-                    "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed = %.2f img/s"
-                    % (pass_id, iters, loss, acc,
-                       len(data) / (time.time() - ts))
+                    "Task:%d Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, "
+                    "Speed = %.2f img/s " % (args.task_index, pass_id, iters,
+                                             loss, acc,
+                                             len(data) / (time.time() - ts))
                 )  # The accuracy is the accumulation of batches, but not the current batch.
 
             pass_elapsed = time.time() - start_time
             pass_train_acc = train_pass_acc.eval()
             pass_test_acc = test(exe)
-            print(
-                "Pass = %d, Training performance = %f imgs/s, Train accuracy = %f, Test accuracy = %f\n"
-                % (pass_id, num_samples / pass_elapsed, pass_train_acc,
-                   pass_test_acc))
+            print("Task:%d Pass = %d, Training performance = %f imgs/s, "
+                  "Train accuracy = %f, Test accuracy = %f\n" %
+                  (args.task_index, pass_id, num_samples / pass_elapsed,
+                   pass_train_acc, pass_test_acc))
 
     if args.local:
         # Parameter initialization
@@ -239,8 +240,6 @@ def main():
 
         t = fluid.DistributeTranspiler()
         t.transpile(
-            optimize_ops,
-            params_grads,
             trainer_id=args.task_index,
             pservers=args.ps_hosts,
             trainers=trainers)

contrib/float16/.gitignore

+*.inference.model

contrib/float16/float16_benchmark.md

+# 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 |

contrib/float16/float16_inference_demo.py

+#   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)

contrib/float16/float16_inference_report.md

+## 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.

python/paddle/fluid/inference_transpiler.py → contrib/float16/float16_transpiler.py

@@ -13,115 +13,13 @@
 # limitations under the License.
 
 import numpy as np
-from framework import Program
-from executor import global_scope
-from . import core
+import paddle.fluid.core as core
+from paddle.fluid.framework import Program
+from paddle.fluid.executor import global_scope
 
 
-class InferenceTranspiler:
+class Float16Transpiler:
     def transpile(self, program, place, scope=None):
-        '''
-        Transpile the program. Support only fuse batch normalization now.
-
-        :param program: program to transpile 
-        :type program: Program
-        :param place: inference place 
-        :type place: Place
-        :param scope: inference scope 
-        :type scope: Scope or None
-        '''
-        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.fuse_batch_norm(program, place, scope)
-
-    def fuse_batch_norm(self, program, place, scope):
-        '''
-        Transpile the program by fused batch normalization.
- 
-        The batch normalization followed the convolution or fully connected layer 
-        can be integrated with them. Doing so will give us a forward acceleration, 
-        especially in environments like mobile or embedded.
-                    
-        For input X:
-        - Conv process:        X = input * W + bias 
-        - Batch norm process:  X' = (X - mean) / std 
-        - Scale Process:       Y = a * X' + b
-
-        After fuse into one operation:
-
-        Y = (input * W + bias - mean) / std * a + b
-          = input * a * W / std + ((bias - mean) / std * a + b)
-
-        The operator transformation is: 
-        - before:
-          - conv->batch_norm->any_other_op (bias == 0)
-          - conv->elementwise_add->batch_norm->any_other_op (bias != 0)
-        - after: 
-          - conv->elementwise_add->any_other_op
-        
-        The transpile stages are:
-        1. insert elementwise_add op when bias == 0.
-        2. fuse the batch_norm's parameters to conv and elementwise_add operators.
-        3. remove batch_norm ops which are not used in any other ops.
-        4. adjust the input of any_other_op to be the output of elementwise_add operator.
-        5. remove unused variables.
-
-        :param program: program to transpile 
-        :type program: Program
-        :param place: inference place 
-        :type place: Place
-        :param scope: inference scope 
-        :type scope: Scope
-        '''
-        self.scope = scope
-        self.place = place
-        self.block = program.block(0)
-        self.input_map = {}  # store the input names should be adjusted 
-
-        i = 0
-        while i < len(self.block.ops):
-            current_op = self.block.ops[i]
-            # TODO(luotao1): consider only conv2d now. fc would be delt later.
-            if current_op.type in ['conv2d']:
-                # TODO(luotao1): consider single chain network now. 
-                # For branch network, we counldn't use block.ops[i + 1] as 
-                # the judgment condition.
-                next_op = self.block.ops[i + 1]
-                # conv2d without bias
-                if (next_op.type == 'batch_norm'):
-                    # insert bias op
-                    bias_op = self._insert_bias_op(i + 1, current_op, next_op)
-                    # fuse batch_norm
-                    self._fuse_param(current_op, next_op, bias_op, 0)
-                    # remove batch_norm_op
-                    self.block.remove_op(i + 2)
-                    i = i + 1
-                # conv2d with bias, the next_op.type is elementwise_add
-                elif (next_op.type == 'elementwise_add'):
-                    next_next_op = self.block.ops[i + 2]
-                    if (next_next_op.type == 'batch_norm'):
-                        # fuse batch_norm
-                        self._fuse_param(current_op, next_next_op, next_op, 1)
-                        # remove batch_norm_op
-                        self.block.remove_op(i + 2)
-                        i = i + 1
-            i = i + 1
-
-        self._adjust_input()
-        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()
-
-    def float16_transpile(self, program, place, scope=None):
         '''
         Transpile the program desc and cast the weights to float16 data type to
         enable float16 inference.
@@ -155,8 +53,15 @@ class InferenceTranspiler:
         :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
@@ -174,101 +79,6 @@ class InferenceTranspiler:
         program = program.clone()
 
     # ====================== private transpiler functions =====================
-
-    def _insert_bias_op(self, index, current_op, bn_op):
-        '''
-        Construct elementwise_add operator for adding bias 
-        and insert it into program.
-        
-        :param index: insert location of bias_op
-        :type index: Int
-        :param current_op: current operator (conv or fc)
-        :type current_op: Operator
-        :param bn_op: batch norm operator
-        :type bn_op: Operator
-        :return: bias_op
-        :rtype: Operator
-        '''
-        # The input of bias_op is current_op's output and Bias of bn_op
-        # The output of bias_op is bn_op's output
-        x_var = self.block.var(current_op.output("Output")[0])
-        y_var = self.block.var(bn_op.input("Bias")[0])
-        out_var = self.block.var(bn_op.output("Y")[0])
-
-        bias_op = self.block.insert_op(
-            index,
-            type="elementwise_add",
-            inputs={"X": x_var,
-                    "Y": y_var},
-            outputs={"Out": out_var},
-            attrs={"axis": 1})  # dim_start=1
-        return bias_op
-
-    def _fuse_param(self, current_op, bn_op, bias_op, with_bias):
-        '''
-        fuse the batch_norm_op' parameters to current_op (conv or fc)
-        
-        :param current_op: current operator (conv or fc)
-        :type current_op: Operator
-        :param bn_op: batch norm operator
-        :type bn_op: Operator
-        :param bias_op: elementwise_add operator for adding bias
-        :type bias_op: Operator
-        :param with_bias: If current operator has bias, with_bias = 1; otherwise 0. 
-        :type with_bias: Int
-        '''
-
-        def _update_param(op, old_param_name, new_param):
-            # For the sake of remaining the original variables the same as before,
-            # create new variables in scope to store the new parameters.
-            old_param_name = old_param_name[0]
-            old_var = self.block.vars[old_param_name]
-            new_param_name = old_param_name + '_fuse_bn'
-            new_var = self.block.create_parameter(
-                name=new_param_name.encode('ascii'),
-                type=old_var.type,
-                dtype=old_var.dtype,
-                shape=old_var.shape)
-            op.rename_input(old_param_name, new_param_name)
-            self.scope.var(new_param_name)
-
-            tensor = self.scope.find_var(new_param_name).get_tensor()
-            tensor.set(np.array(new_param), self.place)
-
-        def _load_param(param_name):
-            return np.array(self.scope.find_var(param_name[0]).get_tensor())
-
-        bias_bn = _load_param(bn_op.input("Bias"))  #Bias
-        scale_bn = _load_param(bn_op.input("Scale"))  #Scale
-        mean_bn = _load_param(bn_op.input("Mean"))  #Mean
-        var_bn = _load_param(bn_op.input("Variance"))  #Variance
-
-        # TODO(luotao1): consider only conv2d now. fc would be delt later.
-        current_param = _load_param(current_op.input("Filter"))
-        std_bn = np.float32(np.sqrt(np.add(var_bn, 1e-5)))
-        tmp = np.float32(np.divide(scale_bn, std_bn))
-
-        # add bias of batch_norm_op to conv2d
-        if with_bias:
-            bias = _load_param(bias_op.input("Y"))
-        else:
-            bias = np.zeros(bias_bn.shape)
-        bias = np.float32(
-            np.add(np.multiply(np.subtract(bias, mean_bn), tmp), bias_bn))
-
-        # re-compute weight of conv2d
-        tmp = tmp.reshape(tmp.shape[0], -1)
-        dst_param = current_param.reshape((tmp.shape[0], -1))
-        dst_param = np.float32(np.multiply(dst_param, tmp))
-        dst_param = dst_param.reshape(current_param.shape)
-
-        # update parameters
-        _update_param(current_op, current_op.input("Filter"), dst_param)
-        _update_param(bias_op, bias_op.input("Y"), bias)
-
-        # collect the renamed input
-        self.input_map[bn_op.output("Y")[0]] = bias_op.output("Out")[0]
-
     def _adjust_input(self, skip=False):
         '''
         Change the input variable name in operators.

contrib/float16/run_float16_demo.sh

+#!/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

doc/fluid/design/dist_train/distributed_traing_review.md

+# 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

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 <https://www.open-mpi.org>`_ 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 <http://www.fabfile.org>`_ 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 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/scripts/cluster_train_v2>`_ .
\ No newline at end of file
+The examples can be found under `cluster_train_v2 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/scripts/cluster_train_v2>`_ .

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();
   }

paddle/cuda/include/hl_base.h

@@ -12,8 +12,7 @@ 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. */
 
-#ifndef HL_BASE_H_
-#define HL_BASE_H_
+#pragma once
 
 #include <cstddef>
 
@@ -207,8 +206,8 @@ typedef struct {
 
 #ifdef __NVCC__
 
-#include "cuda_runtime.h"
-#include "hl_cuda.h"
+#include <cuda_runtime.h>
+#include "paddle/cuda/include/hl_cuda.h"
 #include "paddle/utils/Logging.h"
 
 extern __thread bool g_sync_flag;
@@ -230,6 +229,11 @@ extern __thread cudaStream_t default_stream;
 
 // __shfl has been deprecated as of CUDA 9.0.
 #if CUDA_VERSION < 9000
+template <typename T>
+__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) {
+  return __shfl_down(val, delta);
+}
+
 template <typename T>
 __forceinline__ __device__ T
 __shfl_sync(unsigned, T val, int src_line, int width) {
@@ -243,6 +247,4 @@ __shfl_sync(unsigned, T val, int src_line, int width) {
   mask = __ballot_sync(FULL_WARP_MASK, (predicate))
 #endif
 
-#endif /* __NVCC__ */
-
-#endif /* HL_BASE_H_ */
+#endif  // __NVCC__

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) {

paddle/fluid/framework/block_desc.cc

@@ -143,7 +143,7 @@ OpDesc *BlockDesc::InsertOp(size_t index) {
 }
 
 void BlockDesc::RemoveOp(size_t s, size_t e) {
-  if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) {
+  if (ops_.begin() + s >= ops_.end() || ops_.begin() + e > ops_.end()) {
     return;
   }
   need_update_ = true;

paddle/fluid/framework/details/CMakeLists.txt

@@ -15,12 +15,14 @@ if(WITH_GPU)
             dynload_cuda)
     set(multi_devices_graph_builder_deps nccl_all_reduce_op_handle)
     nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda)
+    nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda)
+
 else()
     set(multi_devices_graph_builder_deps)
     cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim)
+    cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
 endif()
 
-cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
 cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
 
 cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -19,14 +19,12 @@
 namespace paddle {
 namespace framework {
 namespace details {
-BroadcastOpHandle::BroadcastOpHandle(const std::vector<Scope *> &local_scopes,
-                                     const std::vector<platform::Place> &places)
-    : local_scopes_(local_scopes), places_(places) {}
 
 void BroadcastOpHandle::RunImpl() {
-  // the input and output may have dummy var.
-  VarHandle *in_var_handle;
+  if (places_.size() == 1) return;
 
+  // The input and output may have dummy vars.
+  VarHandle *in_var_handle;
   {
     auto in_var_handles = DynamicCast<VarHandle>(inputs_);
     PADDLE_ENFORCE_EQ(in_var_handles.size(), 1,
@@ -55,27 +53,97 @@ void BroadcastOpHandle::RunImpl() {
 
   Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var);
 
-  for (auto *out : out_var_handles) {
-    if (*out == *in_var_handle) {
+  // NOTE: The tensors' Place of input and output must be all on GPU or all on
+  // CPU.
+  for (auto *out_var_handle : out_var_handles) {
+    if (out_var_handle->IsTheSameVar(*in_var_handle)) {
       continue;
     }
-
-    auto &out_p = out->place_;
-    auto *out_var = var_scopes.at(out->scope_idx_)->FindVar(out->name_);
+    auto t_out_p = out_var_handle->place_;
+    auto *out_var = var_scopes.at(out_var_handle->scope_idx_)
+                        ->FindVar(out_var_handle->name_);
     PADDLE_ENFORCE_NOT_NULL(out_var);
-    PADDLE_ENFORCE_EQ(out_p.which(), in_var_handle->place_.which(),
-                      "Places must be all on CPU or all on CUDA.");
-
+    if (platform::is_gpu_place(in_tensor.place())) {
+      PADDLE_ENFORCE(platform::is_gpu_place(t_out_p),
+                     "Places of input and output must be all on GPU.");
+    } else {
+      t_out_p = platform::CPUPlace();
+    }
     VariableVisitor::ShareDimsAndLoD(*in_var, out_var);
-    VariableVisitor::GetMutableTensor(out_var).mutable_data(out_p,
+    VariableVisitor::GetMutableTensor(out_var).mutable_data(t_out_p,
                                                             in_tensor.type());
+  }
+
+  if (platform::is_cpu_place(in_tensor.place())) {
+    for (auto *out_var_handle : out_var_handles) {
+      if (out_var_handle->IsTheSameVar(*in_var_handle)) {
+        continue;
+      }
+      auto &out_p = out_var_handle->place_;
+      auto *out_var = var_scopes.at(out_var_handle->scope_idx_)
+                          ->FindVar(out_var_handle->name_);
+
+      RunAndRecordEvent(out_p, [in_tensor, out_var] {
+        paddle::framework::TensorCopy(
+            in_tensor, platform::CPUPlace(),
+            &VariableVisitor::GetMutableTensor(out_var));
+      });
+    }
+  } else {
+#ifdef PADDLE_WITH_CUDA
+    VarHandle *out_handle = nullptr;
+    int root_id = boost::get<platform::CUDAPlace>(in_tensor.place()).device;
+    std::vector<std::function<void()>> broadcast_calls;
+
+    for (auto out_var_handle : out_var_handles) {
+      Variable *out_var = var_scopes.at(out_var_handle->scope_idx_)
+                              ->FindVar(out_var_handle->name_);
+
+      int dst_id =
+          boost::get<platform::CUDAPlace>(out_var_handle->place_).device;
+
+      auto &nccl_ctx = nccl_ctxs_->at(dst_id);
+
+      void *send_recv_buffer = nullptr;
+      if (root_id == dst_id) {
+        send_recv_buffer = const_cast<void *>(in_tensor.data<void>());
+        out_handle = out_var_handle;
+      } else {
+        send_recv_buffer =
+            VariableVisitor::GetMutableTensor(out_var).mutable_data(
+                out_var_handle->place_);
+      }
+
+      int type = platform::ToNCCLDataType(in_tensor.type());
+      size_t numel = static_cast<size_t>(in_tensor.numel());
+      broadcast_calls.emplace_back(
+          [send_recv_buffer, numel, type, root_id, &nccl_ctx] {
+            PADDLE_ENFORCE(platform::dynload::ncclBcast(
+                send_recv_buffer, numel, static_cast<ncclDataType_t>(type),
+                root_id, nccl_ctx.comm_, nccl_ctx.stream()));
+          });
+    }
 
-    auto dev_ctx = dev_ctxes_.at(out_p);
-    RunAndRecordEvent(out_p, [in_tensor, out_var, dev_ctx, out_p] {
-      paddle::framework::TensorCopy(
-          in_tensor, out_p, *(dev_ctx),
-          &VariableVisitor::GetMutableTensor(out_var));
+    this->RunAndRecordEvent([&] {
+      {
+        platform::NCCLGroupGuard guard;
+        for (auto &call : broadcast_calls) {
+          call();
+        }
+      }
+
+      if (!out_handle->IsTheSameVar(*in_var_handle)) {
+        auto out_var = var_scopes.at(in_var_handle->scope_idx_)
+                           ->FindVar(out_var_handles[0]->name_);
+        paddle::framework::TensorCopy(
+            in_tensor, in_var_handle->place_,
+            *(dev_ctxes_.at(in_var_handle->place_)),
+            &VariableVisitor::GetMutableTensor(out_var));
+      }
     });
+#else
+    PADDLE_THROW("CUDA is not enabled.");
+#endif
   }
 }
 

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -24,14 +24,32 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/nccl_helper.h"
+#endif
+
 namespace paddle {
 namespace framework {
 namespace details {
 
 struct BroadcastOpHandle : public OpHandleBase {
  public:
+#ifdef PADDLE_WITH_CUDA
+  BroadcastOpHandle(const std::vector<Scope *> &local_scopes,
+                    const std::vector<platform::Place> &places,
+                    const platform::NCCLContextMap *nccl_ctxs)
+      : local_scopes_(local_scopes), places_(places), nccl_ctxs_(nccl_ctxs) {
+    if (nccl_ctxs_) {
+      for (auto &p_ctx : nccl_ctxs_->contexts_) {
+        dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get();
+      }
+    }
+  }
+#else
   BroadcastOpHandle(const std::vector<Scope *> &local_scopes,
-                    const std::vector<platform::Place> &places);
+                    const std::vector<platform::Place> &places)
+      : local_scopes_(local_scopes), places_(places) {}
+#endif
 
   std::string Name() const override;
 
@@ -44,6 +62,9 @@ struct BroadcastOpHandle : public OpHandleBase {
  private:
   const std::vector<Scope *> &local_scopes_;
   const std::vector<platform::Place> &places_;
+#ifdef PADDLE_WITH_CUDA
+  const platform::NCCLContextMap *nccl_ctxs_;
+#endif
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/broadcast_op_handle_test.cc

@@ -35,15 +35,25 @@ struct TestBroadcastOpHandle {
   std::unique_ptr<OpHandleBase> op_handle_;
   std::vector<std::unique_ptr<VarHandleBase>> vars_;
   std::vector<p::Place> gpu_list_;
+  bool use_gpu_;
+#ifdef PADDLE_WITH_CUDA
+  std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
+#endif
 
   void WaitAll() {
     for (size_t j = 0; j < ctxs_.size(); ++j) {
       ctxs_[j]->Wait();
     }
+#ifdef PADDLE_WITH_CUDA
+    if (nccl_ctxs_) {
+      nccl_ctxs_->WaitAll();
+    }
+#endif
   }
 
   void InitCtxOnGpu(bool use_gpu) {
-    if (use_gpu) {
+    use_gpu_ = use_gpu;
+    if (use_gpu_) {
 #ifdef PADDLE_WITH_CUDA
       int count = p::GetCUDADeviceCount();
       if (count <= 1) {
@@ -57,6 +67,7 @@ struct TestBroadcastOpHandle {
         gpu_list_.push_back(p);
         ctxs_.emplace_back(new p::CUDADeviceContext(p));
       }
+      nccl_ctxs_.reset(new platform::NCCLContextMap(gpu_list_));
 #else
       PADDLE_THROW("CUDA is not support.");
 #endif
@@ -67,6 +78,9 @@ struct TestBroadcastOpHandle {
         gpu_list_.push_back(p);
         ctxs_.emplace_back(new p::CPUDeviceContext(p));
       }
+#ifdef PADDLE_WITH_CUDA
+      nccl_ctxs_.reset(nullptr);
+#endif
     }
   }
 
@@ -82,7 +96,21 @@ struct TestBroadcastOpHandle {
     }
     param_scopes_[input_scope_idx]->Var("input");
 
-    op_handle_.reset(new BroadcastOpHandle(local_scopes_, gpu_list_));
+    if (use_gpu_) {
+#ifdef PADDLE_WITH_CUDA
+      op_handle_.reset(
+          new BroadcastOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get()));
+#else
+      PADDLE_THROW("CUDA is not support.");
+#endif
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      op_handle_.reset(
+          new BroadcastOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get()));
+#else
+      op_handle_.reset(new BroadcastOpHandle(local_scopes_, gpu_list_));
+#endif
+    }
 
     auto* in_var_handle =
         new VarHandle(1, input_scope_idx, "input", gpu_list_[input_scope_idx]);
@@ -97,7 +125,9 @@ struct TestBroadcastOpHandle {
     op_handle_->AddInput(dummy_var_handle);
 
     for (size_t j = 0; j < gpu_list_.size(); ++j) {
-      op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get());
+      if (!use_gpu_) {
+        op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get());
+      }
       VarHandle* out_var_handle = new VarHandle(2, j, "out", gpu_list_[j]);
       vars_.emplace_back(out_var_handle);
       op_handle_->AddOutput(out_var_handle);

paddle/fluid/framework/details/gather_op_handle.cc

@@ -25,6 +25,7 @@ GatherOpHandle::GatherOpHandle(const std::vector<Scope *> &local_scopes,
     : local_scopes_(local_scopes), places_(places) {}
 
 void GatherOpHandle::RunImpl() {
+  if (places_.size() == 1) return;
   // the input and output may have dummy var.
   auto in_var_handles = DynamicCast<VarHandle>(inputs_);
 
@@ -35,7 +36,6 @@ void GatherOpHandle::RunImpl() {
   VarHandle *out_var_handle;
   {
     auto out_var_handles = DynamicCast<VarHandle>(outputs_);
-
     PADDLE_ENFORCE_EQ(out_var_handles.size(), 1,
                       "The number of output should be one.");
     out_var_handle = out_var_handles.front();
@@ -50,68 +50,62 @@ void GatherOpHandle::RunImpl() {
   auto pre_in_var =
       var_scopes.at(in_0_handle->scope_idx_)->FindVar(in_0_handle->name_);
   PADDLE_ENFORCE_NOT_NULL(pre_in_var);
+
   PADDLE_ENFORCE(pre_in_var->IsType<framework::SelectedRows>(),
                  "Currently, gather_op only can gather SelectedRows.");
 
-  auto pre_place = in_0_handle->place_;
-  PADDLE_ENFORCE_EQ(out_var_handle->place_.which(), pre_place.which(),
-                    "The place of input and output should be the same.");
-
   // Wait input done, this Wait is asynchronous operation
   WaitInputVarGenerated(in_var_handles);
 
+  auto &pre_in_value = pre_in_var->Get<framework::SelectedRows>();
   std::vector<int64_t> out_rows;
   std::vector<Tensor> in_tensors;
-  std::vector<platform::Place> in_places;
 
-  auto &pre_in = pre_in_var->Get<framework::SelectedRows>();
-  // gather the inputs
+  // Gather the inputs
   for (auto *in_handle : in_var_handles) {
-    auto in_p = in_handle->place_;
-    in_places.push_back(in_p);
-    PADDLE_ENFORCE_EQ(in_p.which(), pre_place.which(),
-                      "Places must be all on CPU or all on CUDA.");
     auto *in_var =
         var_scopes.at(in_handle->scope_idx_)->FindVar(in_handle->name_);
-    auto &in_sr = in_var->Get<framework::SelectedRows>();
+    PADDLE_ENFORCE_NOT_NULL(in_var);
+    VariableVisitor::EnforceShapeAndDTypeEQ(*in_var, *pre_in_var);
 
-    PADDLE_ENFORCE_EQ(in_sr.value().type(), pre_in.value().type(),
-                      "The type of input is not consistent.");
-    PADDLE_ENFORCE_EQ(pre_in.height(), in_sr.height(),
-                      "The height of inputs is not consistent.");
-    PADDLE_ENFORCE_EQ(pre_in.GetCompleteDims(), in_sr.GetCompleteDims(),
-                      "The dims of inputs is not consistent.");
+    auto &in_sr_value = in_var->Get<framework::SelectedRows>();
 
-    auto &in_sr_rows = in_sr.rows();
+    auto &in_sr_rows = in_sr_value.rows();
     out_rows.insert(out_rows.end(), in_sr_rows.begin(), in_sr_rows.end());
-
-    in_tensors.emplace_back(in_sr.value());
+    in_tensors.emplace_back(in_sr_value.value());
   }
 
-  // write the output
-  auto &out_place = out_var_handle->place_;
-  auto out_scope_idx = out_var_handle->scope_idx_;
-  auto out_var = var_scopes.at(out_scope_idx)->FindVar(out_var_handle->name_);
+  // NOTE: The Places of all input tensor must be all on CPU or all on GPU.
+  platform::Place t_out_p = out_var_handle->place_;
+  if (platform::is_gpu_place(pre_in_value.place())) {
+    PADDLE_ENFORCE(platform::is_gpu_place(t_out_p),
+                   "Places of input and output must be all on GPU.");
+  } else {
+    t_out_p = platform::CPUPlace();
+  }
 
-  auto out = out_var->GetMutable<framework::SelectedRows>();
-  out->set_height(pre_in.height());
-  out->set_rows(out_rows);
+  auto out_var =
+      var_scopes.at(out_var_handle->scope_idx_)->FindVar(out_var_handle->name_);
+  PADDLE_ENFORCE_NOT_NULL(out_var);
+  auto out_value = out_var->GetMutable<framework::SelectedRows>();
+  out_value->set_height(pre_in_value.height());
+  out_value->set_rows(out_rows);
   size_t rows = out_rows.size();
-  DDim out_dim = pre_in.GetCompleteDims();
+  DDim out_dim = pre_in_value.GetCompleteDims();
   out_dim[0] = static_cast<int64_t>(rows);
-  out->mutable_value()->Resize(out_dim);
-  out->mutable_value()->mutable_data(out_place, pre_in.value().type());
-  Tensor *out_tensor = out->mutable_value();
+  out_value->mutable_value()->Resize(out_dim).mutable_data(
+      t_out_p, pre_in_value.value().type());
+  Tensor *out_tensor = out_value->mutable_value();
 
   // copy
-  auto dev_ctx = dev_ctxes_[out_place];
-  RunAndRecordEvent(out_place, [in_tensors, out_tensor, dev_ctx, out_place] {
+  auto dev_ctx = dev_ctxes_[out_var_handle->place_];
+  RunAndRecordEvent(out_var_handle->place_, [in_tensors, out_tensor, &dev_ctx,
+                                             t_out_p] {
     int s = 0, e = 0;
     for (size_t j = 0; j < in_tensors.size(); ++j) {
       e += in_tensors[j].dims()[0];
       auto sub_out = out_tensor->Slice(s, e);
-      paddle::framework::TensorCopy(in_tensors[j], out_place, *(dev_ctx),
-                                    &sub_out);
+      paddle::framework::TensorCopy(in_tensors[j], t_out_p, *dev_ctx, &sub_out);
       s = e;
     }
   });

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -11,9 +11,11 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
-
 #include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
+#include <utility>
+#include "paddle/fluid/framework/details/broadcast_op_handle.h"
 #include "paddle/fluid/framework/details/computation_op_handle.h"
+#include "paddle/fluid/framework/details/reduce_op_handle.h"
 #include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
 #include "paddle/fluid/framework/details/send_op_handle.h"
 #include "paddle/fluid/framework/scope.h"
@@ -34,8 +36,8 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale,
-    platform::NCCLContextMap *nccl_ctxs)
+    const std::vector<Scope *> &local_scopes,
+    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes),
@@ -105,6 +107,11 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op,
 
 std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
     const ProgramDesc &program) const {
+  std::unordered_map<std::string, proto::VarType::Type> var_types;
+  for (auto *var : program.Block(0).AllVars()) {
+    var_types[var->Name()] = var->GetType();
+  }
+
   auto graph = new SSAGraph();
   SSAGraph &result = *graph;
   std::unordered_set<std::string> og_has_been_broadcast;
@@ -133,12 +140,17 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       is_forwarding = false;
     } else {
       CreateComputationalOps(&result, *op, places_.size());
-      if (!is_forwarding) {
+      if (!is_forwarding && places_.size() > 1) {
         // Currently, we assume that once gradient is generated, it can be
         // broadcast, and each gradient is only broadcast once.
         for (auto &og : op->OutputArgumentNames()) {
           if (IsParameterGradientOnce(og, &og_has_been_broadcast)) {
-            InsertNCCLAllReduceOp(&result, og);
+            if (IsSparseGradient(var_types, og)) {
+              CreateReduceOp(&result, og, 0);
+              CreateBroadcastOp(&result, og, 0);
+            } else {
+              InsertNCCLAllReduceOp(&result, og);
+            }
           }
         }
       }
@@ -165,6 +177,50 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   return std::unique_ptr<SSAGraph>(graph);
 }
 
+bool MultiDevSSAGraphBuilder::IsSparseGradient(
+    const std::unordered_map<std::string, proto::VarType::Type> &var_types,
+    const std::string &og) const {
+  PADDLE_ENFORCE(var_types.count(og) != 0);
+  if (var_types.at(og) == proto::VarType::SELECTED_ROWS) {
+    return true;
+  }
+  return false;
+}
+
+void MultiDevSSAGraphBuilder::CreateBroadcastOp(SSAGraph *result,
+                                                const std::string &p_name,
+                                                size_t src_dev_id) const {
+#ifdef PADDLE_WITH_CUDA
+  auto *op_handle = new BroadcastOpHandle(local_scopes_, places_, nccl_ctxs_);
+#else
+  auto *op_handle = new BroadcastOpHandle(local_scopes_, places_);
+#endif
+
+  result->ops_.emplace_back(op_handle);
+  auto *in = result->vars_.at(src_dev_id).at(p_name).back().get();
+  op_handle->AddInput(in);
+
+  for (size_t i = 0; i < places_.size(); ++i) {
+    auto &vars = result->vars_.at(i).at(p_name);
+    auto &p = places_[i];
+    auto *out_var = new VarHandle(vars.size(), i, p_name, p);
+    vars.emplace_back(out_var);
+    op_handle->AddOutput(out_var);
+#ifndef ADDLE_WITH_CUDA
+    op_handle->SetDeviceContext(p,
+                                platform::DeviceContextPool::Instance().Get(p));
+#endif
+  }
+}
+
+void MultiDevSSAGraphBuilder::CreateComputationalOp(SSAGraph *result,
+                                                    const OpDesc &op,
+                                                    int dev_id) const {
+  result->ops_.emplace_back(
+      new ComputationOpHandle(op, local_scopes_[dev_id], places_[dev_id]));
+  CreateOpHandleIOs(result, op, dev_id);
+}
+
 OpDesc *MultiDevSSAGraphBuilder::GetSendOpDesc(
     const ProgramDesc &program) const {
   for (auto *op : program.Block(0).AllOps()) {
@@ -174,7 +230,6 @@ OpDesc *MultiDevSSAGraphBuilder::GetSendOpDesc(
   }
   return nullptr;
 }
-
 void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp(
     SSAGraph *result, const std::string &og) const {
 #ifdef PADDLE_WITH_CUDA
@@ -247,6 +302,36 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(SSAGraph *result,
   }
 }
 
+VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result,
+                                                   const std::string &og,
+                                                   int dst_dev_id) const {
+#ifdef PADDLE_WITH_CUDA
+  result->ops_.emplace_back(
+      new ReduceOpHandle(local_scopes_, places_, nccl_ctxs_));
+#else
+  result->ops_.emplace_back(new ReduceOpHandle(local_scopes_, places_));
+#endif
+  auto *op_handle = result->ops_.back().get();
+
+  for (size_t i = 0; i < places_.size(); ++i) {
+    auto &vars = result->vars_[i][og];
+#ifndef PADDLE_WITH_CUDA
+    auto &p = places_[i];
+    op_handle->SetDeviceContext(p,
+                                platform::DeviceContextPool::Instance().Get(p));
+#endif
+    PADDLE_ENFORCE(!vars.empty());
+    auto &prev_grad = vars.back();
+    op_handle->AddInput(prev_grad.get());
+  }
+  auto &vars = result->vars_[dst_dev_id][og];
+  auto var =
+      new VarHandle(vars.size() - 1, dst_dev_id, og, places_[dst_dev_id]);
+  vars.emplace_back(var);
+  op_handle->AddOutput(var);
+  return var;
+}
+
 void MultiDevSSAGraphBuilder::CreateSendOp(SSAGraph *result,
                                            const OpDesc &op) const {
   auto &p = places_[0];
@@ -263,6 +348,7 @@ bool MultiDevSSAGraphBuilder::IsScaleLossOp(const OpDesc &op) const {
   return op.OutputArgumentNames().size() == 1 &&
          op.OutputArgumentNames()[0] == GradVarName(loss_var_name_);
 }
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -13,8 +13,8 @@
 // limitations under the License.
 
 #pragma once
-
 #include <string>
+#include <utility>
 #include <vector>
 
 #include "paddle/fluid/framework/details/ssa_graph_builder.h"
@@ -27,6 +27,7 @@ class NCCLContextMap;
 namespace framework {
 class Scope;
 namespace details {
+
 class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
  public:
 #ifdef PADDLE_WITH_CUDA
@@ -34,8 +35,8 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
-                          bool skip_scale_loss,
-                          platform::NCCLContextMap *nccl_ctxs);
+                          platform::NCCLContextMap *nccl_ctxs,
+                          bool use_default_grad_scale);
 #else
   MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
                           const std::string &loss_var_name,
@@ -74,6 +75,10 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                               size_t num_places) const;
 
   void CreateScaleLossGradOp(SSAGraph *result) const;
+  VarHandle *CreateReduceOp(SSAGraph *result, const std::string &og,
+                            int dst_dev_id) const;
+  void CreateComputationalOp(SSAGraph *result, const OpDesc &op,
+                             int dev_id) const;
 
   bool IsParameterGradientOnce(
       const std::string &og,
@@ -81,11 +86,18 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 
   void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const;
 
+  void CreateBroadcastOp(SSAGraph *result, const std::string &p_name,
+                         size_t src_dev_id) const;
+
   /**
    * Get send op in the global block of program.
    * nullptr if not found.
    */
   OpDesc *GetSendOpDesc(const ProgramDesc &program) const;
+
+  bool IsSparseGradient(
+      const std::unordered_map<std::string, proto::VarType::Type> &var_types,
+      const std::string &og) const;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -22,6 +22,7 @@ namespace framework {
 namespace details {
 
 void ReduceOpHandle::RunImpl() {
+  if (places_.size() == 1) return;
   // the input and output may have dummy var.
   auto in_var_handles = DynamicCast<VarHandle>(inputs_);
 
@@ -51,44 +52,48 @@ void ReduceOpHandle::RunImpl() {
 
   // Wait input done, this Wait is asynchronous operation
   WaitInputVarGenerated(in_var_handles);
-  auto pre_place = in_0_handle->place_;
-  std::vector<platform::Place> in_places;
-  auto pre_in_tensor = VariableVisitor::GetMutableTensor(pre_in_var);
-  for (auto *in_handle : in_var_handles) {
-    auto in_p = in_handle->place_;
-    PADDLE_ENFORCE_EQ(in_p.which(), pre_place.which(),
-                      "Places must be all on CPU or all on CUDA.");
-    in_places.emplace_back(in_p);
 
+  // NOTE: The Places of all input tensor must be all on CPU or all on GPU.
+  std::vector<platform::Place> in_places;  // used to get dev_ctx
+  for (auto *in_handle : in_var_handles) {
+    in_places.emplace_back(in_handle->place_);
     auto in_var =
         var_scopes.at(in_handle->scope_idx_)->FindVar(in_handle->name_);
     PADDLE_ENFORCE_NOT_NULL(in_var);
-
-    auto in_tensor = VariableVisitor::GetMutableTensor(in_var);
-    PADDLE_ENFORCE_EQ(in_tensor.type(), pre_in_tensor.type(),
-                      "The type of input is not consistent.");
+    VariableVisitor::EnforceShapeAndDTypeEQ(*pre_in_var, *in_var);
   }
 
   auto out_var =
       var_scopes.at(out_var_handle->scope_idx_)->FindVar(out_var_handle->name_);
   PADDLE_ENFORCE_NOT_NULL(out_var);
 
+  // NOTE: The tensors' Place of input and output must be all on GPU or all on
+  // CPU.
+  auto in_p = VariableVisitor::GetMutableTensor(pre_in_var).place();
+  platform::Place t_out_p;
+  if (platform::is_gpu_place(in_p)) {
+    PADDLE_ENFORCE(platform::is_gpu_place(out_var_handle->place_),
+                   "Places of input and output must be all on GPU.");
+    t_out_p = out_var_handle->place_;
+  } else {
+    t_out_p = platform::CPUPlace();
+  }
+
   if (pre_in_var->IsType<framework::SelectedRows>()) {
     std::vector<const SelectedRows *> in_selected_rows =
         GetInputValues<SelectedRows>(in_var_handles, var_scopes);
 
-    GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_,
-                       out_var_handle->place_,
+    GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_, t_out_p,
                        out_var->GetMutable<framework::SelectedRows>());
   } else {
     std::vector<const LoDTensor *> lod_tensors =
         GetInputValues<LoDTensor>(in_var_handles, var_scopes);
 
-    if (paddle::platform::is_cpu_place(pre_place)) {
+    if (paddle::platform::is_cpu_place(lod_tensors[0]->place())) {
       ReduceLoDTensor func(lod_tensors,
                            out_var->GetMutable<framework::LoDTensor>());
       VisitDataType(ToDataType(lod_tensors[0]->type()), func);
-    } else if (paddle::platform::is_gpu_place(pre_place)) {
+    } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) {
 #ifdef PADDLE_WITH_CUDA
       auto pre_in = pre_in_var->Get<framework::LoDTensor>();
       VariableVisitor::ShareDimsAndLoD(*pre_in_var, out_var);
@@ -96,7 +101,7 @@ void ReduceOpHandle::RunImpl() {
           out_var_handle->place_, pre_in.type());
 
       auto out_p = out_var_handle->place_;
-      int root = boost::get<platform::CUDAPlace>(out_p).device;
+      int root_id = boost::get<platform::CUDAPlace>(out_p).device;
       std::vector<std::function<void()>> all_reduce_calls;
       for (size_t i = 0; i < var_scopes.size(); ++i) {
         auto &p = in_places[i];
@@ -104,23 +109,23 @@ void ReduceOpHandle::RunImpl() {
 
         int dev_id = boost::get<platform::CUDAPlace>(p).device;
         auto &nccl_ctx = nccl_ctxs_->at(dev_id);
-        auto stream = nccl_ctx.stream();
-        auto comm = nccl_ctx.comm_;
 
         void *buffer = const_cast<void *>(lod_tensor.data<void>());
         void *recvbuffer = nullptr;
-        if (root == dev_id) {
+        if (root_id == dev_id) {
           recvbuffer =
               out_var->GetMutable<framework::LoDTensor>()->mutable_data(
                   out_var_handle->place_);
         }
 
         int type = platform::ToNCCLDataType(lod_tensor.type());
-        all_reduce_calls.emplace_back([=] {
-          PADDLE_ENFORCE(platform::dynload::ncclReduce(
-              buffer, recvbuffer, static_cast<size_t>(lod_tensor.numel()),
-              static_cast<ncclDataType_t>(type), ncclSum, root, comm, stream));
-        });
+        size_t numel = static_cast<size_t>(lod_tensor.numel());
+        all_reduce_calls.emplace_back(
+            [buffer, recvbuffer, type, numel, root_id, &nccl_ctx] {
+              PADDLE_ENFORCE(platform::dynload::ncclReduce(
+                  buffer, recvbuffer, numel, static_cast<ncclDataType_t>(type),
+                  ncclSum, root_id, nccl_ctx.comm_, nccl_ctx.stream()));
+            });
       }
 
       this->RunAndRecordEvent([&] {
@@ -130,7 +135,7 @@ void ReduceOpHandle::RunImpl() {
         }
       });
 #else
-      PADDLE_THROW("CUDA is not support.");
+      PADDLE_THROW("CUDA is not enabled.");
 #endif
     } else {
       PADDLE_THROW("Place should be CPUPlace or CUDAPlace.");

paddle/fluid/framework/details/reduce_op_handle.h

@@ -55,7 +55,7 @@ struct ReduceOpHandle : public OpHandleBase {
 
   std::string Name() const override;
 
-  bool IsMultiDeviceTransfer() override { return false; };
+  bool IsMultiDeviceTransfer() override { return true; };
 
  protected:
   void RunImpl() override;

paddle/fluid/framework/details/var_handle.h

@@ -62,7 +62,7 @@ struct VarHandle : public VarHandleBase {
   std::string name_;
   platform::Place place_;
 
-  bool operator==(const VarHandle& o) const {
+  bool IsTheSameVar(const VarHandle& o) const {
     return o.generated_op_ == generated_op_ && o.name_ == name_ &&
            o.scope_idx_ == scope_idx_;
   }

paddle/fluid/framework/details/variable_visitor.cc

@@ -88,6 +88,52 @@ void VariableVisitor::ShareDimsAndLoD(const Variable& src, Variable* trg) {
   VisitVariable(src, &visitor);
 }
 
+struct EnforceShapeAndDTypeEQVisitor {
+  const Variable* trg_;
+
+  void operator()(const LoDTensor& src) {
+    auto& tensor = trg_->Get<LoDTensor>();
+    PADDLE_ENFORCE_EQ(
+        src.place().which(), tensor.place().which(),
+        "The Places of the two Variable must be all on CPU or all on GPU.");
+    PADDLE_ENFORCE_EQ(src.type(), tensor.type(),
+                      "The dtype of the two Variable is not equal.");
+    PADDLE_ENFORCE_EQ(src.dims(), tensor.dims(),
+                      "The dims of the two Variable is not equal.");
+    PADDLE_ENFORCE_EQ(src.lod(), tensor.lod(),
+                      "The lod of the two Variable is not equal.");
+    PADDLE_ENFORCE_EQ(src.layout(), tensor.layout(),
+                      "The layout of the two Variable's tensor is not equal.");
+  }
+
+  void operator()(const SelectedRows& src) {
+    auto& selected_rows = trg_->Get<SelectedRows>();
+    PADDLE_ENFORCE_EQ(
+        src.place().which(), selected_rows.place().which(),
+        "The Places of the two Variable must be all on CPU or all on GPU.");
+    PADDLE_ENFORCE_EQ(src.value().type(), selected_rows.value().type(),
+                      "The dtype of the two Variable is not equal.");
+    PADDLE_ENFORCE_EQ(src.value().layout(), selected_rows.value().layout(),
+                      "The layout of the two Variable's tensor is not equal.");
+    PADDLE_ENFORCE_EQ(src.height(), selected_rows.height(),
+                      "The height of the two Variable is not equal.");
+    PADDLE_ENFORCE_EQ(src.GetCompleteDims(), selected_rows.GetCompleteDims(),
+                      "The dims of the two Variable is not equal.");
+  }
+
+  template <typename T>
+  void operator()(const T&) {
+    PADDLE_ENFORCE("EnforceShapeAndDTypeEQ is not supported by type %s",
+                   typeid(T).name());
+  }
+};
+
+void VariableVisitor::EnforceShapeAndDTypeEQ(const Variable& var1,
+                                             const Variable& var2) {
+  EnforceShapeAndDTypeEQVisitor visitor{&var1};
+  VisitVariable(var2, &visitor);
+}
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/variable_visitor.h

@@ -26,6 +26,9 @@ class VariableVisitor {
   static Tensor &GetMutableTensor(Variable *var);
 
   static void ShareDimsAndLoD(const Variable &src, Variable *trg);
+
+  static void EnforceShapeAndDTypeEQ(const Variable &var1,
+                                     const Variable &var2);
 };
 
 }  // namespace details

paddle/fluid/framework/executor.cc

@@ -348,8 +348,12 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
       }
     }
   }
+  platform::DeviceContextPool::Instance().Get(place_)->Wait();
   if (create_vars && create_local_scope) {
     scope->DeleteScope(local_scope);
+  } else {
+    // Delete the local scopes created in operators.
+    scope->DropKids();
   }
   if (FLAGS_benchmark) {
     VLOG(2) << "-------------------------------------------------------";

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<std::string>& 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;

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<std::string> 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.

paddle/fluid/framework/parallel_executor.cc

@@ -99,7 +99,7 @@ ParallelExecutor::ParallelExecutor(
 #ifdef PADDLE_WITH_CUDA
   details::MultiDevSSAGraphBuilder builder(
       member_->places_, loss_var_name, params, member_->local_scopes_,
-      use_default_grad_scale, member_->nccl_ctxs_.get());
+      member_->nccl_ctxs_.get(), use_default_grad_scale);
 #else
   details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
                                            params, member_->local_scopes_,

paddle/fluid/inference/tensorrt/CMakeLists.txt

-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)

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

+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)

paddle/fluid/inference/tensorrt/convert/activation_op.cc

+/* 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<nvinfer1::ITensor*>(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

paddle/fluid/inference/tensorrt/convert/conv2d_op.cc

+/* 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

paddle/fluid/inference/tensorrt/convert/mul_op.cc

+/* 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

paddle/fluid/inference/tensorrt/convert/op_converter.h

+/* 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 <string>
+#include <unordered_map>
+#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 <typename T>
+  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<std::string, OpConverter*> 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<Converter__>(#op_type__); \
+    }                                                          \
+  };                                                           \
+  trt_##op_type__##_converter trt_##op_type__##_converter__;
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tensorrt/convert/test_activation_op.cc

+/* 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 <gtest/gtest.h>
+#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<framework::LoDTensor>();
+  x_tensor->Resize({1, 1});
+  std::vector<float> init;
+  init.push_back(input);
+  framework::TensorFromVector(init, ctx, x_tensor);
+
+  auto out_var = scope.Var("Out");
+  auto out_tensor = out_var->GetMutable<framework::LoDTensor>();
+  out_tensor->Resize({1, 1});
+  out_tensor->mutable_data<float>(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<float> 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

paddle/fluid/inference/tensorrt/convert/test_op_converter.cc

+/* 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 <gtest/gtest.h>
+#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

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<int>(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

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<void*> buffers_;
   // max data size for the buffers.
   std::unordered_map<std::string /*name*/, size_t /*max size*/> buffer_sizes_;
+  std::unordered_map<std::string /*name*/, nvinfer1::ITensor* /*ITensor*/>
+      itensor_map_;
 
   // TensorRT related internal members
   template <typename T>

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));
 

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<float>(&input, {FLAGS_batch_size, 3, 32, 32},
-                     static_cast<float>(0), static_cast<float>(1));
+  if (FLAGS_data_set == "cifar10") {
+    SetupTensor<float>(&input, {FLAGS_batch_size, 3, 32, 32},
+                       static_cast<float>(0), static_cast<float>(1));
+  } else if (FLAGS_data_set == "imagenet") {
+    SetupTensor<float>(&input, {FLAGS_batch_size, 3, 224, 224},
+                       static_cast<float>(0), static_cast<float>(1));
+  } else {
+    LOG(FATAL) << "Only cifar10 or imagenet is supported.";
+  }
+
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&input);
 
   paddle::framework::LoDTensor output1;
-  std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
-  cpu_fetchs1.push_back(&output1);
-
-  // Run inference on CPU
-  LOG(INFO) << "--- CPU Runs: ---";
-  TestInference<paddle::platform::CPUPlace, false, true>(
-      dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat);
-  LOG(INFO) << output1.dims();
+  if (!FLAGS_skip_cpu) {
+    std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
+    cpu_fetchs1.push_back(&output1);
+
+    // Run inference on CPU
+    LOG(INFO) << "--- CPU Runs: ---";
+    LOG(INFO) << "Batch size is " << FLAGS_batch_size;
+    TestInference<paddle::platform::CPUPlace, false, true>(
+        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<paddle::platform::CUDAPlace, false, true>(
       dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat);
   LOG(INFO) << output2.dims();
 
-  CheckError<float>(output1, output2);
+  if (!FLAGS_skip_cpu) {
+    CheckError<float>(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<paddle::framework::LoDTensor*> 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<paddle::platform::CUDAPlace, false, true>(
-        fp16_dirname, cpu_feeds, cpu_fetchs3, FLAGS_repeat);
+        FLAGS_fp16_dirname, cpu_feeds, cpu_fetchs3, FLAGS_repeat);
 
     CheckError<float>(output2, output3);
   }

paddle/fluid/operators/adadelta_op.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using Tensor = framework::Tensor;
 class AdadeltaOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -55,6 +56,12 @@ class AdadeltaOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("AvgSquaredGradOut", param_dim);
     ctx->SetOutputDim("AvgSquaredUpdateOut", param_dim);
   }
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("Param")->type());
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker {

paddle/fluid/operators/adagrad_op.cc

@@ -23,6 +23,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using Tensor = framework::Tensor;
 class AdagradOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -56,6 +57,12 @@ class AdagradOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("ParamOut", param_dims);
     ctx->SetOutputDim("MomentOut", param_dims);
   }
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("Param")->type());
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class AdagradOpMaker : public framework::OpProtoAndCheckerMaker {

paddle/fluid/operators/adam_op.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using Tensor = framework::Tensor;
 class AdamOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -69,6 +70,12 @@ class AdamOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("Moment1Out", param_dims);
     ctx->SetOutputDim("Moment2Out", param_dims);
   }
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("Param")->type());
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class AdamOpMaker : public framework::OpProtoAndCheckerMaker {

paddle/fluid/operators/adamax_op.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using Tensor = framework::Tensor;
 class AdamaxOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -63,6 +64,12 @@ class AdamaxOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("MomentOut", param_dims);
     ctx->SetOutputDim("InfNormOut", param_dims);
   }
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("Param")->type());
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker {

paddle/fluid/operators/batch_norm_mkldnn_op.cc

+/* 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 <typename T>
+using EigenArrayMap =
+    Eigen::Map<Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using ConstEigenArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using EigenVectorArrayMap = Eigen::Map<Eigen::Array<T, Eigen::Dynamic, 1>>;
+template <typename T>
+using ConstEigenVectorArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, 1>>;
+
+namespace {
+template <typename T>
+struct bn_type_traits {
+  using op_type = T;
+  using op_desc = typename op_type::desc;
+  using op_prim = typename op_type::primitive_desc;
+};
+
+template <typename T, typename Container>
+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 <typename Op, typename... Args>
+void run_batch_norm_op(Args &&... args) {
+  Op batch_norm_op{args...};
+
+  std::vector<mkldnn::primitive> pipeline;
+  pipeline.push_back(batch_norm_op);
+  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+}
+
+template <typename T>
+inline void *cast_const_to_void(const T *t) {
+  return static_cast<void *>(const_cast<T *>(t));
+}
+}  // namespace
+
+template <typename T>
+class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto data_layout_str = ctx.Attr<std::string>("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<float>("epsilon");
+    const float momentum = ctx.Attr<float>("momentum");
+    const bool is_test = ctx.Attr<bool>("is_test");
+
+    const auto *x = ctx.Input<Tensor>("X");
+    const auto *mean = ctx.Input<Tensor>("Mean");
+    const auto *variance = ctx.Input<Tensor>("Variance");
+
+    auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    auto mkldnn_engine = dev_ctx.GetEngine();
+
+    auto *y = ctx.Output<Tensor>("Y");
+    auto *mean_out = ctx.Output<Tensor>("MeanOut");
+    auto *variance_out = ctx.Output<Tensor>("VarianceOut");
+    auto *batch_mean = ctx.Output<Tensor>("SavedMean");
+    auto *batch_variance = ctx.Output<Tensor>("SavedVariance");
+
+    const auto *scale = ctx.Input<Tensor>("Scale");
+    const auto *shift = ctx.Input<Tensor>("Bias");
+
+    y->mutable_data<T>(ctx.GetPlace());
+    mean_out->mutable_data<T>(ctx.GetPlace());
+    variance_out->mutable_data<T>(ctx.GetPlace());
+
+    if (!is_test) {
+      batch_mean->mutable_data<T>(ctx.GetPlace());
+      batch_variance->mutable_data<T>(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<T>())};
+    auto dst = mkldnn::memory{dst_pd, y->data<T>()};
+
+    unsigned flags = mkldnn::use_scale_shift;
+    if (is_test) flags |= mkldnn::use_global_stats;
+
+    using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
+    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<T> scaleshift_data;
+    scaleshift_data.reserve(scaleshift_size);
+
+    copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(),
+                    shift->data<T>() + 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<T>())};
+
+      auto variance_memory =
+          mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(),
+                         cast_const_to_void(variance->data<T>())};
+
+      run_batch_norm_op<typename bn_fwd_types::op_type>(
+          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<T>())};
+
+      auto variance_memory =
+          mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(),
+                         cast_const_to_void(batch_variance->data<T>())};
+
+      run_batch_norm_op<bn_fwd_types::op_type>(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<T> saved_mean_e(
+          batch_mean->mutable_data<T>(ctx.GetPlace()), ic);
+      EigenVectorArrayMap<T> saved_variance_e(
+          batch_variance->mutable_data<T>(ctx.GetPlace()), ic);
+      saved_mean_e.setZero();
+      saved_variance_e.setZero();
+
+      const unsigned int x_arr_size = in * ic;
+      ConstEigenArrayMap<T> x_arr(x->data<T>(), 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<T> mean_arr{mean->data<T>(), ic};
+      ConstEigenVectorArrayMap<T> variance_arr{variance->data<T>(), ic};
+
+      EigenVectorArrayMap<T> running_mean_arr(
+          mean_out->mutable_data<T>(ctx.GetPlace()), ic);
+      EigenVectorArrayMap<T> running_var_arr(
+          variance_out->mutable_data<T>(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 <typename T>
+class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext &ctx) const override {
+    auto data_layout_str = ctx.Attr<std::string>("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<MKLDNNDeviceContext>();
+    auto mkldnn_engine = dev_ctx.GetEngine();
+
+    const float epsilon = ctx.Attr<float>("epsilon");
+
+    const auto *x = ctx.Input<Tensor>("X");
+    const auto *scale = ctx.Input<Tensor>("Scale");
+    const auto *shift = ctx.Input<Tensor>("Bias");
+    const auto *batch_mean = ctx.Input<Tensor>("SavedMean");
+    const auto *batch_variance = ctx.Input<Tensor>("SavedVariance");
+
+    const auto *diff_y = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    auto *diff_x = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *diff_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
+    auto *diff_shift = ctx.Output<Tensor>(framework::GradVarName("Bias"));
+
+    diff_x->mutable_data<T>(ctx.GetPlace());
+    diff_scale->mutable_data<T>(ctx.GetPlace());
+    diff_shift->mutable_data<T>(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<mkldnn::batch_normalization_backward>;
+    using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
+
+    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<T>())};
+
+    auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(),
+                               cast_const_to_void(batch_mean->data<T>())};
+
+    auto variance =
+        mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(),
+                       cast_const_to_void(batch_variance->data<T>())};
+
+    auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine},
+                                   cast_const_to_void(diff_y->data<T>())};
+
+    const unsigned int ic = dims[1];
+
+    const size_t scaleshift_size = 2 * ic;
+
+    std::vector<T> scaleshift_data;
+    scaleshift_data.reserve(scaleshift_size);
+    copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(),
+                    shift->data<T>() + ic, &scaleshift_data);
+
+    auto scaleshift_memory = mkldnn::memory{
+        batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()};
+
+    std::vector<T> diff_scaleshift_data;
+    diff_scaleshift_data.reserve(scaleshift_size);
+    copy_to_weights(diff_scale->data<T>(), diff_scale->data<T>() + ic,
+                    diff_shift->data<T>(), diff_shift->data<T>() + 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<void *>(diff_x->data<T>())};
+
+    run_batch_norm_op<bn_bwd_types::op_type>(
+        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<T>());
+    std::copy(std::next(it, ic), std::end(diff_scaleshift_data),
+              diff_shift->data<T>());
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace,
+                   ops::BatchNormMKLDNNOpKernel<float>);
+REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace,
+                   ops::BatchNormMKLDNNGradOpKernel<float>);

paddle/fluid/operators/batch_norm_op.cc

@@ -15,6 +15,9 @@ limitations under the License. */
 #include "paddle/fluid/operators/batch_norm_op.h"
 #include <string>
 #include "paddle/fluid/framework/data_layout.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
 
 namespace paddle {
 namespace operators {
@@ -87,9 +90,13 @@ class BatchNormOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext &ctx) const override {
     auto input_data_type =
         framework::ToDataType(ctx.Input<Tensor>("X")->type());
-    // For float or float16 input tensor, the type of the scale, bias, mean,
-    // and var tensors should both be float.
+    // By default, the type of the scale, bias, mean,
+    // and var tensors should both be float. (For float or float16 input tensor)
+    // or double (For double input tensor).
     auto bn_param_type = framework::proto::VarType::FP32;
+    if (input_data_type == framework::proto::VarType::FP64) {
+      bn_param_type = framework::proto::VarType::FP64;
+    }
     PADDLE_ENFORCE_EQ(bn_param_type,
                       framework::ToDataType(ctx.Input<Tensor>("Scale")->type()),
                       "Scale input should be of float type");
@@ -102,7 +109,18 @@ class BatchNormOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(bn_param_type, framework::ToDataType(
                                          ctx.Input<Tensor>("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_);
   }
 };
 
@@ -147,6 +165,9 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
               "Variance of the current mini batch, "
               "will apply to output when training")
         .AsIntermediate();
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Batch Normalization.
 
@@ -345,8 +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<Tensor>("X")->type()), ctx.GetPlace(),
+        layout, library_);
   }
 };
 
@@ -470,6 +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"));
 
@@ -492,8 +525,9 @@ REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker,
 REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp);
 
 REGISTER_OP_CPU_KERNEL(
-    batch_norm,
-    ops::BatchNormKernel<paddle::platform::CPUDeviceContext, float>);
+    batch_norm, ops::BatchNormKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::BatchNormKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     batch_norm_grad,
-    ops::BatchNormGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::BatchNormGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::BatchNormGradKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/batch_norm_op.cu.cc

@@ -287,6 +287,8 @@ namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(
     batch_norm, ops::BatchNormKernel<plat::CUDADeviceContext, float>,
+    ops::BatchNormKernel<plat::CUDADeviceContext, double>,
     ops::BatchNormKernel<plat::CUDADeviceContext, plat::float16>);
 REGISTER_OP_CUDA_KERNEL(
-    batch_norm_grad, ops::BatchNormGradKernel<plat::CUDADeviceContext, float>);
+    batch_norm_grad, ops::BatchNormGradKernel<plat::CUDADeviceContext, float>,
+    ops::BatchNormGradKernel<plat::CUDADeviceContext, double>);

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 <typename T>
   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 <typename T>
 void BeamSearchDecodeFunctor::operator()() const {
   BeamSearchDecoder<T> 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 <>

paddle/fluid/operators/bilinear_tensor_product_op.h

@@ -16,7 +16,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/operators/math/blas.h"
 
 namespace paddle {
 namespace operators {
@@ -61,9 +61,9 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
       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<DeviceContext, T>(dev_ctx, CblasNoTrans, CblasNoTrans,
-                                   batch_size, y_dim, x_dim, 1, x->data<T>(),
-                                   weight_mat.data<T>(), 0, left_mul.data<T>());
+      math::GetBlas<DeviceContext, T>(dev_ctx).GEMM(
+          CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, x->data<T>(),
+          weight_mat.data<T>(), 0, left_mul.data<T>());
       output_col_vec.device(place) =
           (left_mul_mat * y_mat).sum(Eigen::DSizes<int, 1>(1));
     }
@@ -125,6 +125,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
       set_zero(dev_ctx, d_y, static_cast<T>(0));
     }
 
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+
     // Caculate the Output(X@Grad) and Output(Y@Grad).
     if (d_x || d_y) {
       Eigen::DSizes<int, 2> bcast_for_x(1, y_dim);
@@ -138,18 +140,16 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
               output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
                   .broadcast(bcast_for_x) *
               y_mat;
-          math::gemm<DeviceContext, T>(
-              dev_ctx, CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1,
-              y_scale.data<T>(), weight_i.data<T>(), 1, d_x->data<T>());
+          blas.GEMM(CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1,
+                    y_scale.data<T>(), weight_i.data<T>(), 1, d_x->data<T>());
         }
         if (d_y) {
           x_scale_mat.device(place) =
               output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
                   .broadcast(bcast_for_y) *
               x_mat;
-          math::gemm<DeviceContext, T>(
-              dev_ctx, CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1,
-              x_scale.data<T>(), weight_i.data<T>(), 1, d_y->data<T>());
+          blas.GEMM(CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1,
+                    x_scale.data<T>(), weight_i.data<T>(), 1, d_y->data<T>());
         }
       }
     }
@@ -166,9 +166,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
             output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
                 .broadcast(bcast_for_weight) *
             x_mat;
-        math::gemm<DeviceContext, T>(dev_ctx, CblasTrans, CblasNoTrans, x_dim,
-                                     y_dim, batch_size, 1, x_scale.data<T>(),
-                                     y->data<T>(), 0, d_weight_i.data<T>());
+        blas.GEMM(CblasTrans, CblasNoTrans, x_dim, y_dim, batch_size, 1,
+                  x_scale.data<T>(), y->data<T>(), 0, d_weight_i.data<T>());
       }
     }
 

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -366,7 +366,8 @@ REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace,
 
 REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace,
                    paddle::operators::CUDNNConvOpKernel<float>,
-                   paddle::operators::CUDNNConvOpKernel<double>);
+                   paddle::operators::CUDNNConvOpKernel<double>,
+                   paddle::operators::CUDNNConvOpKernel<plat::float16>);
 REGISTER_OP_KERNEL(conv3d_grad, CUDNN, plat::CUDAPlace,
                    paddle::operators::CUDNNConvGradOpKernel<float>,
                    paddle::operators::CUDNNConvGradOpKernel<double>);

paddle/fluid/operators/conv_op.h

@@ -17,9 +17,9 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/depthwise_conv.h"
 #include "paddle/fluid/operators/math/im2col.h"
-#include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/vol2col.h"
 
 namespace paddle {
@@ -161,6 +161,7 @@ class GemmConvKernel : public framework::OpKernel<T> {
     math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
 
     auto& dev_ctx = context.template device_context<DeviceContext>();
+    auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
     for (int i = 0; i < batch_size; i++) {
       Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
       Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
@@ -186,8 +187,7 @@ class GemmConvKernel : public framework::OpKernel<T> {
         // gemm
         Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
         Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
-        math::matmul<DeviceContext, T>(dev_ctx, filter_slice, false, col_matrix,
-                                       false, T(1.0), &out_slice, T(0.0));
+        blas.MatMul(filter_slice, col_matrix, &out_slice);
       }
     }
   }
@@ -274,6 +274,7 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
 
     math::SetConstant<DeviceContext, T> set_zero;
     auto& dev_ctx = context.template device_context<DeviceContext>();
+    auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
 
     if (input_grad) {
       input_grad->mutable_data<T>(context.GetPlace());
@@ -303,9 +304,7 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
             col_matrix.ShareDataWith(in_grad_slice);
             col_matrix.Resize(col_matrix_shape);
           }
-          math::matmul<DeviceContext, T>(dev_ctx, filter_slice, true,
-                                         out_grad_slice, false, T(1.0),
-                                         &col_matrix, T(0.0));
+          blas.MatMul(filter_slice, true, out_grad_slice, false, &col_matrix);
 
           if (is_expand && data_dim == 2U) {
             col2im(dev_ctx, col, dilations, strides,
@@ -352,9 +351,8 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
           // gemm
           Tensor filter_grad_slice =
               filter_grad_.Slice(g * out_step, (g + 1) * out_step);
-          math::matmul<DeviceContext, T>(dev_ctx, out_grad_slice, false,
-                                         col_matrix, true, T(1.0),
-                                         &filter_grad_slice, T(1.0));
+          blas.MatMul(out_grad_slice, false, col_matrix, true,
+                      &filter_grad_slice);
         }
       }
     }

paddle/fluid/operators/conv_transpose_op.h

@@ -16,8 +16,8 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/im2col.h"
-#include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/vol2col.h"
 
 namespace paddle {
@@ -118,6 +118,7 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
     output->mutable_data<T>(context.GetPlace());
     math::SetConstant<DeviceContext, T> set_zero;
     auto& dev_ctx = context.template device_context<DeviceContext>();
+    auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
     set_zero(dev_ctx, output, static_cast<T>(0));
 
     math::Col2ImFunctor<math::ColFormat::kCFO, DeviceContext, T> col2im;
@@ -134,9 +135,7 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
 
       // col_matrix = filter * input_batch
       // of shape (c * k_h * k_w, h * w) or (c * k_d * k_h * k_w, d * h * w)
-      math::matmul<DeviceContext, T>(dev_ctx, filter, true, input_batch, false,
-                                     static_cast<T>(1.0), &col_matrix,
-                                     static_cast<T>(0.0));
+      blas.MatMul(filter, true, input_batch, false, &col_matrix);
 
       if (data_dim == 2U) {
         // col2im: col_matrix -> dy
@@ -213,6 +212,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
     // im2col + gemm (similar to conv-forward)
     // input need to compute gradient
     auto& dev_ctx = context.template device_context<DeviceContext>();
+    auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
     if (input_grad || filter_grad) {
       Tensor col;
       col.mutable_data<T>(col_shape, context.GetPlace());
@@ -267,9 +267,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
           // or
           // (m, c * k_d * k_h * k_w) * (c * k_d * k_h * k_w, d * h * w) -> (m,
           // d, h, w)
-          math::matmul<DeviceContext, T>(
-              dev_ctx, filter, false, col_matrix, false, static_cast<T>(1.0),
-              &input_grad_batch, static_cast<T>(0.0));
+          blas.MatMul(filter, false, col_matrix, false, &input_grad_batch);
         }
         if (filter_grad) {
           // input batch
@@ -279,9 +277,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
           // or
           // (m, d * h * w) * (d * h * w, c * k_d * k_h * k_w) -> (m, c * k_d *
           // k_h * k_w)
-          math::matmul<DeviceContext, T>(dev_ctx, in_batch, false, col_matrix,
-                                         true, static_cast<T>(1.0),
-                                         &filter_grad_, static_cast<T>(1.0));
+          blas.MatMul(in_batch, false, col_matrix, true, &filter_grad_);
         }
       }
     }

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<true>);
 REGISTER_OPERATOR(cross_entropy_grad, ops::CrossEntropyGradientOp);
-REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<float>,
-                       ops::CrossEntropyOpKernel<double>);
+REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<CPUCtx, float>,
+                       ops::CrossEntropyOpKernel<CPUCtx, double>);
 REGISTER_OP_CPU_KERNEL(cross_entropy_grad,
-                       ops::CrossEntropyGradientOpKernel<float>,
-                       ops::CrossEntropyGradientOpKernel<double>);
+                       ops::CrossEntropyGradientOpKernel<CPUCtx, float>,
+                       ops::CrossEntropyGradientOpKernel<CPUCtx, double>);

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 <typename T>
-__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 <typename T>
-__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 <typename T>
-class CrossEntropyOpCUDAKernel : public framework::OpKernel<T> {
- 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<Tensor>("X");
-    const Tensor* label = ctx.Input<Tensor>("Label");
-    Tensor* y = ctx.Output<Tensor>("Y");
-    y->mutable_data<T>(ctx.GetPlace());
-
-    math::CrossEntropyFunctor<platform::CUDADeviceContext, T>()(
-        ctx.template device_context<platform::CUDADeviceContext>(), y, x, label,
-        ctx.Attr<bool>("soft_label"));
-  }
-};
-
-template <typename T>
-class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel<T> {
- 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<Tensor>("X");
-    const Tensor* label = ctx.Input<Tensor>("Label");
-    Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    dx->mutable_data<T>(ctx.GetPlace());
-
-    const T* dy_data =
-        ctx.Input<Tensor>(framework::GradVarName("Y"))->data<T>();
-    T* dx_data = dx->mutable_data<T>(ctx.GetPlace());
-    const T* x_data = x->data<T>();
-
-    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<platform::CUDADeviceContext>();
-    auto stream = dev_ctx.stream();
-
-    if (ctx.Attr<bool>("soft_label")) {
-      auto* label_data = label->data<T>();
-      SoftCrossEntropyGradientKernel<T><<<grid, block, 0, stream>>>(
-          dx_data, dy_data, x_data, label_data, batch_size, class_num);
-    } else {
-      math::SetConstant<platform::CUDADeviceContext, T> functor;
-      functor(dev_ctx, dx, 0);
-      auto* label_data = label->data<int64_t>();
-      grid = (batch_size + block - 1) / block;
-      CrossEntropyGradientKernel<T><<<grid, block, 0, stream>>>(
-          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<float>,
-                        ops::CrossEntropyOpCUDAKernel<double>);
+using CUDACtx = paddle::platform::CUDADeviceContext;
+REGISTER_OP_CUDA_KERNEL(cross_entropy,
+                        ops::CrossEntropyOpKernel<CUDACtx, float>,
+                        ops::CrossEntropyOpKernel<CUDACtx, double>);
 REGISTER_OP_CUDA_KERNEL(cross_entropy_grad,
-                        ops::CrossEntropyGradientOpCUDAKernel<float>,
-                        ops::CrossEntropyGradientOpCUDAKernel<double>);
+                        ops::CrossEntropyGradientOpKernel<CUDACtx, float>,
+                        ops::CrossEntropyGradientOpKernel<CUDACtx, double>);

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 <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename T>
+template <typename DeviceContext, typename T>
 class CrossEntropyOpKernel : public framework::OpKernel<T> {
  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<Tensor>("X");
-    const Tensor* labels = ctx.Input<Tensor>("Label");
-    Tensor* y = ctx.Output<Tensor>("Y");
+    auto* x = ctx.Input<Tensor>("X");
+    auto* labels = ctx.Input<Tensor>("Label");
+    auto* y = ctx.Output<Tensor>("Y");
     y->mutable_data<T>(ctx.GetPlace());
 
-    math::CrossEntropyFunctor<platform::CPUDeviceContext, T>()(
-        ctx.template device_context<platform::CPUDeviceContext>(), y, x, labels,
+    math::CrossEntropyFunctor<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), y, x, labels,
         ctx.Attr<bool>("soft_label"));
   }
 };
 
 template <typename T>
+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 <typename T>
+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<T>(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 <typename DeviceContext, typename T>
 class CrossEntropyGradientOpKernel : public framework::OpKernel<T> {
  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<Tensor>("X");
-    const Tensor* dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
-    const Tensor* label = ctx.Input<Tensor>("Label");
-    Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    T* dx_data = dx->mutable_data<T>(ctx.GetPlace());
+    auto* x = ctx.Input<Tensor>("X");
+    auto* dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());
 
     int64_t class_num = x->dims()[1];
     if (ctx.Attr<bool>("soft_label")) {
-      auto x_mat = EigenMatrix<T>::From(*x);
-      auto dy_mat = EigenMatrix<T>::From(*dy);
-      auto lbl_mat = EigenMatrix<T>::From(*label);
-      auto dx_mat = EigenMatrix<T>::From(*dx);
-
-      dx_mat.device(*ctx.template device_context<platform::CPUDeviceContext>()
-                         .eigen_device()) =
-          -(lbl_mat *
-            dy_mat.broadcast(Eigen::DSizes<int64_t, 2>(1, class_num)) / x_mat);
+      XeSoftlabelGradFunctor<T> functor(dx_data, dy->data<T>(), x->data<T>(),
+                                        label->data<T>(),
+                                        static_cast<size_t>(class_num));
+      platform::ForRange<DeviceContext> for_range(
+          ctx.template device_context<DeviceContext>(),
+          static_cast<size_t>(dx->numel()));
+      for_range(functor);
     } else {
-      int64_t batch_size = x->dims()[0];
-      const T* dy_data = dy->data<T>();
-      const T* x_data = x->data<T>();
-      const int64_t* label_data = label->data<int64_t>();
-
-      math::SetConstant<platform::CPUDeviceContext, T> functor;
-      functor(ctx.template device_context<platform::CPUDeviceContext>(), 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<T>()(-dy_data[i] / x_data[index]);
-      }
+      XeGradFunctor<T> functor(dx_data, dy->data<T>(), x->data<T>(),
+                               label->data<int64_t>(),
+                               static_cast<size_t>(class_num));
+      platform::ForRange<DeviceContext> for_range(
+          ctx.template device_context<DeviceContext>(),
+          static_cast<size_t>(dy->numel()));
+      for_range(functor);
     }
   }
 };

paddle/fluid/operators/decayed_adagrad_op.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using Tensor = framework::Tensor;
 class DecayedAdagradOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -51,6 +52,12 @@ class DecayedAdagradOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("ParamOut", param_dims);
     ctx->SetOutputDim("MomentOut", param_dims);
   }
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("Param")->type());
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class DecayedAdagradOpMaker : public framework::OpProtoAndCheckerMaker {

paddle/fluid/operators/detail/grpc_server.cc

@@ -211,6 +211,11 @@ void AsyncGRPCServer::WaitClientGet(int count) {
   }
 }
 
+void AsyncGRPCServer::WaitServerReady() {
+  std::unique_lock<std::mutex> 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<std::mutex> lock(this->mutex_ready_);
+    ready_ = 1;
+  }
+  condition_ready_.notify_all();
   // wait server
   server_->Wait();
   t_send_->join();
@@ -261,8 +272,8 @@ void AsyncGRPCServer::ShutdownQueue() {
 // This URL explains why shutdown is complicate:
 void AsyncGRPCServer::ShutDown() {
   is_shut_down_ = true;
-  server_->Shutdown();
   ShutdownQueue();
+  server_->Shutdown();
 }
 
 void AsyncGRPCServer::TryToRegisterNewSendOne() {

paddle/fluid/operators/detail/grpc_server.h

@@ -45,10 +45,10 @@ 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) {}
 
   ~AsyncGRPCServer() {}
-
+  void WaitServerReady();
   void RunSyncUpdate();
 
   // functions to sync server barrier status.
@@ -120,6 +120,10 @@ 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

paddle/fluid/operators/elementwise_op_function.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #ifdef __NVCC__
 #include <cuda.h>
 #include <thrust/iterator/iterator_adaptor.h>
+#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 <typename T>
-__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 <typename T, typename DX_OP, typename DY_OP>
 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;
     }

paddle/fluid/operators/ftrl_op.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using Tensor = framework::Tensor;
 class FTRLOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -53,6 +54,12 @@ class FTRLOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("SquaredAccumOut", param_dim);
     ctx->SetOutputDim("LinearAccumOut", param_dim);
   }
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("Param")->type());
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class FTRLOpMaker : public framework::OpProtoAndCheckerMaker {

paddle/fluid/operators/gen_nccl_id_op.cc

@@ -71,7 +71,7 @@ class GenNCCLIdOp : public framework::OperatorBase {
   void GetIdByServer(framework::Scope* scope,
                      const platform::DeviceContext& dev_ctx) const {
     std::string endpoint = Attr<std::string>("endpoint");
-    rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, true));
+    rpc_service_ = new detail::AsyncGRPCServer(endpoint, true);
     framework::ProgramDesc empty_program;
     framework::Executor executor(dev_ctx.GetPlace());
     rpc_service_->SetScope(scope);
@@ -79,21 +79,21 @@ class GenNCCLIdOp : public framework::OperatorBase {
     rpc_service_->SetProgram(&empty_program);
     rpc_service_->SetExecutor(&executor);
 
-    server_thread_.reset(new std::thread(std::bind(
-        &detail::AsyncGRPCServer::RunSyncUpdate, rpc_service_.get())));
+    server_thread_.reset(new std::thread(
+        std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, rpc_service_)));
     rpc_service_->SetCond(0);
     VLOG(3) << "start getting nccl id from trainer 0...";
     auto recv = rpc_service_->Get();
     VLOG(3) << "got nccl id and stop server...";
-    // rpc_service_->SetCond(1);
-    // rpc_service_->ShutDown();
-    rpc_service->Push(LISTEN_TERMINATE_MESSAGE);
+    rpc_service_->ShutDown();
     VLOG(3) << "rpc server stopped";
     // TODO(wuyi): reinit nccl communicators
+    server_thread_->join();
+    delete rpc_service_;
   }
 
  protected:
-  mutable std::shared_ptr<detail::AsyncGRPCServer> rpc_service_;
+  mutable detail::AsyncGRPCServer* rpc_service_ = nullptr;
   mutable std::shared_ptr<std::thread> server_thread_;
 };
 

paddle/fluid/operators/gru_unit_op.h

@@ -14,11 +14,10 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/fluid/operators/activation_op.h"
-#include "paddle/fluid/operators/math/math_function.h"
-
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/activation_op.h"
+#include "paddle/fluid/operators/math/blas.h"
 
 namespace paddle {
 namespace operators {
@@ -87,10 +86,10 @@ class GRUUnitKernel : public framework::OpKernel<T> {
     const T* weight_data = weight->data<T>();
     T* gate_data = gate->data<T>();
     T* reset_hidden_prev_data = reset_hidden_prev->data<T>();
-    math::gemm<DeviceContext, T>(
-        context.template device_context<DeviceContext>(), 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<DeviceContext, T>(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<int, 2> extents({{batch_size, frame_size}});
@@ -103,11 +102,10 @@ class GRUUnitKernel : public framework::OpKernel<T> {
                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<DeviceContext, T>(
-        context.template device_context<DeviceContext>(), 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<int, 2> c_offsets({{0, frame_size * 2}});
     ActCompute(context.Attr<int>("activation"), place,
@@ -188,11 +186,11 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
     ActGradCompute(context.Attr<int>("activation"), place, c, c,
                    d_g.slice(c_offsets, extents), d_h * u);
     // backward for reset_hidden_prev
-    math::gemm<DeviceContext, T>(
-        context.template device_context<DeviceContext>(), 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<DeviceContext, T>(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<int>("gate_activation"), place, r, r,
                    d_g.slice(r_offsets, extents), d_r_h_p * h_p);
@@ -200,18 +198,15 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
     if (weight_grad) {
       T* weight_grad_data = weight_grad->mutable_data<T>(context.GetPlace());
       // backward for state_weight
-      math::gemm<DeviceContext, T>(
-          context.template device_context<DeviceContext>(), 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<DeviceContext, T>(
-          context.template device_context<DeviceContext>(), 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 +214,9 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
           hidden_prev_grad->mutable_data<T>(context.GetPlace());
       auto d_h_p = EigenMatrix<T>::From(*hidden_prev_grad);
       d_h_p.device(place) = d_r_h_p * r + d_h * (u.constant(T(1)) - u);
-      math::gemm<DeviceContext, T>(
-          context.template device_context<DeviceContext>(), 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) {

paddle/fluid/operators/layer_norm_op.h

@@ -15,8 +15,8 @@ limitations under the License. */
 #pragma once
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
-
 #include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
 namespace paddle {
@@ -46,9 +46,9 @@ class RowwiseMean2D<platform::CUDADeviceContext, T> {
   }
   void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, framework::Tensor* out) {
-    math::gemv<platform::CUDADeviceContext, T>(
-        context, false, left_, right_, 1., input.data<T>(), divisor_.data<T>(),
-        0., out->data<T>());
+    math::GetBlas<platform::CUDADeviceContext, T>(context).GEMV(
+        false, left_, right_, 1., input.data<T>(), divisor_.data<T>(), 0.,
+        out->data<T>());
   }
 
  private:
@@ -93,9 +93,9 @@ class ColwiseSum2D<platform::CUDADeviceContext, T> {
 
   void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, framework::Tensor* out) {
-    math::gemv<platform::CUDADeviceContext, T>(
-        context, true, left_, right_, 1., input.data<T>(), divisor_.data<T>(),
-        0., out->data<T>());
+    math::GetBlas<platform::CUDADeviceContext, T>(context).GEMV(
+        true, left_, right_, 1., input.data<T>(), divisor_.data<T>(), 0.,
+        out->data<T>());
   }
 
  private:

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<detail::AsyncGRPCServer> 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 +325,12 @@ 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<int>(::getpid()));
+  SavePort(file_path);
   if (sync_mode) {
     RunSyncLoop(&executor, program, &recv_scope, prefetch_block);
   } else {

paddle/fluid/operators/listen_and_serv_op.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <stdint.h>
+#include <atomic>
 #include <ostream>
 #include <string>
 
@@ -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<detail::AsyncGRPCServer> rpc_service_;
   mutable std::shared_ptr<std::thread> server_thread_;
+  // FIXME(wuyi): it's static so that the operator can be cloned.
+  static std::atomic_int selected_port_;
 };
 
 }  // namespace operators

paddle/fluid/operators/lod_reset_op.h

@@ -46,8 +46,7 @@ class LoDResetKernel : public framework::OpKernel<T> {
         auto* lod = lod_t->data<int>();
         if (platform::is_gpu_place(ctx.GetPlace())) {
           framework::Tensor lod_cpu;
-          framework::TensorCopy(*lod_t, platform::CPUPlace(),
-                                ctx.device_context(), &lod_cpu);
+          framework::TensorCopySync(*lod_t, platform::CPUPlace(), &lod_cpu);
           lod = lod_cpu.data<int>();
         }
         level0 = std::vector<int>(lod, lod + lod_t->numel());

paddle/fluid/operators/lstm_op.h

@@ -15,9 +15,9 @@ limitations under the License. */
 #pragma once
 #include <string>
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/detail/activation_functions.h"
 #include "paddle/fluid/operators/math/lstm_compute.h"
-#include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/sequence2batch.h"
 
 namespace paddle {
@@ -114,6 +114,7 @@ class LSTMKernel : public framework::OpKernel<T> {
     auto cand_act = math::detail::GetActivationType(
         ctx.Attr<std::string>("candidate_activation"));
 
+    auto blas = math::GetBlas<DeviceContext, T>(device_ctx);
     for (size_t n = 0; n < num_batch; n++) {
       int bstart = static_cast<int>(batch_starts[n]);
       int bend = static_cast<int>(batch_starts[n + 1]);
@@ -129,9 +130,8 @@ class LSTMKernel : public framework::OpKernel<T> {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_hidden_t = batch_hidden.Slice(pre_h_start, pre_h_end);
-        math::matmul<DeviceContext, T>(device_ctx, pre_hidden_t, false, *weight,
-                                       false, static_cast<T>(1.0), &gate_t,
-                                       static_cast<T>(1.0));
+        blas.MatMul(pre_hidden_t, false, *weight, false, static_cast<T>(1.0),
+                    &gate_t, static_cast<T>(1.0));
       } else if (hidden_t0) {
         // If n == 0 and there is no initialized hidden state, that is to say
         // the H0 is zeros, the calculation W_h * H0 will be skiped.
@@ -143,9 +143,8 @@ class LSTMKernel : public framework::OpKernel<T> {
         Tensor ordered_h0;
         ReorderInitState<DeviceContext, T>(device_ctx, *hidden_t0, order,
                                            &ordered_h0, true);
-        math::matmul<DeviceContext, T>(device_ctx, ordered_h0, false, *weight,
-                                       false, static_cast<T>(1.0), &gate_t,
-                                       static_cast<T>(1.0));
+        blas.MatMul(ordered_h0, false, *weight, false, static_cast<T>(1.0),
+                    &gate_t, static_cast<T>(1.0));
       }
 
       lstm_value.gate_value = gate_t.data<T>();
@@ -282,6 +281,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
 
     auto batch_starts = batch_gate->lod()[0];
     size_t num_batch = batch_starts.size() - 1;
+    auto blas = math::GetBlas<DeviceContext, T>(device_ctx);
     for (int n = static_cast<int>(num_batch) - 1; n >= 0; n--) {
       int bstart = static_cast<int>(batch_starts[n]);
       int bend = static_cast<int>(batch_starts[n + 1]);
@@ -320,29 +320,25 @@ class LSTMGradKernel : public framework::OpKernel<T> {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_hidden_g = batch_hidden_g.Slice(pre_h_start, pre_h_end);
-        math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight, true,
-                                       static_cast<T>(1.0), &pre_hidden_g,
-                                       static_cast<T>(1.0));
+        blas.MatMul(gate_g, false, *weight, true, static_cast<T>(1.0),
+                    &pre_hidden_g, static_cast<T>(1.0));
         if (weight_g) {
           /* backward weight */
           auto pre_hidden = batch_hidden.Slice(pre_h_start, pre_h_end);
-          math::matmul<DeviceContext, T>(device_ctx, pre_hidden, true, gate_g,
-                                         false, static_cast<T>(1.0), weight_g,
-                                         static_cast<T>(1.0));
+          blas.MatMul(pre_hidden, true, gate_g, false, static_cast<T>(1.0),
+                      weight_g, static_cast<T>(1.0));
         }
       } else {
         if (h0 && weight_g) {
           ReorderInitState<DeviceContext, T>(device_ctx, *h0, order,
                                              &ordered_h0, true);
-          math::matmul<DeviceContext, T>(device_ctx, ordered_h0, true, gate_g,
-                                         false, static_cast<T>(1.0), weight_g,
-                                         static_cast<T>(1.0));
+          blas.MatMul(ordered_h0, true, gate_g, false, static_cast<T>(1.0),
+                      weight_g, static_cast<T>(1.0));
         }
         if (h0 && h0_g) {
           ordered_h0_g.mutable_data<T>(h0_g->dims(), ctx.GetPlace());
-          math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight,
-                                         true, static_cast<T>(1.0),
-                                         &ordered_h0_g, static_cast<T>(0.0));
+          blas.MatMul(gate_g, false, *weight, true, static_cast<T>(1.0),
+                      &ordered_h0_g, static_cast<T>(0.0));
         }
       }
     }

paddle/fluid/operators/lstmp_op.h

@@ -14,15 +14,14 @@ limitations under the License. */
 
 #pragma once
 #include <string>
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/activation_op.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/detail/activation_functions.h"
 #include "paddle/fluid/operators/math/lstm_compute.h"
-#include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/sequence2batch.h"
 
-#include "paddle/fluid/framework/eigen.h"
-#include "paddle/fluid/framework/op_registry.h"
-
 namespace paddle {
 namespace operators {
 
@@ -143,7 +142,7 @@ class LSTMPKernel : public framework::OpKernel<T> {
     auto proj_act = math::detail::GetActivationType(
         ctx.Attr<std::string>("proj_activation"));
     auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
-
+    auto blas = math::GetBlas<DeviceContext, T>(device_ctx);
     for (size_t n = 0; n < num_batch; n++) {
       int bstart = static_cast<int>(batch_starts[n]);
       int bend = static_cast<int>(batch_starts[n + 1]);
@@ -160,9 +159,8 @@ class LSTMPKernel : public framework::OpKernel<T> {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_proj_t = batch_proj.Slice(pre_h_start, pre_h_end);
-        math::matmul<DeviceContext, T>(device_ctx, pre_proj_t, false, *weight,
-                                       false, static_cast<T>(1.0), &gate_t,
-                                       static_cast<T>(1.0));
+        blas.MatMul(pre_proj_t, false, *weight, false, static_cast<T>(1.0),
+                    &gate_t, static_cast<T>(1.0));
       } else if (hidden_t0) {
         // If n == 0 and there is no initialized hidden state, that is to say
         // the H0 is zeros, the calculation W_h * H0 will be skiped.
@@ -176,16 +174,14 @@ class LSTMPKernel : public framework::OpKernel<T> {
         ordered_proj0->mutable_data<T>(ctx.GetPlace());
         ReorderInitState<DeviceContext, T>(device_ctx, *hidden_t0, order,
                                            &ordered_h0, true);
-        math::matmul<DeviceContext, T>(device_ctx, ordered_h0, false,
-                                       *proj_weight, false, static_cast<T>(1.0),
-                                       ordered_proj0, static_cast<T>(0.0));
+        blas.MatMul(ordered_h0, false, *proj_weight, false, static_cast<T>(1.0),
+                    ordered_proj0, static_cast<T>(0.0));
         if (proj_act != math::detail::ActivationType::kIdentity) {
           auto proj0_dev = EigenMatrix<T>::From(*ordered_proj0);
           ActCompute(cell_act, place, proj0_dev, proj0_dev);
         }
-        math::matmul<DeviceContext, T>(device_ctx, *ordered_proj0, false,
-                                       *weight, false, static_cast<T>(1.0),
-                                       &gate_t, static_cast<T>(1.0));
+        blas.MatMul(*ordered_proj0, false, *weight, false, static_cast<T>(1.0),
+                    &gate_t, static_cast<T>(1.0));
       }
 
       lstmp_value.gate_value = gate_t.data<T>();
@@ -196,9 +192,8 @@ class LSTMPKernel : public framework::OpKernel<T> {
           device_ctx, lstmp_value, frame_size, cur_batch_size, gate_act,
           cell_act, cand_act);
       lstmp_value.prev_state_value = lstmp_value.state_value;
-      math::matmul<DeviceContext, T>(device_ctx, hidden_t, false, *proj_weight,
-                                     false, static_cast<T>(1.0), &proj_t,
-                                     static_cast<T>(0.0));
+      blas.MatMul(hidden_t, false, *proj_weight, false, static_cast<T>(1.0),
+                  &proj_t, static_cast<T>(0.0));
       if (proj_act != math::detail::ActivationType::kIdentity) {
         auto proj_t_dev = EigenMatrix<T>::From(proj_t);
         ActCompute(cell_act, place, proj_t_dev, proj_t_dev);
@@ -361,6 +356,7 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
 
     auto batch_starts = batch_gate->lod()[0];
     size_t num_batch = batch_starts.size() - 1;
+    auto blas = math::GetBlas<DeviceContext, T>(device_ctx);
     for (int n = static_cast<int>(num_batch) - 1; n >= 0; n--) {
       int bstart = static_cast<int>(batch_starts[n]);
       int bend = static_cast<int>(batch_starts[n + 1]);
@@ -375,15 +371,13 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
       }
       /* hidden state backwarad */
       Tensor out_g = batch_hidden_g.Slice(bstart, bend);
-      math::matmul<DeviceContext, T>(device_ctx, proj_g, false, *proj_weight,
-                                     true, static_cast<T>(1.0), &out_g,
-                                     static_cast<T>(0.0));
+      blas.MatMul(proj_g, false, *proj_weight, true, static_cast<T>(1.0),
+                  &out_g, static_cast<T>(0.0));
       /* projection weight backward*/
       if (proj_weight_g) {
         Tensor hidden_t = batch_hidden->Slice(bstart, bend);
-        math::matmul<DeviceContext, T>(device_ctx, hidden_t, true, proj_g,
-                                       false, static_cast<T>(1.0),
-                                       proj_weight_g, static_cast<T>(1.0));
+        blas.MatMul(hidden_t, true, proj_g, false, static_cast<T>(1.0),
+                    proj_weight_g, static_cast<T>(1.0));
       }
 
       Tensor gate = batch_gate->Slice(bstart, bend);
@@ -419,24 +413,21 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_proj_g = batch_proj_g.Slice(pre_h_start, pre_h_end);
-        math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight, true,
-                                       static_cast<T>(1.0), &pre_proj_g,
-                                       static_cast<T>(1.0));
+        blas.MatMul(gate_g, false, *weight, true, static_cast<T>(1.0),
+                    &pre_proj_g, static_cast<T>(1.0));
         if (weight_g) {
           /* weight backward*/
           auto pre_proj = batch_proj.Slice(pre_h_start, pre_h_end);
-          math::matmul<DeviceContext, T>(device_ctx, pre_proj, true, gate_g,
-                                         false, static_cast<T>(1.0), weight_g,
-                                         static_cast<T>(1.0));
+          blas.MatMul(pre_proj, true, gate_g, false, static_cast<T>(1.0),
+                      weight_g, static_cast<T>(1.0));
         }
       } else {
         if (h0 && weight_g) {
           ReorderInitState<DeviceContext, T>(device_ctx, *h0, order,
                                              &ordered_h0, true);
           if (weight_g) {
-            math::matmul<DeviceContext, T>(device_ctx, *ordered_proj0, true,
-                                           gate_g, false, static_cast<T>(1.0),
-                                           weight_g, static_cast<T>(1.0));
+            blas.MatMul(*ordered_proj0, true, gate_g, false,
+                        static_cast<T>(1.0), weight_g, static_cast<T>(1.0));
           }
         }
         if (h0 && (h0_g || proj_weight_g)) {
@@ -444,9 +435,8 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
           Tensor proj0_g;
           proj0_g.Resize({in_dims[0], proj_weight->dims()[1]});
           proj0_g.mutable_data<T>(ctx.GetPlace());
-          math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight,
-                                         true, static_cast<T>(1.0), &proj0_g,
-                                         static_cast<T>(0.0));
+          blas.MatMul(gate_g, false, *weight, true, static_cast<T>(1.0),
+                      &proj0_g, static_cast<T>(0.0));
           if (proj_act != math::detail::ActivationType::kIdentity) {
             auto proj0_dev = EigenMatrix<T>::From(*ordered_proj0);
             auto proj0_g_dev = EigenMatrix<T>::From(proj0_g);
@@ -454,14 +444,12 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
                            proj0_g_dev);
           }
           if (h0_g) {
-            math::matmul<DeviceContext, T>(
-                device_ctx, proj0_g, false, *proj_weight, true,
-                static_cast<T>(1.0), &ordered_h0_g, static_cast<T>(0.0));
+            blas.MatMul(proj0_g, false, *proj_weight, true, static_cast<T>(1.0),
+                        &ordered_h0_g, static_cast<T>(0.0));
           }
           if (proj_weight_g) {
-            math::matmul<DeviceContext, T>(device_ctx, ordered_h0, true,
-                                           proj0_g, false, static_cast<T>(1.0),
-                                           proj_weight_g, static_cast<T>(1.0));
+            blas.MatMul(ordered_h0, true, proj0_g, false, static_cast<T>(1.0),
+                        proj_weight_g, static_cast<T>(1.0));
           }
         }
       }

paddle/fluid/operators/math/CMakeLists.txt

@@ -41,7 +41,8 @@ math_library(depthwise_conv)
 math_library(gru_compute DEPS activation_functions math_function)
 math_library(im2col)
 math_library(lstm_compute DEPS activation_functions)
-math_library(math_function DEPS cblas)
+cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context)
+math_library(math_function DEPS blas)
 math_library(maxouting)
 math_library(pooling)
 math_library(selected_rows_functor DEPS selected_rows math_function)

paddle/fluid/operators/math/blas.cc

+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/math/blas.h"
+namespace paddle {
+namespace operators {
+namespace math {
+// Do nothing. Blas is a header only library.
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/blas.h

+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/tensor.h"
+
+#ifdef PADDLE_WITH_MKLML
+#include <mkl_cblas.h>
+#include <mkl_lapacke.h>
+#include <mkl_vml_functions.h>
+#endif
+
+#ifdef PADDLE_USE_OPENBLAS
+#include <cblas.h>
+#include <lapacke.h>
+#endif
+
+#ifndef LAPACK_FOUND
+extern "C" {
+#include <cblas.h>  // NOLINT
+int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda,
+                   int* ipiv);
+int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda,
+                   int* ipiv);
+int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda,
+                   const int* ipiv);
+int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda,
+                   const int* ipiv);
+}
+#endif
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <typename DeviceContext>
+class Blas {
+ public:
+  explicit Blas(const DeviceContext& context) : context_(context) {}
+
+  template <typename T>
+  void GEMM(CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K,
+            T alpha, const T* A, const T* B, T beta, T* C) const;
+
+  template <typename T>
+  void GEMM(bool transA, bool transB, int M, int N, int K, T alpha, const T* A,
+            int lda, const T* B, int ldb, T beta, T* C, int ldc) const;
+
+  template <typename T>
+  void MatMul(const framework::Tensor& mat_a, bool trans_a,
+              const framework::Tensor& mat_b, bool trans_b, T alpha,
+              framework::Tensor* mat_out, T beta) const;
+
+  template <typename T>
+  void MatMul(const framework::Tensor& mat_a, bool trans_a,
+              const framework::Tensor& mat_b, bool trans_b,
+              framework::Tensor* mat_out) const {
+    MatMul(mat_a, trans_a, mat_b, trans_b, static_cast<T>(1.0), mat_out,
+           static_cast<T>(0.0));
+  }
+
+  template <typename T>
+  void MatMul(const framework::Tensor& mat_a, const framework::Tensor& mat_b,
+              framework::Tensor* mat_out) const {
+    this->template MatMul<T>(mat_a, false, mat_b, false, mat_out);
+  }
+
+  template <typename T>
+  void AXPY(int n, T alpha, const T* x, T* y) const;
+
+  template <typename T>
+  void GEMV(bool trans_a, int M, int N, T alpha, const T* A, const T* B, T beta,
+            T* C) const;
+
+  template <typename T>
+  void BatchedGEMM(CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N,
+                   int K, T alpha, const T* A, const T* B, T beta, T* C,
+                   int batchCount, int64_t strideA, int64_t strideB) const;
+
+ private:
+  const DeviceContext& context_;
+};
+
+template <typename DeviceContext, typename T>
+class BlasT : private Blas<DeviceContext> {
+ public:
+  using Blas<DeviceContext>::Blas;
+
+  template <typename... ARGS>
+  void GEMM(ARGS... args) const {
+    Base()->template GEMM<T>(args...);
+  }
+
+  template <typename... ARGS>
+  void MatMul(ARGS... args) const {
+    Base()->template MatMul<T>(args...);
+  }
+
+  template <typename... ARGS>
+  void AXPY(ARGS... args) const {
+    Base()->template AXPY<T>(args...);
+  }
+
+  template <typename... ARGS>
+  void GEMV(ARGS... args) const {
+    Base()->template GEMV<T>(args...);
+  }
+
+  template <typename... ARGS>
+  void BatchedGEMM(ARGS... args) const {
+    Base()->template BatchedGEMM<T>(args...);
+  }
+
+ private:
+  const Blas<DeviceContext>* Base() const {
+    return static_cast<const Blas<DeviceContext>*>(this);
+  }
+};
+
+template <typename DeviceContext, typename T>
+inline BlasT<DeviceContext, T> GetBlas(
+    const framework::ExecutionContext& exe_ctx) {
+  return BlasT<DeviceContext, T>(
+      exe_ctx.template device_context<DeviceContext>());
+}
+
+template <typename DeviceContext, typename T>
+inline BlasT<DeviceContext, T> GetBlas(const DeviceContext& dev_ctx) {
+  return BlasT<DeviceContext, T>(dev_ctx);
+}
+
+}  // 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

paddle/fluid/operators/math/blas_impl.cu.h

+//   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 <typename T>
+struct CUBlas;
+
+template <>
+struct CUBlas<float> {
+  template <typename... ARGS>
+  static void GEMM(ARGS... args) {
+    PADDLE_ENFORCE(platform::dynload::cublasSgemm(args...));
+  }
+
+  template <typename... ARGS>
+  static void AXPY(ARGS... args) {
+    PADDLE_ENFORCE(platform::dynload::cublasSaxpy(args...));
+  }
+
+  template <typename... ARGS>
+  static void GEMV(ARGS... args) {
+    PADDLE_ENFORCE(platform::dynload::cublasSgemv(args...));
+  }
+
+  template <typename... ARGS>
+  static void GEMM_BATCH(ARGS... args) {
+#if CUDA_VERSION >= 8000
+    PADDLE_ENFORCE(platform::dynload::cublasSgemmStridedBatched(args...));
+#else
+    PADDLE_THROW("SgemmStridedBatched is not supported on cuda <= 7.5");
+#endif
+  }
+};
+
+template <>
+struct CUBlas<double> {
+  template <typename... ARGS>
+  static void GEMM(ARGS... args) {
+    PADDLE_ENFORCE(platform::dynload::cublasDgemm(args...));
+  }
+
+  template <typename... ARGS>
+  static void AXPY(ARGS... args) {
+    PADDLE_ENFORCE(platform::dynload::cublasDaxpy(args...));
+  }
+
+  template <typename... ARGS>
+  static void GEMV(ARGS... args) {
+    PADDLE_ENFORCE(platform::dynload::cublasDgemv(args...));
+  }
+
+  template <typename... ARGS>
+  static void GEMM_BATCH(ARGS... args) {
+#if CUDA_VERSION >= 8000
+    PADDLE_ENFORCE(platform::dynload::cublasDgemmStridedBatched(args...));
+#else
+    PADDLE_THROW("DgemmStridedBatched is not supported on cuda <= 7.5");
+#endif
+  }
+};
+
+template <>
+struct CUBlas<platform::float16> {
+  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<const __half *>(alpha),
+                                       reinterpret_cast<const __half *>(A), lda,
+                                       reinterpret_cast<const __half *>(B), ldb,
+                                       reinterpret_cast<const __half *>(beta),
+                                       reinterpret_cast<__half *>(C), ldc));
+  }
+
+  template <typename... ARGS>
+  static void GEMM_BATCH(ARGS... args) {
+#if CUDA_VERSION >= 8000
+    PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched(args...));
+#else
+    PADDLE_THROW("HgemmStridedBatched is not supported on cuda <= 7.5");
+#endif
+  }
+};
+
+template <>
+template <typename T>
+void Blas<platform::CUDADeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
+                                             CBLAS_TRANSPOSE transB, int M,
+                                             int N, int K, T alpha, const T *A,
+                                             const T *B, 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<T>::GEMM(context_.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha,
+                  B, ldb, A, lda, &beta, C, N);
+}
+
+template <>
+template <>
+inline void Blas<platform::CUDADeviceContext>::GEMM(
+    CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K,
+    platform::float16 alpha, const platform::float16 *A,
+    const platform::float16 *B, 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<float>(alpha);
+  float h_beta = static_cast<float>(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<platform::float16>::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 <typename T>
+void Blas<platform::CUDADeviceContext>::GEMM(bool transA, bool transB, int M,
+                                             int N, int K, T alpha, const T *A,
+                                             int lda, const T *B, int ldb,
+                                             T beta, T *C, int ldc) const {
+  // Note that cublas follows fortran order, so the order is different from
+  // the cblas convention.
+  cublasOperation_t cuTransA = transA ? CUBLAS_OP_T : CUBLAS_OP_N;
+  cublasOperation_t cuTransB = transB ? CUBLAS_OP_T : CUBLAS_OP_N;
+  CUBlas<T>::GEMM(context_.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha,
+                  B, ldb, A, lda, &beta, C, ldc);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CUDADeviceContext>::AXPY(int n, T alpha, const T *x,
+                                             T *y) const {
+  CUBlas<T>::AXPY(context_.cublas_handle(), n, &alpha, x, 1, y, 1);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CUDADeviceContext>::GEMV(bool trans_a, int M, int N,
+                                             T alpha, const T *A, const T *B,
+                                             T beta, T *C) const {
+  cublasOperation_t cuTransA = !trans_a ? CUBLAS_OP_T : CUBLAS_OP_N;
+
+  CUBlas<T>::GEMV(context_.cublas_handle(), cuTransA, N, M, &alpha, A, N, B, 1,
+                  &beta, C, 1);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CUDADeviceContext>::BatchedGEMM(
+    CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K,
+    T alpha, const T *A, const T *B, T beta, T *C, int batchCount,
+    int64_t strideA, int64_t strideB) 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;
+  int ldc = N;
+  cublasOperation_t cuTransA =
+      (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
+  cublasOperation_t cuTransB =
+      (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
+  const int64_t strideC = M * N;
+
+  CUBlas<T>::GEMM_BATCH(context_.cublas_handle(), cuTransB, cuTransA, N, M, K,
+                        &alpha, B, ldb, strideB, A, lda, strideA, &beta, C, ldc,
+                        strideC, batchCount);
+}
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/blas_impl.h

+//   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 <vector>
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <typename T>
+struct CBlas;
+
+template <>
+struct CBlas<float> {
+  template <typename... ARGS>
+  static void GEMM(ARGS... args) {
+    cblas_sgemm(args...);
+  }
+
+  template <typename... ARGS>
+  static void AXPY(ARGS... args) {
+    cblas_saxpy(args...);
+  }
+
+  template <typename... ARGS>
+  static void GEMV(ARGS... args) {
+    cblas_sgemv(args...);
+  }
+
+#ifdef PADDLE_WITH_MKLML
+  template <typename... ARGS>
+  static void GEMM_BATCH(ARGS... args) {
+    cblas_sgemm_batch(args...);
+  }
+#endif
+};
+
+template <>
+struct CBlas<double> {
+  template <typename... ARGS>
+  static void GEMM(ARGS... args) {
+    cblas_dgemm(args...);
+  }
+
+  template <typename... ARGS>
+  static void AXPY(ARGS... args) {
+    cblas_daxpy(args...);
+  }
+
+  template <typename... ARGS>
+  static void GEMV(ARGS... args) {
+    cblas_dgemv(args...);
+  }
+
+#ifdef PADDLE_WITH_MKLML
+  template <typename... ARGS>
+  static void GEMM_BATCH(ARGS... args) {
+    cblas_dgemm_batch(args...);
+  }
+#endif
+};
+
+template <>
+struct CBlas<platform::float16> {
+  static void GEMM(...) { PADDLE_THROW("float16 GEMM not supported on CPU"); }
+#ifdef PADDLE_WITH_MKLML
+  static void GEMM_BATCH(...) {
+    PADDLE_THROW("float16 GEMM_BATCH not supported on CPU");
+  }
+#endif
+};
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
+                                            CBLAS_TRANSPOSE transB, int M,
+                                            int N, int K, T alpha, const T *A,
+                                            const T *B, T beta, T *C) const {
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
+  int ldc = N;
+  CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
+                 beta, C, ldc);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::GEMM(bool transA, bool transB, int M,
+                                            int N, int K, T alpha, const T *A,
+                                            int lda, const T *B, int ldb,
+                                            T beta, T *C, int ldc) const {
+  CBlas<T>::GEMM(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans,
+                 transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A,
+                 lda, B, ldb, beta, C, ldc);
+}
+
+template <typename DeviceContext>
+template <typename T>
+void Blas<DeviceContext>::MatMul(const framework::Tensor &mat_a, bool trans_a,
+                                 const framework::Tensor &mat_b, bool trans_b,
+                                 T alpha, framework::Tensor *mat_out,
+                                 T beta) const {
+  auto dim_a = mat_a.dims();
+  auto dim_b = mat_b.dims();
+  auto dim_out = mat_out->dims();
+  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+                 "The input and output of matmul be matrix");
+  PADDLE_ENFORCE(
+      mat_a.place() == mat_b.place() && mat_a.place() == mat_out->place(),
+      "The places of matrices must be same");
+
+  int M = dim_out[0];
+  int N = dim_out[1];
+  int K = !trans_a ? dim_a[1] : dim_a[0];
+
+  CBLAS_TRANSPOSE transA = !trans_a ? CblasNoTrans : CblasTrans;
+  CBLAS_TRANSPOSE transB = !trans_b ? CblasNoTrans : CblasTrans;
+
+  this->GEMM(transA, transB, M, N, K, alpha, mat_a.data<T>(), mat_b.data<T>(),
+             beta, mat_out->data<T>());
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::AXPY(int n, T alpha, const T *x,
+                                            T *y) const {
+  CBlas<T>::AXPY(n, alpha, x, 1, y, 1);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::GEMV(bool trans_a, int M, int N, T alpha,
+                                            const T *A, const T *B, T beta,
+                                            T *C) const {
+  CBLAS_TRANSPOSE transA = !trans_a ? CblasNoTrans : CblasTrans;
+  CBlas<T>::GEMV(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::BatchedGEMM(
+    CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K,
+    T alpha, const T *A, const T *B, T beta, T *C, int batchCount,
+    int64_t strideA, int64_t strideB) const {
+#ifdef PADDLE_WITH_MKLML
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
+  int ldc = N;
+  auto a_array = std::vector<const T *>(batchCount);
+  auto b_array = std::vector<const T *>(batchCount);
+  auto c_array = std::vector<T *>(batchCount);
+  for (int k = 0; k < batchCount; ++k) {
+    a_array[k] = &A[k * strideA];
+    b_array[k] = &B[k * strideB];
+    c_array[k] = &C[k * M * N];
+  }
+
+  CBlas<T>::GEMM_BATCH(CblasRowMajor, &transA, &transB, &M, &N, &K, &alpha,
+                       a_array.data(), &lda, b_array.data(), &ldb, &beta,
+                       c_array.data(), &ldc, 1 /* group_count */, &batchCount);
+#else
+  for (int k = 0; k < batchCount; ++k) {
+    const float *Ak = &A[k * strideA];
+    const float *Bk = &B[k * strideB];
+    float *Ck = &C[k * M * N];
+    this->template GEMM<T>(transA, transB, M, N, K, alpha, Ak, Bk, beta, Ck);
+  }
+#endif
+}
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/context_project.h

@@ -17,8 +17,8 @@ limitations under the License. */
 #include <algorithm>
 #include <vector>
 #include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/im2col.h"
-#include "paddle/fluid/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {
@@ -211,6 +211,7 @@ class ContextProjectGradFunctor {
     int input_row_begin, input_row_end;
     int sequence_height, sequence_width;
     sequence_width = in.dims()[1];
+    auto blas = math::GetBlas<DeviceContext, T>(context);
 
     if (input_grad) {
       for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
@@ -262,8 +263,8 @@ class ContextProjectGradFunctor {
               Tensor out_t_sub = out_t.Slice(k * context_length,
                                              k * context_length + padding_size);
               Tensor w_sub = padding_data->Slice(k, k + padding_size);
-              axpy<DeviceContext, T>(context, w_sub.numel(), static_cast<T>(1),
-                                     out_t_sub.data<T>(), w_sub.data<T>());
+              blas.AXPY(w_sub.numel(), static_cast<T>(1), out_t_sub.data<T>(),
+                        w_sub.data<T>());
             }
           }
           if (down_pad > 0) {
@@ -294,8 +295,8 @@ class ContextProjectGradFunctor {
                   (down_pad_begin_row + t) * context_length);
               Tensor w_sub = padding_data->Slice(
                   up_pad + padding_idx, up_pad + padding_idx + padding_size);
-              axpy<DeviceContext, T>(context, w_sub.numel(), static_cast<T>(1),
-                                     out_t_sub.data<T>(), w_sub.data<T>());
+              blas.AXPY(w_sub.numel(), static_cast<T>(1), out_t_sub.data<T>(),
+                        w_sub.data<T>());
             }
           }
           out_t.Resize({sequence_height, context_length * sequence_width});

paddle/fluid/operators/math/sequence_padding_test.cc

@@ -41,7 +41,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
   if (paddle::platform::is_cpu_place(*place)) {
     seq = cpu_seq;
   } else {
-    TensorCopy(cpu_seq, *place, *context, &seq);
+    TensorCopySync(cpu_seq, *place, &seq);
     seq.set_lod(lod);
   }
 
@@ -64,7 +64,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
   if (paddle::platform::is_cpu_place(*place)) {
     cpu_seq_back = seq_back;
   } else {
-    TensorCopy(seq_back, paddle::platform::CPUPlace(), *context, &cpu_seq_back);
+    TensorCopySync(seq_back, paddle::platform::CPUPlace(), &cpu_seq_back);
     cpu_seq_back.set_lod(lod);
   }