Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 9e3a9eb2 编写于 作者: Q qijun
浏览文件
操作

merge baidu/develop

上级 d649dbf4 c40707b6
隐藏空白更改
内联 并排
Showing with 2583 addition and 655 deletion
Dockerfile
浏览文件 @ 9e3a9eb2
......@@ -25,7 +25,7 @@ COPY ./paddle/scripts/docker/root/ /root/
RUN apt-get update && \
apt-get install -y \
git python-pip python-dev openssh-server bison \
wget unzip tar xz-utils bzip2 gzip coreutils ntp \
wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
curl sed grep graphviz libjpeg-dev zlib1g-dev \
python-numpy python-matplotlib gcc g++ \
automake locales clang-format-3.8 swig doxygen cmake \
......
cmake/configure.cmake
浏览文件 @ 9e3a9eb2
......@@ -102,12 +102,19 @@ if(WITH_GOLANG)
message(FATAL_ERROR "no glide executeble found: $ENV{GOPATH}/bin/glide")
endif()
add_custom_target(go_vendor)
add_custom_command(TARGET go_vendor
# this command will only run when the file it depends is missing
# or has changed, or the output is missing.
add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide
COMMAND env GOPATH=${GOPATH} ${GLIDE} install
COMMAND touch ${CMAKE_BINARY_DIR}/glide
DEPENDS ${PROJ_ROOT}/go/glide.lock
WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go"
)
add_dependencies(go_vendor go_path)
)
# depends on the custom command which outputs
# ${CMAKE_BINARY_DIR}/glide, the custom command does not need to
# run every time this target is built.
add_custom_target(go_vendor DEPENDS ${CMAKE_BINARY_DIR}/glide go_path)
endif()
endif(WITH_GOLANG)
cmake/cpplint.cmake
浏览文件 @ 9e3a9eb2
......@@ -27,7 +27,8 @@ set(IGNORE_PATTERN
.*cblas\\.h.*
.*\\.pb\\.txt
.*LtrDataProvider.*
.*MultiDataProvider.*)
.*MultiDataProvider.*
.*pb.*)
# add_style_check_target
#
......@@ -52,14 +53,13 @@ macro(add_style_check_target TARGET_NAME)
endif()
endforeach()
if(LINT MATCHES ON)
# cpplint code style
get_filename_component(base_filename ${filename} NAME)
set(CUR_GEN ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.cpplint)
add_custom_command(OUTPUT ${CUR_GEN}
PRE_BUILD
COMMAND env ${py_env} "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
"--filter=${STYLE_FILTER}"
"--write-success=${CUR_GEN}" ${filename}
DEPENDS ${filename}
add_custom_command(TARGET ${TARGET_NAME} PRE_BUILD
COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
"--filter=${STYLE_FILTER}"
"--write-success=${CUR_GEN}" ${filename}
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
endif()
endforeach()
......
cmake/generic.cmake
浏览文件 @ 9e3a9eb2
......@@ -104,6 +104,7 @@ function(merge_static_libs TARGET_NAME)
foreach(lib ${libs})
list(APPEND libs_deps ${${lib}_LIB_DEPENDS})
endforeach()
list(REMOVE_DUPLICATES libs_deps)
if(APPLE) # Use OSX's libtool to merge archives
# To produce a library we need at least one source file.
......@@ -127,7 +128,7 @@ function(merge_static_libs TARGET_NAME)
# Get the file names of the libraries to be merged
set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
endforeach()
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles})
else() # general UNIX: use "ar" to extract objects and re-add to a common lib
......@@ -145,11 +146,11 @@ function(merge_static_libs TARGET_NAME)
DEPENDS ${lib} ${objdir}
WORKING_DIRECTORY ${objdir})
# Empty dummy source file that goes into merged library
set(mergebase ${lib}.mergebase.c)
add_custom_command(OUTPUT ${mergebase}
COMMAND ${CMAKE_COMMAND} -E touch ${mergebase}
DEPENDS ${objlistfile})
# Empty dummy source file that goes into merged library
set(mergebase ${lib}.mergebase.c)
add_custom_command(OUTPUT ${mergebase}
COMMAND ${CMAKE_COMMAND} -E touch ${mergebase}
DEPENDS ${objlistfile})
list(APPEND mergebases "${mergebase}")
endforeach()
......@@ -184,6 +185,10 @@ function(cc_library TARGET_NAME)
add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
endif()
# cpplint code style
add_style_check_target(${TARGET_NAME} ${cc_library_SRCS})
else(cc_library_SRCS)
if (cc_library_DEPS)
merge_static_libs(${TARGET_NAME} ${cc_library_DEPS})
......@@ -337,7 +342,7 @@ function(go_test TARGET_NAME)
string(REPLACE "${PADDLE_GO_PATH}" "" CMAKE_CURRENT_SOURCE_REL_DIR ${CMAKE_CURRENT_SOURCE_DIR})
add_custom_target(${TARGET_NAME} ALL DEPENDS go_vendor ${go_test_DEPS})
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} test
COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} test -race
-c -o "${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}"
".${CMAKE_CURRENT_SOURCE_REL_DIR}"
WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go")
......
doc/api/v2/config/layer.rst
浏览文件 @ 9e3a9eb2
......@@ -474,6 +474,11 @@ prelu
.. autoclass:: paddle.v2.layer.prelu
:noindex:
gated_unit
-----------
.. autoclass:: paddle.v2.layer.gated_unit
:noindex:
Detection output Layer
======================
......
go/pserver/client/client_test.go
浏览文件 @ 9e3a9eb2
......@@ -164,7 +164,7 @@ func testClient(t *testing.T, c *client.Client) {
wg.Add(1)
go func(gs []pserver.Gradient) {
err = c.SendGrads(gs)
err := c.SendGrads(gs)
if err != nil {
t.Fatal(err)
}
......
paddle/api/ConfigParser.cpp
浏览文件 @ 9e3a9eb2
......@@ -64,11 +64,7 @@ ModelConfig* TrainerConfig::getModelConfig() const {
ParameterConfig::ParameterConfig() : m(new ParameterConfigPrivate()) {}
ParameterConfig::~ParameterConfig() {
if (m) {
delete m;
}
}
ParameterConfig::~ParameterConfig() { delete m; }
ParameterConfig* ParameterConfig::createParameterConfigFromParameterSharedPtr(
void* ptr) {
......@@ -98,11 +94,7 @@ void* ParameterConfig::getRawPtr() { return m->getConfigPtr(); }
OptimizationConfig::OptimizationConfig() : m(new OptimizationConfigPrivate()) {}
OptimizationConfig::~OptimizationConfig() {
if (m) {
delete m;
}
}
OptimizationConfig::~OptimizationConfig() { delete m; }
std::string OptimizationConfig::toProtoString() {
return m->getConfig().SerializeAsString();
......
paddle/api/ParameterOptimizer.cpp
浏览文件 @ 9e3a9eb2
......@@ -53,11 +53,7 @@ struct ParameterTraverseCallbackPrivate {
ParameterOptimizer::ParameterOptimizer() : m(new ParameterOptimizerPrivate()) {}
ParameterOptimizer::~ParameterOptimizer() {
if (m) {
delete m;
}
}
ParameterOptimizer::~ParameterOptimizer() { delete m; }
ParameterOptimizer* ParameterOptimizer::create(OptimizationConfig* config) {
CHECK(config != nullptr);
......@@ -104,11 +100,7 @@ std::vector<int> ParameterOptimizer::getParameterTypes() const {
ParameterTraverseCallback::ParameterTraverseCallback()
: m(new ParameterTraverseCallbackPrivate()) {}
ParameterTraverseCallback::~ParameterTraverseCallback() {
if (m) {
delete m;
}
}
ParameterTraverseCallback::~ParameterTraverseCallback() { delete m; }
void ParameterTraverseCallback::apply(const std::vector<Vector*>& vecs,
const ParameterConfig& conf,
......
paddle/api/Vector.cpp
浏览文件 @ 9e3a9eb2
......@@ -171,11 +171,7 @@ struct VectorPrivate {
Vector::Vector() : m(new VectorPrivate()) {}
Vector::~Vector() {
if (m) {
delete m;
}
}
Vector::~Vector() { delete m; }
Vector* Vector::createZero(size_t sz, bool useGpu) {
auto retVec = new Vector();
......
paddle/framework/CMakeLists.txt
浏览文件 @ 9e3a9eb2
......@@ -11,8 +11,10 @@ proto_library(op_proto SRCS op_proto.proto DEPS attr_type)
cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
proto_library(op_desc SRCS op_desc.proto DEPS attr_type)
cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
cc_library(operator SRCS operator.cc DEPS op_desc device_context)
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto op_desc)
cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry operator)
py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.proto)
......@@ -21,4 +23,5 @@ add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch
add_dependencies(framework_py_proto framework_py_proto_init)
proto_library(net_proto SRCS net_proto.proto DEPS op_proto)
cc_library(net SRCS net.cc DEPS net_proto)
cc_library(net SRCS net.cc DEPS operator net_proto op_registry)
cc_test(net_op_test SRCS net_op_test.cc DEPS net)
paddle/framework/ddim.cc
浏览文件 @ 9e3a9eb2
/* 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/ddim.h"
#include "paddle/framework/enforce.h"
namespace paddle {
namespace framework {
///@cond HIDDEN
/// @cond HIDDEN
template <int i>
Dim<i> make_dim(const int* d) {
......@@ -51,7 +64,7 @@ void make_ddim(DDim& ddim, const int* dims, int n) {
}
}
///@endcond
/// @endcond
DDim make_ddim(std::initializer_list<int> dims) {
DDim result(make_dim(0));
......@@ -65,11 +78,11 @@ DDim make_ddim(const std::vector<int>& dims) {
return result;
}
///@cond HIDDEN
/// @cond HIDDEN
// XXX For some reason, putting this in an anonymous namespace causes errors
class DynamicMutableIndexer : public boost::static_visitor<int&> {
public:
DynamicMutableIndexer(int idx) : idx_(idx) {}
explicit DynamicMutableIndexer(int idx) : idx_(idx) {}
template <int D>
int& operator()(Dim<D>& dim) const {
......@@ -82,7 +95,7 @@ class DynamicMutableIndexer : public boost::static_visitor<int&> {
class DynamicConstIndexer : public boost::static_visitor<int> {
public:
DynamicConstIndexer(int idx) : idx_(idx) {}
explicit DynamicConstIndexer(int idx) : idx_(idx) {}
template <int D>
int operator()(const Dim<D>& dim) const {
......@@ -93,7 +106,7 @@ class DynamicConstIndexer : public boost::static_visitor<int> {
int idx_;
};
///@endcond
/// @endcond
int& DDim::operator[](int idx) {
return boost::apply_visitor(DynamicMutableIndexer(idx), var);
......@@ -156,11 +169,11 @@ int get(const DDim& ddim, int idx) { return ddim[idx]; }
void set(DDim& ddim, int idx, int value) { ddim[idx] = value; }
///@cond HIDDEN
/// @cond HIDDEN
struct VectorizeVisitor : public boost::static_visitor<> {
std::vector<int>& vector;
VectorizeVisitor(std::vector<int>& v) : vector(v) {}
explicit VectorizeVisitor(std::vector<int>& v) : vector(v) {}
template <typename T>
void operator()(const T& t) {
......@@ -170,7 +183,7 @@ struct VectorizeVisitor : public boost::static_visitor<> {
void operator()(const Dim<1>& t) { vector.push_back(t.head); }
};
///@endcond
/// @endcond
std::vector<int> vectorize(const DDim& ddim) {
std::vector<int> result;
......@@ -188,7 +201,7 @@ ssize_t product(const DDim& ddim) {
return result;
}
///\cond HIDDEN
/// \cond HIDDEN
struct ArityVisitor : boost::static_visitor<int> {
template <int D>
......@@ -197,15 +210,15 @@ struct ArityVisitor : boost::static_visitor<int> {
}
};
///\endcond
/// \endcond
int arity(const DDim& d) { return boost::apply_visitor(ArityVisitor(), d); }
///\cond HIDDEN
/// \cond HIDDEN
struct DDimPrinter : boost::static_visitor<void> {
std::ostream& os;
DDimPrinter(std::ostream& os_) : os(os_) {}
explicit DDimPrinter(std::ostream& os_) : os(os_) {}
template <typename T>
void operator()(const T& t) {
......@@ -213,7 +226,7 @@ struct DDimPrinter : boost::static_visitor<void> {
}
};
///\endcond
/// \endcond
std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
DDimPrinter printer(os);
......@@ -221,16 +234,5 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
return os;
}
template <int NDIMS>
Eigen::DSizes<Eigen::DenseIndex, NDIMS> ToEigenDSizes(const DDim& dims) {
int rank = arity(dims);
PADDLE_ENFORCE(rank == NDIMS, "DDim and NDIMS must be same");
Eigen::DSizes<Eigen::DenseIndex, NDIMS> dsizes;
for (int d = 0; d < rank; d++) {
dsizes[d] = dims[d];
}
return dsizes;
}
} // namespace framework
} // namespace paddle
paddle/framework/ddim.h
浏览文件 @ 9e3a9eb2
/* 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 <boost/variant.hpp>
#include <initializer_list>
#include <stdexcept>
#include <vector>
#include "paddle/framework/dim.h"
#include "paddle/framework/enforce.h"
#include "unsupported/Eigen/CXX11/Tensor"
namespace paddle {
......@@ -28,7 +42,7 @@ struct DDim {
DDim() : var(Dim<1>()) {}
template <int D>
DDim(const Dim<D>& in) : var(in) {}
explicit DDim(const Dim<D>& in) : var(in) {}
template <int D>
DDim& operator=(const Dim<D>& in) {
......@@ -93,7 +107,15 @@ int arity(const DDim& ddim);
std::ostream& operator<<(std::ostream&, const DDim&);
template <int NDIMS>
Eigen::DSizes<Eigen::DenseIndex, NDIMS> ToEigenDSizes(const DDim& dims);
Eigen::DSizes<Eigen::DenseIndex, NDIMS> ToEigenDSizes(const DDim& dims) {
int rank = arity(dims);
PADDLE_ENFORCE(rank == NDIMS, "DDim and NDIMS must be same");
Eigen::DSizes<Eigen::DenseIndex, NDIMS> dsizes;
for (int d = 0; d < rank; d++) {
dsizes[d] = dims[d];
}
return dsizes;
}
} // namespace framework
} // namespace paddle
......
paddle/framework/net.cc
浏览文件 @ 9e3a9eb2
/*
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/net.h"
namespace paddle {
namespace framework {
PlainNet::PlainNet(const NetDesc& def) {}
void PlainNet::InferShape(Scope* scope) {
void PlainNet::CompleteAddOp() {
std::unordered_set<std::string> input_set;
std::unordered_set<std::string> output_set;
std::unordered_set<std::string> temp_output;
for (auto& op : ops_) {
op.InferShape();
for (auto& ipt : op->inputs_) {
if (!Contains(output_set, ipt)) { // Not other op's output
input_set.insert(ipt);
} else {
temp_output.insert(ipt);
}
}
for (auto& opt : op->outputs_) {
output_set.insert(opt);
}
}
}
inputs_.reserve(input_set.size());
std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs_));
void PlainNet::Run(std::shared_ptr<Scope> scope, DeviceContext* ctx) {
for (auto& op : ops_) {
op.Run(ctx);
outputs_.reserve(output_set.size());
std::vector<int> tmp_index;
tmp_index.reserve(temp_output.size());
int idx = 0;
for (auto& opt : output_set) {
if (Contains(temp_output, opt)) {
tmp_index.push_back(idx);
}
outputs_.push_back(opt);
++idx;
}
attrs_["temporary_index"] = tmp_index;
add_op_done_ = true;
}
} // namespace framework
} // namespace paddle
paddle/framework/net.h
浏览文件 @ 9e3a9eb2
/* 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
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
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. */
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/op_desc.pb.h>
#include <paddle/framework/operator.h>
#include "paddle/framework/net_proto.pb.h"
#include "paddle/framework/op_proto.pb.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/scope.h"
#include "paddle/platform/device_context.h"
namespace paddle {
namespace framework {
using namespace paddle::platform;
// operator's index stored in a network.
typedef int OpIndex;
/**
* NOTE following codes are some definitions of unimplemented concepts.
* We write some basic implementation to make Net compilable. These APIs will
* keep updating if the concepts related are implemented.
*/
struct OpDesc;
struct OpAttrs {};
class Operator {
public:
Operator(const OpDesc &def) {}
void InferShape() {}
void Run(DeviceContext *ctx) {}
};
/**
* @brief Network that manage the operators it has.
* @brief Network is also a type of Operator
*
* It will manage the operators it has.
*
* Network is the container and controller of a set of operators, user can build
* a real network from a NetDesc which is a protobuf message and use
* Network.Run() * to run all the operators in the network.
* Network is the container and controller of a set of operators.
* A network object knows all Operators belonging to this network. Variables,
* which are inputs and outputs of these operators, are created and managed by a
* hierarchy of Scope objects.
*
* This is the base class of network, all the networks should implement the apis
* This is the base class of network, all the networks should implement the APIs
* it defines.
*/
class Net {
class Net : public OperatorBase {
public:
/**
* @brief Infer shapes of all inputs and outputs of operators.
*/
virtual void InferShape(Scope *scope) = 0;
/**
* @brief Run the network.
*
* Run all the operators and return success(true) or not, with all the
* variables are located in `scope`. `context` describes the detail execution
* environment for ops. `begin` and `end` specify the scope of `ops_` to run,
* If no positive indexes are provided, all operators in `ops_` will run.
*/
virtual void Run(std::shared_ptr<Scope> scope, DeviceContext *ctx) = 0;
/**
* @brief Add an Operator according to `def`.
*/
virtual OpIndex AddOp(const OpProto &def) = 0;
/**
* @brief Add optimizer operators acctording to `attrs`.
*/
virtual void AddOptimizerOps(const OpAttrs &attrs) = 0;
/**
* @brief Add backward operators.
*/
virtual void AddBackwardOps() = 0;
/**
* @brief Create a network.
*/
static std::unique_ptr<Net> Create(const NetDesc &def = NetDesc());
virtual ~Net() {}
virtual void AddOp(const OperatorPtr& op) = 0;
virtual void CompleteAddOp() = 0;
};
using NetPtr = std::shared_ptr<Net>;
/**
* @brief a basic implementation of Net.
*
......@@ -103,18 +55,14 @@ class Net {
class PlainNet : public Net {
public:
/**
* @brief Initialize a PlainNet.
*
* Initialize from a network describe by `def`. NetDesc is the definition of
* a network.
*/
PlainNet(const NetDesc &def);
/**
* Infer all the operators' input and output varialbes' shapes, will be called
* Infer all the operators' input and output variables' shapes, will be called
* before every mini-batch
*/
virtual void InferShape(Scope *scope) override;
void InferShape(const ScopePtr& scope) const override {
for (auto& op : ops_) {
op->InferShape(scope);
}
}
/**
* @brief Run the network.
......@@ -123,48 +71,32 @@ class PlainNet : public Net {
* scope will be used instead. If no OpContext is provicded, default context
* will be used.
*/
virtual void Run(std::shared_ptr<Scope> scope, DeviceContext *ctx) override;
void Run(const ScopePtr& scope,
const platform::DeviceContext& dev_ctx) const override {
for (auto& op : ops_) {
op->Run(scope, dev_ctx);
}
}
/**
* @brief Add an operator to this network.
* @brief Add an operator by ptr
*/
virtual OpIndex AddOp(const OpProto &def) override;
void AddOp(const OperatorPtr& op) override {
PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
ops_.push_back(op);
}
/**
* @brief Add all optimizer operators related into the network.
*/
virtual void AddOptimizerOps(const OpAttrs &attrs) override;
void CompleteAddOp() override;
/**
* @brief Add all backward operators related into the network.
*/
virtual void AddBackwardOps() override;
virtual ~PlainNet() override {}
protected:
/**
* @brief Build the network.
*
* Create operators accordding to `def`, will be called by the constructor.
*/
void BuildNet(const NetDesc &def);
/**
* @brief Add an operator into this network.
*
* Add a operator which is identified as `type` and has attributes described
* in `attrs`, the `inputs` are the keys of readonly input variables,
* `outputs` are keys of mutable output variables. An `OpIndex` will be
* returned to indicate the offset of the new operator in `ops_`.
*/
OpIndex AddOp(const std::string &type, const std::vector<std::string> &inputs,
const std::vector<std::string> &outputs,
const OpAttrs &attrs = OpAttrs());
std::vector<OperatorPtr> ops_;
private:
// the operators owned by `Network`.
std::vector<Operator> ops_;
bool add_op_done_{false};
template <typename T, typename KeyType>
static bool Contains(T container, KeyType key) {
return container.find(key) != container.end();
}
};
} // namespace framework
......
paddle/framework/net_op_test.cc 0 → 100644
浏览文件 @ 9e3a9eb2
#include <gtest/gtest.h>
#include <paddle/framework/net.h>
#include <paddle/framework/op_registry.h>
#include <paddle/framework/operator.h>
namespace pd = paddle::framework;
static int infer_shape_cnt = 0;
static int run_cnt = 0;
class TestOp : public pd::OperatorBase {
public:
void InferShape(const paddle::framework::ScopePtr& scope) const override {
++infer_shape_cnt;
}
void Run(const paddle::framework::ScopePtr& scope,
const paddle::platform::DeviceContext& dev_ctx) const override {
++run_cnt;
}
};
template <typename T>
void AssertSameVectorWithoutOrder(const std::vector<T>& expected,
const std::vector<T>& actual) {
ASSERT_EQ(expected.size(), actual.size());
std::unordered_set<T> expected_set;
for (auto& tmp : expected) {
expected_set.insert(tmp);
}
for (auto& act : actual) {
ASSERT_NE(expected_set.end(), expected_set.find(act));
}
}
TEST(OpKernel, all) {
auto net = std::make_shared<paddle::framework::PlainNet>();
ASSERT_NE(net, nullptr);
auto op1 = std::make_shared<TestOp>();
op1->inputs_ = {"x", "w1", "b1"};
op1->outputs_ = {"y"};
net->AddOp(op1);
auto op2 = std::make_shared<TestOp>();
op2->inputs_ = {"y", "w2", "b2"};
op2->outputs_ = {"z"};
net->AddOp(op2);
net->CompleteAddOp();
AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, net->inputs_);
AssertSameVectorWithoutOrder({"y", "z"}, net->outputs_);
auto tmp_idx_iter = net->attrs_.find("temporary_index");
ASSERT_NE(net->attrs_.end(), tmp_idx_iter);
auto& tmp_idx = boost::get<std::vector<int>>(tmp_idx_iter->second);
ASSERT_EQ(1UL, tmp_idx.size());
ASSERT_EQ("y", net->outputs_[tmp_idx[0]]);
auto scope = std::make_shared<pd::Scope>();
paddle::platform::CPUDeviceContext dev_ctx;
net->InferShape(scope);
net->Run(scope, dev_ctx);
ASSERT_EQ(2, infer_shape_cnt);
ASSERT_EQ(2, run_cnt);
ASSERT_THROW(net->AddOp(op2), paddle::framework::EnforceNotMet);
}
paddle/framework/op_proto.proto
浏览文件 @ 9e3a9eb2
......@@ -34,6 +34,11 @@ message AttrProto {
// Supported attribute comments. It helps 3rd-party language generate doc-string.
required string comment = 3;
// If that attribute is generated, it means the Paddle third language
// binding has responsibility to fill that attribute. End-User should
// not set that attribute.
optional bool generated = 4 [default=false];
}
// Input or output message for 3rd-party language binding.
......@@ -45,6 +50,40 @@ message VarProto {
// The comment for that input. It helps 3rd-party language generate doc-string.
required string comment = 2;
// Is that input/output could be a list or not.
// If so, that Op should write a attributed named `input_format` or
// `output_format`.
//
// e.g.
// If the op is a fc op, the inputs are `X`, `W`, `b`. The `X` and `W`
// could be multiple, so the multiple of `X` and `W` is True, and OpDesc
// will hold a attribute of them.
//
// The Op desc of same fc could be
// {
// "type": "fc",
// "input": ["X1", "X2", "W1", "W2", "b"],
// "output": "fc.out",
// "attrs" : {
// "input_format": [0, 2, 4, 5]
// }
// }
//
optional bool multiple = 3 [default=false];
// It marks that output is a temporary output. That output is not used by
// user, but used by other op internally as input. If other op is not use
// that output, it could be optimized early.
//
// Attribute temporary_index will be set in OpDesc if there is some
// outputs are temporary.
//
// output = [ "xxx.out1", "xxx.tmp", "xxx.out2"],
// attrs = {
// "temporary_index": [1]
// }
optional bool temporary = 4 [default=false];
}
// Op protocol message for 3rd-party language binding.
......
paddle/framework/op_registry.cc
浏览文件 @ 9e3a9eb2
/* 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>
namespace paddle {
......@@ -33,4 +47,4 @@ void AttrTypeHelper::SetAttrType<std::vector<std::string>>(AttrProto* attr) {
attr->set_type(paddle::framework::AttrType::STRINGS);
}
} // namespace framework
} // namespace paddle
\ No newline at end of file
} // namespace paddle
paddle/framework/op_registry.h
浏览文件 @ 9e3a9eb2
......@@ -2,6 +2,8 @@
#include <algorithm>
#include <type_traits>
#include <unordered_map>
#include <unordered_set>
#include "paddle/framework/attr_checker.h"
#include "paddle/framework/op_desc.pb.h"
#include "paddle/framework/op_proto.pb.h"
......@@ -59,25 +61,52 @@ class OpProtoAndCheckerMaker {
OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
: proto_(proto), op_checker_(op_checker) {}
~OpProtoAndCheckerMaker() { CheckNoDuplicatedAttrs(); }
protected:
void AddInput(const std::string& name, const std::string& comment) {
void AddInput(const std::string& name, const std::string& comment,
bool multiple = false) {
auto input = proto_->mutable_inputs()->Add();
*input->mutable_name() = name;
*input->mutable_comment() = comment;
input->set_multiple(multiple);
if (multiple) {
SetHasMultipleInput();
}
}
void AddOutput(const std::string& name, const std::string& comment) {
void AddInputs(const std::string& name, const std::string& comment) {
AddInput(name, comment, true);
}
void AddOutput(const std::string& name, const std::string& comment,
bool temporary = false, bool multiple = false) {
auto output = proto_->mutable_outputs()->Add();
*output->mutable_name() = name;
*output->mutable_comment() = comment;
output->set_multiple(multiple);
if (multiple) {
SetHasMultipleOutput();
}
output->set_temporary(temporary);
if (temporary) {
SetHasTemporaryOutput();
}
}
void AddOutputs(const std::string& name, const std::string& comment,
bool temporary = false) {
AddOutput(name, comment, temporary, true);
}
template <typename T>
TypedAttrChecker<T>& AddAttr(const std::string& name,
const std::string& comment) {
const std::string& comment,
bool generated = false) {
auto attr = proto_->mutable_attrs()->Add();
*attr->mutable_name() = name;
*attr->mutable_comment() = comment;
attr->set_generated(generated);
AttrTypeHelper::SetAttrType<T>(attr);
return op_checker_->AddAttrChecker<T>(name);
}
......@@ -86,8 +115,70 @@ class OpProtoAndCheckerMaker {
*(proto_->mutable_comment()) = comment;
}
private:
void SetHasMultiple(const std::string& in_out, bool* flag) {
if (!*flag) {
AddAttr<std::vector<int>>(in_out + "_format",
"The multiple index of " + in_out +
"\n"
R"DOC(
This attribute is used by Paddle core framework. Paddle's Op support each input
or output could be a list of variable. This attribute is used to show how that
list organized.
e.g.
input = ["a", "b", "c", "d", "e", "f"]
input_format = [0, 4, 5, 6]
means
The number of all input variables this op is six, and they are segmented into
three inputs.
The first input is input[0:4], second is input[4:5], third is input[5:6].
)DOC",
/*generated*/ true);
*flag = true;
}
}
void SetHasMultipleInput() { SetHasMultiple("input", &has_multiple_input_); }
void SetHasMultipleOutput() {
SetHasMultiple("output", &has_multiple_output_);
}
void SetHasTemporaryOutput() {
if (!has_temporary_output_) {
AddAttr<std::vector<int>>("temporary_index",
R"DOC(The temporary index of output.
Not all output of Paddle Op is used by user. For faster computation, each op
could output some its internal state to other op, other op could take that
output to make compute faster.
Add a mark to which output is temporary is helpful for future optimization.
)DOC",
/*generated*/ true)
.SetDefault(std::vector<int>());
has_temporary_output_ = true;
}
}
void CheckNoDuplicatedAttrs() {
std::unordered_set<std::string> names;
size_t cnt = 0;
for (auto& attr : proto_->attrs()) {
names.insert(attr.name());
++cnt;
}
PADDLE_ENFORCE(names.size() == cnt,
"Cannot register two attribute in same name!");
}
OpProto* proto_;
OpAttrChecker* op_checker_;
bool has_multiple_input_{false};
bool has_multiple_output_{false};
bool has_temporary_output_{false};
};
class OpRegistry {
......@@ -107,10 +198,10 @@ class OpRegistry {
op_type, op_proto.InitializationErrorString());
}
static OperatorBase* CreateOp(const OpDesc& op_desc) {
static OperatorPtr CreateOp(const OpDesc& op_desc) {
std::string op_type = op_desc.type();
OperatorBase* op = creators().at(op_type)();
op->desc_ = op_desc;
OperatorPtr op(creators().at(op_type)());
op->type_ = op_desc.type();
op->inputs_.reserve((size_t)op_desc.inputs_size());
std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
std::back_inserter(op->inputs_));
......@@ -125,17 +216,17 @@ class OpRegistry {
return op;
}
static std::unordered_map<std::string, OpProto>& protos() {
static std::unordered_map<std::string, OpProto> protos_;
return protos_;
};
private:
static std::unordered_map<std::string, OpCreator>& creators() {
static std::unordered_map<std::string, OpCreator> creators_;
return creators_;
}
static std::unordered_map<std::string, OpProto>& protos() {
static std::unordered_map<std::string, OpProto> protos_;
return protos_;
};
static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
return op_checkers_;
......@@ -150,12 +241,18 @@ class OpRegisterHelper {
}
};
/**
* check if MACRO is used in GLOBAL NAMESPACE.
*/
#define STATIC_ASSERT_GLOBAL_NAMESPACE(uniq_name, msg) \
struct __test_global_namespace_##uniq_name##__ {}; \
static_assert(std::is_same<::__test_global_namespace_##uniq_name##__, \
__test_global_namespace_##uniq_name##__>::value, \
msg)
/**
* Macro to Register Operator.
*/
#define REGISTER_OP(__op_type, __op_class, __op_maker_class) \
STATIC_ASSERT_GLOBAL_NAMESPACE(__reg_op__##__op_type, \
"REGISTER_OP must be in global namespace"); \
......@@ -163,9 +260,12 @@ class OpRegisterHelper {
__op_register_##__op_type##__(#__op_type); \
int __op_register_##__op_type##_handle__() { return 0; }
#define REGISTER_OP_KERNEL(type, GPU_OR_CPU, PlaceType, KernelType) \
/**
* Macro to Register OperatorKernel.
*/
#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, KernelType) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_op_kernel_##type##_##GPU_OR_CPU##__, \
__reg_op_kernel_##type##_##DEVICE_TYPE##__, \
"REGISTER_OP_KERNEL must be in global namespace"); \
struct __op_kernel_register__##type##__ { \
__op_kernel_register__##type##__() { \
......@@ -176,7 +276,7 @@ class OpRegisterHelper {
} \
}; \
static __op_kernel_register__##type##__ __reg_kernel_##type##__; \
int __op_kernel_register_##type##_handle_##GPU_OR_CPU##__() { return 0; }
int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
#define REGISTER_OP_GPU_KERNEL(type, KernelType) \
REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, KernelType)
......@@ -184,6 +284,10 @@ class OpRegisterHelper {
#define REGISTER_OP_CPU_KERNEL(type, KernelType) \
REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, KernelType)
/**
* Macro to mark what Operator and Kernel we will use and tell the compiler to
* link them into target.
*/
#define USE_OP_WITHOUT_KERNEL(op_type) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__use_op_without_kernel_##op_type, \
......@@ -201,15 +305,16 @@ class OpRegisterHelper {
__attribute__((unused)) = \
__op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__()
#ifdef PADDLE_ONLY_CPU
#define USE_OP(op_type) \
// use Operator with only cpu kernel.
#define USE_OP_CPU(op_type) \
USE_OP_WITHOUT_KERNEL(op_type); \
USE_OP_KERNEL(op_type, CPU);
USE_OP_KERNEL(op_type, CPU)
#ifdef PADDLE_ONLY_CPU
#define USE_OP(op_type) USE_OP_CPU(op_type)
#else
#define USE_OP(op_type) \
USE_OP_WITHOUT_KERNEL(op_type); \
USE_OP_KERNEL(op_type, CPU); \
#define USE_OP(op_type) \
USE_OP_CPU(op_type); \
USE_OP_KERNEL(op_type, GPU)
#endif
......
paddle/framework/op_registry_test.cc
浏览文件 @ 9e3a9eb2
......@@ -5,9 +5,9 @@ namespace paddle {
namespace framework {
class CosineOp : public OperatorBase {
public:
void Run(const std::shared_ptr<Scope>& scope,
void Run(const ScopePtr& scope,
const platform::DeviceContext& dev_ctx) const override {}
void InferShape(const std::shared_ptr<Scope>& scope) const override {}
void InferShape(const ScopePtr& scope) const override {}
};
class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
......@@ -25,8 +25,8 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
class MyTestOp : public OperatorBase {
public:
void InferShape(const std::shared_ptr<Scope>& scope) const override {}
void Run(const std::shared_ptr<Scope>& scope,
void InferShape(const ScopePtr& scope) const override {}
void Run(const ScopePtr& scope,
const platform::DeviceContext& dev_ctx) const override {}
public:
......@@ -36,8 +36,9 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
public:
MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("input", "input of cosine op");
AddOutput("output", "output of cosine op");
AddInputs("input", "input of cosine op");
AddOutput("output", "output of cosine op",
/*temporary*/ true);
auto my_checker = [](int i) {
PADDLE_ENFORCE(i % 2 == 0, "'test_attr' must be even!");
};
......@@ -66,7 +67,7 @@ TEST(OpRegistry, CreateOp) {
attr->set_type(paddle::framework::AttrType::FLOAT);
attr->set_f(scale);
paddle::framework::OperatorBase* op =
paddle::framework::OperatorPtr op =
paddle::framework::OpRegistry::CreateOp(op_desc);
auto scope = std::make_shared<paddle::framework::Scope>();
paddle::platform::CPUDeviceContext dev_ctx;
......@@ -88,7 +89,7 @@ TEST(OpRegistry, IllegalAttr) {
bool caught = false;
try {
paddle::framework::OperatorBase* op __attribute__((unused)) =
paddle::framework::OperatorPtr op __attribute__((unused)) =
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (paddle::framework::EnforceNotMet err) {
caught = true;
......@@ -109,7 +110,7 @@ TEST(OpRegistry, DefaultValue) {
ASSERT_TRUE(op_desc.IsInitialized());
paddle::framework::OperatorBase* op =
paddle::framework::OperatorPtr op =
paddle::framework::OpRegistry::CreateOp(op_desc);
auto scope = std::make_shared<paddle::framework::Scope>();
paddle::platform::CPUDeviceContext dev_ctx;
......@@ -117,16 +118,25 @@ TEST(OpRegistry, DefaultValue) {
ASSERT_EQ(op->GetAttr<float>("scale"), 1.0);
}
static void SetInputFormat(paddle::framework::OpDesc* desc) {
auto attr = desc->add_attrs();
attr->set_name("input_format");
attr->set_type(paddle::framework::INTS);
attr->mutable_ints()->Add(0);
attr->mutable_ints()->Add(1);
}
TEST(OpRegistry, CustomChecker) {
paddle::framework::OpDesc op_desc;
op_desc.set_type("my_test_op");
op_desc.add_inputs("ii");
op_desc.add_outputs("oo");
SetInputFormat(&op_desc);
// attr 'test_attr' is not set
bool caught = false;
try {
paddle::framework::OperatorBase* op __attribute__((unused)) =
paddle::framework::OperatorPtr op __attribute__((unused)) =
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (paddle::framework::EnforceNotMet err) {
caught = true;
......@@ -145,7 +155,7 @@ TEST(OpRegistry, CustomChecker) {
attr->set_i(3);
caught = false;
try {
paddle::framework::OperatorBase* op __attribute__((unused)) =
paddle::framework::OperatorPtr op __attribute__((unused)) =
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (paddle::framework::EnforceNotMet err) {
caught = true;
......@@ -163,7 +173,8 @@ TEST(OpRegistry, CustomChecker) {
attr->set_name("test_attr");
attr->set_type(paddle::framework::AttrType::INT);
attr->set_i(4);
paddle::framework::OperatorBase* op =
SetInputFormat(&op_desc);
paddle::framework::OperatorPtr op =
paddle::framework::OpRegistry::CreateOp(op_desc);
paddle::platform::CPUDeviceContext dev_ctx;
auto scope = std::make_shared<paddle::framework::Scope>();
......
paddle/framework/operator.cc
浏览文件 @ 9e3a9eb2
......@@ -34,7 +34,7 @@ DeviceType* OpKernel::KernelContext::get_eigen_device<platform::GPUPlace>()
std::string OperatorBase::DebugString() const {
std::stringstream ss;
ss << "=================\n";
ss << "type = " << desc_.type() << "\n";
ss << "type = " << type_ << "\n";
ss << "inputs = [";
for (auto& ipt : inputs_) {
ss << ipt << ", ";
......@@ -54,4 +54,4 @@ std::string OperatorBase::DebugString() const {
}
} // namespace framework
} // namespace paddle
\ No newline at end of file
} // namespace paddle
paddle/framework/operator.h
浏览文件 @ 9e3a9eb2
......@@ -45,7 +45,7 @@ struct EigenDeviceConverter<platform::GPUPlace> {
#endif
class OperatorBase;
using OperatorPtr = std::shared_ptr<OperatorBase>;
/**
* OperatorBase has the basic element that Net will call to do computation.
* Only CreateOperator from OpRegistry will new Operator directly. User
......@@ -71,17 +71,14 @@ class OperatorBase {
/// InferShape infer the size of Variables used by this Operator with
/// information inside scope
virtual void InferShape(const std::shared_ptr<Scope>& scope) const = 0;
virtual void InferShape(const ScopePtr& scope) const = 0;
/// Net will call this function to Run an op.
virtual void Run(const std::shared_ptr<Scope>& scope,
virtual void Run(const ScopePtr& scope,
const platform::DeviceContext& dev_ctx) const = 0;
protected:
std::string Type() const { return desc_.type(); }
public:
OpDesc desc_;
std::string type_;
std::vector<std::string> inputs_;
std::vector<std::string> outputs_;
AttributeMap attrs_;
......@@ -97,7 +94,7 @@ class OpKernel {
*/
class KernelContext {
public:
KernelContext(const OperatorBase* op, const std::shared_ptr<Scope>& scope,
KernelContext(const OperatorBase* op, const ScopePtr& scope,
const platform::DeviceContext& device_context)
: op_(*op), scope_(scope), device_context_(device_context) {}
......@@ -115,7 +112,7 @@ class OpKernel {
DeviceType* get_eigen_device() const;
const OperatorBase& op_;
const std::shared_ptr<Scope>& scope_;
const ScopePtr& scope_;
const platform::DeviceContext& device_context_;
};
......@@ -160,9 +157,9 @@ class OperatorWithKernel : public OperatorBase {
using OpKernelMap =
std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;
void Run(const std::shared_ptr<Scope>& scope,
void Run(const ScopePtr& scope,
const platform::DeviceContext& dev_ctx) const final {
auto& opKernel = AllOpKernels().at(Type()).at(OpKernelKey(dev_ctx));
auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
opKernel->Compute(OpKernel::KernelContext(this, scope, dev_ctx));
}
......
paddle/framework/operator_test.cc
浏览文件 @ 9e3a9eb2
......@@ -19,14 +19,18 @@ limitations under the License. */
namespace paddle {
namespace framework {
class OperatorTest : public OperatorBase {
static int op_run_num = 0;
class OpWithoutKernelTest : public OperatorBase {
public:
void Init() override { x = 1; }
void InferShape(const std::shared_ptr<Scope>& scope) const override {}
void Run(const std::shared_ptr<Scope>& scope,
void InferShape(const ScopePtr& scope) const override {}
void Run(const ScopePtr& scope,
const platform::DeviceContext& dev_ctx) const override {
float scale = GetAttr<float>("scale");
ASSERT_NEAR(scale, 3.14, 1e-5);
op_run_num++;
ASSERT_EQ((int)inputs_.size(), 1);
ASSERT_EQ((int)outputs_.size(), 1);
ASSERT_NEAR(GetAttr<float>("scale"), 3.14, 1e-5);
ASSERT_EQ(scope->GetVariable(inputs_[0]), nullptr);
ASSERT_EQ(x, 1);
ASSERT_NE(scope->GetVariable(outputs_[0]), nullptr);
......@@ -36,19 +40,61 @@ class OperatorTest : public OperatorBase {
float x = 0;
};
class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
public:
OpeWithoutKernelTestProtoAndCheckerMaker(OpProto* proto,
OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("input", "input of test op");
AddOutput("output", "output of test op");
AddAttr<float>("scale", "scale of cosine op");
AddComment("This is test op");
}
};
} // namespace framework
} // namespace paddle
REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest,
paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
TEST(OperatorBase, all) {
paddle::framework::OpDesc op_desc;
op_desc.set_type("test_operator");
*op_desc.mutable_inputs()->Add() = "IN1";
*op_desc.mutable_outputs()->Add() = "OUT1";
auto attr = op_desc.mutable_attrs()->Add();
attr->set_name("scale");
attr->set_type(paddle::framework::AttrType::FLOAT);
attr->set_f(3.14);
paddle::platform::CPUDeviceContext device_context;
auto scope = std::make_shared<paddle::framework::Scope>();
paddle::framework::OperatorPtr op =
paddle::framework::OpRegistry::CreateOp(op_desc);
scope->CreateVariable("OUT1");
ASSERT_EQ(paddle::framework::op_run_num, 0);
op->Run(scope, device_context);
ASSERT_EQ(paddle::framework::op_run_num, 1);
}
namespace paddle {
namespace framework {
class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
public:
OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("input", "input of test op");
AddOutput("output", "output of test op");
AddAttr<float>("scale", "scale of cosine op")
.SetDefault(1.0)
.LargerThan(0.0);
AddAttr<float>("scale", "scale of cosine op");
AddComment("This is test op");
}
};
static int cpu_kernel_run_num = 0;
class OpWithKernelTest : public OperatorWithKernel {
protected:
void InferShape(const std::vector<const Tensor*>& inputs,
......@@ -58,10 +104,10 @@ class OpWithKernelTest : public OperatorWithKernel {
class CPUKernelTest : public OpKernel {
public:
void Compute(const KernelContext& context) const {
float scale = context.op_.GetAttr<float>("scale");
ASSERT_NEAR(scale, 3.14, 1e-5);
std::cout << "this is cpu kernel" << std::endl;
std::cout << context.op_.DebugString() << std::endl;
cpu_kernel_run_num++;
ASSERT_EQ((int)context.op_.inputs_.size(), 1);
ASSERT_EQ((int)context.op_.outputs_.size(), 1);
ASSERT_NEAR(context.op_.GetAttr<float>("scale"), 3.14, 1e-5);
}
};
......@@ -73,9 +119,7 @@ REGISTER_OP(op_with_kernel, paddle::framework::OpWithKernelTest,
REGISTER_OP_CPU_KERNEL(op_with_kernel, paddle::framework::CPUKernelTest);
TEST(OpKernel, all) {
using namespace paddle::framework;
OpDesc op_desc;
paddle::framework::OpDesc op_desc;
op_desc.set_type("op_with_kernel");
*op_desc.mutable_inputs()->Add() = "IN1";
*op_desc.mutable_outputs()->Add() = "OUT1";
......@@ -85,10 +129,11 @@ TEST(OpKernel, all) {
attr->set_f(3.14);
paddle::platform::CPUDeviceContext cpu_device_context;
auto scope = std::make_shared<Scope>();
auto scope = std::make_shared<paddle::framework::Scope>();
OperatorBase* op = paddle::framework::OpRegistry::CreateOp(op_desc);
paddle::framework::OperatorPtr op =
paddle::framework::OpRegistry::CreateOp(op_desc);
ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0);
op->Run(scope, cpu_device_context);
delete op;
ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
}
paddle/framework/scope.h
浏览文件 @ 9e3a9eb2
......@@ -23,6 +23,9 @@ limitations under the License. */
namespace paddle {
namespace framework {
class Scope;
using ScopePtr = std::shared_ptr<Scope>;
/**
* @brief Scope that manage all variables.
*
......@@ -41,7 +44,7 @@ class Scope {
/**
* @brief Initialize a Scope with parent.
*/
explicit Scope(const std::shared_ptr<Scope>& parent) : parent_(parent) {}
explicit Scope(const ScopePtr& parent) : parent_(parent) {}
/**
* @brief Create Variable
......@@ -88,7 +91,7 @@ class Scope {
private:
std::unordered_map<std::string, std::unique_ptr<Variable>> vars_;
std::shared_ptr<Scope> parent_{nullptr};
ScopePtr parent_{nullptr};
};
} // namespace framework
......
paddle/framework/tensor.h
浏览文件 @ 9e3a9eb2
......@@ -15,8 +15,8 @@ limitations under the License. */
#pragma once
#include <cstdint>
#include <cstring>
#include <memory>
#include <type_traits>
#include "paddle/framework/ddim.h"
#include "paddle/framework/enforce.h"
#include "paddle/framework/tensor_types.h"
......@@ -29,68 +29,78 @@ namespace framework {
class Tensor {
public:
Tensor() : offset_(0) {}
Tensor() : numel_(0), offset_(0) {}
explicit Tensor(const DDim& dims) : dims_(dims), offset_(0) {}
Tensor& operator=(const Tensor& src) = delete;
template <typename T>
const T* data() const {
CheckDims<T>();
return reinterpret_cast<const T*>(
reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
}
T* data() const {
PADDLE_ENFORCE(
holder_ != nullptr,
"Tenosr has not been initialized. Call Tensor::mutable_data first.");
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->Ptr()) +
template <typename T>
T* raw_data() const {
CheckDims<T>();
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
template <typename T, // must be POD types
typename std::enable_if<std::is_pod<T>::value>::type* = nullptr>
template <typename T>
T* mutable_data(DDim dims, paddle::platform::Place place) {
if (holder_ == nullptr ||
!(holder_->Place() ==
place) /* some versions of boost::variant don't have operator!= */
|| holder_->Size() < product(dims) * sizeof(T) + offset_) {
holder_.reset(new PlaceholderImpl<T>(place, product(dims) * sizeof(T)));
dims_ = dims;
offset_ = 0;
}
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->Ptr()) +
offset_);
set_dims(dims);
return mutable_data<T>(place);
}
template <typename T, // must be POD types
typename std::enable_if<std::is_pod<T>::value>::type* = nullptr>
template <typename T>
T* mutable_data(paddle::platform::Place place) {
PADDLE_ENFORCE(numel_ > 0,
"Tensor::numel_ must be larger than zero to call "
"Tensor::mutable_data. Call Tensor::set_dim first.");
if (holder_ == nullptr ||
!(holder_->Place() ==
!(holder_->place() ==
place) /* some versions of boost::variant don't have operator!= */
|| holder_->Size() < product(dims_) * sizeof(T) + offset_) {
holder_.reset(new PlaceholderImpl<T>(place, product(dims_) * sizeof(T)));
|| holder_->size() < numel_ * sizeof(T) + offset_) {
#ifdef __CUDACC__
switch (place.which()) {
case 0:
holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
boost::get<platform::GPUPlace>(place), numel_ * sizeof(T)));
break;
case 1:
holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
boost::get<platform::CPUPlace>(place), numel_ * sizeof(T)));
break;
}
#else
holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
boost::get<platform::CPUPlace>(place), numel_ * sizeof(T)));
#endif
offset_ = 0;
}
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->Ptr()) +
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
size_t NumElements() const { return product(dims_); }
template <typename T, size_t NDIMS>
typename TTypes<T, NDIMS>::Tensor shaped(DDim new_dims) {
Eigen::array<Eigen::DenseIndex, NDIMS> dims =
paddle::framework::ToEigenDSizes<NDIMS>(new_dims);
return typename TTypes<T, NDIMS>::Tensor(data<T>(), dims);
return typename TTypes<T, NDIMS>::Tensor(raw_data<T>(), dims);
}
template <typename T, size_t NDIMS>
typename TTypes<T, NDIMS>::Tensor tensor() {
return typename TTypes<T, NDIMS>::Tensor(
data<T>(), paddle::framework::ToEigenDSizes<NDIMS>(dims_));
raw_data<T>(), paddle::framework::ToEigenDSizes<NDIMS>(dims_));
}
// flat to rank = 1
template <typename T>
typename TTypes<T>::Flat flat() {
return shaped<T, 1>(make_ddim({static_cast<int>(NumElements())}));
return shaped<T, 1>(make_ddim({static_cast<int>(numel_)}));
}
// to TensorType Vec
......@@ -106,6 +116,13 @@ class Tensor {
}
// const versions of all the methods above.
template <typename T, size_t NDIMS>
typename TTypes<T, NDIMS>::Tensor shaped(DDim new_dims) const {
Eigen::array<Eigen::DenseIndex, NDIMS> dims =
paddle::framework::ToEigenDSizes<NDIMS>(new_dims);
return typename TTypes<T, NDIMS>::Tensor(data<T>(), dims);
}
template <typename T, size_t NDIMS>
typename TTypes<T, NDIMS>::ConstantTensor tensor() const {
return typename TTypes<T, NDIMS>::Tensor(
......@@ -114,7 +131,7 @@ class Tensor {
template <typename T>
typename TTypes<T>::ConstFlat flat() const {
return shaped<T, 1>(make_ddim({static_cast<int>(NumElements())}));
return shaped<T, 1>(make_ddim({static_cast<int>(numel_)}));
}
template <typename T>
......@@ -127,17 +144,30 @@ class Tensor {
return tensor<T, 2>();
}
template <typename T>
void ShareDataFrom(const Tensor& src) {
PADDLE_ENFORCE(src.holder_ != nullptr,
"Can not share data from an uninitialized tensor.");
src.CheckDims<T>();
holder_ = src.holder_;
dims_ = src.dims_;
set_dims(src.dims());
offset_ = src.offset_;
}
template <typename T>
void CopyFrom(const Tensor& src, paddle::platform::Place dst_place) {
PADDLE_ENFORCE(platform::is_cpu_place(src.holder_->place()) &&
platform::is_cpu_place(dst_place),
"Tensor::CopyFrom only support CPU now.");
src.CheckDims<T>();
size_t size = src.numel_ * sizeof(T);
set_dims(src.dims());
const void* src_ptr = static_cast<const void*>(src.data<T>());
void* dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
memcpy(dst_ptr, src_ptr, size);
}
template <typename T>
Tensor Slice(const int& begin_idx, const int& end_idx) const {
PADDLE_ENFORCE(holder_ != nullptr,
"The sliced tenosr has not been initialized.");
CheckDims<T>();
PADDLE_ENFORCE(begin_idx >= 0 && end_idx <= dims_[0],
"Slice index is less than zero or out of bound.");
PADDLE_ENFORCE(begin_idx < end_idx,
......@@ -150,12 +180,21 @@ class Tensor {
}
Tensor dst;
dst.holder_ = holder_;
dst.dims_ = dims_;
dst.dims_[0] = end_idx - begin_idx;
dst.offset_ = offset_ + begin_idx * base * holder_->TypeSize();
DDim dst_dims = dims_;
dst_dims[0] = end_idx - begin_idx;
dst.set_dims(dst_dims);
dst.offset_ = offset_ + begin_idx * base * sizeof(T);
return dst;
}
void set_dims(const DDim& dims) {
if (dims == dims_) {
return;
}
dims_ = dims;
numel_ = product(dims_);
}
DDim dims() const { return dims_; }
private:
......@@ -163,45 +202,54 @@ class Tensor {
// parameter of Variable.
struct Placeholder {
virtual ~Placeholder() {}
virtual void* Ptr() const = 0;
virtual paddle::platform::Place Place() const = 0;
virtual size_t Size() const = 0;
virtual size_t TypeSize() const = 0;
virtual void* ptr() const = 0;
virtual paddle::platform::Place place() const = 0;
virtual size_t size() const = 0;
};
template <typename T>
template <typename T, typename PlaceType>
struct PlaceholderImpl : public Placeholder {
private:
template <typename PType>
class Deleter {
public:
Deleter(platform::Place place) : place_(place) {}
Deleter(PType place) : place_(place) {}
void operator()(T* ptr) {
paddle::memory::Free(place_, static_cast<void*>(ptr));
}
private:
paddle::platform::Place place_;
PType place_;
};
public:
PlaceholderImpl(paddle::platform::Place place, size_t size)
PlaceholderImpl(PlaceType place, size_t size)
: ptr_(static_cast<T*>(paddle::memory::Alloc(place, size)),
Deleter(place)),
Deleter<PlaceType>(place)),
place_(place),
size_(size) {}
virtual void* Ptr() const { return static_cast<void*>(ptr_.get()); }
virtual size_t Size() const { return size_; }
virtual paddle::platform::Place Place() const { return place_; }
virtual size_t TypeSize() const { return sizeof(T); }
virtual void* ptr() const { return static_cast<void*>(ptr_.get()); }
virtual size_t size() const { return size_; }
virtual paddle::platform::Place place() const { return place_; }
std::unique_ptr<T, Deleter> ptr_;
std::unique_ptr<T, Deleter<PlaceType>> ptr_;
paddle::platform::Place place_; // record the place of ptr_.
size_t size_; // size of the memory block.
};
template <typename T>
inline void CheckDims() const {
PADDLE_ENFORCE(holder_ != nullptr,
"Tenosr holds no memory. Call Tensor::mutable_data first.");
PADDLE_ENFORCE(holder_->size() >= numel_ * sizeof(T) + offset_,
"Tensor's dims_ is out of bound. Call Tensor::mutable_data "
"first to re-allocate memory.");
}
std::shared_ptr<Placeholder> holder_; // holds the memory block if allocated.
DDim dims_;
size_t numel_; // cache of `product(dims_)`
size_t offset_; // marks the begin of tensor data area.
};
......
paddle/framework/tensor_test.cc
浏览文件 @ 9e3a9eb2
......@@ -18,7 +18,8 @@
TEST(Tensor, Dims) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor tt(make_ddim({2, 3, 4}));
Tensor tt;
tt.set_dims(make_ddim({2, 3, 4}));
DDim dims = tt.dims();
ASSERT_EQ(arity(dims), 3);
for (int i = 0; i < 3; ++i) {
......@@ -35,7 +36,7 @@ TEST(Tensor, DataAssert) {
} catch (paddle::framework::EnforceNotMet err) {
caught = true;
std::string msg =
"Tenosr has not been initialized. Call Tensor::mutable_data first.";
"Tenosr holds no memory. Call Tensor::mutable_data first.";
const char* what = err.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
......@@ -104,19 +105,18 @@ TEST(Tensor, ShareDataFrom) {
// Try to share data form uninitialized tensor
bool caught = false;
try {
dst_tensor.ShareDataFrom(src_tensor);
dst_tensor.ShareDataFrom<float>(src_tensor);
} catch (EnforceNotMet err) {
caught = true;
std::string msg = "Can not share data from an uninitialized tensor.";
const char* what = err.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
std::string msg = "Tenosr holds no memory. Call Tensor::mutable_data
first."; const char* what = err.what(); for (size_t i = 0; i < msg.length();
++i) { ASSERT_EQ(what[i], msg[i]);
}
}
ASSERT_TRUE(caught);
src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), CPUPlace());
dst_tensor.ShareDataFrom(src_tensor);
dst_tensor.ShareDataFrom<int>(src_tensor);
ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
}
......@@ -124,7 +124,7 @@ TEST(Tensor, ShareDataFrom) {
Tensor src_tensor;
Tensor dst_tensor;
src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), GPUPlace());
dst_tensor.ShareDataFrom(src_tensor);
dst_tensor.ShareDataFrom<int>(src_tensor);
ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
}
}
......@@ -135,7 +135,7 @@ TEST(Tensor, Slice) {
{
Tensor src_tensor;
src_tensor.mutable_data<int>(make_ddim({5, 3, 4}), CPUPlace());
Tensor slice_tensor = src_tensor.Slice(1, 3);
Tensor slice_tensor = src_tensor.Slice<int>(1, 3);
DDim slice_dims = slice_tensor.dims();
ASSERT_EQ(arity(slice_dims), 3);
EXPECT_EQ(slice_dims[0], 2);
......@@ -158,7 +158,7 @@ TEST(Tensor, Slice) {
{
Tensor src_tensor;
src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
Tensor slice_tensor = src_tensor.Slice(2, 6);
Tensor slice_tensor = src_tensor.Slice<double>(2, 6);
DDim slice_dims = slice_tensor.dims();
ASSERT_EQ(arity(slice_dims), 2);
EXPECT_EQ(slice_dims[0], 4);
......@@ -178,4 +178,29 @@ TEST(Tensor, Slice) {
}
}
TEST(Tensor, CopyFrom) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor src_tensor;
int* src_ptr = src_tensor.mutable_data<int>(make_ddim({3, 3}), CPUPlace());
int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9};
memcpy(src_ptr, arr, 9 * sizeof(int));
Tensor dst_tensor;
dst_tensor.CopyFrom<int>(src_tensor, CPUPlace());
const int* dst_ptr = dst_tensor.data<int>();
ASSERT_NE(src_ptr, dst_ptr);
for (size_t i = 0; i < 9; ++i) {
EXPECT_EQ(src_ptr[i], dst_ptr[i]);
}
Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
dst_tensor.CopyFrom<int>(slice_tensor, CPUPlace());
const int* slice_ptr = slice_tensor.data<int>();
dst_ptr = dst_tensor.data<int>();
ASSERT_NE(dst_ptr, slice_ptr);
for (size_t i = 0; i < 3; ++i) {
EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
}
}
*/
\ No newline at end of file
paddle/function/GemmConvOp.cpp
浏览文件 @ 9e3a9eb2
......@@ -117,8 +117,7 @@ public:
ConvFunctionBase::init(config);
}
virtual void check(const BufferArgs& inputs,
const BufferArgs& outputs) override {
void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
const TensorShape& input = inputs[0].shape();
const TensorShape& filter = inputs[1].shape();
const TensorShape& output = outputs[0].shape();
......@@ -217,8 +216,7 @@ public:
ConvFunctionBase::init(config);
}
virtual void check(const BufferArgs& inputs,
const BufferArgs& outputs) override {
void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
const TensorShape& output = inputs[0].shape();
const TensorShape& filter = inputs[1].shape();
const TensorShape& input = outputs[0].shape();
......@@ -311,8 +309,7 @@ public:
ConvFunctionBase::init(config);
}
virtual void check(const BufferArgs& inputs,
const BufferArgs& outputs) override {
void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
const TensorShape& output = inputs[0].shape();
const TensorShape& input = inputs[1].shape();
const TensorShape& filter = outputs[0].shape();
......
paddle/function/NaiveConvOp.cpp
浏览文件 @ 9e3a9eb2
......@@ -90,8 +90,7 @@ public:
ConvFunctionBase::init(config);
}
virtual void check(const BufferArgs& inputs,
const BufferArgs& outputs) override {
void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
const TensorShape& input = inputs[0].shape();
const TensorShape& filter = inputs[1].shape();
const TensorShape& output = outputs[0].shape();
......
paddle/gserver/dataproviders/DataProvider.h
浏览文件 @ 9e3a9eb2
......@@ -205,10 +205,8 @@ public:
hl_destroy_event(hlEvent_);
hlEvent_ = NULL;
}
if (batchData_) {
delete batchData_;
batchData_ = NULL;
}
delete batchData_;
batchData_ = NULL;
}
void setDataBatch(DataBatch* batchData) { batchData_ = batchData; }
......
paddle/gserver/gradientmachines/NeuralNetwork.cpp
浏览文件 @ 9e3a9eb2
......@@ -403,7 +403,7 @@ public:
: layerName_(layerName) {
addEvaluator(std::move(evaluator));
}
virtual void eval(const NeuralNetwork& nn) override {
void eval(const NeuralNetwork& nn) override {
const LayerPtr& layer = nn.getLayer(layerName_);
CHECK(layer) << "Nonexisted layer: " << layerName_ << " in submodel "
<< nn.getName();
......
paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
浏览文件 @ 9e3a9eb2
......@@ -636,7 +636,7 @@ void lenToStarts(std::vector<int>& starts) {
}
starts.back() = pos;
}
}
} // namespace
void RecurrentGradientMachine::calcSequenceStartPositions() {
std::vector<int> starts(commonSeqInfo_.size() + 1);
......
paddle/gserver/layers/AgentLayer.cpp
浏览文件 @ 9e3a9eb2
......@@ -124,7 +124,7 @@ void copyElements(const IVector& srcVec,
dest[index[i]] = src[i];
}
}
}
} // namespace
void GatherAgentLayer::forwardIds(PassType passType) {
IVectorPtr realId = realLayers_[0]->getOutputLabel();
......
paddle/math/Storage.cpp
浏览文件 @ 9e3a9eb2
......@@ -32,9 +32,7 @@ static InitFunction __init_storage_engine([]() { StorageEngine::singleton(); },
StorageEngine::StorageEngine() : cpuAllocator_(nullptr) {}
StorageEngine::~StorageEngine() {
if (cpuAllocator_) {
delete cpuAllocator_;
}
delete cpuAllocator_;
for (auto it : gpuAllocator_) {
delete it;
}
......
paddle/memory/CMakeLists.txt
浏览文件 @ 9e3a9eb2
add_subdirectory(detail)
cc_library(memory SRCS memory.cc)
cc_library(paddle_memory
DEPS
memory meta_data
meta_cache memory_block
buddy_allocator system_allocator)
cc_test(memory_test SRCS memory_test.cc DEPS place paddle_memory)
paddle/memory/detail/CMakeLists.txt
浏览文件 @ 9e3a9eb2
if(${WITH_GPU})
nv_library(system_allocator SRCS system_allocator.cc DEPS gflags)
nv_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator gflags)
nv_library(system_allocator SRCS system_allocator.cc DEPS gflags cpu_info gpu_info)
else(${WITH_GPU})
cc_library(system_allocator SRCS system_allocator.cc DEPS gflags)
cc_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator gflags)
cc_library(system_allocator SRCS system_allocator.cc DEPS gflags cpu_info)
endif(${WITH_GPU})
cc_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator)
cc_library(meta_data SRCS meta_data.cc)
cc_library(meta_cache SRCS meta_cache.cc)
cc_library(memory_block SRCS memory_block.cc)
cc_library(buddy_allocator SRCS buddy_allocator.cc DEPS glog)
paddle/memory/detail/buddy_allocator.cc
浏览文件 @ 9e3a9eb2
......@@ -12,22 +12,317 @@
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/memory/detail/buddy_allocator.h"
#include "glog/logging.h"
namespace paddle {
namespace memory {
namespace detail {
BuddyAllocator::BuddyAllocator(size_t pool_size, size_t max_pools,
SystemAllocator* system_allocator)
: pool_size_(pool_size),
max_pools_(max_pools),
system_allocator_(system_allocator) {
PADDLE_ASSERT(pool_size > 0);
PADDLE_ASSERT(max_pools > 0);
PADDLE_ASSERT(system_allocator != nullptr);
BuddyAllocator::BuddyAllocator(SystemAllocator* system_allocator,
size_t min_chunk_size, size_t max_chunk_size)
: min_chunk_size_(min_chunk_size),
max_chunk_size_(max_chunk_size),
cache_(system_allocator->UseGpu()),
system_allocator_(std::move(system_allocator)) {}
BuddyAllocator::~BuddyAllocator() {
DLOG(INFO) << "BuddyAllocator Disconstructor makes sure that all of these "
"have actually been freed";
while (!pool_.empty()) {
auto block = static_cast<MemoryBlock*>(std::get<2>(*pool_.begin()));
DLOG(INFO) << "Free from block (" << block << ", " << max_chunk_size_
<< ")";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
pool_.erase(pool_.begin());
}
}
inline size_t align(size_t size, size_t alignment) {
size_t remaining = size % alignment;
return remaining == 0 ? size : size + (alignment - remaining);
}
void* BuddyAllocator::Alloc(size_t unaligned_size) {
// adjust allocation alignment
size_t size = align(unaligned_size + sizeof(Metadata), min_chunk_size_);
// acquire the allocator lock
std::lock_guard<std::mutex> lock(mutex_);
DLOG(INFO) << "Allocate " << unaligned_size << " bytes from chunk size "
<< size;
// if the allocation is huge, send directly to the system allocator
if (size > max_chunk_size_) {
DLOG(INFO) << "Allocate from system allocator.";
return SystemAlloc(size);
}
// query and allocate from the existing chunk
auto it = FindExistChunk(size);
// refill the pool if failure
if (it == pool_.end()) {
it = RefillPool();
// if still failure, fail fatally
if (it == pool_.end()) {
return nullptr;
}
} else {
DLOG(INFO) << "Allocation from existing memory block " << std::get<2>(*it)
<< " at address "
<< reinterpret_cast<MemoryBlock*>(std::get<2>(*it))->data();
}
total_used_ += size;
total_free_ -= size;
// split the allocation and return data for use
return reinterpret_cast<MemoryBlock*>(SplitToAlloc(it, size))->data();
}
void BuddyAllocator::Free(void* p) {
// Point back to metadata
auto block = static_cast<MemoryBlock*>(p)->metadata();
// Acquire the allocator lock
std::lock_guard<std::mutex> lock(mutex_);
DLOG(INFO) << "Free from address " << block;
if (block->type(cache_) == MemoryBlock::HUGE_CHUNK) {
DLOG(INFO) << "Free directly from system allocator";
system_allocator_->Free(block, block->total_size(cache_),
block->index(cache_));
// Invalidate GPU allocation from cache
cache_.invalidate(block);
return;
}
block->mark_as_free(cache_);
total_used_ -= block->total_size(cache_);
total_free_ += block->total_size(cache_);
// Trying to merge the right buddy
if (block->has_right_buddy(cache_)) {
DLOG(INFO) << "Merging this block " << block << " with its right buddy "
<< block->right_buddy(cache_);
auto right_buddy = block->right_buddy(cache_);
if (right_buddy->type(cache_) == MemoryBlock::FREE_CHUNK) {
// Take away right buddy from pool
pool_.erase(IndexSizeAddress(right_buddy->index(cache_),
right_buddy->total_size(cache_),
right_buddy));
// merge its right buddy to the block
block->merge(cache_, right_buddy);
}
}
// Trying to merge the left buddy
if (block->has_left_buddy(cache_)) {
DLOG(INFO) << "Merging this block " << block << " with its left buddy "
<< block->left_buddy(cache_);
auto left_buddy = block->left_buddy(cache_);
if (left_buddy->type(cache_) == MemoryBlock::FREE_CHUNK) {
// Take away right buddy from pool
pool_.erase(IndexSizeAddress(left_buddy->index(cache_),
left_buddy->total_size(cache_), left_buddy));
// merge the block to its left buddy
left_buddy->merge(cache_, block);
block = left_buddy;
}
}
// Dumping this block into pool
DLOG(INFO) << "Inserting free block (" << block << ", "
<< block->total_size(cache_) << ")";
pool_.insert(
IndexSizeAddress(block->index(cache_), block->total_size(cache_), block));
// Clean up if existing too much free memory
// Prefer freeing fallback allocation first
CleanIdleFallBackAlloc();
// Free normal allocation
CleanIdleNormalAlloc();
}
size_t BuddyAllocator::Used() { return total_used_; }
void* BuddyAllocator::SystemAlloc(size_t size) {
size_t index = 0;
void* p = system_allocator_->Alloc(index, size);
DLOG(INFO) << "Allocated " << p << " from system allocator.";
if (p == nullptr) return nullptr;
static_cast<MemoryBlock*>(p)->init(cache_, MemoryBlock::HUGE_CHUNK, index,
size, nullptr, nullptr);
return static_cast<MemoryBlock*>(p)->data();
}
BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
#ifndef PADDLE_ONLY_CPU
if (system_allocator_->UseGpu()) {
if ((total_used_ + total_free_) == 0) {
// Compute the maximum allocation size for the first allocation.
max_chunk_size_ = platform::GpuMaxChunkSize();
}
}
#endif // PADDLE_ONLY_CPU
// Allocate a new maximum sized block
size_t index = 0;
void* p = system_allocator_->Alloc(index, max_chunk_size_);
if (p == nullptr) return pool_.end();
DLOG(INFO) << "Creating and inserting new block " << p
<< " from system allocator";
static_cast<MemoryBlock*>(p)->init(cache_, MemoryBlock::FREE_CHUNK, index,
max_chunk_size_, nullptr, nullptr);
// gpu fallback allocation
if (system_allocator_->UseGpu() &&
static_cast<MemoryBlock*>(p)->index(cache_) == 1) {
fallback_alloc_count_++;
}
total_free_ += max_chunk_size_;
// dump the block into pool
return pool_.insert(IndexSizeAddress(index, max_chunk_size_, p)).first;
}
BuddyAllocator::PoolSet::iterator BuddyAllocator::FindExistChunk(size_t size) {
size_t index = 0;
while (1) {
auto it = pool_.lower_bound(IndexSizeAddress(index, size, nullptr));
// no match chunk memory
if (it == pool_.end()) return it;
if (std::get<0>(*it) > index) {
// find suitable one
if (std::get<1>(*it) >= size) {
return it;
}
// update and continue
index = std::get<0>(*it);
continue;
}
return it;
}
}
void* BuddyAllocator::SplitToAlloc(BuddyAllocator::PoolSet::iterator it,
size_t size) {
auto block = static_cast<MemoryBlock*>(std::get<2>(*it));
pool_.erase(it);
DLOG(INFO) << "Split block (" << block << ", " << block->total_size(cache_)
<< ") into";
block->split(cache_, size);
DLOG(INFO) << "Left block (" << block << ", " << block->total_size(cache_)
<< ")";
block->set_type(cache_, MemoryBlock::ARENA_CHUNK);
// the rest of memory if exist
if (block->has_right_buddy(cache_)) {
if (block->right_buddy(cache_)->type(cache_) == MemoryBlock::FREE_CHUNK) {
DLOG(INFO) << "Insert right block (" << block->right_buddy(cache_) << ", "
<< block->right_buddy(cache_)->total_size(cache_) << ")";
pool_.insert(
IndexSizeAddress(block->right_buddy(cache_)->index(cache_),
block->right_buddy(cache_)->total_size(cache_),
block->right_buddy(cache_)));
}
}
return block;
}
void BuddyAllocator::CleanIdleFallBackAlloc() {
// If fallback allocation does not exist, return directly
if (!fallback_alloc_count_) return;
for (auto pool = pool_.rbegin(); pool != pool_.rend();) {
// If free memory block less than max_chunk_size_, return directly
if (std::get<1>(*pool) < max_chunk_size_) return;
MemoryBlock* block = static_cast<MemoryBlock*>(std::get<2>(*pool));
// If no GPU fallback allocator, return
if (!system_allocator_->UseGpu() || block->index(cache_) == 0) {
return;
}
DLOG(INFO) << "Return block " << block << " to fallback allocator.";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
pool = PoolSet::reverse_iterator(pool_.erase(std::next(pool).base()));
total_free_ -= max_chunk_size_;
fallback_alloc_count_--;
// If no fall allocation exists, return directly
if (!fallback_alloc_count_) return;
}
}
void BuddyAllocator::CleanIdleNormalAlloc() {
auto shall_free_alloc = [&]() -> bool {
// free all fallback allocations
if (fallback_alloc_count_ > 0) {
return true;
}
// keep 2x overhead if we haven't fallen back
if ((total_used_ + max_chunk_size_) * 2 < total_free_) {
return true;
}
return false;
};
if (!shall_free_alloc()) return;
for (auto pool = pool_.rbegin(); pool != pool_.rend();) {
// If free memory block less than max_chunk_size_, return directly
if (std::get<1>(*pool) < max_chunk_size_) return;
MemoryBlock* block = static_cast<MemoryBlock*>(std::get<2>(*pool));
DLOG(INFO) << "Return block " << block << " to base allocator.";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
pool = PoolSet::reverse_iterator(pool_.erase(std::next(pool).base()));
total_free_ -= max_chunk_size_;
if (!shall_free_alloc()) return;
}
}
} // namespace detail
......
paddle/memory/detail/buddy_allocator.h
浏览文件 @ 9e3a9eb2
......@@ -14,9 +14,16 @@
#pragma once
#include "paddle/memory/detail/meta_cache.h"
#include "paddle/memory/detail/meta_data.h"
#include "paddle/memory/detail/system_allocator.h"
#include "paddle/platform/assert.h"
#include "paddle/platform/cpu_info.h"
#include "paddle/platform/gpu_info.h"
#include <mutex>
#include <set>
#include <unordered_map>
#include <vector>
namespace paddle {
......@@ -25,61 +32,80 @@ namespace detail {
class BuddyAllocator {
public:
BuddyAllocator(size_t pool_size, size_t max_pools,
SystemAllocator* system_allocator);
BuddyAllocator(SystemAllocator* system_allocator, size_t min_chunk_size,
size_t max_chunk_size);
~BuddyAllocator();
void* Alloc(size_t size);
public:
void* Alloc(size_t unaligned_size);
void Free(void*);
size_t Used();
public:
// Disable copy and assignment
BuddyAllocator(const BuddyAllocator&) = delete;
BuddyAllocator& operator=(const BuddyAllocator&) = delete;
private:
struct Block {
size_t size_;
Block* left_; // left buddy
Block* right_; // right buddy
};
// Tuple (allocator index, memory size, memory address)
using IndexSizeAddress = std::tuple<size_t, size_t, void*>;
// Each element in PoolSet is a free allocation
using PoolSet = std::set<IndexSizeAddress>;
// Initially, there is only one pool. If a Alloc founds not enough
// memory from that pool, and there has not been max_num_pools_,
// create a new pool by calling system_allocator_.Alloc(pool_size_).
std::vector<void*> pools_;
/*! \brief Allocate fixed-size memory from system */
void* SystemAlloc(size_t size);
size_t pool_size_; // the size of each pool;
size_t max_num_pools_; // the size of all pools;
/*! \brief If existing chunks are not suitable, refill pool */
PoolSet::iterator RefillPool();
SystemAllocator* system_allocator_;
/**
* \brief Find the suitable chunk from existing pool and split
* it to left and right buddies
*
* \param it the iterator of pool list
* \param size the size of allocation
*
* \return the left buddy address
*/
void* SplitToAlloc(PoolSet::iterator it, size_t size);
std::mutex mutex_;
/*! \brief Find the existing chunk which used to allocation */
PoolSet::iterator FindExistChunk(size_t size);
// Disable copy and assignment.
BuddyAllocator(const BuddyAllocator&) = delete;
BuddyAllocator& operator=(const BuddyAllocator&) = delete;
};
/*! \brief Clean idle fallback allocation */
void CleanIdleFallBackAlloc();
/*! \brief Clean idle normal allocation */
void CleanIdleNormalAlloc();
BuddyAllocator<CPUAllocator>* GetCPUBuddyAllocator() {
static BuddyAllocator<CPUAllocator>* a = nullptr;
if (a == nullptr) {
a = new BuddyAllocator<CPUAllocator>();
}
return a;
}
#ifndef PADDLE_ONLY_CPU // The following code are for CUDA.
BuddyAllocator<GPUAllocator>* GetGPUBuddyAllocator(int gpu_id) {
static BuddyAllocator<GPUAllocator>** as = NULL;
if (as == NULL) {
int gpu_num = platform::GetDeviceCount();
as = new BuddyAllocator<GPUAllocator>*[gpu_num];
for (int gpu = 0; gpu < gpu_num; gpu++) {
as[gpu] = new BuddyAllocator<GPUAllocator>();
}
}
return as[gpu_id];
}
#endif // PADDLE_ONLY_CPU
private:
size_t total_used_ = 0; // the total size of used memory
size_t total_free_ = 0; // the total size of free memory
size_t min_chunk_size_; // the minimum size of each chunk
size_t max_chunk_size_; // the maximum size of each chunk
private:
/**
* \brief A list of free allocation
*
* \note Only store free chunk memory in pool
*/
PoolSet pool_;
/*! Record fallback allocation count for auto-scaling */
size_t fallback_alloc_count_ = 0;
private:
/*! Unify the metadata format between GPU and CPU allocations */
MetadataCache cache_;
private:
/*! Allocate CPU/GPU memory from system */
SystemAllocator* system_allocator_;
std::mutex mutex_;
};
} // namespace detail
} // namespace memory
......
paddle/memory/detail/memory_block.cc 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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/memory/detail/memory_block.h"
#include "paddle/memory/detail/meta_cache.h"
#include "paddle/memory/detail/meta_data.h"
#include "paddle/platform/assert.h"
namespace paddle {
namespace memory {
namespace detail {
void MemoryBlock::init(MetadataCache& cache, Type t, size_t index, size_t size,
void* left_buddy, void* right_buddy) {
cache.store(this, Metadata(t, index, size - sizeof(Metadata), size,
static_cast<MemoryBlock*>(left_buddy),
static_cast<MemoryBlock*>(right_buddy)));
}
MemoryBlock::Type MemoryBlock::type(MetadataCache& cache) const {
return cache.load(this).type;
}
size_t MemoryBlock::size(MetadataCache& cache) const {
return cache.load(this).size;
}
size_t MemoryBlock::total_size(MetadataCache& cache) const {
return cache.load(this).total_size;
}
MemoryBlock* MemoryBlock::left_buddy(MetadataCache& cache) const {
return cache.load(this).left_buddy;
}
MemoryBlock* MemoryBlock::right_buddy(MetadataCache& cache) const {
return cache.load(this).right_buddy;
}
void MemoryBlock::split(MetadataCache& cache, size_t size) {
// make sure the split fits
PADDLE_ASSERT(total_size(cache) >= size);
// bail out if there is no room for another partition
if (total_size(cache) - size <= sizeof(Metadata)) {
return;
}
// find the position of the split
void* right_partition = reinterpret_cast<uint8_t*>(this) + size;
size_t remaining_size = total_size(cache) - size;
// Add the new block as a buddy
auto metadata = cache.load(this);
// Write the metadata for the new block
auto new_block_right_buddy = metadata.right_buddy;
cache.store(
static_cast<MemoryBlock*>(right_partition),
Metadata(FREE_CHUNK, index(cache), remaining_size - sizeof(Metadata),
remaining_size, this, new_block_right_buddy));
metadata.right_buddy = static_cast<MemoryBlock*>(right_partition);
metadata.size = size - sizeof(Metadata);
metadata.total_size = size;
cache.store(this, metadata);
// Write metadata for the new block's right buddy
if (new_block_right_buddy != nullptr) {
auto buddy_metadata = cache.load(new_block_right_buddy);
buddy_metadata.left_buddy = static_cast<MemoryBlock*>(right_partition);
cache.store(new_block_right_buddy, buddy_metadata);
}
}
void MemoryBlock::merge(MetadataCache& cache, MemoryBlock* right_buddy) {
// only free blocks can be merged
PADDLE_ASSERT(type(cache) == FREE_CHUNK);
PADDLE_ASSERT(right_buddy->type(cache) == FREE_CHUNK);
auto metadata = cache.load(this);
// link this->buddy's buddy
metadata.right_buddy = right_buddy->right_buddy(cache);
// link buddy's buddy -> this
if (metadata.right_buddy != nullptr) {
auto buddy_metadata = cache.load(metadata.right_buddy);
buddy_metadata.left_buddy = this;
cache.store(metadata.right_buddy, buddy_metadata);
}
metadata.size += right_buddy->total_size(cache);
metadata.total_size += right_buddy->total_size(cache);
cache.store(this, metadata);
cache.store(right_buddy, Metadata(INVALID_CHUNK, 0, 0, 0, nullptr, nullptr));
}
void MemoryBlock::mark_as_free(MetadataCache& cache) {
// check for double free or corruption
PADDLE_ASSERT(type(cache) != FREE_CHUNK);
PADDLE_ASSERT(type(cache) != INVALID_CHUNK);
set_type(cache, FREE_CHUNK);
}
void MemoryBlock::set_type(MetadataCache& cache, Type t) {
auto metadata = cache.load(this);
metadata.type = t;
cache.store(this, metadata);
}
bool MemoryBlock::has_left_buddy(MetadataCache& cache) const {
return left_buddy(cache) != nullptr;
}
bool MemoryBlock::has_right_buddy(MetadataCache& cache) const {
return right_buddy(cache) != nullptr;
}
size_t MemoryBlock::index(MetadataCache& cache) const {
return cache.load(this).index;
}
void* MemoryBlock::data() const {
return const_cast<Metadata*>(reinterpret_cast<const Metadata*>(this)) + 1;
}
MemoryBlock* MemoryBlock::metadata() const {
return const_cast<MemoryBlock*>(reinterpret_cast<const MemoryBlock*>(
reinterpret_cast<const Metadata*>(this) - 1));
}
} // namespace detail
} // namespace memory
} // namespace paddle
paddle/memory/detail/memory_block.h 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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 <cstddef>
namespace paddle {
namespace memory {
namespace detail {
// Forward Declarations
class MetadataCache;
/*! \brief A class used to interpret the contents of a memory block */
class MemoryBlock {
public:
enum Type {
FREE_CHUNK, // memory is free and idle
ARENA_CHUNK, // memory is being occupied
HUGE_CHUNK, // memory is out of management
INVALID_CHUNK // memory is invalid
};
public:
void init(MetadataCache& cache, Type t, size_t index, size_t size,
void* left_buddy, void* right_buddy);
public:
/*! \brief The type of the allocation */
Type type(MetadataCache& cache) const;
/*! \brief The size of the data region */
size_t size(MetadataCache& cache) const;
/*! \brief An index to track the allocator */
size_t index(MetadataCache& cache) const;
/*! \brief The total size of the block */
size_t total_size(MetadataCache& cache) const;
/*! \brief Check the left buddy of the block */
bool has_left_buddy(MetadataCache& cache) const;
/*! \brief Check the right buddy of the block */
bool has_right_buddy(MetadataCache& cache) const;
/*! \brief Get the left buddy */
MemoryBlock* left_buddy(MetadataCache& cache) const;
/*! \brief Get the right buddy */
MemoryBlock* right_buddy(MetadataCache& cache) const;
public:
/*! \brief Split the allocation into left/right blocks */
void split(MetadataCache& cache, size_t size);
/*! \brief Merge left and right blocks together */
void merge(MetadataCache& cache, MemoryBlock* right_buddy);
/*! \brief Mark the allocation as free */
void mark_as_free(MetadataCache& cache);
/*! \brief Change the type of the allocation */
void set_type(MetadataCache& cache, Type t);
public:
/*! \brief Get a pointer to the memory block's data */
void* data() const;
/*! \brief Get a pointer to the memory block's metadata */
MemoryBlock* metadata() const;
public:
static size_t overhead();
};
} // namespace detail
} // namespace memory
} // namespace paddle
paddle/memory/detail/meta_cache.cc 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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/memory/detail/meta_cache.h"
#include "paddle/memory/detail/memory_block.h"
#include "paddle/platform/assert.h"
namespace paddle {
namespace memory {
namespace detail {
MetadataCache::MetadataCache(bool uses_gpu) : uses_gpu_(uses_gpu) {}
Metadata MetadataCache::load(const MemoryBlock* block) {
if (uses_gpu_) {
auto existing_metadata = cache_.find(block);
PADDLE_ASSERT(existing_metadata->second.check_guards());
return existing_metadata->second;
} else {
PADDLE_ASSERT(reinterpret_cast<const Metadata*>(block)->check_guards());
return *reinterpret_cast<const Metadata*>(block);
}
}
void MetadataCache::store(MemoryBlock* block,
const Metadata& original_metadata) {
auto metadata = original_metadata;
metadata.update_guards();
if (uses_gpu_) {
cache_[block] = metadata;
} else {
*reinterpret_cast<Metadata*>(block) = metadata;
}
}
void MetadataCache::invalidate(MemoryBlock* block) {
if (uses_gpu_) {
cache_.erase(block);
}
}
} // namespace detail
} // namespace memory
} // namespace paddle
paddle/memory/detail/meta_cache.h 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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/memory/detail/memory_block.h"
#include "paddle/memory/detail/meta_data.h"
#include <unordered_map>
namespace paddle {
namespace memory {
namespace detail {
/**
* \brief A cache for accessing memory block meta-data that may be expensive
* to access directly.
*
* \note This class exists to unify the metadata format between GPU and CPU
* allocations. It should be removed when the CPU can access all GPU
* allocations directly via UVM.
*/
class MetadataCache {
public:
MetadataCache(bool uses_gpu);
public:
/*! \brief Load the associated metadata for the specified memory block. */
Metadata load(const MemoryBlock*);
/*! \brief Store the associated metadata for the specified memory block. */
void store(MemoryBlock*, const Metadata&);
/*! \brief Indicate that the specified metadata will no longer be used. */
void invalidate(MemoryBlock*);
public:
MetadataCache(const MetadataCache&) = delete;
MetadataCache& operator=(const MetadataCache&) = delete;
private:
bool uses_gpu_;
private:
typedef std::unordered_map<const MemoryBlock*, Metadata> MetadataMap;
private:
MetadataMap cache_;
};
} // namespace detail
} // namespace memory
} // namespace paddle
paddle/memory/detail/meta_data.cc 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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/memory/detail/meta_data.h"
#include <functional>
namespace paddle {
namespace memory {
namespace detail {
Metadata::Metadata(MemoryBlock::Type t, size_t i, size_t s, size_t ts,
MemoryBlock* l, MemoryBlock* r)
: type(t),
index(i),
size(s),
total_size(ts),
left_buddy(l),
right_buddy(r) {}
Metadata::Metadata()
: type(MemoryBlock::INVALID_CHUNK),
index(0),
size(0),
total_size(0),
left_buddy(nullptr),
right_buddy(nullptr) {}
template <class T>
inline void hash_combine(std::size_t& seed, const T& v) {
std::hash<T> hasher;
seed ^= hasher(v) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
}
inline size_t hash(const Metadata* metadata, size_t initial_seed) {
size_t seed = initial_seed;
hash_combine(seed, (size_t)metadata->type);
hash_combine(seed, metadata->index);
hash_combine(seed, metadata->size);
hash_combine(seed, metadata->total_size);
hash_combine(seed, metadata->left_buddy);
hash_combine(seed, metadata->right_buddy);
return seed;
}
void Metadata::update_guards() {
guard_begin = hash(this, 1);
guard_end = hash(this, 2);
}
bool Metadata::check_guards() const {
return guard_begin == hash(this, 1) && guard_end == hash(this, 2);
}
} // namespace detail
} // namespace memory
} // namespace paddle
paddle/memory/detail/meta_data.h 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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/memory/detail/memory_block.h"
#include <stddef.h>
namespace paddle {
namespace memory {
namespace detail {
class Metadata {
public:
Metadata(MemoryBlock::Type t, size_t i, size_t s, size_t ts, MemoryBlock* l,
MemoryBlock* r);
Metadata();
public:
/*! \brief Update the guards when metadata is changed */
void update_guards();
/*! \brief Check consistency to previous modification */
bool check_guards() const;
public:
// TODO(gangliao): compress this
// clang-format off
size_t guard_begin = 0;
MemoryBlock::Type type = MemoryBlock::INVALID_CHUNK;
size_t index = 0;
size_t size = 0;
size_t total_size = 0;
MemoryBlock* left_buddy = nullptr;
MemoryBlock* right_buddy = nullptr;
size_t guard_end = 0;
// clang-format on
};
} // namespace detail
} // namespace memory
} // namespace paddle
paddle/memory/detail/system_allocator.cc
浏览文件 @ 9e3a9eb2
......@@ -13,76 +13,128 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/memory/detail/system_allocator.h"
#include "paddle/platform/assert.h"
#include "paddle/platform/error.h"
#include "paddle/platform/gpu_info.h"
#include <stdlib.h> // for malloc and free
#include <sys/mman.h> // for mlock and munlock
#include "gflags/gflags.h"
#include "paddle/platform/assert.h"
#include "paddle/platform/cuda.h"
// If use_pinned_memory is true, CPUAllocator calls mlock, which
// returns pinned and locked memory as staging areas for data exchange
// between host and device. Allocates too much would reduce the amount
// of memory available to the system for paging. So, by default, we
// should set false to use_pinned_memory.
DEFINE_bool(use_pinned_memory, false,
"If set, allocate cpu/gpu pinned memory.");
DEFINE_bool(use_pinned_memory, false, "If set, allocate cpu pinned memory.");
namespace paddle {
namespace memory {
namespace detail {
void* CPUAllocator::Alloc(size_t size) {
void* CPUAllocator::Alloc(size_t& index, size_t size) {
// According to http://www.cplusplus.com/reference/cstdlib/malloc/,
// malloc might not return nullptr if size is zero, but the returned
// pointer shall not be dereferenced -- so we make it nullptr.
if (size <= 0) return nullptr;
index = 0; // unlock memory
void* p = malloc(size);
if (p != nullptr && FLAGS_use_pinned_memory) {
mlock(p, size);
if (p != nullptr) {
if (FLAGS_use_pinned_memory) {
index = 1;
mlock(p, size); // lock memory
}
}
return p;
}
void CPUAllocator::Free(void* p, size_t size) {
if (p != nullptr && FLAGS_use_pinned_memory) {
void CPUAllocator::Free(void* p, size_t size, size_t index) {
if (p != nullptr && index == 1) {
munlock(p, size);
}
free(p);
}
bool CPUAllocator::UseGpu() const { return false; }
#ifndef PADDLE_ONLY_CPU
void* GPUAllocator::Alloc(size_t size) {
void* GPUAllocator::Alloc(size_t& index, size_t size) {
// CUDA documentation doesn't explain if cudaMalloc returns nullptr
// if size is 0. We just make sure it does.
if (size <= 0) {
return nullptr;
}
if (size <= 0) return nullptr;
size_t available = 0;
size_t capacity = 0;
paddle::platform::GpuMemoryUsage(available, capacity);
// Reserve memory for page tables, etc.
size_t reserving = capacity - paddle::platform::GpuMaxAllocSize();
size_t usable = available > reserving ? available - reserving : 0;
// If remaining size no less than expected size, using general
// cudaMalloc to allocate GPU memory.
void* p = 0;
cudaError_t result =
FLAGS_use_pinned_memory ? cudaMallocHost(&p, size) : cudaMalloc(&p, size);
if (result != cudaSuccess) {
cudaGetLastError(); // clear error if there is any.
if (size <= usable) {
cudaError_t result = cudaMalloc(&p, size);
if (result == cudaSuccess) {
index = 0;
gpu_alloc_size_ += size;
return p;
}
}
// If remaining size less than expected size or cudaMalloc failed,
// cudaMallocHost will be considered as a fallback allocator.
//
// NOTE: here, we use GpuMaxAllocSize() as the maximum memory size
// of host fallback allocation. Allocates too much would reduce
// the amount of memory available to the underlying system for paging.
usable = paddle::platform::GpuMaxAllocSize() - fallback_alloc_size_;
if (size > usable) return nullptr;
cudaError_t result = cudaMallocHost(&p, size);
if (result == cudaSuccess) {
index = 1;
fallback_alloc_size_ += size;
return p;
}
return result == cudaSuccess ? p : nullptr;
return nullptr;
}
void GPUAllocator::Free(void* p, size_t size) {
void GPUAllocator::Free(void* p, size_t size, size_t index) {
cudaError_t err;
if (index == 0) {
PADDLE_ASSERT(gpu_alloc_size_ >= size);
gpu_alloc_size_ -= size;
err = cudaFree(p);
} else {
PADDLE_ASSERT(fallback_alloc_size_ >= size);
fallback_alloc_size_ -= size;
err = cudaFreeHost(p);
}
// Purposefully allow cudaErrorCudartUnloading, because
// that is returned if you ever call cudaFree after the
// driver has already shutdown. This happens only if the
// process is terminating, in which case we don't care if
// cudaFree succeeds.
cudaError_t err = FLAGS_use_pinned_memory ? cudaFreeHost(p) : cudaFree(p);
if (err != cudaErrorCudartUnloading) {
platform::throw_on_error(err, "cudaFree{Host} failed");
platform::throw_on_error(err,
"cudaFree{Host} failed in GPUAllocator::Free.");
}
}
bool GPUAllocator::UseGpu() const { return true; }
#endif // PADDLE_ONLY_CPU
} // namespace detail
......
paddle/memory/detail/system_allocator.h
浏览文件 @ 9e3a9eb2
......@@ -20,31 +20,36 @@ namespace paddle {
namespace memory {
namespace detail {
// SystemAllocator is the parent class of CPUAllocator and
// GPUAllocator. A BuddyAllocator object uses a SystemAllocator*
// pointing to the underlying system allocator. An alternative to
// this class hierarchy is to pass a system allocator class to
// BuddyAllocator as a template parameter. This approach makes
// BuddyAllocator a class template, and it's very complicated
// algorithm would make the buddy_allocator.h messy.
/**
* \brief SystemAllocator is the parent class of CPUAllocator and GPUAllocator.
* A BuddyAllocator object uses a SystemAllocator* pointing to the
* underlying system allocator.
*/
class SystemAllocator {
public:
virtual ~SystemAllocator() {}
virtual void* Alloc(size_t size) = 0;
virtual void Free(void* p, size_t size) = 0;
virtual void* Alloc(size_t& index, size_t size) = 0;
virtual void Free(void* p, size_t size, size_t index) = 0;
virtual bool UseGpu() const = 0;
};
class CPUAllocator : public SystemAllocator {
public:
virtual void* Alloc(size_t size);
virtual void Free(void* p, size_t size);
virtual void* Alloc(size_t& index, size_t size);
virtual void Free(void* p, size_t size, size_t index);
virtual bool UseGpu() const;
};
#ifndef PADDLE_ONLY_CPU
class GPUAllocator : public SystemAllocator {
public:
virtual void* Alloc(size_t size);
virtual void Free(void* p, size_t size);
virtual void* Alloc(size_t& index, size_t size);
virtual void Free(void* p, size_t size, size_t index);
virtual bool UseGpu() const;
private:
size_t gpu_alloc_size_ = 0;
size_t fallback_alloc_size_ = 0;
};
#endif // PADDLE_ONLY_CPU
......
paddle/memory/detail/system_allocator_test.cc
浏览文件 @ 9e3a9eb2
......@@ -25,7 +25,8 @@ DECLARE_bool(use_pinned_memory);
void TestAllocator(paddle::memory::detail::SystemAllocator& a, size_t size) {
bool freed = false;
{
void* p = a.Alloc(size);
size_t index;
void* p = a.Alloc(index, size);
if (size > 0) {
EXPECT_NE(p, nullptr);
} else {
......@@ -35,7 +36,7 @@ void TestAllocator(paddle::memory::detail::SystemAllocator& a, size_t size) {
int* i = static_cast<int*>(p);
std::shared_ptr<int> ptr(i, [&](void* p) {
freed = true;
a.Free(p, size);
a.Free(p, size, index);
});
}
EXPECT_TRUE(freed);
......@@ -56,14 +57,7 @@ TEST(CPUAllocator, LockMem) {
}
#ifndef PADDLE_ONLY_CPU
TEST(GPUAllocator, NoStaging) {
FLAGS_use_pinned_memory = false;
paddle::memory::detail::GPUAllocator a;
TestAllocator(a, 2048);
TestAllocator(a, 0);
}
TEST(GPUAllocator, Staging) {
FLAGS_use_pinned_memory = true;
TEST(GPUAllocator, Alloc) {
paddle::memory::detail::GPUAllocator a;
TestAllocator(a, 2048);
TestAllocator(a, 0);
......
paddle/memory/memory.cc
浏览文件 @ 9e3a9eb2
......@@ -17,43 +17,67 @@ limitations under the License. */
#include "paddle/memory/detail/system_allocator.h"
#include "paddle/platform/assert.h"
#include <boost/variant.hpp>
namespace paddle {
namespace memory {
void* Alloc(platform::Place pl, size_t size) {
#ifndef PADDLE_ONLY_CPU
if (paddle::platform::is_gpu_place(pl)) {
size_t gpu_id = boost::get<platform::GPUPlace>(pl).device;
return detail::GetGPUBuddyAllocator(gpu_id)->Alloc(size);
detail::BuddyAllocator* GetCPUBuddyAllocator() {
static detail::BuddyAllocator* a = nullptr;
if (a == nullptr) {
a = new detail::BuddyAllocator(new detail::CPUAllocator,
platform::CpuMinChunkSize(),
platform::CpuMaxChunkSize());
}
#endif // PADDLE_ONLY_CPU
PADDLE_ASSERT(paddle::platform::is_cpu_place(pl));
return detail::GetCPUBuddyAllocator()->Alloc(size);
return a;
}
void Free(paddle::platform::Place pl, void* p) {
#ifndef PADDLE_ONLY_CPU
if (paddle::platform::is_gpu_place(pl)) {
size_t gpu_id = boost::get<platform::GPUPlace>(pl).device;
detail::GetGPUBuddyAllocator(gpu_id)->Free(p);
}
#endif // PADDLE_ONLY_CPU
PADDLE_ASSERT(paddle::platform::is_cpu_place(pl));
detail::GetCPUBuddyAllocator()->Free(p);
template <>
void* Alloc<platform::CPUPlace>(platform::CPUPlace place, size_t size) {
return GetCPUBuddyAllocator()->Alloc(size);
}
template <>
void Free<platform::CPUPlace>(platform::CPUPlace place, void* p) {
GetCPUBuddyAllocator()->Free(p);
}
template <>
size_t Used<platform::CPUPlace>(platform::CPUPlace place) {
return GetCPUBuddyAllocator()->Used();
}
size_t Used(paddle::platform::Place pl) {
#ifndef PADDLE_ONLY_CPU
if (paddle::platform::is_gpu_place(pl)) {
size_t gpu_id = boost::get<platform::GPUPlace>(pl).device;
return detail::GetGPUBuddyAllocator(gpu_id)->Used();
detail::BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
static detail::BuddyAllocator** as = NULL;
if (as == NULL) {
int gpu_num = platform::GetDeviceCount();
as = new detail::BuddyAllocator*[gpu_num];
for (int gpu = 0; gpu < gpu_num; gpu++) {
platform::SetDeviceId(gpu);
as[gpu] = new detail::BuddyAllocator(new detail::GPUAllocator,
platform::GpuMinChunkSize(),
platform::GpuMaxChunkSize());
}
}
#endif // PADDLE_ONLY_CPU
PADDLE_ASSERT(paddle::platform::is_cpu_place(pl));
return detail::GetCPUBuddyAllocator()->Used();
return as[gpu_id];
}
template <>
void* Alloc<platform::GPUPlace>(platform::GPUPlace place, size_t size) {
return GetGPUBuddyAllocator(place.device)->Alloc(size);
}
template <>
void Free<platform::GPUPlace>(platform::GPUPlace place, void* p) {
GetGPUBuddyAllocator(place.device)->Free(p);
}
template <>
size_t Used<platform::GPUPlace>(platform::GPUPlace place) {
return GetGPUBuddyAllocator(place.device)->Used();
}
#endif // PADDLE_ONLY_CPU
} // namespace memory
} // namespace paddle
paddle/memory/memory.h
浏览文件 @ 9e3a9eb2
......@@ -19,9 +19,14 @@ limitations under the License. */
namespace paddle {
namespace memory {
void* Alloc(paddle::platform::Place, size_t);
void Free(paddle::platform::Place, void*);
size_t Used(paddle::platform::Place);
template <class Place>
void* Alloc(Place, size_t);
template <class Place>
void Free(Place, void*);
template <class Place>
size_t Used(Place);
} // namespace memory
} // namespace paddle
paddle/memory/memory_test.cc 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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/memory/memory.h"
#include "paddle/memory/detail/memory_block.h"
#include "paddle/memory/detail/meta_data.h"
#include "paddle/platform/cpu_info.h"
#include "paddle/platform/gpu_info.h"
#include "paddle/platform/place.h"
#include <gtest/gtest.h>
#include <unordered_map>
inline bool is_aligned(void const *p) {
return 0 == (reinterpret_cast<uintptr_t>(p) & 0x3);
}
size_t align(size_t size, paddle::platform::CPUPlace place) {
size += sizeof(paddle::memory::detail::Metadata);
size_t alignment = paddle::platform::CpuMinChunkSize();
size_t remaining = size % alignment;
return remaining == 0 ? size : size + (alignment - remaining);
}
TEST(BuddyAllocator, CPUAllocation) {
void *p = nullptr;
EXPECT_EQ(p, nullptr);
paddle::platform::CPUPlace cpu;
p = paddle::memory::Alloc(cpu, 4096);
EXPECT_NE(p, nullptr);
paddle::memory::Free(cpu, p);
}
TEST(BuddyAllocator, CPUMultAlloc) {
paddle::platform::CPUPlace cpu;
std::unordered_map<void *, size_t> ps;
size_t total_size = paddle::memory::Used(cpu);
EXPECT_EQ(total_size, 0UL);
for (auto size :
{128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) {
ps[paddle::memory::Alloc(cpu, size)] = size;
// Buddy Allocator doesn't manage too large memory chunk
if (paddle::memory::Used(cpu) == total_size) continue;
size_t aligned_size = align(size, cpu);
total_size += aligned_size;
EXPECT_EQ(total_size, paddle::memory::Used(cpu));
}
for (auto p : ps) {
EXPECT_EQ(is_aligned(p.first), true);
paddle::memory::Free(cpu, p.first);
// Buddy Allocator doesn't manage too large memory chunk
if (paddle::memory::Used(cpu) == total_size) continue;
size_t aligned_size = align(p.second, cpu);
total_size -= aligned_size;
EXPECT_EQ(total_size, paddle::memory::Used(cpu));
}
}
#ifndef PADDLE_ONLY_CPU
size_t align(size_t size, paddle::platform::GPUPlace place) {
size += sizeof(paddle::memory::detail::Metadata);
size_t alignment = paddle::platform::GpuMinChunkSize();
size_t remaining = size % alignment;
return remaining == 0 ? size : size + (alignment - remaining);
}
TEST(BuddyAllocator, GPUAllocation) {
void *p = nullptr;
EXPECT_EQ(p, nullptr);
paddle::platform::GPUPlace gpu(0);
p = paddle::memory::Alloc(gpu, 4096);
EXPECT_NE(p, nullptr);
paddle::memory::Free(gpu, p);
}
TEST(BuddyAllocator, GPUMultAlloc) {
paddle::platform::GPUPlace gpu;
std::unordered_map<void *, size_t> ps;
size_t total_size = paddle::memory::Used(gpu);
EXPECT_EQ(total_size, 0UL);
for (auto size :
{128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) {
ps[paddle::memory::Alloc(gpu, size)] = size;
// Buddy Allocator doesn't manage too large memory chunk
if (paddle::memory::Used(gpu) == total_size) continue;
size_t aligned_size = align(size, gpu);
total_size += aligned_size;
EXPECT_EQ(total_size, paddle::memory::Used(gpu));
}
for (auto p : ps) {
EXPECT_EQ(is_aligned(p.first), true);
paddle::memory::Free(gpu, p.first);
// Buddy Allocator doesn't manage too large memory chunk
if (paddle::memory::Used(gpu) == total_size) continue;
size_t aligned_size = align(p.second, gpu);
total_size -= aligned_size;
EXPECT_EQ(total_size, paddle::memory::Used(gpu));
}
}
#endif // PADDLE_ONLY_CPU
paddle/operators/CMakeLists.txt
浏览文件 @ 9e3a9eb2
if(WITH_GPU)
nv_library(add_op SRCS add_op.cc add_op.cu DEPS operator op_registry glog ddim)
nv_library(add_op SRCS add_op.cc add_op.cu DEPS operator op_registry ddim glog paddle_memory)
else()
cc_library(add_op SRCS add_op.cc DEPS operator op_registry glog ddim)
cc_library(add_op SRCS add_op.cc DEPS operator op_registry ddim glog paddle_memory)
endif()
cc_test(add_op_test SRCS add_op_test.cc DEPS add_op)
paddle/operators/add_op.cc
浏览文件 @ 9e3a9eb2
/* 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/operators/add_op.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/tensor.h"
......@@ -17,8 +31,7 @@ protected:
"Inputs/Outputs of AddOp must all be set");
PADDLE_ENFORCE(inputs[0]->dims() == inputs[1]->dims(),
"Two input of Add Op's dimension must be same.");
// Need set dims in Tensor
// outputs[0]->set_dims(inputs[0]->dims())
outputs[0]->set_dims(inputs[0]->dims());
}
};
......@@ -42,4 +55,4 @@ The equation is: Out = X + Y
REGISTER_OP(add_two, paddle::operators::AddOp, paddle::operators::AddOpMaker);
typedef paddle::operators::AddKernel<::paddle::platform::CPUPlace, float>
AddKernel_CPU_float;
REGISTER_OP_CPU_KERNEL(add_two, AddKernel_CPU_float);
\ No newline at end of file
REGISTER_OP_CPU_KERNEL(add_two, AddKernel_CPU_float);
paddle/operators/add_op.h
浏览文件 @ 9e3a9eb2
/* 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 "glog/logging.h"
#include "paddle/framework/operator.h"
//#include "paddle/operators/add_op_functor.h"
namespace paddle {
namespace operators {
......
paddle/operators/add_op_test.cc
浏览文件 @ 9e3a9eb2
/* 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 <gtest/gtest.h>
#define private public
#include <paddle/framework/op_registry.h>
......
paddle/optimizer/parameter_optimizer_test.cpp
浏览文件 @ 9e3a9eb2
/*
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 "parameter_optimizer.h"
#include <cmath>
#include <map>
......@@ -5,21 +21,18 @@
#include "gtest/gtest.h"
#include "lr_policy.h"
using namespace paddle;
using namespace paddle::optimizer;
Tensor* FillTensor(size_t size) {
Tensor* param = new Tensor(size);
Tensor& p = *param;
paddle::optimizer::Tensor* FillTensor(size_t size) {
paddle::optimizer::Tensor* param = new paddle::optimizer::Tensor(size);
paddle::optimizer::Tensor& p = *param;
for (size_t i = 0; i < p.size(); ++i) {
p[i] = (float)rand() / (float)RAND_MAX;
}
return param;
}
Tensor* FixedTensor(size_t size) {
Tensor* param = new Tensor(size);
Tensor& p = *param;
paddle::optimizer::Tensor* FixedTensor(size_t size) {
paddle::optimizer::Tensor* param = new paddle::optimizer::Tensor(size);
paddle::optimizer::Tensor& p = *param;
for (size_t i = 0; i < p.size(); ++i) {
p[i] = i;
}
......@@ -28,7 +41,8 @@ Tensor* FixedTensor(size_t size) {
class OptimizerTest : public testing::Test {
public:
// init tensor shape
virtual ~OptimizerTest() {}
// init paddle::optimizer::Tensor shape
const size_t kSize = 5;
virtual void SetUp() {
......@@ -38,34 +52,36 @@ public:
virtual void TearDown() {}
void CreateSGD() {
Tensor* parameter = FixedTensor(kSize);
config_.set_optimizer(OptimizerConfig::SGD);
paddle::optimizer::Tensor* parameter = FixedTensor(kSize);
config_.set_optimizer(paddle::OptimizerConfig::SGD);
config_.mutable_sgd()->set_momentum(0.0);
config_.mutable_sgd()->set_decay(0.0);
config_.mutable_sgd()->set_nesterov(false);
config_.set_lr_policy(OptimizerConfig::Const);
config_.set_lr_policy(paddle::OptimizerConfig::Const);
config_.mutable_const_lr()->set_learning_rate(0.1);
std::string str = config_.SerializeAsString();
ParameterOptimizer* opt = ParameterOptimizer::Create(str, parameter);
paddle::optimizer::ParameterOptimizer* opt =
paddle::optimizer::ParameterOptimizer::Create(str, parameter);
opts_.push_back(opt);
}
void CreateAdam() {
Tensor* parameter = FixedTensor(kSize);
config_.set_optimizer(OptimizerConfig::Adam);
paddle::optimizer::Tensor* parameter = FixedTensor(kSize);
config_.set_optimizer(paddle::OptimizerConfig::Adam);
config_.mutable_adam()->set_beta_1(0.9);
config_.mutable_adam()->set_beta_2(0.1);
config_.mutable_adam()->set_epsilon(1e-3);
config_.mutable_adam()->set_decay(0.0);
config_.set_lr_policy(OptimizerConfig::Const);
config_.set_lr_policy(paddle::OptimizerConfig::Const);
config_.mutable_const_lr()->set_learning_rate(0.1);
std::string str = config_.SerializeAsString();
ParameterOptimizer* opt = ParameterOptimizer::Create(str, parameter);
paddle::optimizer::ParameterOptimizer* opt =
paddle::optimizer::ParameterOptimizer::Create(str, parameter);
opts_.push_back(opt);
}
void TestGetWeight() {
Tensor* p = FixedTensor(kSize);
paddle::optimizer::Tensor* p = FixedTensor(kSize);
for (size_t i = 0; i < opts_.size(); ++i) {
int s = 0;
float* newp = (float*)opts_[i]->get_weight(&s);
......@@ -76,7 +92,7 @@ public:
}
void TestUpdate() {
Tensor* g = FixedTensor(kSize);
paddle::optimizer::Tensor* g = FixedTensor(kSize);
for (size_t i = 0; i < opts_.size(); ++i) {
opts_[i]->Update(g);
}
......@@ -91,8 +107,8 @@ public:
}
private:
std::vector<ParameterOptimizer*> opts_;
OptimizerConfig config_;
std::vector<paddle::optimizer::ParameterOptimizer*> opts_;
paddle::OptimizerConfig config_;
};
TEST_F(OptimizerTest, TestGetWeight) { TestGetWeight(); }
......
paddle/optimizer/serialization_test.cpp
浏览文件 @ 9e3a9eb2
/*
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 "serialization.h"
#include "gtest/gtest.h"
using namespace paddle;
using namespace paddle::optimizer;
TEST(TensorToProto, Case1) {
Tensor t(3), t1(3);
paddle::optimizer::Tensor t(3), t1(3);
for (size_t i = 0; i < t.size(); ++i) {
t[i] = i;
t1[i] = 0;
}
TensorProto proto;
TensorToProto(t, &proto);
ProtoToTensor(proto, &t1);
paddle::TensorProto proto;
paddle::optimizer::TensorToProto(t, &proto);
paddle::optimizer::ProtoToTensor(proto, &t1);
for (size_t i = 0; i < t1.size(); ++i) {
EXPECT_EQ(t1[i], t[i]);
}
......
paddle/platform/CMakeLists.txt
浏览文件 @ 9e3a9eb2
add_subdirectory(dynload)
cc_library(cpu_info SRCS cpu_info.cc DEPS gflags glog)
cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info)
nv_test(cuda_test SRCS cuda_test.cu)
nv_library(gpu_info SRCS gpu_info.cc DEPS gflags)
cc_library(place SRCS place.cc)
cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
add_subdirectory(dynload)
IF(WITH_GPU)
set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
ELSE()
......@@ -12,4 +15,4 @@ ELSE()
ENDIF()
cc_library(device_context SRCS device_context.cc DEPS place eigen3 ${GPU_CTX_DEPS})
nv_test(device_context_test SRCS device_context_test.cc DEPS device_context glog gflags)
nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)
paddle/platform/cpu_info.cc 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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/platform/cpu_info.h"
#ifdef __APPLE__
#include <sys/sysctl.h>
#include <sys/types.h>
#else
#include <unistd.h>
#endif
#include "gflags/gflags.h"
#include "paddle/platform/error.h"
DEFINE_double(fraction_of_cpu_memory_to_use, 1,
"Default use 100% of CPU memory for PaddlePaddle,"
"reserve the rest for page tables, etc");
namespace paddle {
namespace platform {
inline size_t CpuTotalPhysicalMemory() {
#ifdef __APPLE__
int mib[2];
mib[0] = CTL_HW;
mib[1] = HW_MEMSIZE;
int64_t size = 0;
size_t len = sizeof(size);
if (sysctl(mib, 2, &size, &len, NULL, 0) == 0) return (size_t)size;
return 0L;
#else
int64_t pages = sysconf(_SC_PHYS_PAGES);
int64_t page_size = sysconf(_SC_PAGE_SIZE);
return pages * page_size;
#endif
}
size_t CpuMaxAllocSize() {
// For distributed systems, it requires configuring and limiting
// the fraction of memory to use.
return FLAGS_fraction_of_cpu_memory_to_use * CpuTotalPhysicalMemory();
}
size_t CpuMinChunkSize() {
// Allow to allocate the minimum chunk size is 4 KB.
return 1 << 12;
}
size_t CpuMaxChunkSize() {
// Allow to allocate the maximum chunk size is roughly 3% of CPU memory.
return CpuMaxAllocSize() / 32;
}
} // namespace platform
} // namespace paddle
paddle/operators/add_op_functor.h → paddle/platform/cpu_info.h
浏览文件 @ 9e3a9eb2
......@@ -14,22 +14,19 @@ limitations under the License. */
#pragma once
#include "paddle/framework/tensor_types.h"
#include "unsupported/Eigen/CXX11/Tensor"
#include <stddef.h>
namespace paddle {
namespace operators {
namespace functor {
template <typename Device, typename T>
struct Add {
void Operator()(const Device& d,
typename TTypes<T>::ConstTensor input1,
typename TTypes<T>::ConstTensor input2,
typename TTypes<T>::Tensor output) {
output.device(d) = input1 + input2;
}
};
} // namespace functor
} // namespace operators
namespace platform {
//! Get the maximum allocation size for a machine.
size_t CpuMaxAllocSize();
//! Get the minimum chunk size for buddy allocator.
size_t CpuMinChunkSize();
//! Get the maximum chunk size for buddy allocator.
size_t CpuMaxChunkSize();
} // namespace platform
} // namespace paddle
paddle/platform/cpu_info_test.cc 0 → 100644
浏览文件 @ 9e3a9eb2
#include "paddle/platform/cpu_info.h"
#include "paddle/string/printf.h"
#include <ostream>
#include <sstream>
#include "gflags/gflags.h"
#include "glog/logging.h"
#include "gtest/gtest.h"
DECLARE_double(fraction_of_cpu_memory_to_use);
TEST(CpuMemoryUsage, Print) {
std::stringstream ss;
size_t memory_size = paddle::platform::CpuMaxAllocSize() / 1024 / 1024 / 1024;
float use_percent = FLAGS_fraction_of_cpu_memory_to_use * 100;
std::cout << paddle::string::Sprintf("\n%.2f %% of CPU Memory Usage: %d GB\n",
use_percent, memory_size)
<< std::endl;
}
paddle/platform/cuda_test.cu 已删除 100644 → 0
浏览文件 @ d649dbf4
#include <cuda_runtime.h>
#include <stdio.h>
#include "gtest/gtest.h"
#define CHECK_ERR(x) \
if (x != cudaSuccess) { \
fprintf(stderr, \
"%s in %s at line %d\n", \
cudaGetErrorString(err), \
__FILE__, \
__LINE__); \
exit(-1); \
}
__global__ void vecAdd(float *d_A, float *d_B, float *d_C, int n) {
int i = blockDim.x * blockIdx.x + threadIdx.x;
if (i < n) {
d_C[i] = d_A[i] + d_B[i];
}
}
TEST(Cuda, Equality) {
int n = 10;
// Memory allocation for h_A, h_B and h_C (in the host)
float h_A[10] = {1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 0.0};
float h_B[10] = {0.0, 9.0, 8.0, 7.0, 6.0, 5.0, 4.0, 3.0, 2.0, 1.0};
float h_C[10];
float *d_A, *d_B, *d_C;
cudaError_t err;
// Memory allocation for d_A, d_B and d_C (in the device)
err = cudaMalloc((void **)&d_A, sizeof(float) * n);
CHECK_ERR(err);
err = cudaMalloc((void **)&d_B, sizeof(float) * n);
CHECK_ERR(err);
err = cudaMalloc((void **)&d_C, sizeof(float) * n);
CHECK_ERR(err);
// Copying memory to device
err = cudaMemcpy(d_A, h_A, sizeof(float) * n, cudaMemcpyHostToDevice);
CHECK_ERR(err);
err = cudaMemcpy(d_B, h_B, sizeof(float) * n, cudaMemcpyHostToDevice);
CHECK_ERR(err);
// Calling the kernel
vecAdd<<<ceil(n / 256.0), 256>>>(d_A, d_B, d_C, n);
// Copying results back to host
err = cudaMemcpy(h_C, d_C, sizeof(float) * n, cudaMemcpyDeviceToHost);
CHECK_ERR(err);
EXPECT_EQ(h_C[0], 1.0);
for (int i = 1; i < n - 1; ++i) {
EXPECT_EQ(h_C[i], 11.0);
}
EXPECT_EQ(h_C[9], 1.0);
}
paddle/platform/device_context.h
浏览文件 @ 9e3a9eb2
......@@ -13,10 +13,11 @@ limitations under the License. */
#include "paddle/framework/enforce.h"
#ifndef PADDLE_ONLY_CPU
#include "paddle/platform/cuda.h"
#include "paddle/platform/dynload/cublas.h"
#include "paddle/platform/dynload/cudnn.h"
#include "paddle/platform/dynload/curand.h"
#include "paddle/platform/error.h"
#include "paddle/platform/gpu_info.h"
#define EIGEN_USE_GPU
#endif
#include <paddle/platform/place.h>
......
paddle/platform/error.h 0 → 100644
浏览文件 @ 9e3a9eb2
#pragma once
#include <sstream>
#include <stdexcept>
#include <string>
#ifndef PADDLE_ONLY_CPU
#include <cublas_v2.h>
#include <cudnn.h>
#include <curand.h>
#include <thrust/system/cuda/error.h>
#include <thrust/system_error.h>
#endif // PADDLE_ONLY_CPU
namespace paddle {
namespace platform {
#ifndef PADDLE_ONLY_CPU
inline void throw_on_error(cudaError_t e, const char* message) {
if (e) {
throw thrust::system_error(e, thrust::cuda_category(), message);
}
}
inline void throw_on_error(curandStatus_t stat, const char* message) {
if (stat != CURAND_STATUS_SUCCESS) {
throw thrust::system_error(cudaErrorLaunchFailure, thrust::cuda_category(),
message);
}
}
inline void throw_on_error(cudnnStatus_t stat, const char* message) {
std::stringstream ss;
if (stat == CUDNN_STATUS_SUCCESS) {
return;
} else {
ss << cudnnGetErrorString(stat);
ss << ", " << message;
throw std::runtime_error(ss.str());
}
}
inline void throw_on_error(cublasStatus_t stat, const char* message) {
std::stringstream ss;
if (stat == CUBLAS_STATUS_SUCCESS) {
return;
} else if (stat == CUBLAS_STATUS_NOT_INITIALIZED) {
ss << "CUBLAS: not initialized";
} else if (stat == CUBLAS_STATUS_ALLOC_FAILED) {
ss << "CUBLAS: alloc failed";
} else if (stat == CUBLAS_STATUS_INVALID_VALUE) {
ss << "CUBLAS: invalid value";
} else if (stat == CUBLAS_STATUS_ARCH_MISMATCH) {
ss << "CUBLAS: arch mismatch";
} else if (stat == CUBLAS_STATUS_MAPPING_ERROR) {
ss << "CUBLAS: mapping error";
} else if (stat == CUBLAS_STATUS_EXECUTION_FAILED) {
ss << "CUBLAS: execution failed";
} else if (stat == CUBLAS_STATUS_INTERNAL_ERROR) {
ss << "CUBLAS: internal error";
} else if (stat == CUBLAS_STATUS_NOT_SUPPORTED) {
ss << "CUBLAS: not supported";
} else if (stat == CUBLAS_STATUS_LICENSE_ERROR) {
ss << "CUBLAS: license error";
}
ss << ", " << message;
throw std::runtime_error(ss.str());
}
inline void throw_on_error(cublasStatus_t stat) {
const char* message = "";
throw_on_error(stat, message);
}
#endif // PADDLE_ONLY_CPU
inline void throw_on_error(int stat, const char* message) {
if (stat) {
throw std::runtime_error(message + (", stat = " + std::to_string(stat)));
}
}
} // namespace platform
} // namespace paddle
paddle/platform/gpu_info.cc 0 → 100644
浏览文件 @ 9e3a9eb2
/* 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/platform/gpu_info.h"
#include "gflags/gflags.h"
#include "paddle/platform/error.h"
DEFINE_double(fraction_of_gpu_memory_to_use, 0.95,
"Default use 95% of GPU memory for PaddlePaddle,"
"reserve the rest for page tables, etc");
namespace paddle {
namespace platform {
int GetDeviceCount() {
int count;
throw_on_error(
cudaGetDeviceCount(&count),
"cudaGetDeviceCount failed in paddle::platform::GetDeviceCount");
return count;
}
int GetCurrentDeviceId() {
int device_id;
throw_on_error(
cudaGetDevice(&device_id),
"cudaGetDevice failed in paddle::platform::GetCurrentDeviceId");
return device_id;
}
void SetDeviceId(int id) {
throw_on_error(cudaSetDevice(id),
"cudaSetDevice failed in paddle::platform::SetDeviceId");
}
void GpuMemoryUsage(size_t& available, size_t& total) {
throw_on_error(cudaMemGetInfo(&available, &total),
"cudaMemGetInfo failed in paddle::platform::GetMemoryUsage");
}
size_t GpuMaxAllocSize() {
size_t total = 0;
size_t available = 0;
GpuMemoryUsage(available, total);
// Reserve the rest for page tables, etc.
return static_cast<size_t>(total * FLAGS_fraction_of_gpu_memory_to_use);
}
size_t GpuMinChunkSize() {
// Allow to allocate the minimum chunk size is 256 bytes.
return 1 << 8;
}
size_t GpuMaxChunkSize() {
size_t total = 0;
size_t available = 0;
GpuMemoryUsage(available, total);
// Reserving the rest memory for page tables, etc.
size_t reserving = (1 - FLAGS_fraction_of_gpu_memory_to_use) * total;
// If available less than minimum chunk size, no usable memory exists.
available = std::max(available, GpuMinChunkSize()) - GpuMinChunkSize();
// If available less than reserving, no usable memory exists.
size_t usable = std::max(available, reserving) - reserving;
return usable;
}
} // namespace platform
} // namespace paddle
paddle/platform/cuda.h → paddle/platform/gpu_info.h
浏览文件 @ 9e3a9eb2
......@@ -16,33 +16,31 @@ limitations under the License. */
#ifndef PADDLE_ONLY_CPU
#include <thrust/system/cuda/error.h>
#include <thrust/system_error.h>
#include <stddef.h>
namespace paddle {
namespace platform {
inline void throw_on_error(cudaError_t e, const char* message) {
if (e) {
throw thrust::system_error(e, thrust::cuda_category(), message);
}
}
inline int GetDeviceCount(void) {
int count;
throw_on_error(cudaGetDeviceCount(&count), "cudaGetDeviceCount failed");
return count;
}
inline int GetCurrentDeviceId(void) {
int device_id;
throw_on_error(cudaGetDevice(&device_id), "cudaGetDevice failed");
return device_id;
}
inline void SetDeviceId(int device_id) {
throw_on_error(cudaSetDevice(device_id), "cudaSetDevice failed");
}
//! Get the total number of GPU devices in system.
int GetDeviceCount();
//! Get the current GPU device id in system.
int GetCurrentDeviceId();
//! Set the GPU device id for next execution.
void SetDeviceId(int device_id);
//！Get the memory usage of current GPU device.
void GpuMemoryUsage(size_t& available, size_t& total);
//! Get the maximum allocation size of current GPU device.
size_t GpuMaxAllocSize();
//! Get the minimum chunk size for GPU buddy allocator.
size_t GpuMinChunkSize();
//! Get the maximum chunk size for GPU buddy allocator.
size_t GpuMaxChunkSize();
} // namespace platform
} // namespace paddle
......
paddle/platform/place.cc
浏览文件 @ 9e3a9eb2
/* 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/platform/place.h"
namespace paddle {
......@@ -7,7 +21,7 @@ namespace detail {
class PlacePrinter : public boost::static_visitor<> {
public:
PlacePrinter(std::ostream &os) : os_(os) {}
explicit PlacePrinter(std::ostream &os) : os_(os) {}
void operator()(const CPUPlace &) { os_ << "CPUPlace"; }
void operator()(const GPUPlace &p) { os_ << "GPUPlace(" << p.device << ")"; }
......
paddle/pybind/CMakeLists.txt
浏览文件 @ 9e3a9eb2
cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python)
cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python add_op)
paddle/pybind/pybind.cc
浏览文件 @ 9e3a9eb2
......@@ -13,12 +13,18 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include <Python.h>
#include <paddle/framework/op_registry.h>
#include <paddle/framework/scope.h>
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include <fstream>
#include <vector>
namespace py = pybind11;
namespace pd = paddle::framework;
USE_OP(add_two);
PYBIND11_PLUGIN(core) {
py::module m("core", "C++ core of Paddle Paddle");
......@@ -43,5 +49,20 @@ All parameter, weight, gradient are variables in Paddle.
&pd::Scope::CreateVariable,
py::return_value_policy::reference);
//! @note: Be careful! PyBind will return std::string as an unicode, not
//! Python str. If you want a str object, you should cast them in Python.
m.def("get_all_op_protos", []() -> std::vector<std::string> {
auto& protos = pd::OpRegistry::protos();
std::vector<std::string> ret_values;
for (auto it = protos.begin(); it != protos.end(); ++it) {
PADDLE_ENFORCE(it->second.IsInitialized(),
"OpProto must all be initialized");
ret_values.emplace_back();
PADDLE_ENFORCE(it->second.SerializeToString(&ret_values.back()),
"Serialize OpProto Error. This could be a bug of Paddle.");
}
return ret_values;
});
return m.ptr();
}
paddle/string/piece.h
浏览文件 @ 9e3a9eb2
......@@ -35,7 +35,7 @@ public:
// We provide non-explicit singleton constructors so users can
// pass in a "const char*" or a "string" wherever a "Piece"
// is expected. These contructors ensure that if data_ is NULL,
// is expected. These constructors ensure that if data_ is NULL,
// size_ is 0.
Piece();
Piece(const char* d, size_t n);
......
paddle/trainer/TrainerConfigHelper.cpp
浏览文件 @ 9e3a9eb2
......@@ -62,11 +62,7 @@ TrainerConfigHelper::TrainerConfigHelper(const TrainerConfig &config)
m->conf = config;
}
TrainerConfigHelper::~TrainerConfigHelper() {
if (m) {
delete m;
}
}
TrainerConfigHelper::~TrainerConfigHelper() { delete m; }
const TrainerConfig &TrainerConfigHelper::getConfig() const { return m->conf; }
......
paddle/utils/DynamicLoader.h
浏览文件 @ 9e3a9eb2
......@@ -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 DYNAMIC_LOAD_H_
#define DYNAMIC_LOAD_H_
#pragma once
#include <dlfcn.h>
#include <memory>
......@@ -59,5 +58,3 @@ void GetWarpCTCDsoHandle(void** dso_handle);
*
*/
void GetLapackDsoHandle(void** dso_handle);
#endif // DYNAMIC_LOAD_H_
paddle/utils/ThreadLocal.h
浏览文件 @ 9e3a9eb2
......@@ -51,7 +51,7 @@ template <class T>
class ThreadLocal {
public:
ThreadLocal() {
CHECK(pthread_key_create(&threadSpecificKey_, dataDestructor) == 0);
CHECK_EQ(pthread_key_create(&threadSpecificKey_, dataDestructor), 0);
}
~ThreadLocal() { pthread_key_delete(threadSpecificKey_); }
......@@ -65,7 +65,7 @@ public:
if (!p && createLocal) {
p = new T();
int ret = pthread_setspecific(threadSpecificKey_, p);
CHECK(ret == 0);
CHECK_EQ(ret, 0);
}
return p;
}
......@@ -79,7 +79,7 @@ public:
if (T* q = get(false)) {
dataDestructor(q);
}
CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
CHECK_EQ(pthread_setspecific(threadSpecificKey_, p), 0);
}
/**
......@@ -112,7 +112,7 @@ private:
template <class T>
class ThreadLocalD {
public:
ThreadLocalD() { CHECK(pthread_key_create(&threadSpecificKey_, NULL) == 0); }
ThreadLocalD() { CHECK_EQ(pthread_key_create(&threadSpecificKey_, NULL), 0); }
~ThreadLocalD() {
pthread_key_delete(threadSpecificKey_);
for (auto t : threadMap_) {
......@@ -127,7 +127,7 @@ public:
T* p = (T*)pthread_getspecific(threadSpecificKey_);
if (!p) {
p = new T();
CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
CHECK_EQ(pthread_setspecific(threadSpecificKey_, p), 0);
updateMap(p);
}
return p;
......@@ -141,7 +141,7 @@ public:
if (T* q = (T*)pthread_getspecific(threadSpecificKey_)) {
dataDestructor(q);
}
CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
CHECK_EQ(pthread_setspecific(threadSpecificKey_, p), 0);
updateMap(p);
}
......
python/paddle/trainer_config_helpers/layers.py
浏览文件 @ 9e3a9eb2
......@@ -126,6 +126,7 @@ __all__ = [
'row_conv_layer',
'dropout_layer',
'prelu_layer',
'gated_unit_layer',
]
......@@ -5862,7 +5863,7 @@ def prelu_layer(input,
:rtype: LayerOutput
"""
assert isinstance(input, LayerOutput), 'prelu_layer only accepts one input'
assert isinstance(input, LayerOutput), 'prelu_layer accepts only one input.'
assert isinstance(param_attr, ParameterAttribute)
l = Layer(
......@@ -5876,3 +5877,96 @@ def prelu_layer(input,
layer_type=LayerType.PRELU,
parents=input,
size=l.config.size)
@wrap_name_default()
@layer_support(ERROR_CLIPPING, DROPOUT)
@wrap_act_default(act=LinearActivation())
def gated_unit_layer(input,
size,
act=None,
name=None,
gate_attr=None,
gate_param_attr=None,
gate_bias_attr=True,
inproj_attr=None,
inproj_param_attr=None,
inproj_bias_attr=True,
layer_attr=None):
"""
The gated unit layer implements a simple gating mechanism over the input.
The input :math:`X` is first projected into a new space :math:`X'`, and
it is also used to produce a gate weight :math:`\sigma`. Element-wise
prodict between :match:`X'` and :math:`\sigma` is finally returned.
Reference:
Language Modeling with Gated Convolutional Networks
https://arxiv.org/abs/1612.08083
.. math::
y=\\text{act}(X \cdot W + b)\otimes \sigma(X \cdot V + c)
The example usage is:
.. code-block:: python
gated_unit = gated_unit_layer(size=128, input=input_layer))
:param input: input for this layer.
:type input: LayerOutput
:param size: output size of the gated unit.
:type size: int
:param act: activation type of the projected input.
:type act: BaseActivation
:param name: name of this layer.
:type name: basestring
:param gate_attr: Attributes to tune the gate output, for example, error
clipping threshold, dropout and so on. See ExtraLayerAttribute for
more details.
:type gate_attr: ExtraLayerAttribute|None
:param gate_param_attr: Attributes to tune the learnable projected matrix
parameter of the gate.
:type gate_param_attr: ParameterAttribute|None
:param gate_bias_attr: Attributes to tune the learnable bias of the gate.
:type gate_bias_attr: ParameterAttribute|None
:param inproj_attr: Attributes to the tune the projected input, for
example, error clipping threshold, dropout and so on. See
ExtraLayerAttribute for more details.
:type inproj_attr: ExtraLayerAttribute|None
:param inproj_param_attr: Attributes to tune the learnable parameter of
the projection of input.
:type inproj_param_attr: ParameterAttribute|None
:param inproj_bias_attr: Attributes to tune the learnable bias of
projection of the input.
:type inproj_bias_attr: ParameterAttribute|None
:param layer_attr: Attributes to tune the final output of the gated unit,
for example, error clipping threshold, dropout and so on. See
ExtraLayerAttribute for more details.
:type layer_attr: ExtraLayerAttribute|None
:return: LayerOutput object.
:rtype: LayerOutput
"""
assert isinstance(
input, LayerOutput), 'The gated linear unit accepts only one input.'
input_proj = fc_layer(
input=input,
name="%s_input_proj" % name,
size=size,
act=act,
layer_attr=inproj_attr,
param_attr=inproj_param_attr,
bias_attr=inproj_bias_attr)
gate = fc_layer(
size=size,
name="%s_gate" % name,
act=SigmoidActivation(),
input=input,
layer_attr=gate_attr,
param_attr=gate_param_attr,
bias_attr=gate_bias_attr)
return mixed_layer(
name="%s_gated_act" % name,
input=dotmul_operator(input_proj, gate),
layer_attr=layer_attr)
python/paddle/trainer_config_helpers/tests/configs/file_list.sh
浏览文件 @ 9e3a9eb2
......@@ -7,6 +7,6 @@ test_rnn_group shared_fc shared_lstm shared_gru test_cost_layers_with_weight
test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
test_recursive_topology)
test_recursive_topology test_gated_unit_layer)
export whole_configs=(test_split_datasource)
python/paddle/trainer_config_helpers/tests/configs/protostr/test_gated_unit_layer.protostr 0 → 100644
浏览文件 @ 9e3a9eb2
type: "nn"
layers {
name: "input"
type: "data"
size: 256
active_type: ""
}
layers {
name: "__gated_unit_layer_0___input_proj"
type: "fc"
size: 512
active_type: "tanh"
inputs {
input_layer_name: "input"
input_parameter_name: "___gated_unit_layer_0___input_proj.w0"
}
bias_parameter_name: "___gated_unit_layer_0___input_proj.wbias"
error_clipping_threshold: 100.0
}
layers {
name: "__gated_unit_layer_0___gate"
type: "fc"
size: 512
active_type: "sigmoid"
inputs {
input_layer_name: "input"
input_parameter_name: "___gated_unit_layer_0___gate.w0"
}
bias_parameter_name: "___gated_unit_layer_0___gate.wbias"
error_clipping_threshold: 100.0
}
layers {
name: "__gated_unit_layer_0___gated_act"
type: "mixed"
size: 512
active_type: ""
inputs {
input_layer_name: "__gated_unit_layer_0___input_proj"
}
inputs {
input_layer_name: "__gated_unit_layer_0___gate"
}
error_clipping_threshold: 100.0
operator_confs {
type: "dot_mul"
input_indices: 0
input_indices: 1
input_sizes: 512
input_sizes: 512
output_size: 512
dotmul_scale: 1
}
}
parameters {
name: "___gated_unit_layer_0___input_proj.w0"
size: 131072
initial_mean: 0.0
initial_std: 0.0001
dims: 256
dims: 512
initial_strategy: 0
initial_smart: false
}
parameters {
name: "___gated_unit_layer_0___input_proj.wbias"
size: 512
initial_mean: 0.0
initial_std: 1
dims: 1
dims: 512
initial_strategy: 0
initial_smart: false
}
parameters {
name: "___gated_unit_layer_0___gate.w0"
size: 131072
initial_mean: 0.0
initial_std: 0.0001
dims: 256
dims: 512
initial_strategy: 0
initial_smart: false
}
parameters {
name: "___gated_unit_layer_0___gate.wbias"
size: 512
initial_mean: 0.0
initial_std: 1
dims: 1
dims: 512
initial_strategy: 0
initial_smart: false
}
input_layer_names: "input"
output_layer_names: "__gated_unit_layer_0___gated_act"
sub_models {
name: "root"
layer_names: "input"
layer_names: "__gated_unit_layer_0___input_proj"
layer_names: "__gated_unit_layer_0___gate"
layer_names: "__gated_unit_layer_0___gated_act"
input_layer_names: "input"
output_layer_names: "__gated_unit_layer_0___gated_act"
is_recurrent_layer_group: false
}
python/paddle/trainer_config_helpers/tests/configs/test_gated_unit_layer.py 0 → 100644
浏览文件 @ 9e3a9eb2
from paddle.trainer_config_helpers import *
data = data_layer(name='input', size=256)
glu = gated_unit_layer(
size=512,
input=data,
act=TanhActivation(),
gate_attr=ExtraLayerAttribute(error_clipping_threshold=100.0),
gate_param_attr=ParamAttr(initial_std=1e-4),
gate_bias_attr=ParamAttr(initial_std=1),
inproj_attr=ExtraLayerAttribute(error_clipping_threshold=100.0),
inproj_param_attr=ParamAttr(initial_std=1e-4),
inproj_bias_attr=ParamAttr(initial_std=1),
layer_attr=ExtraLayerAttribute(error_clipping_threshold=100.0))
outputs(glu)
python/paddle/v2/__init__.py
浏览文件 @ 9e3a9eb2
......@@ -20,7 +20,6 @@ import trainer
import event
import data_type
import topology
import data_feeder
import networks
import evaluator
from . import dataset
......@@ -31,7 +30,6 @@ import op
import pooling
import inference
import networks
import py_paddle.swig_paddle as api
import minibatch
import plot
import image
......@@ -47,7 +45,6 @@ __all__ = [
'data_type',
'attr',
'pooling',
'data_feeder',
'dataset',
'reader',
'topology',
......@@ -61,6 +58,7 @@ __all__ = [
def init(**kwargs):
import py_paddle.swig_paddle as api
args = []
args_dict = {}
# NOTE: append arguments if they are in ENV
......
python/paddle/v2/data_feeder.py
浏览文件 @ 9e3a9eb2
......@@ -11,7 +11,6 @@
# 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 py_paddle import DataProviderConverter
import collections
import paddle.trainer.PyDataProvider2 as pydp2
......
python/paddle/v2/event.py
浏览文件 @ 9e3a9eb2
......@@ -9,8 +9,6 @@ There are:
* BeginPass
* EndPass
"""
import py_paddle.swig_paddle as api
__all__ = [
'EndIteration', 'BeginIteration', 'BeginPass', 'EndPass', 'TestResult'
]
......@@ -18,6 +16,7 @@ __all__ = [
class WithMetric(object):
def __init__(self, evaluator):
import py_paddle.swig_paddle as api
if not isinstance(evaluator, api.Evaluator):
raise TypeError("Evaluator should be api.Evaluator type")
self.__evaluator__ = evaluator
......
python/paddle/v2/framework/create_op_creation_methods.py 0 → 100644
浏览文件 @ 9e3a9eb2
import paddle.v2.framework.core as core
import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2
def get_all_op_protos():
protostrs = core.get_all_op_protos()
ret_values = []
for pbstr in protostrs:
op_proto = op_proto_pb2.OpProto.FromString(str(pbstr))
ret_values.append(op_proto)
return ret_values
python/paddle/v2/framework/tests/CMakeLists.txt
浏览文件 @ 9e3a9eb2
add_python_test(test_framework test_protobuf.py test_scope.py
test_default_scope_funcs.py)
test_default_scope_funcs.py test_op_creation_methods.py)
python/paddle/v2/framework/tests/test_op_creation_methods.py 0 → 100644
浏览文件 @ 9e3a9eb2
import unittest
import paddle.v2.framework.create_op_creation_methods as creation
class TestOpCreationsMethods(unittest.TestCase):
def test_all_protos(self):
all_protos = creation.get_all_op_protos()
self.assertNotEqual(0, len(all_protos))
for each in all_protos:
self.assertTrue(each.IsInitialized())
if __name__ == "__main__":
unittest.main()
python/paddle/v2/inference.py
浏览文件 @ 9e3a9eb2
import numpy
import py_paddle.swig_paddle as api
import collections
import topology
import minibatch
from data_feeder import DataFeeder
__all__ = ['infer', 'Inference']
......@@ -28,6 +26,7 @@ class Inference(object):
"""
def __init__(self, output_layer, parameters):
import py_paddle.swig_paddle as api
topo = topology.Topology(output_layer)
gm = api.GradientMachine.createFromConfigProto(
topo.proto(), api.CREATE_MODE_TESTING, [api.PARAMETER_VALUE])
......@@ -40,6 +39,7 @@ class Inference(object):
self.__data_types__ = topo.data_type()
def iter_infer(self, input, feeding=None):
from data_feeder import DataFeeder
feeder = DataFeeder(self.__data_types__, feeding)
batch_size = len(input)
......
python/paddle/v2/optimizer.py
浏览文件 @ 9e3a9eb2
import py_paddle.swig_paddle as swig_api
import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils
import paddle.trainer_config_helpers.optimizers as v1_optimizers
"""
......@@ -18,6 +16,7 @@ __all__ = [
class Optimizer(object):
def __init__(self, **kwargs):
import py_paddle.swig_paddle as swig_api
if 'batch_size' in kwargs:
del kwargs['batch_size'] # not important for python library.
......@@ -268,6 +267,7 @@ ModelAverage = v1_optimizers.ModelAverage
L2Regularization = v1_optimizers.L2Regularization
if __name__ == '__main__':
import py_paddle.swig_paddle as swig_api
swig_api.initPaddle('--use_gpu=false')
for opt in [
Momentum(), Adam(), Adamax(), AdaGrad(), DecayedAdaGrad(),
......
python/paddle/v2/parameters.py
浏览文件 @ 9e3a9eb2
import numpy as np
import py_paddle.swig_paddle as api
from paddle.proto.ParameterConfig_pb2 import ParameterConfig
import paddle.trainer.config_parser as cp
import struct
......@@ -124,6 +123,7 @@ class Parameters(object):
:return: parameter value
:rtype: np.ndarray
"""
import py_paddle.swig_paddle as api
shape = self.get_shape(key)
if len(self.__gradient_machines__) == 0:
......@@ -223,7 +223,7 @@ class Parameters(object):
:type gradient_machine: api.GradientMachine
:return:
"""
import py_paddle.swig_paddle as api
if not isinstance(gradient_machine, api.GradientMachine):
raise ValueError("gradient_machine should be api.GradientMachine")
......@@ -359,6 +359,7 @@ def __copy_parameter_to_gradient_machine__(gradient_machine, name, arr):
:return:
:rtype: api.Parameter
"""
import py_paddle.swig_paddle as api
param = __get_parameter_in_gradient_machine__(gradient_machine, name)
vec = param.getBuf(api.PARAMETER_VALUE)
assert isinstance(vec, api.Vector)
......
python/paddle/v2/trainer.py
浏览文件 @ 9e3a9eb2
......@@ -2,12 +2,6 @@
Module Trainer
"""
import collections
import gzip
import os
import py_paddle.swig_paddle as api
from data_feeder import DataFeeder
from topology import Topology
from . import event as v2_event
from . import optimizer as v2_optimizer
......@@ -59,6 +53,7 @@ class SGD(object):
if not isinstance(update_equation, v2_optimizer.Optimizer):
raise TypeError("update equation parameter must be "
"paddle.v2.optimizer.Optimizer")
import py_paddle.swig_paddle as api
topology = Topology(cost, extra_layers=extra_layers)
self.__optimizer__ = update_equation
self.__topology__ = topology
......@@ -124,6 +119,8 @@ class SGD(object):
:type feeding: dict|list
:return:
"""
import py_paddle.swig_paddle as api
from data_feeder import DataFeeder
if event_handler is None:
event_handler = default_event_handler
__check_train_args__(**locals())
......@@ -187,6 +184,8 @@ class SGD(object):
:type feeding: dict
:return:
"""
import py_paddle.swig_paddle as api
from data_feeder import DataFeeder
feeder = DataFeeder(self.__data_types__, feeding)
evaluator = self.__gradient_machine__.makeEvaluator()
out_args = api.Arguments.createArguments(0)
......
python/setup.py.in
浏览文件 @ 9e3a9eb2
......@@ -19,7 +19,8 @@ setup_requires=["requests",
"recordio",
"matplotlib",
"rarfile",
"scipy>=0.19.0"]
"scipy>=0.19.0",
"nltk"]
if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
setup_requires+=["opencv-python"]
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或