m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge branch 'develop' of https://github.com/paddlepaddle/paddle into restructure-attribute

Browse Source
cblas_new
Yi Wang 8 years ago
parent cedc2e645e ef8de51587
commit 949e8f53f9
44 changed files with 626 additions and 87 deletions
Show Stats Download Patch File Download Diff File

4
CMakeLists.txt
Unescape View File

@ -36,8 +36,8 @@ include(simd)
################################ Configurations #######################################
option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." ${AVX_FOUND})
option(WITH_MKLML "Compile PaddlePaddle with mklml package." ${AVX_FOUND})
option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." OFF)
option(WITH_MKLML "Compile PaddlePaddle with mklml package." OFF)
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" ON)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)

16
cmake/generic.cmake
Unescape View File

@ -187,7 +187,13 @@ function(cc_library TARGET_NAME)
endif()
# cpplint code style
add_style_check_target(${TARGET_NAME} ${cc_library_SRCS})
foreach(source_file ${cc_library_SRCS})
string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
endif()
endforeach()
add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS})
else(cc_library_SRCS)
if (cc_library_DEPS)
@ -239,6 +245,14 @@ function(nv_library TARGET_NAME)
add_dependencies(${TARGET_NAME} ${nv_library_DEPS})
target_link_libraries(${TARGET_NAME} ${nv_library_DEPS})
endif()
# cpplint code style
foreach(source_file ${nv_library_SRCS})
string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
endif()
endforeach()
add_style_check_target(${TARGET_NAME} ${nv_library_SRCS} ${nv_library_HEADERS})
else(nv_library_SRCS)
if (nv_library_DEPS)
merge_static_libs(${TARGET_NAME} ${nv_library_DEPS})

9
paddle/framework/ddim.h
Unescape View File

@ -25,18 +25,15 @@ limitations under the License. */
namespace paddle {
namespace framework {
namespace {
typedef boost::variant<Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>,
Dim<8>, Dim<9>>
DDimVar;
}
/**
* \brief A dynamically sized dimension.
*
* The number of dimensions must be between [1, 9].
*/
struct DDim {
typedef boost::variant<Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>,
Dim<8>, Dim<9>>
DDimVar;
DDimVar var;
DDim() : var(Dim<1>()) {}

14
paddle/framework/grad_op_builder.h
Unescape View File

@ -1,3 +1,17 @@
/* 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. */
#pragma once
#include "paddle/framework/operator.h"

6
paddle/framework/op_registry.h
Unescape View File

@ -271,7 +271,7 @@ class OpRegistry {
static std::unordered_map<std::string, OpProto>& protos() {
static std::unordered_map<std::string, OpProto> protos_;
return protos_;
};
}
static std::unordered_map<std::string, std::string>& grad_ops() {
static std::unordered_map<std::string, std::string> grad_ops_;
@ -293,7 +293,7 @@ class OpRegistry {
static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
return op_checkers_;
};
}
static void GenerateTempVariableName(OperatorBase* op) {
static std::atomic<size_t> gUniqId(0UL);
@ -310,7 +310,7 @@ class OpRegistry {
template <typename OpType, typename ProtoMakerType>
class OpRegisterHelper {
public:
OpRegisterHelper(const char* op_type) {
explicit OpRegisterHelper(const char* op_type) {
OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
}
};

2
paddle/framework/operator.h
Unescape View File

@ -284,7 +284,7 @@ class OperatorWithKernel : public OperatorBase {
platform::Place place_;
OpKernelKey() = default;
OpKernelKey(const platform::DeviceContext& dev_ctx) {
explicit OpKernelKey(const platform::DeviceContext& dev_ctx) {
place_ = dev_ctx.GetPlace();
}

11
paddle/framework/pybind.cc
Unescape View File

@ -105,7 +105,16 @@ PYBIND11_PLUGIN(core) {
.def("set", PyCUDATensorSetFromArray<float>)
.def("set", PyCUDATensorSetFromArray<int>)
#endif
.def("shape", [](Tensor &self) { return vectorize(self.dims()); });
.def("shape", [](Tensor &self) { return vectorize(self.dims()); })
.def("set_float_element",
[](Tensor &self, size_t offset, float f) {
// TODO(yuyang18): Only support GPU now.
self.data<float>()[offset] = f;
})
.def("get_float_element", [](Tensor &self, size_t offset) -> float {
// TODO(yuyang18): Only support GPU now.
return self.data<float>()[offset];
});
py::class_<Variable>(m, "Variable", R"DOC(Variable Class.

26
paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
Unescape View File

@ -967,8 +967,9 @@ void RecurrentGradientMachine::generateSequence() {
size_t numSequences = getGenBatchSize();
resizeBootFrame(numSequences);
// We create only two sub-network in generation for alternate use.
// Thus, we can reduce total memory of output_ in layer forward.
// We create only two sub-network in generation, one stores states of all
// layers in previous time step and the other storing the states at current
// time step.
resizeOrCreateFrames(2);
// outFrameLines_.size() > 1UL
@ -1001,10 +1002,9 @@ void RecurrentGradientMachine::generateSequence() {
// init outArg
size_t resultNum = generator_.config.num_results_per_sample();
IVector::resizeOrCreate(
generator_.outArg.ids,
generator_.config.max_num_frames() * numSequences * resultNum,
false);
size_t maxGenWordCount =
generator_.config.max_num_frames() * numSequences * resultNum;
IVector::resizeOrCreate(generator_.outArg.ids, maxGenWordCount, false);
if (resultNum > 1) {
CHECK_LE(resultNum, static_cast<size_t>(generator_.config.beam_size()));
Matrix::resizeOrCreate(generator_.outArg.in,
@ -1012,6 +1012,11 @@ void RecurrentGradientMachine::generateSequence() {
/* width */ resultNum,
false,
/* useGpu */ false);
Matrix::resizeOrCreate(generator_.outArg.value,
/* height */ maxGenWordCount,
/* width */ 1,
false,
/* useGpu */ false);
}
ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
numSequences + 1,
@ -1313,13 +1318,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
starts[0] = 0;
if (numResults > 1) {
real* probs = generator_.outArg.in->getData();
real* idsProb = generator_.outArg.value->getData();
size_t curPos = 0;
for (size_t i = 0; i < finalPaths_.size(); ++i) {
for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
Path& path = finalPaths_[i][j];
generator_.ids.push_back(path.ids.size()); // sequence size
size_t genLen = path.ids.size();
generator_.ids.push_back(genLen); // sequence size
generator_.ids.insert(
generator_.ids.end(), path.ids.begin(), path.ids.end());
generator_.ids.push_back(-1); // end of sequence
memcpy(idsProb + curPos, path.idsProb.data(), sizeof(real) * genLen);
curPos += genLen;
idsProb[curPos++] = -1.0;
probs[i * numResults + j] = path.logProb;
if (!j && dataArgsSize_) {

12
paddle/gserver/gradientmachines/RecurrentGradientMachine.h
Unescape View File

@ -189,6 +189,11 @@ public:
*/
std::vector<int> ids;
/**
* @brief idsProb, log probability of each generated words.
*/
std::vector<real> idsProb;
/**
* @brief logProb, current probability of path.
*/
@ -228,11 +233,13 @@ public:
*/
Path(Path& old, int newId, real logProb, int machineId, int topIndex)
: ids(old.ids),
idsProb(old.idsProb),
logProb(old.logProb + logProb),
machineId(machineId),
topIndex(topIndex),
seqId(old.seqId) {
ids.push_back(newId);
idsProb.push_back(logProb);
if (!old.probHistory.empty()) {
this->probHistory = old.probHistory;
// probHistory store current prob, not sum
@ -411,8 +418,9 @@ protected:
struct Generator {
GeneratorConfig config;
std::vector<int> ids; // store generated sequences
Argument outArg; // final output argument
std::vector<int> ids; // store generated sequences
std::vector<real> idsProb; // log probability of each generated word
Argument outArg; // final output argument
};
bool generating_;
Generator generator_;

14
paddle/gserver/tests/LayerGradUtil.cpp
Unescape View File

@ -400,7 +400,6 @@ void initDataLayer(TestConfig testConf,
const std::vector<int>& labelSeqStartPositions =
testConf.inputDefs[i].labelSeqStartPositions;
if (labelSeqStartPositions.size() != 0) {
CHECK(!sequenceStartPositions);
CHECK_GE(static_cast<int>(labelSeqStartPositions.size()), 2);
sequenceStartPositions =
@ -410,6 +409,19 @@ void initDataLayer(TestConfig testConf,
useGpu);
data.sequenceStartPositions = sequenceStartPositions;
}
const std::vector<int>& labelSubSeqStartPositions =
testConf.inputDefs[i].labelSubSeqStartPositions;
if (labelSubSeqStartPositions.size() != 0) {
CHECK_GE(static_cast<int>(labelSubSeqStartPositions.size()), 2);
subSequenceStartPositions =
ICpuGpuVector::create(labelSubSeqStartPositions.size(), useGpu);
subSequenceStartPositions->copyFrom(labelSubSeqStartPositions.data(),
labelSubSeqStartPositions.size(),
useGpu);
data.subSequenceStartPositions = subSequenceStartPositions;
}
break;
}
default:

5
paddle/gserver/tests/LayerGradUtil.h
Unescape View File

@ -67,6 +67,7 @@ struct InputDef {
bool isStatic;
std::vector<int> labelInitValue;
std::vector<int> labelSeqStartPositions;
std::vector<int> labelSubSeqStartPositions;
MatrixPtr selfDefinedData;
InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) {
@ -81,8 +82,10 @@ struct InputDef {
InputDef(InputType type,
string nameIn,
MatrixPtr selfDefinedData,
std::vector<int> selfDefinedSeqStartPos = {})
std::vector<int> selfDefinedSeqStartPos = {},
std::vector<int> selfDefinedSubSeqStartPos = {})
: labelSeqStartPositions(selfDefinedSeqStartPos),
labelSubSeqStartPositions(selfDefinedSubSeqStartPos),
selfDefinedData(selfDefinedData) {
inputType = type;
name = nameIn;

2
paddle/memory/detail/buddy_allocator.h
Unescape View File

@ -39,7 +39,7 @@ class BuddyAllocator {
public:
void* Alloc(size_t unaligned_size);
void Free(void*);
void Free(void* ptr);
size_t Used();
public:

8
paddle/memory/detail/meta_cache.h
Unescape View File

@ -33,17 +33,17 @@ namespace detail {
*/
class MetadataCache {
public:
MetadataCache(bool uses_gpu);
explicit MetadataCache(bool uses_gpu);
public:
/*! \brief Load the associated metadata for the specified memory block. */
Metadata load(const MemoryBlock*);
Metadata load(const MemoryBlock* memory_block);
/*! \brief Store the associated metadata for the specified memory block. */
void store(MemoryBlock*, const Metadata&);
void store(MemoryBlock* memory_block, const Metadata& meta_data);
/*! \brief Indicate that the specified metadata will no longer be used. */
void invalidate(MemoryBlock*);
void invalidate(MemoryBlock* memory_block);
public:
MetadataCache(const MetadataCache&) = delete;

2
paddle/memory/memory.h
Unescape View File

@ -68,7 +68,7 @@ class PODDeleter {
static_assert(std::is_pod<T>::value, "T must be POD");
public:
PODDeleter(Place place) : place_(place) {}
explicit PODDeleter(Place place) : place_(place) {}
void operator()(T* ptr) { Free(place_, static_cast<void*>(ptr)); }
private:

14
paddle/operators/add_op.cu
Unescape View File

@ -1,3 +1,17 @@
/* 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. */
#define EIGEN_USE_GPU
#include "paddle/framework/op_registry.h"
#include "paddle/operators/add_op.h"

16
paddle/operators/cross_entropy_op.cu
Unescape View File

@ -1,5 +1,19 @@
/* 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. */
#define EIGEN_USE_GPU
#include "paddle/operators/cross_entropy_op.h"
REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
ops::OnehotCrossEntropyOpKernel<ops::GPUPlace, float>);
ops::OnehotCrossEntropyOpKernel<ops::GPUPlace, float>);

16
paddle/operators/fill_zeros_like_op.cu
Unescape View File

@ -1,6 +1,20 @@
/* 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. */
#include "paddle/framework/op_registry.h"
#include "paddle/operators/fill_zeros_like_op.h"
REGISTER_OP_GPU_KERNEL(
fill_zeros_like,
paddle::operators::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);
paddle::operators::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);

16
paddle/operators/mean_op.cu
Unescape View File

@ -1,6 +1,20 @@
/* 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. */
#define EIGEN_USE_GPU
#include "paddle/operators/mean_op.h"
REGISTER_OP_GPU_KERNEL(mean, ops::MeanKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::GPUPlace, float>);

2
paddle/operators/mul_op.cu
Unescape View File

@ -15,4 +15,4 @@
#define EIGEN_USE_GPU
#include "paddle/operators/mul_op.h"
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);

20
paddle/operators/recurrent_op.h
Unescape View File

@ -19,7 +19,7 @@
namespace paddle {
namespace operators {
using namespace paddle::framework;
using namespace paddle::framework; // NOLINT
namespace rnn {
@ -94,7 +94,7 @@ void InitArgument(const ArgumentName& name, Argument* arg);
}; // namespace rnn
// The sequence format in RecurrentOp is Tensor<seq_len, batch_size, dim> now.
// TODO:
// TODO(Yan Chunwei):
// 1. No-padding computing for sequences with indifinite length in one batch.
// 2. Hierarchical RNN for sequence with sub-sequence.
// 3. Internal Memory.
@ -172,12 +172,10 @@ public:
/**
* InferShape must be called before Run.
*/
virtual void InferShape(const Scope& scope) const override {
alg_.InferShape(scope);
}
void InferShape(const Scope& scope) const override { alg_.InferShape(scope); }
virtual void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {
alg_.Run(scope, dev_ctx);
}
@ -194,12 +192,10 @@ public:
/**
* InferShape must be called before Run.
*/
virtual void InferShape(const Scope& scope) const override {
alg_.InferShape(scope);
}
void InferShape(const Scope& scope) const override { alg_.InferShape(scope); }
virtual void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {
alg_.Run(scope, dev_ctx);
}

14
paddle/operators/rowwise_add_op.cu
Unescape View File

@ -1,3 +1,17 @@
/* 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. */
#define EIGEN_USE_GPU
#include "paddle/operators/rowwise_add_op.h"

16
paddle/operators/sgd_op.cu
Unescape View File

@ -1,4 +1,18 @@
/* 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. */
#define EIGEN_USE_GPU
#include "paddle/operators/sgd_op.h"
REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);

14
paddle/operators/sigmoid_op.cu
Unescape View File

@ -1,3 +1,17 @@
/* 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. */
#define EIGEN_USE_GPU
#include "paddle/operators/sigmoid_op.h"

17
paddle/operators/softmax_op.cu
Unescape View File

@ -1,6 +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. */
#define EIGEN_USE_GPU
#include "paddle/framework/op_registry.h"
#include "paddle/operators/softmax_op.h"
REGISTER_OP_GPU_KERNEL(softmax, ops::SoftmaxKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(softmax_grad, ops::SoftmaxGradKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(softmax_grad,
ops::SoftmaxGradKernel<ops::GPUPlace, float>);

8
paddle/platform/device_context.h
Unescape View File

@ -40,7 +40,7 @@ class DeviceContext {
class CPUDeviceContext : public DeviceContext {
public:
CPUDeviceContext();
CPUDeviceContext(CPUPlace);
explicit CPUDeviceContext(CPUPlace);
virtual ~CPUDeviceContext() {}
Eigen::DefaultDevice* eigen_device() const;
@ -55,7 +55,7 @@ class CPUDeviceContext : public DeviceContext {
class CUDADeviceContext : public DeviceContext {
public:
explicit CUDADeviceContext(GPUPlace);
CUDADeviceContext(GPUPlace); // NOLINT
virtual ~CUDADeviceContext();
/*! \brief Wait for all operations completion in the stream. */
@ -69,10 +69,10 @@ class CUDADeviceContext : public DeviceContext {
// clang-format off
/*! \brief Return cublas handle in the device context. */
cublasHandle_t cublas_handle ();
cublasHandle_t cublas_handle();
/*! \brief Return cudnn handle in the device context. */
cudnnHandle_t cudnn_handle ();
cudnnHandle_t cudnn_handle();
/*! \brief Return curand handle in the device context. */
curandGenerator_t curand_generator();

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

Write Preview
Loading…
Cancel
Save