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

revert-13637-optimize-opyreader
Dang Qingqing 7 years ago
commit 8a850e21ad

@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
### Latest PaddlePaddle Release: [Fluid 0.14.0](https://github.com/PaddlePaddle/Paddle/tree/v0.14.0)
### Latest PaddlePaddle Release: [Fluid 0.15.0](https://github.com/PaddlePaddle/Paddle/tree/v0.15.0)
### Install Latest Stable Release:
```
# Linux CPU
@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==0.14.0.post85
## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/0.14.0/new_docs/beginners_guide/install/install_doc.html) on our website.
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/install/install_doc.html) on our website.
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/0.14.0/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/0.14.0/new_docs/beginners_guide/index.html) documentation.
We provide [English](http://paddlepaddle.org/documentation/docs/en/0.15.0/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/index.html) documentation.
- [Deep Learning 101](https://github.com/PaddlePaddle/book)
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/0.14.0/new_docs/user_guides/howto/training/cluster_howto.html)
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/user_guides/howto/training/cluster_howto.html)
You can run distributed training jobs on MPI clusters.
- [Python API](http://paddlepaddle.org/documentation/api/zh/0.14.0/fluid.html)
- [Python API](http://paddlepaddle.org/documentation/api/zh/0.15.0/fluid.html)
Our new API enables much shorter programs.
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/0.14.0/new_docs/advanced_usage/development/contribute_to_paddle.html)
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/advanced_usage/development/contribute_to_paddle.html)
We appreciate your contributions!

@ -91,7 +91,8 @@ def dist_transpile(trainer_id, args, train_prog, startup_prog):
program=train_prog,
pservers=pserver_endpoints,
trainers=trainers,
sync_mode=not args.async_mode)
sync_mode=not args.async_mode,
startup_program=startup_prog)
if training_role == "PSERVER":
pserver_program = t.get_pserver_program(current_endpoint)
pserver_startup_program = t.get_startup_program(

File diff suppressed because it is too large Load Diff

@ -1,24 +1,23 @@
# PaddlePaddle发行规范
PaddlePaddle使用git-flow branching model做分支管理,使用[Semantic Versioning](http://semver.org/)标准表示PaddlePaddle版本号。
PaddlePaddle使用Trunk Based Development,使用[Semantic Versioning](http://semver.org/)标准表示PaddlePaddle版本号。
PaddlePaddle每次发新的版本遵循以下流程:
1. 从`develop`分支派生出新的分支,分支名为`release/版本号`。例如,`release/0.10.0`
1. 将新分支的版本打上tagtag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`,第二个为`0.10.0rc2`,依次类推。
1. 对这个版本的提交,做如下几个操作:
* 使用Regression Test List作为检查列表测试本次release的正确性。
* 如果失败,记录下所有失败的例子,在这个`release/版本号`分支中修复所有bug后Patch号加一到第二步
* 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。
* 将这个版本的python wheel包发布到pypi。
* 更新Docker镜像参考后面的操作细节
1. 第三步完成后,将`release/版本号`分支合入master分支将master分支的合入commit打上tagtag为`版本号`。同时再将`master`分支合入`develop`分支。
1. 协同完成Release Note的书写。
2. 将新分支的版本打上tagtag为`版本号rc-Patch号`。例如第一个tag为`0.10.0-rc0`。
3. 新分支一般不接受新的feature和优化。QA在release分支上进行测试。研发基于最新的develop开发。
4. QA和研发发现的bug在develop上修复验证后cherry-pick修复到release分支。直到release分支相对稳定。
5. 如果有需要在release分支最新代码上打上新的tag比如`0.10.0-rc1`让更多的用户加入测试。重复3-4步。
6. release分支稳定后打上正式的release tag比如`0.10.0`。
7. 将这个版本的python wheel包发布到pypi。
8. 更新Docker镜像参考后面的操作细节
需要注意的是:
* `release/版本号`分支一旦建立,一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭方便测试人员测试PaddlePaddle的行为。
* 在`release/版本号`分支存在的时候如果有bugfix的行为需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。
* bug修复需要先在develop上进行然后进入release分支。而不是直接在release分支上开发。
* release分支原则上只接受修复类的修改不接受新feature。
## 发布wheel包到pypi
@ -61,24 +60,21 @@ docker push [镜像]:[version]
## PaddlePaddle 分支规范
PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范并适应github的特性做了一些区别。
* PaddlePaddle的主版本库遵循[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范。其中:
* `master`分支为稳定(stable branch)版本分支。每一个`master`分支的版本都是经过单元测试和回归测试的版本。
* `develop`分支为开发(develop branch)版本分支。每一个`develop`分支的版本都经过单元测试,但并没有经过回归测试。
* `release/版本号`分支为每一次Release时建立的临时分支。在这个阶段的代码正在经历回归测试。
PaddlePaddle开发过程使用[Trunk Based Development](https://trunkbaseddevelopment.com/) 开发规范。
* 其他用户的fork版本库并不需要严格遵守[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范但所有fork的版本库的所有分支都相当于特性分支。
* 建议开发者fork的版本库使用`develop`分支同步主版本库的`develop`分支
* 建议开发者fork的版本库中再基于`develop`版本fork出自己的功能分支。
* 当功能分支开发完毕后向PaddlePaddle的主版本库提交`Pull Reuqest`,进而进行代码评审。
* 在评审过程中,开发者修改自己的代码,可以继续在自己的功能分支提交代码。
* `develop`分支为开发(develop branch)版本分支。每一个`develop`分支的版本都经过单元测试。并且会经过模型回归测试。
* `release/版本号`分支为每一次Release时建立的临时分支。release分支主要用于测试bug修复和最终发版。
* `master`分支因为历史原因,已经废弃。
* BugFix分支也是在开发者自己的fork版本库维护与功能分支不同的是BugFix分支需要分别给主版本库的`master`、`develop`与可能有的`release/版本号`分支,同时提起`Pull Request`。
* 其他开发者fork的feature branch。
* 建议开发者的feature branch需要同步主版本库的`develop`分支。
* 建议开发者的feature branch需要基于主版本库中的`develop`分支。
* 当feature branch开发完毕后向PaddlePaddle的主版本库提交`Pull Reuqest`,进而进行代码评审。
* 在评审过程中开发者修改自己的代码可以继续在自己的feature branch提交代码。
## PaddlePaddle回归测试列表
本列表说明PaddlePaddle发版之前需要测试的功能点。
TODO
### PaddlePaddle Book中所有章节

@ -4,26 +4,21 @@ PaddlePaddle manages its branches using "git-flow branching model", and [Semanti
Each time we release a new PaddlePaddle version, we should follow the below steps:
1. Fork a new branch from `develop` named `release/[version]`, e.g. `release/0.10.0`.
1. Push a new tag on the release branch, the tag name should be like `[version]rc.patch`. The
first tag should be `0.10.0rc1`, and the second should be `0.10.0.rc2` and so on.
1. After that, we should do:
* Run all regression test on the Regression Test List (see PaddlePaddle TeamCity CI), to confirm
that this release has no major bugs.
* If regression test fails, we must fix those bugs and create a new `release/[version]`
branch from previous release branch.
* Modify `python/setup.py.in`, change the version number and change `ISTAGED` to `True`.
* Publish PaddlePaddle release wheel packages to pypi (see below instructions for detail).
* Update the Docker images (see below instructions for detail).
1. After above step, merge `release/[version]` branch to master and push a tag on the master commit,
then merge `master` to `develop`.
1. Update the Release Note.
***NOTE:***
* Do ***NOT*** merge commits from develop branch to release branches to keep the release branch contain
features only for current release, so that we can test on that version.
* If we want to fix bugs on release branches, we must merge the fix to master, develop and release branch.
1. Create a new release branch from `develop`named `release/[version]`. E.g.`release/0.10.0`
2. Create a new tag for the release branch, tag format: `version-rc.Patch`. E.g. the first tag is `0.10.0-rc0`
3. New release branch normally doesn't accept new features or optimizations. QA will test on the release branch. Developer should develop based on `develop` branch.
4. If QA or Developer find bugs. They should first fix and verify on `develop` branch. Then cherry-pick the fix to the release branch. Wait until the release branch is stable.
5. If necessary, create a new tag on the relese branch, e.g. `0.10.0-rc1`. Involve more users to try it and repeat step 3-4.
6. After release branch is stableCreate the official release tagsuch as `0.10.0`.
7. Release the python wheel package to pypi.
8. Update the docker image (More details below).
NOTE:
* bug fix should happen on `develop` branch, then cherry-pick to relese branch. Avoid developing directly on release branch.
* release normally only accept bug fixes. Don't add new features.
## Publish Wheel Packages to pypi
@ -97,26 +92,22 @@ You can then checkout the latest pushed tags at https://hub.docker.com/r/paddlep
## Branching Model
We use [git-flow](http://nvie.com/posts/a-successful-git-branching-model/) as our branching model,
with some modifications:
* `master` branch is the stable branch. Each version on the master branch is tested and guaranteed.
* `develop` branch is for development. Each commit on develop branch has passed CI unit test, but no
regression tests are run.
* `release/[version]` branch is used to publish each release. Latest release version branches have
bugfix only for that version, but no feature updates.
* Developer forks are not required to follow
[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)
branching model, all forks is like a feature branch.
* Advise: developer fork's develop branch is used to sync up with main repo's develop branch.
* Advise: developer use it's fork's develop branch to for new branch to start developing.
* Use that branch on developer's fork to create pull requests and start reviews.
* developer can push new commits to that branch when the pull request is open.
* Bug fixes are also started from developers forked repo. And, bug fixes branch can merge to
`master`, `develop` and `releases`.
PaddlePaddle uses [Trunk Based Development](https://trunkbaseddevelopment.com/) as our branching model.
* `develop` branch is used for development. Each comment to `develop` branc goes through unit tests and model regression tests.
* `release/[version]` branch is used for each release. Release branch is used for tests, bug fix and evetual release.
* `master` branch as been deprecated for historical reasons
* Developer's feature branch。
* Developer's feature branch should sync with upstream `develop` branch.
* Developer's feature branch should be forked from upstream `develop` branch.
* After feature branch is ready, create a `Pull Request` against the Paddle repo and go through code review.
* In the review process, develop modify codes and push to their own feature branch.
## PaddlePaddle Regression Test List
TODO
### All Chapters of PaddlePaddle Book
We need to guarantee that all the chapters of PaddlePaddle Book can run correctly. Including

@ -100,7 +100,7 @@ paddle.fluid.layers.gru_unit ArgSpec(args=['input', 'hidden', 'size', 'param_att
paddle.fluid.layers.linear_chain_crf ArgSpec(args=['input', 'label', 'param_attr'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.crf_decoding ArgSpec(args=['input', 'param_attr', 'label'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None))
@ -142,7 +142,7 @@ paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 's
paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))

@ -56,9 +56,9 @@ else()
cc_test(mixed_vector_test SRCS mixed_vector_test.cc DEPS place memory device_context tensor)
endif()
if (NOT WIN32)
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio)
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio version)
else()
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto)
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto version)
endif (NOT WIN32)
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
@ -116,7 +116,11 @@ cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope gl
endif(NOT WIN32)
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
cc_library(version SRCS version.cc)
cc_test(version_test SRCS version_test.cc DEPS version)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)

@ -442,8 +442,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
use_gpu = nccl_ctxs_ != nullptr;
#endif
if (use_gpu ||
strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
if (use_gpu && strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
// Insert BCast Ops
for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
auto &to_bcast_set = bcast_var_name_set[dev_id];

@ -16,6 +16,13 @@ syntax = "proto2";
option optimize_for = LITE_RUNTIME;
package paddle.framework.proto;
// Any incompatible changes to ProgramDesc and its dependencies should
// raise the version defined version.h.
//
// Serailization and Deserialization codes should be modified in a way
// that supports old versions following the version and compatibility policy.
message Version { optional int64 version = 1 [ default = 0 ]; }
enum AttrType {
INT = 0;
FLOAT = 1;
@ -180,4 +187,8 @@ message BlockDesc {
// for more details.
// TODO(panyx0718): A model can have multiple programs. Need a
// way to distinguish them. Maybe ID or name?
message ProgramDesc { repeated BlockDesc blocks = 1; }
message ProgramDesc {
repeated BlockDesc blocks = 1;
optional Version version = 2;
}

@ -19,7 +19,7 @@ function(pass_library TARGET DEST)
endfunction()
cc_library(node SRCS node.cc DEPS proto_desc)
cc_library(graph SRCS graph.cc DEPS node)
cc_library(graph SRCS graph.cc DEPS node pretty_log)
cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
cc_library(graph_traits SRCS graph_traits.cc DEPS graph)
@ -28,6 +28,9 @@ cc_library(graph_pattern_detector SRCS graph_pattern_detector.cc DEPS graph grap
pass_library(graph_to_program_pass base)
pass_library(graph_viz_pass base)
pass_library(fc_fuse_pass inference)
if(WITH_MKLDNN)
pass_library(conv_relu_mkldnn_fuse_pass inference)
endif()
pass_library(attention_lstm_fuse_pass inference)
pass_library(infer_clean_graph_pass inference)
pass_library(fc_lstm_fuse_pass inference)
@ -42,3 +45,6 @@ cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_r
cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph_to_program_pass)
cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector)
cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
if(WITH_MKLDNN)
cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
endif()

@ -0,0 +1,90 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h"
#include <string>
#include <vector>
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("conv_relu_mkldnn_fuse", graph.get());
std::unordered_set<Node*> nodes2delete;
GraphPatternDetector gpd;
auto* conv_input = gpd.mutable_pattern()
->NewNode("conv_relu_mkldnn_fuse/conv_input")
->AsInput()
->assert_is_op_input("conv2d", "Input");
patterns::ConvReLU conv_relu_pattern(gpd.mutable_pattern(),
"conv_relu_mkldnn_fuse");
conv_relu_pattern(conv_input);
int found_conv_relu_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle ConvReLU fuse";
GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight,
conv_relu_pattern); // Filter
GET_IR_NODE_FROM_SUBGRAPH(conv_bias, conv_bias, conv_relu_pattern); // Bias
GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_relu_pattern); // tmp
GET_IR_NODE_FROM_SUBGRAPH(conv, conv, conv_relu_pattern); // CONV op
GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, conv_relu_pattern); // Out
GET_IR_NODE_FROM_SUBGRAPH(relu, relu, conv_relu_pattern); // ReLU op
// Create an ConvReLU Node.
OpDesc desc;
std::string conv_relu_i_in = subgraph.at(conv_input)->Name();
std::string conv_relu_w_in = conv_weight->Name();
std::string conv_relu_b_in = conv_bias->Name();
std::string conv_relu_out = relu_out->Name();
desc.SetInput("Input", std::vector<std::string>({conv_relu_i_in}));
desc.SetInput("Filter", std::vector<std::string>({conv_relu_w_in}));
desc.SetInput("Bias", std::vector<std::string>({conv_relu_b_in}));
desc.SetOutput("Out", std::vector<std::string>({conv_relu_out}));
desc.SetType("conv2d");
for (auto& attr : conv->Op()->GetAttrMap()) {
desc.SetAttr(attr.first, attr.second);
}
desc.SetAttr("fuse_relu", true);
auto conv_relu_node = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(graph.get(), {conv, relu, conv_out});
PADDLE_ENFORCE(subgraph.count(conv_input));
IR_NODE_LINK_TO(subgraph.at(conv_input), conv_relu_node);
IR_NODE_LINK_TO(conv_weight, conv_relu_node);
IR_NODE_LINK_TO(conv_bias, conv_relu_node);
IR_NODE_LINK_TO(conv_relu_node, relu_out);
found_conv_relu_count++;
};
gpd(graph.get(), handler);
AddStatis(found_conv_relu_count);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(conv_relu_mkldnn_fuse_pass,
paddle::framework::ir::ConvReLUFusePass);

@ -0,0 +1,39 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
/*
* Fuse the CONV and ReLU to a ConvReLUOp.
*/
class ConvReLUFusePass : public FusePassBase {
public:
virtual ~ConvReLUFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
};
} // namespace ir
} // namespace framework
} // namespace paddle

@ -0,0 +1,108 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h"
#include <gtest/gtest.h>
namespace paddle {
namespace framework {
namespace ir {
void SetOp(ProgramDesc* prog, const std::string& type,
const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
if (type == "conv2d") {
op->SetAttr("use_mkldnn", true);
op->SetInput("Input", {inputs[0]});
op->SetInput("Filter", {inputs[1]});
op->SetInput("Bias", {inputs[2]});
} else if (type == "relu") {
op->SetInput("X", inputs);
}
op->SetOutput("Out", outputs);
}
// a->OP0->b
// b->OP1->c
// (c, weights, bias)->conv->f
// (f)->relu->g
ProgramDesc BuildProgramDesc() {
ProgramDesc prog;
for (auto& v :
std::vector<std::string>({"a", "b", "c", "weights", "bias", "f", "g"})) {
auto* var = prog.MutableBlock(0)->Var(v);
var->SetType(proto::VarType::SELECTED_ROWS);
if (v == "weights" || v == "bias") {
var->SetPersistable(true);
}
}
SetOp(&prog, "OP0", std::vector<std::string>({"a"}),
std::vector<std::string>({"b"}));
SetOp(&prog, "OP1", std::vector<std::string>({"b"}),
std::vector<std::string>({"c"}));
SetOp(&prog, "conv2d", std::vector<std::string>({"c", "weights", "bias"}),
std::vector<std::string>({"f"}));
SetOp(&prog, "relu", std::vector<std::string>({"f"}),
std::vector<std::string>({"g"}));
return prog;
}
TEST(ConvReLUFusePass, basic) {
auto prog = BuildProgramDesc();
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
auto pass = PassRegistry::Instance().Get("conv_relu_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
// Remove 3 Nodes: CONV, RELU, conv_out
// Add 1 Node: ConvReLU
EXPECT_EQ(original_nodes_num - 2, current_nodes_num);
// Assert conv_relu op in newly generated graph
int conv_relu_count = 0;
for (auto* node : graph->Nodes()) {
if (node->IsOp() && node->Op()->Type() == "conv2d") {
if (node->Op()->HasAttr("use_mkldnn")) {
bool use_mkldnn = boost::get<bool>(node->Op()->GetAttr("use_mkldnn"));
if (use_mkldnn) {
if (node->Op()->HasAttr("fuse_relu")) {
bool fuse_relu = boost::get<bool>(node->Op()->GetAttr("fuse_relu"));
if (fuse_relu) {
++conv_relu_count;
}
}
}
}
}
}
EXPECT_EQ(conv_relu_count, 1);
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(conv_relu_mkldnn_fuse_pass);

@ -51,7 +51,7 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
if (with_fc_bias) {
// Add FC-bias with LSTM-bias and create a new weight
PADDLE_ENFORCE(scope);
const std::string& new_bias_var = name_scope + "_bias.new";
const std::string& new_bias_var = patterns::UniqueKey("NewBias");
auto* bias_var = scope->Var(new_bias_var);
PADDLE_ENFORCE(bias_var);
auto* bias_tensor = bias_var->GetMutable<framework::LoDTensor>();
@ -120,7 +120,6 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_IR_NODE_FROM_SUBGRAPH(lstm, lstm, lstm_pattern);
GET_IR_NODE_FROM_SUBGRAPH(Weight, Weight, lstm_pattern);
GET_IR_NODE_FROM_SUBGRAPH(Bias, Bias, lstm_pattern);
@ -136,7 +135,7 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
fc_bias);
// Remove unneeded nodes.
std::unordered_set<const Node*> marked_nodes(
{mul, lstm, elementwise_add});
{mul, lstm, elementwise_add, fc_bias});
GraphSafeRemoveNodes(graph, marked_nodes);
} else {
GET_IR_NODE_FROM_SUBGRAPH(fc_out, mul_out, fc_pattern);

@ -21,12 +21,17 @@
#include "paddle/fluid/framework/ir/graph_traits.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/string/pretty_log.h"
#include "paddle/fluid/string/printf.h"
namespace paddle {
namespace framework {
namespace ir {
using string::PrettyLogEndl;
using string::PrettyLog;
using string::Style;
size_t PDPattern::id_ = 0UL;
PDNode* PDPattern::NewNode(const std::string& name) {
@ -83,7 +88,7 @@ void GraphPatternDetector::operator()(Graph* graph,
ValidateByNodeRole(&subgraphs);
if (subgraphs.empty()) return;
LOG(INFO) << "detect " << subgraphs.size() << " subgraph matches the pattern";
PrettyLogEndl(Style::detail(), "--- detect %d subgraphs", subgraphs.size());
int id = 0;
for (auto& g : subgraphs) {
VLOG(3) << "optimizing #" << id++ << " subgraph";
@ -517,6 +522,39 @@ bool VarLinksFromOp(Node* node, const std::string& op_type) {
return false;
}
PDNode* patterns::ConvReLU::operator()(
paddle::framework::ir::PDNode* conv_input) {
// Create Operators
conv_input->assert_is_op_input("conv2d", "Input");
auto* conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
auto* relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu");
// Create variables
// Filter
auto* conv_weight_var = pattern->NewNode(conv_weight_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("conv2d", "Filter");
// Bias
auto* conv_bias_var = pattern->NewNode(conv_bias_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("conv2d", "Bias");
// intermediate variable, will be removed in the IR after fuse.
auto* conv_out_var = pattern->NewNode(conv_out_repr())
->AsIntermediate()
->assert_is_only_output_of_op("conv2d")
->assert_is_op_input("relu");
// output
auto* relu_out_var = pattern->NewNode(relu_out_repr())
->AsOutput()
->assert_is_op_output("relu");
conv_op->LinksFrom({conv_input, conv_weight_var, conv_bias_var})
.LinksTo({conv_out_var});
relu_op->LinksFrom({conv_out_var}).LinksTo({relu_out_var});
return relu_out_var;
}
PDNode* patterns::FC::operator()(paddle::framework::ir::PDNode* x,
bool with_bias) {
// Create shared nodes.

@ -360,6 +360,28 @@ struct PatternBase {
size_t id_;
};
// CONV with ReLU
// op: conv + relu
// named nodes:
// conv_input, conv_weight,
// conv_bias, conv_out, conv,
// relu_out, relu
struct ConvReLU : public PatternBase {
ConvReLU(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "conv_relu") {}
PDNode* operator()(PDNode* conv_input);
// declare operator node's name
PATTERN_DECL_NODE(conv);
PATTERN_DECL_NODE(relu);
// declare variable node's name
PATTERN_DECL_NODE(conv_weight);
PATTERN_DECL_NODE(conv_bias);
PATTERN_DECL_NODE(conv_out);
PATTERN_DECL_NODE(relu_out);
};
// FC with bias
// op: mul + elementwise_add
// named nodes:

@ -21,6 +21,7 @@ limitations under the License. */
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/var_type.h"
#include "paddle/fluid/framework/version.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/memory/memory.h"
@ -251,8 +252,8 @@ void AppendLoD(LoD *lod, const LoD &lod_length) {
void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
const platform::DeviceContext &dev_ctx) {
{ // the 1st field, uint32_t version for LoDTensor
constexpr uint32_t version = 0;
os.write(reinterpret_cast<const char *>(&version), sizeof(version));
os.write(reinterpret_cast<const char *>(&kCurTensorVersion),
sizeof(kCurTensorVersion));
}
{
// the 2st field, LoD information
@ -281,6 +282,8 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
// the 1st field, unit32_t version for LoDTensor
uint32_t version;
is.read(reinterpret_cast<char *>(&version), sizeof(version));
PADDLE_ENFORCE(framework::IsTensorVersionSupported(version),
"tensor version %u is not supported.", version);
PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
}
{

@ -464,35 +464,35 @@ class RuntimeInferShapeContext : public InferShapeContext {
: op_(op), scope_(scope) {}
bool HasInput(const std::string& name) const override {
if (!op_.HasInputs(name)) {
// has only one input
const auto& ins = op_.Inputs();
auto it = ins.find(name);
if (it == ins.end()) {
return false;
}
auto& ins = Inputs(name);
size_t length = ins.size();
if (length == 0) {
const auto& in = it->second;
if (in.size() == 0 || in[0] == kEmptyVarName) {
return false;
}
PADDLE_ENFORCE_EQ(length, 1UL,
PADDLE_ENFORCE_EQ(in.size(), 1UL,
"Input %s should not have more than one inputs", name);
auto ipt = ins[0];
auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
return var != nullptr;
return scope_.FindVar(in[0]) != nullptr;
}
bool HasOutput(const std::string& name) const override {
if (!op_.HasOutputs(name)) {
// has only one output
const auto& outs = op_.Outputs();
auto it = outs.find(name);
if (it == outs.end()) {
return false;
}
auto& outs = Outputs(name);
size_t length = outs.size();
if (length == 0) {
const auto& out = it->second;
if (out.size() == 0 || out[0] == kEmptyVarName) {
return false;
}
PADDLE_ENFORCE_EQ(length, 1UL,
"Output %s should not have more than one inputs", name);
auto ipt = outs[0];
auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
return var != nullptr;
PADDLE_ENFORCE_EQ(out.size(), 1UL,
"Output %s should not have more than one outputs", name);
return scope_.FindVar(out[0]) != nullptr;
}
bool HasInputs(const std::string& name) const override {

@ -352,7 +352,10 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
ParallelExecutor::~ParallelExecutor() {
if (member_->own_local_scope_) {
for (size_t i = 1; i < member_->local_scopes_.size(); ++i) {
member_->global_scope_->DeleteScope(member_->local_scopes_[i]);
Scope *local_scope = member_->local_scopes_[i];
if (member_->global_scope_->HasKid(local_scope)) {
member_->global_scope_->DeleteScope(local_scope);
}
}
}
}

@ -15,6 +15,7 @@ limitations under the License. */
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/version.h"
namespace paddle {
namespace framework {
@ -38,7 +39,10 @@ proto::ProgramDesc *ProgramDesc::Proto() {
return &desc_;
}
int64_t ProgramDesc::Version() const { return desc_.version().version(); }
ProgramDesc::ProgramDesc() {
desc_.mutable_version()->set_version(kCurProgramVersion);
auto *block = desc_.mutable_blocks()->Add();
block->set_idx(kRootBlockIndex);
block->set_parent_idx(kNoneBlockIndex);

@ -57,6 +57,8 @@ class ProgramDesc {
proto::ProgramDesc *Proto();
int64_t Version() const;
// The output variable of feed_op is referenced as feed_target.
// This function is used to collect the output variable's name of all
// feed_ops.

@ -87,8 +87,17 @@ TEST(ProgramDesc, copy_ctor) {
ASSERT_EQ(op_origin->Inputs(), op_copy->Inputs());
ASSERT_EQ(op_origin->Outputs(), op_copy->Outputs());
ASSERT_EQ(op_copy->Proto()->SerializeAsString(),
op_origin->Proto()->SerializeAsString());
ASSERT_EQ(op_origin->Proto()->attrs().size(),
op_copy->Proto()->attrs().size());
for (auto it = op_origin->Proto()->attrs().begin();
it != op_origin->Proto()->attrs().end(); ++it) {
for (auto it_2 = op_copy->Proto()->attrs().begin();
it_2 != op_copy->Proto()->attrs().end(); ++it_2) {
if (it->name() == it_2->name()) {
ASSERT_TRUE(it_2->SerializeAsString() == it->SerializeAsString());
}
}
}
if (op->Type() == "op_with_subblock") {
ASSERT_EQ(1, op->GetBlockAttrId("sub_block"));

@ -56,5 +56,76 @@ struct RWLock {
};
#endif
class RWLockGuard {
public:
enum Status { kUnLock, kWRLock, kRDLock };
RWLockGuard(RWLock* rw_lock, Status init_status)
: lock_(rw_lock), status_(Status::kUnLock) {
switch (init_status) {
case Status::kRDLock: {
RDLock();
break;
}
case Status::kWRLock: {
WRLock();
break;
}
case Status::kUnLock: {
break;
}
}
}
void WRLock() {
switch (status_) {
case Status::kUnLock: {
lock_->WRLock();
status_ = Status::kWRLock;
break;
}
case Status::kWRLock: {
break;
}
case Status::kRDLock: {
PADDLE_THROW(
"Please unlock read lock first before invoking write lock.");
break;
}
}
}
void RDLock() {
switch (status_) {
case Status::kUnLock: {
lock_->RDLock();
status_ = Status::kRDLock;
break;
}
case Status::kRDLock: {
break;
}
case Status::kWRLock: {
PADDLE_THROW(
"Please unlock write lock first before invoking read lock.");
break;
}
}
}
void UnLock() {
if (status_ != Status::kUnLock) {
lock_->UNLock();
status_ = Status::kUnLock;
}
}
~RWLockGuard() { UnLock(); }
private:
RWLock* lock_;
Status status_;
};
} // namespace framework
} // namespace paddle

@ -72,6 +72,12 @@ void Scope::DropKids() {
kids_.clear();
}
bool Scope::HasKid(const Scope* scope) const {
std::unique_lock<std::mutex> lock(mutex_);
auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
return it != this->kids_.end();
}
std::vector<std::string> Scope::LocalVarNames() const {
std::unique_lock<std::mutex> lock(mutex_);
std::vector<std::string> known_vars;

@ -71,6 +71,9 @@ class Scope {
/// Drop all kids scopes belonged to this scope.
void DropKids();
/// Find if a scope exists in the kid scopes
bool HasKid(const Scope* scope) const;
// enumerate all the variables current contains.
std::vector<std::string> LocalVarNames() const;

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save