remove conflict

mobile_baidu
chengduoZH 8 years ago
commit acc3278811

@ -126,7 +126,7 @@ include(external/swig) # download, build, install swig
include(external/warpctc) # download, build, install warpctc
include(external/any) # download libn::any
include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/pybind11) # download pybind11
include(external/nccl)
include(cudnn) # set cudnn libraries, must before configure

@ -79,9 +79,8 @@ if(NOT DEFINED IOS_ARCH)
# FIXME(liuyiqun): support "armv7;armv7s;arm64" future
set(IOS_ARCH "arm64")
elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
set(IOS_ARCH "i386;x86_64")
elseif(IOS_PLATFORM STREQUAL "WATCHOS")
set(IOS_ARCH armv7k)
# FIXME(liuyiqun): support "i386;x86_64" future
set(IOS_ARCH "x86_64")
endif()
endif()
set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string "Build architecture for iOS")

@ -1,3 +1,21 @@
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
if(NOT WITH_GPU)
return()
endif()
include(ExternalProject)
set(NCCL_SOURCE_DIR ${THIRD_PARTY_PATH}/nccl)

@ -1,11 +1,11 @@
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

@ -1,8 +1,26 @@
INCLUDE(ExternalProject)
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
SET(PYBIND_SOURCE_DIR ${THIRD_PARTY_PATH}/pybind)
if(NOT WITH_PYTHON)
return()
endif()
include(ExternalProject)
INCLUDE_DIRECTORIES(${PYBIND_SOURCE_DIR}/src/extern_pybind/include)
set(PYBIND_SOURCE_DIR ${THIRD_PARTY_PATH}/pybind)
include_directories(${PYBIND_SOURCE_DIR}/src/extern_pybind/include)
ExternalProject_Add(
extern_pybind
@ -17,14 +35,12 @@ ExternalProject_Add(
TEST_COMMAND ""
)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
if(${CMAKE_VERSION} VERSION_LESS "3.3.0")
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/pybind_dummy.c)
file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";")
file(WRITE ${dummyfile} "const char * dummy_pybind = \"${dummyfile}\";")
add_library(pybind STATIC ${dummyfile})
else()
add_library(pybind INTERFACE)
endif()
add_dependencies(pybind extern_pybind)
LIST(APPEND external_project_dependencies pybind)

@ -1,11 +1,11 @@
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

@ -1,11 +1,11 @@
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

@ -1,27 +1,28 @@
# This file is use to check all support level of AVX on your machine
# so that PaddlePaddle can unleash the vectorization power of muticore.
INCLUDE(CheckCXXSourceRuns)
INCLUDE(CheckCXXSourceCompiles)
include(CheckCXXSourceRuns)
include(CheckCXXSourceCompiles)
IF(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
if(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
set(MMX_FLAG "-mmmx")
set(SSE2_FLAG "-msse2")
set(SSE3_FLAG "-msse3")
SET(AVX_FLAG "-mavx")
SET(AVX2_FLAG "-mavx2")
ELSEIF(MSVC)
set(AVX_FLAG "-mavx")
set(AVX2_FLAG "-mavx2")
elseif(MSVC)
set(MMX_FLAG "/arch:MMX")
set(SSE2_FLAG "/arch:SSE2")
set(SSE3_FLAG "/arch:SSE3")
SET(AVX_FLAG "/arch:AVX")
SET(AVX2_FLAG "/arch:AVX2")
ENDIF()
endif()
set(CMAKE_REQUIRED_FLAGS_RETAINED ${CMAKE_REQUIRED_FLAGS})
# Check MMX
set(CMAKE_REQUIRED_FLAGS ${MMX_FLAG})
set(MMX_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <mmintrin.h>
int main()
@ -32,6 +33,7 @@ int main()
# Check SSE2
set(CMAKE_REQUIRED_FLAGS ${SSE2_FLAG})
set(SSE2_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <emmintrin.h>
int main()
@ -42,6 +44,7 @@ int main()
# Check SSE3
set(CMAKE_REQUIRED_FLAGS ${SSE3_FLAG})
set(SSE3_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <pmmintrin.h>
int main()
@ -55,6 +58,7 @@ int main()
# Check AVX
set(CMAKE_REQUIRED_FLAGS ${AVX_FLAG})
set(AVX_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <immintrin.h>
int main()
@ -67,6 +71,7 @@ int main()
# Check AVX 2
set(CMAKE_REQUIRED_FLAGS ${AVX2_FLAG})
set(AVX2_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <immintrin.h>
int main()

@ -145,7 +145,7 @@ PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以
Jupyter Notebook是一个开源的web程序大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。
PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Nodebook。
PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Notebook。
如果您想要更深入了解deep learningPaddlePaddle Book一定是您最好的选择。
我们提供可以直接运行PaddlePaddle Book的Docker镜像直接运行

@ -29,32 +29,32 @@ add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER}
add_dependencies(paddle_capi paddle_proto)
# TODO: paddle_capi_whole will be removed.
set(PADDLE_CAPI_LAYERS_LIBS
paddle_function
paddle_gserver)
if(MOBILE_INFERENCE)
set(PADDLE_CAPI_INFER_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_function
paddle_gserver
paddle_proto)
set(PADDLE_CAPI_ENGINE_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_proto)
else()
set(PADDLE_CAPI_INFER_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_function
paddle_gserver
paddle_proto
paddle_pserver
paddle_network)
set(PADDLE_CAPI_ENGINE_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_proto
paddle_pserver
paddle_network)
endif()
set(PADDLE_CAPI_INFER_LIBS ${PADDLE_CAPI_LAYERS_LIBS} ${PADDLE_CAPI_ENGINE_LIBS})
cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
# Link the static library for inference
cc_library(paddle_capi_engine DEPS paddle_capi paddle_utils paddle_parameter paddle_math paddle_cuda paddle_proto)
cc_library(paddle_capi_layers DEPS paddle_function paddle_gserver)
cc_library(paddle_capi_engine DEPS paddle_capi ${PADDLE_CAPI_ENGINE_LIBS})
cc_library(paddle_capi_layers DEPS ${PADDLE_CAPI_LAYERS_LIBS})
# Link the shared library for inference
if(NOT IOS)

@ -23,6 +23,7 @@ limitations under the License. */
#include "paddle/framework/feed_fetch_type.h"
#include "paddle/framework/lod_rank_table.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/lod_tensor_array.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/scope.h"
@ -73,6 +74,8 @@ static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
var->GetMutable<std::vector<framework::Scope>>();
} else if (var_type == VarDesc::LOD_RANK_TABLE) {
var->GetMutable<LoDRankTable>();
} else if (var_type == VarDesc::LOD_TENSOR_ARRAY) {
var->GetMutable<LoDTensorArray>();
} else {
PADDLE_THROW(
"Variable type %d is not in "

@ -109,6 +109,11 @@ message LoDTensorDesc {
optional int32 lod_level = 2 [ default = 0 ];
}
message LoDTensorArrayDesc {
required TensorDesc tensor = 1;
optional int32 lod_level = 2 [ default = 0 ];
}
message VarDesc {
enum VarType {
LOD_TENSOR = 1;
@ -117,11 +122,13 @@ message VarDesc {
FETCH_LIST = 4;
STEP_SCOPES = 5;
LOD_RANK_TABLE = 6;
LOD_TENSOR_ARRAY = 7;
}
required string name = 1;
required VarType type = 2;
optional LoDTensorDesc lod_tensor = 3;
optional TensorDesc selected_rows = 4;
optional LoDTensorArrayDesc tensor_array = 6;
optional bool persistable = 5 [ default = false ];
}

@ -33,10 +33,15 @@ void LoDRankTable::Reset(const LoD& lod, size_t level) {
item.length = vec[i + 1] - vec[i];
items_.emplace_back(item);
}
std::sort(items_.begin(), items_.end(),
[](const TableItem& a, const TableItem& b) {
return a.length > b.length;
});
// NOTE(yuyang18):
//
// The time complexity of stable_sort is O(N*log(N)) if additional memory is
// available. It is easy to debug and unit test when using `stable_sort`
// instead of `sort`. Also, the items of a rank table will not be too large.
std::stable_sort(items_.begin(), items_.end(),
[](const TableItem& a, const TableItem& b) {
return a.length > b.length;
});
}
} // namespace framework

@ -135,5 +135,43 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin,
PADDLE_ENFORCE_LT(begin, end, "Cannot shrink, the result tensor is empty.");
ShareDataWith(Slice(begin, end));
}
void GetFineGrainedLoDLength(const LoD& lod, size_t start_idx, size_t end_idx,
std::vector<std::vector<size_t>>* lod_length,
size_t* start_offset) {
lod_length->clear();
PADDLE_ENFORCE(start_idx < lod.size() - 1,
"start_idx should be >= 0 and < lod.size() - 1.");
PADDLE_ENFORCE(end_idx < lod.size(),
"end_idx should be >= 0 and < lod.size().");
PADDLE_ENFORCE_LE(start_idx, end_idx,
"start_idx should be less than end_idx.");
for (size_t level_idx = 0; level_idx < lod.size(); ++level_idx) {
std::vector<size_t> level_lens;
for (size_t i = start_idx; i < end_idx; ++i) {
level_lens.push_back(lod[level_idx][i + 1] - lod[level_idx][i]);
}
lod_length->emplace_back(level_lens);
start_idx = lod[level_idx][start_idx];
end_idx = lod[level_idx][end_idx];
}
*start_offset = start_idx;
}
void AppendLoD(LoD* lod, const std::vector<std::vector<size_t>>& lod_length) {
PADDLE_ENFORCE_EQ(
lod->size(), lod_length.size(),
"The lod_length should has the same size with the appended lod.");
for (size_t i = 0; i < lod->size(); ++i) {
auto& level = (*lod)[i];
if (level.empty()) {
level.push_back(0);
}
for (size_t len : lod_length[i]) {
level.push_back(level.back() + len);
}
}
}
} // namespace framework
} // namespace paddle

@ -181,5 +181,11 @@ LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level,
return tensor;
}
void GetFineGrainedLoDLength(const LoD& lod, size_t start_idx, size_t end_idx,
std::vector<std::vector<size_t>>* lod_length,
size_t* start_offset);
void AppendLoD(LoD* lod, const std::vector<std::vector<size_t>>& lod_length);
} // namespace framework
} // namespace paddle

@ -0,0 +1,23 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "paddle/framework/lod_tensor.h"
namespace paddle {
namespace framework {
using LoDTensorArray = std::vector<LoDTensor>;
}
} // namespace paddle

@ -144,5 +144,47 @@ TEST(LodExpand, test) {
}
}
TEST(LoD, GetFineGrainedLoDLength) {
LoD lod;
lod.push_back(std::vector<size_t>{0, 2, 4, 5});
lod.push_back(std::vector<size_t>{0, 1, 6, 8, 10, 11});
lod.push_back(
std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20, 24, 26, 29});
std::vector<std::vector<size_t>> lod_length;
size_t start_offset;
paddle::framework::GetFineGrainedLoDLength(lod, 1, 2, &lod_length,
&start_offset);
std::vector<std::vector<size_t>> expected;
expected.push_back(std::vector<size_t>{2});
expected.push_back(std::vector<size_t>{2, 2});
expected.push_back(std::vector<size_t>{2, 3, 4, 2});
EXPECT_EQ(lod_length, expected);
EXPECT_EQ(start_offset, 15UL);
}
TEST(LoD, AppendLoD) {
std::vector<std::vector<size_t>> lod_lens;
lod_lens.push_back(std::vector<size_t>{2});
lod_lens.push_back(std::vector<size_t>{2, 2});
lod_lens.push_back(std::vector<size_t>{2, 3, 4, 2});
LoD origin;
origin.push_back(std::vector<size_t>{0, 2});
origin.push_back(std::vector<size_t>{0, 1, 6});
origin.push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15});
paddle::framework::AppendLoD(&origin, lod_lens);
LoD expected;
expected.push_back(std::vector<size_t>{0, 2, 4});
expected.push_back(std::vector<size_t>{0, 1, 6, 8, 10});
expected.push_back(
std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20, 24, 26});
EXPECT_EQ(origin, expected);
}
} // namespace framework
} // namespace paddle

@ -37,13 +37,27 @@ std::vector<int64_t> VarDescBind::Shape() const {
DataType VarDescBind::GetDataType() const { return tensor_desc().data_type(); }
void VarDescBind::SetLoDLevel(int32_t lod_level) {
PADDLE_ENFORCE(desc_.type() == VarDesc::LOD_TENSOR);
desc_.mutable_lod_tensor()->set_lod_level(lod_level);
switch (desc_.type()) {
case VarDesc::LOD_TENSOR:
desc_.mutable_lod_tensor()->set_lod_level(lod_level);
break;
case VarDesc::LOD_TENSOR_ARRAY:
desc_.mutable_tensor_array()->set_lod_level(lod_level);
break;
default:
PADDLE_THROW("Tensor type=%d does not support LoDLevel", desc_.type());
}
}
int32_t VarDescBind::GetLodLevel() const {
PADDLE_ENFORCE(desc_.type() == VarDesc::LOD_TENSOR);
return desc_.lod_tensor().lod_level();
switch (desc_.type()) {
case VarDesc::LOD_TENSOR:
return desc_.lod_tensor().lod_level();
case VarDesc::LOD_TENSOR_ARRAY:
return desc_.tensor_array().lod_level();
default:
PADDLE_THROW("Tensor type=%d does not support LoDLevel", desc_.type());
}
}
const TensorDesc &VarDescBind::tensor_desc() const {
@ -53,6 +67,8 @@ const TensorDesc &VarDescBind::tensor_desc() const {
return desc_.selected_rows();
case VarDesc::LOD_TENSOR:
return desc_.lod_tensor().tensor();
case VarDesc::LOD_TENSOR_ARRAY:
return desc_.tensor_array().tensor();
default:
PADDLE_THROW("Unexpected branch.");
}
@ -66,6 +82,8 @@ TensorDesc *VarDescBind::mutable_tensor_desc() {
return desc_.mutable_selected_rows();
case VarDesc::LOD_TENSOR:
return desc_.mutable_lod_tensor()->mutable_tensor();
case VarDesc::LOD_TENSOR_ARRAY:
return desc_.mutable_tensor_array()->mutable_tensor();
default:
PADDLE_THROW("Unexpected branch.");
}

@ -0,0 +1,154 @@
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "MKLDNNAddtoLayer.h"
using namespace mkldnn; // NOLINT
namespace paddle {
REGISTER_LAYER(mkldnn_addto, MKLDNNAddtoLayer);
bool MKLDNNAddtoLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
if (!MKLDNNLayer::init(layerMap, parameterMap)) {
return false;
}
layerSize_ = getSize();
for (size_t i = 0; i < inputLayers_.size(); i++) {
CHECK_EQ(layerSize_, inputLayers_[i]->getSize()) << "input size must equal";
}
if (biasParameter_.get() != NULL) {
biases_ =
std::unique_ptr<Weight>(new Weight(1, layerSize_, biasParameter_, 0));
}
return true;
}
void MKLDNNAddtoLayer::reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
CHECK_EQ(layerSize_, getSize()) << "this layer size can not be changed";
reshapeInput(bs, ih, iw);
ic = inputLayers_[0]->getSize() / ih / iw;
CHECK_EQ((size_t)ic * ih * iw, inputLayers_[0]->getSize());
CHECK_EQ(inputElemenCnt_, (size_t)bs * ic * ih * iw);
for (size_t i = 0; i < inputLayers_.size(); i++) {
CHECK_EQ(int64_t(bs), inputLayers_[i]->getOutput().getBatchSize());
CHECK_EQ(layerSize_, inputLayers_[i]->getSize());
}
oc = ic;
oh = ih;
ow = iw;
reshapeOutput(oh, ow);
resizeOutput(bs, oc * oh * ow);
printSizeInfo();
}
void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
if (biases_) {
LOG(FATAL) << "not implemented yet";
}
resetFwdBuffers(inVals_, out);
in = inVals_[0];
std::shared_ptr<sum::primitive_desc> fwdPD;
resetFwdPD(fwdPD, inVals_, out);
resetFwdPipeline(pipeline, fwdPD, inVals_, out);
}
void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
resetBwdBuffers(inGrads_, out);
in = inGrads_[0];
// backward only need share output grad to input grad
for (size_t i = 0; i < inGrads_.size(); i++) {
if (inGrads_[i] != nullptr) {
inGrads_[i] = out;
inputLayers_[i]->getOutputGrad()->setData(inGrads_[i]->getData());
}
}
}
void MKLDNNAddtoLayer::updateWeights(const UpdateCallback& callback) {
if (biases_ && biases_->getWGrad()) {
biases_->getParameterPtr()->incUpdate(callback);
}
}
void MKLDNNAddtoLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out) {
inputs.resize(inputLayers_.size());
for (size_t i = 0; i < inputs.size(); i++) {
resetInValue(inputs[i], nullptr, i);
CHECK(inputs[i]);
inputs[i]->downSpatial();
}
for (size_t i = 1; i < inputs.size(); i++) {
CHECK_PRIMITIVE_DESC_EQ(inputs[i], inputs[0]->getPrimitiveDesc());
}
resetOutValue(out, inputs[0]->getPrimitiveDesc());
}
void MKLDNNAddtoLayer::resetFwdPD(std::shared_ptr<sum::primitive_desc>& pd,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr out) {
std::vector<double> scales(inputs.size(), 1.0);
std::vector<memory::primitive_desc> srcPDs;
for (size_t i = 0; i < inputs.size(); i++) {
srcPDs.push_back(inputs[i]->getPrimitiveDesc());
}
CHECK(out);
pd.reset(new sum::primitive_desc(out->getMemoryDesc(), scales, srcPDs));
CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc());
}
void MKLDNNAddtoLayer::resetFwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<sum::primitive_desc>& pd,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out) {
std::vector<primitive::at> srcs;
for (size_t i = 0; i < inputs.size(); i++) {
srcs.push_back(*(inputs[i]));
}
fwd_.reset(new sum(*pd, srcs, *out));
pipeline.push_back(*fwd_);
}
void MKLDNNAddtoLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out) {
CHECK(outVal_);
resetOutGrad(out, outVal_->getPrimitiveDesc());
CHECK(out);
inputs.resize(inputLayers_.size());
for (size_t i = 0; i < inputs.size(); i++) {
resetInGrad(inputs[i], inVal_->getPrimitiveDesc(), i);
CHECK_PRIMITIVE_DESC_EQ(inputs[i], out->getPrimitiveDesc());
}
}
} // namespace paddle

@ -0,0 +1,110 @@
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "MKLDNNLayer.h"
#include "mkldnn.hpp"
namespace paddle {
/**
* @brief A subclass of MKLDNNLayer Addto layer.
*
* The config file api is mkldnn_addto
*/
class MKLDNNAddtoLayer : public MKLDNNLayer {
protected:
std::vector<MKLDNNMatrixPtr> inVals_;
std::vector<MKLDNNMatrixPtr> inGrads_;
// layer size == ic * ih * iw == oc * oh *ow, and can not be changed
size_t layerSize_;
// TODO(TJ): this part has not been optimized by MKL-DNN
std::unique_ptr<Weight> biases_;
public:
explicit MKLDNNAddtoLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
~MKLDNNAddtoLayer() {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
void resetFwd(std::vector<mkldnn::primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void resetBwd(std::vector<mkldnn::primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void updateWeights(const UpdateCallback& callback) override;
void printValueFormat() override {
for (size_t i = 0; i < inVals_.size(); ++i) {
VLOG(MKLDNN_FMTS) << i << " input: " << inVals_[i]->getFormat() << " >>>";
}
if (outVal_) {
VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> ";
}
if (extOutVal_) {
VLOG(MKLDNN_FMTS) << extOutVal_->getFormat();
}
}
void printGradFormat() override {
if (extOutGrad_) {
VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat();
}
if (outGrad_) {
VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< ";
}
for (size_t i = 0; i < inGrads_.size(); ++i) {
VLOG(MKLDNN_FMTS) << i << " input: " << inGrads_[i]->getFormat() << "<<<";
}
}
protected:
/**
* Forward functions: reset buffers(inputs, output, bias),
* reset primitive descriptor,
* reset pipeline.
*/
void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out);
void resetFwdPD(std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr out);
void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out);
/**
* Backward functions: reset buffers(inputs, output, bias)
*/
void resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out);
};
} // namespace paddle

@ -77,7 +77,7 @@ void MKLDNNLayer::forward(PassType passType) {
needResetBwd_ = true;
}
if (inputLayers_[0]->getType() == "data") {
if (inputLayers_[0]->getType() == "data" && inputLayers_.size() == 1) {
// Update input value data when input layer is "data" type,
// since the input value data address might be changed.
CHECK(extInVal_);
@ -171,14 +171,16 @@ void MKLDNNLayer::resetWithMatrix(MKLDNNMatrixPtr& dnn,
}
void MKLDNNLayer::resetInValue(
MKLDNNMatrixPtr& in, const std::shared_ptr<memory::primitive_desc>& intPD) {
MKLDNNMatrixPtr& in,
const std::shared_ptr<memory::primitive_desc>& intPD,
size_t inputIdx) {
cvtInVal_ = nullptr;
extInVal_ = nullptr;
in = nullptr;
CHECK_GT(bs_ * ic_ * ih_ * iw_, 0);
auto extPD = MKLDNNMatrix::createPrimitiveDesc(
{bs_, ic_, ih_, iw_}, format::nchw, engine_);
const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
const MatrixPtr& inMat = inputLayers_[inputIdx]->getOutputValue();
in = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
CHECK_EQ(inputIsOnlyMKLDNN(), in != nullptr);
if (in == nullptr || in->getFormat() == format::nc) {
@ -216,11 +218,12 @@ void MKLDNNLayer::resetOutValue(MKLDNNMatrixPtr& out,
}
void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in,
memory::primitive_desc intPD) {
memory::primitive_desc intPD,
size_t inputIdx) {
cvtInGrad_ = nullptr;
extInGrad_ = nullptr;
in = nullptr;
LayerPtr& input = inputLayers_[0];
LayerPtr& input = inputLayers_[inputIdx];
if (input->getOutputGrad() == nullptr) {
// no need input grad
return;
@ -245,7 +248,6 @@ void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in,
return;
}
// need create reorder
// TODO(TJ): add macro definition to simplify it
CHECK(extInVal_ != nullptr && isPaddleFormat(extInVal_->getFormat()))
<< "should have external input value and the format must be nchw(nc)";
extInGrad_ = MKLDNNMatrix::create(extInVal_->getPrimitiveDesc(), inMat);

@ -199,7 +199,8 @@ protected:
*/
void resetInValue(
MKLDNNMatrixPtr& in,
const std::shared_ptr<mkldnn::memory::primitive_desc>& intPD = nullptr);
const std::shared_ptr<mkldnn::memory::primitive_desc>& intPD = nullptr,
size_t inputIdx = 0);
/**
* reset output value from internal primitive desc.
@ -212,7 +213,9 @@ protected:
* reset input grad from internal primitive desc.
* reset both internal and external buffer and create reorder if necessary.
*/
void resetInGrad(MKLDNNMatrixPtr& in, mkldnn::memory::primitive_desc intPD);
void resetInGrad(MKLDNNMatrixPtr& in,
mkldnn::memory::primitive_desc intPD,
size_t inputIdx = 0);
/**
* reset output grad from internal primitive desc.

@ -98,8 +98,19 @@ void SubSequenceLayer::forward(PassType passType) {
CHECK_EQ(numSequences2, numSequences3);
MatrixPtr inputValue = input.value;
IVectorPtr offsetValue = offsetSeq.ids;
IVectorPtr sizeValue = sizeSeq.ids;
IVectorPtr offsetValue;
IVectorPtr sizeValue;
if (useGpu_) {
// copy to cpu
IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false);
IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false);
offsetValue->copyFrom(*offsetSeq.ids);
sizeValue->copyFrom(*sizeSeq.ids);
} else {
offsetValue = offsetSeq.ids;
sizeValue = sizeSeq.ids;
}
CHECK_EQ(offsetValue->getSize(), numSequences1);
CHECK_EQ(sizeValue->getSize(), numSequences1);
@ -176,8 +187,21 @@ void SubSequenceLayer::backward(const UpdateCallback& callback) {
size_t numSequences1 = startPositions1->getSize() - 1;
const int* starts1 = startPositions1->getData();
IVectorPtr offsetValue = getInput(1).ids;
IVectorPtr sizeValue = getInput(2).ids;
const Argument& offsetSeq = getInput(1);
const Argument& sizeSeq = getInput(2);
IVectorPtr offsetValue;
IVectorPtr sizeValue;
if (useGpu_) {
// copy to cpu
IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false);
IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false);
offsetValue->copyFrom(*offsetSeq.ids);
sizeValue->copyFrom(*sizeSeq.ids);
} else {
offsetValue = offsetSeq.ids;
sizeValue = sizeSeq.ids;
}
int* offsets = offsetValue->getData();
int* sizes = sizeValue->getData();

@ -132,7 +132,7 @@ void MKLDNNTester::checkForward() {
VLOG(MKLDNN_TESTS) << "Check Forward";
printTopDatas();
double delta =
compareMatrix(dnnLayer_->getOutputValue(), refLayer_->getOutputValue());
compareMatrix(refLayer_->getOutputValue(), dnnLayer_->getOutputValue());
EXPECT_LE(fabs(delta), eps_);
}
@ -147,7 +147,7 @@ void MKLDNNTester::checkBackwardData() {
VLOG(MKLDNN_ALL) << "Reference Backward Result: InputGrad " << i;
printMatrix(refDiff);
double delta = compareMatrix(dnnDiff, refDiff);
double delta = compareMatrix(refDiff, dnnDiff);
EXPECT_LE(fabs(delta), eps_);
if (isBN) {
// the other two inputs in batch norm are for moving mean and var
@ -177,7 +177,7 @@ void MKLDNNTester::checkBackwardWgts() {
<< parameters_[REF][i]->getName();
printVector(ref);
double delta = compareVector(dnn, ref);
double delta = compareVector(ref, dnn);
EXPECT_LE(fabs(delta), eps_);
}

@ -271,20 +271,53 @@ TEST(MKLDNNLayer, BatchNormLayer) {
testBatchNormLayer({16, 32, 16, 16});
}
struct testActDesc {
struct testImageDesc {
int bs, ic, ih, iw;
};
static void getAddtoConfig(TestConfig& cfg, const testActDesc& pm) {
static void getAddtoConfig(TestConfig& cfg,
const testImageDesc& pm,
const size_t nInputs = 1) {
cfg.biasSize = 0;
cfg.layerConfig.set_type("addto");
size_t layerSize = pm.ic * pm.ih * pm.iw;
cfg.layerConfig.set_size(layerSize);
cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0});
cfg.layerConfig.add_inputs();
cfg.layerConfig.set_active_type("relu");
for (size_t i = 0; i < nInputs; ++i) {
std::stringstream ss;
ss << "layer_" << i;
cfg.inputDefs.push_back({INPUT_DATA, ss.str(), layerSize, 0});
LayerInputConfig* input = cfg.layerConfig.add_inputs();
ImageConfig* img_conf = input->mutable_image_conf();
img_conf->set_channels(pm.ic);
img_conf->set_img_size_y(pm.ih);
img_conf->set_img_size(pm.iw);
}
}
void testAddtoLayer(const testImageDesc& pm, const size_t nInputs) {
CHECK_GE(nInputs, 1);
TestConfig dnnConfig;
getAddtoConfig(dnnConfig, pm, nInputs);
dnnConfig.layerConfig.set_type("mkldnn_addto");
// TODO(TJ): test with bias
for (auto withBias : {false}) {
if (withBias) {
dnnConfig.biasSize = pm.ic * pm.ih * pm.iw;
} else {
dnnConfig.biasSize = 0;
}
RUN_MKLDNN_TEST_LAYER(dnnConfig, "addto", pm)
}
}
TEST(MKLDNNLayer, AddtoLayer) {
testAddtoLayer({16, 5, 14, 14}, 1);
testAddtoLayer({8, 10, 8, 8}, 2);
testAddtoLayer({4, 12, 1, 1}, 3);
}
void testActivation(std::string actType, const testActDesc& pm) {
void testActivation(std::string actType, const testImageDesc& pm) {
// TODO(TJ): remove me when paddle support elu activation
if (actType == "mkldnn_elu") {
return;

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

Loading…
Cancel
Save