提交 de8a8fee 编写于 作者: L liaogang

ENH: Refine Tensor and And CopyFrom

上级 1294b3c5
......@@ -3,7 +3,7 @@ cc_library(ddim SRCS ddim.cc DEPS eigen3)
cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
nv_test(dim_test SRCS dim_test.cu DEPS ddim)
cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory)
cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context)
cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
......
/* 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/memcpy.h"
namespace paddle {
namespace framework {
template <typename T>
inline void Tensor::check_memory_size() const {
PADDLE_ENFORCE(holder_ != nullptr,
"Tenosr holds no memory. Call Tensor::mutable_data first.");
PADDLE_ENFORCE(holder_->size() >= product(dims_) * sizeof(T) + offset_,
"Tensor's dims_ is out of bound. Call Tensor::mutable_data "
"first to re-allocate memory.");
}
template <typename T>
inline const T* Tensor::data() const {
check_memory_size<T>();
return reinterpret_cast<const T*>(
reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
}
template <typename T>
inline T* Tensor::data() {
check_memory_size<T>();
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
template <typename T>
inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
Resize(dims);
return mutable_data<T>(place);
}
template <typename T>
inline T* Tensor::mutable_data(platform::Place place) {
PADDLE_ENFORCE(product(dims_) > 0,
"Tensor's numel must be larger than zero to call "
"Tensor::mutable_data. Call Tensor::set_dim first.");
/* some versions of boost::variant don't have operator!= */
size_t size = product(dims_) * sizeof(T);
if (holder_ == nullptr || !(holder_->place() == place) ||
holder_->size() < size + offset_) {
if (platform::is_cpu_place(place)) {
holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
boost::get<platform::CPUPlace>(place), size));
}
#ifndef PADDLE_ONLY_CPU
else if (platform::is_gpu_place(place)) {
holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
boost::get<platform::GPUPlace>(place), size));
}
#endif
offset_ = 0;
}
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
template <typename T>
inline void Tensor::ShareDataWith(const Tensor& src) {
src.check_memory_size<T>();
*this = src;
}
template <typename T>
inline void Tensor::CopyFrom(const Tensor& src,
const platform::CPUDeviceContext& ctx) {
src.check_memory_size<T>();
Resize(src.dims());
auto src_place = src.holder_->place();
auto src_ptr = static_cast<const void*>(src.data<T>());
auto dst_place = ctx.GetPlace();
auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
auto size = product(src.dims_) * sizeof(T);
if (platform::is_cpu_place(src_place)) {
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
boost::get<platform::CPUPlace>(src_place), src_ptr, size);
}
#ifndef PADDLE_ONLY_CPU
else if (platform::is_gpu_place(src_place)) {
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
}
#endif
}
#ifndef PADDLE_ONLY_CPU
template <typename T>
inline void Tensor::CopyFrom(const Tensor& src,
const platform::CUDADeviceContext& ctx) {
src.check_memory_size<T>();
Resize(src.dims());
auto src_place = src.holder_->place();
auto src_ptr = static_cast<const void*>(src.data<T>());
auto dst_place = ctx.GetPlace();
auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
auto size = product(src.dims_) * sizeof(T);
if (platform::is_cpu_place(src_place)) {
memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
boost::get<platform::CPUPlace>(src_place), src_ptr, size,
ctx.stream());
} else if (platform::is_gpu_place(src_place)) {
memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
boost::get<platform::GPUPlace>(src_place), src_ptr, size,
ctx.stream());
}
}
#endif
template <typename T>
inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
check_memory_size<T>();
PADDLE_ENFORCE(begin_idx >= 0, "Slice begin index is less than zero.");
PADDLE_ENFORCE(end_idx <= dims_[0], "Slice end index is out of bound.");
PADDLE_ENFORCE(begin_idx < end_idx,
"Begin index must be less than end index.");
PADDLE_ENFORCE(dims_[0] != 1, "Can not slice a tensor with dims_[0] = 1.");
int base = product(dims_) / dims_[0];
Tensor dst;
dst.holder_ = holder_;
DDim dst_dims = dims_;
dst_dims[0] = end_idx - begin_idx;
dst.Resize(dst_dims);
dst.offset_ = offset_ + begin_idx * base * sizeof(T);
return dst;
}
inline void Tensor::Resize(const DDim& dims) { dims_ = dims; }
inline const DDim& Tensor::dims() const { return dims_; }
} // namespace framework
} // namespace paddle
......@@ -12,7 +12,7 @@
See the License for the specific language governing permissions and
limitations under the License. */
#include <paddle/framework/tensor.h>
#include "paddle/framework/tensor.h"
namespace paddle {
namespace framework {}
......
......@@ -20,6 +20,7 @@ limitations under the License. */
#include <typeindex>
#include "paddle/framework/ddim.h"
#include "paddle/memory/memory.h"
#include "paddle/platform/device_context.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/place.h"
#include "unsupported/Eigen/CXX11/Tensor"
......@@ -31,9 +32,11 @@ template <bool less, size_t i, typename... args>
struct CastToPyBufferImpl;
} // namespace details
} // namespace pybind
namespace framework {
class Tensor {
public:
template <bool less, size_t i, typename... args>
friend struct paddle::pybind::details::CastToPyBufferImpl;
......@@ -46,106 +49,84 @@ class Tensor {
public:
Tensor() : offset_(0) {}
/*! Return a pointer to mutable memory block. */
template <typename T>
const T* data() const {
EnforceSufficientMemory<T>();
return reinterpret_cast<const T*>(
reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
}
inline T* data();
/*! Return a pointer to constant memory block. */
template <typename T>
T* data() {
EnforceSufficientMemory<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>
T* mutable_data(DDim dims, platform::Place place) {
Resize(dims);
return mutable_data<T>(place);
}
template <typename T, // must be POD types
typename std::enable_if<std::is_pod<T>::value>::type* = nullptr>
T* mutable_data(platform::Place place) {
PADDLE_ENFORCE(product(dims_) > 0,
"Tensor's numel must be larger than zero to call "
"Tensor::mutable_data. Call Tensor::set_dim first.");
if (holder_ == nullptr ||
!(holder_->place() ==
place) /* some versions of boost::variant don't have operator!= */
|| holder_->size() < product(dims_) * sizeof(T) + offset_) {
if (platform::is_cpu_place(place)) {
holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
boost::get<platform::CPUPlace>(place), product(dims_) * sizeof(T)));
} else if (platform::is_gpu_place(place)) {
#ifdef PADDLE_ONLY_CPU
PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
#else
holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
boost::get<platform::GPUPlace>(place), product(dims_) * sizeof(T)));
#endif
} else {
PADDLE_THROW("Unknown 'place'.");
}
offset_ = 0;
}
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
inline const T* data() const;
/**
* @brief Return a pointer to mutable memory block.
* @note If not exist, then allocation.
*/
template <typename T>
inline T* mutable_data(platform::Place place);
/**
* @brief Return a pointer to mutable memory block.
*
* @param[in] dims The dimensions of the memory block.
* @param[in] place The place of the memory block.
*
* @note If not exist, then allocation.
*/
template <typename T>
inline T* mutable_data(DDim dims, platform::Place place);
/*! Return the dimensions of the memory block. */
inline const DDim& dims() const;
/*! Resize the dimensions of the memory block. */
inline void Resize(const DDim& dims);
/*! The internal of two tensors share the same memory block. */
template <typename T>
inline void ShareDataWith(const Tensor& src);
/**
* @brief Copy the content of external tensor to a new place.
*
* @param[in] src The external tensor.
* @param[in] ctx The device context contains place where to store.
*
* @note CopyFrom supports CPU <-> GPU, GPU <-> GPU.
*/
template <typename T>
void ShareDataWith(const Tensor& src) {
src.EnforceSufficientMemory<T>();
*this = src;
}
inline void CopyFrom(const Tensor& src,
const platform::CPUDeviceContext& ctx);
#ifndef PADDLE_ONLY_CPU
template <typename T>
void CopyFrom(const Tensor& src, 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.EnforceSufficientMemory<T>();
size_t size = product(src.dims_) * sizeof(T);
Resize(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);
}
inline void CopyFrom(const Tensor& src,
const platform::CUDADeviceContext& ctx);
#endif
/**
* @brief Return the slice of the tensor.
*
* @param[in] begin_idx The begin index of the slice.
* @param[in] end_idx The end index of the slice.
*/
template <typename T>
Tensor Slice(const int& begin_idx, const int& end_idx) const {
EnforceSufficientMemory<T>();
PADDLE_ENFORCE(begin_idx >= 0, "Slice begin index is less than zero.");
PADDLE_ENFORCE(end_idx <= dims_[0], "Slice end index is out of bound.");
PADDLE_ENFORCE(begin_idx < end_idx,
"Begin index must be less than end index.");
PADDLE_ENFORCE(dims_[0] != 1, "Can not slice a tensor with dims_[0] = 1.");
int base = product(dims_) / dims_[0];
Tensor dst;
dst.holder_ = holder_;
DDim dst_dims = dims_;
dst_dims[0] = end_idx - begin_idx;
dst.Resize(dst_dims);
dst.offset_ = offset_ + begin_idx * base * sizeof(T);
return dst;
}
void Resize(const DDim& dims) { dims_ = dims; }
const DDim& dims() const { return dims_; }
inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
private:
// Placeholder hides type T, so it doesn't appear as a template
// parameter of Variable.
template <typename T>
inline void check_memory_size() const;
private:
/**
* @note Placeholder hides type T, so it doesn't appear as a template
* parameter of Variable.
*/
struct Placeholder {
virtual ~Placeholder() {}
virtual void* ptr() const = 0;
virtual platform::Place place() const = 0;
virtual size_t size() const = 0;
virtual std::type_index type() const = 0;
virtual platform::Place place() const = 0;
};
template <typename T, typename PlaceType>
......@@ -156,33 +137,38 @@ class Tensor {
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 platform::Place place() const { return place_; }
virtual void* ptr() const { return static_cast<void*>(ptr_.get()); }
virtual std::type_index type() const { return std::type_index(typeid(T)); }
/*! the pointer of memory block. */
std::unique_ptr<T, memory::PODDeleter<T, PlaceType>> ptr_;
platform::Place place_; // record the place of ptr_.
size_t size_; // size of the memory block.
/*! the place of memory block. */
platform::Place place_;
/*! the size of memory block. */
size_t size_;
};
template <typename T>
inline void EnforceSufficientMemory() const {
PADDLE_ENFORCE(holder_ != nullptr,
"Tenosr holds no memory. Call Tensor::mutable_data first.");
PADDLE_ENFORCE(holder_->size() >= product(dims_) * 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.
/*! holds the memory block if allocated. */
std::shared_ptr<Placeholder> holder_;
/*! points to dimensions of memory block. */
DDim dims_;
// A PlaceHolder may be shared by more than one tensor. Some of them may be
// slices of the others. So the offset_ is introduced here to indicate the
// byte offset between PlaceHolder::ptr_ and where tensor's data really
// begins.
/**
* @brief A PlaceHolder may be shared by more than one tensor.
*
* @note Some of them may be slices of the others. So the offset_
* is introduced here to indicate the byte offset between
* PlaceHolder::ptr_ and where the tensor data really begins.
*/
size_t offset_;
};
} // namespace framework
} // namespace paddle
#include "paddle/framework/detail/tensor-inl.h"
......@@ -72,7 +72,8 @@ TEST(Tensor, MutableData) {
p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), CPUPlace());
EXPECT_EQ(p1, p2);
}
#ifdef __CUDACC__
#ifndef PADDLE_ONLY_CPU
{
Tensor src_tensor;
float* p1 = nullptr;
......@@ -123,7 +124,7 @@ TEST(Tensor, ShareDataWith) {
ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
}
#ifdef __CUDACC__
#ifndef PADDLE_ONLY_CPU
{
Tensor src_tensor;
Tensor dst_tensor;
......@@ -160,7 +161,7 @@ TEST(Tensor, Slice) {
EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address);
}
#ifdef __CUDACC__
#ifndef PADDLE_ONLY_CPU
{
Tensor src_tensor;
src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
......@@ -188,13 +189,18 @@ TEST(Tensor, Slice) {
TEST(Tensor, CopyFrom) {
using namespace paddle::framework;
using namespace paddle::platform;
{
Tensor src_tensor;
Tensor dst_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());
auto cpu_ctx = paddle::platform::CPUDeviceContext();
dst_tensor.CopyFrom<int>(src_tensor, cpu_ctx);
const int* dst_ptr = dst_tensor.data<int>();
ASSERT_NE(src_ptr, dst_ptr);
for (size_t i = 0; i < 9; ++i) {
......@@ -202,11 +208,12 @@ TEST(Tensor, CopyFrom) {
}
Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
dst_tensor.CopyFrom<int>(slice_tensor, CPUPlace());
dst_tensor.CopyFrom<int>(slice_tensor, cpu_ctx);
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]);
}
}
}
......@@ -87,7 +87,7 @@ class CUDADeviceContext : public DeviceContext {
"cudaStreamSynchronize failed");
}
cudaStream_t stream() { return stream_; }
cudaStream_t stream() const { return stream_; }
Eigen::GpuDevice* eigen_device() const { return eigen_device_.get(); }
......
Markdown is supported
0% .
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