提交 e615c84a 编写于 作者: Z zchen0211

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

...@@ -62,6 +62,24 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place, ...@@ -62,6 +62,24 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
} }
} }
template <>
void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
void* dst,
platform::GPUPlace src_place,
const void* src, size_t num) {
platform::SetDeviceId(src_place.device);
platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost);
}
template <>
void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
void* dst,
platform::CPUPlace src_place,
const void* src, size_t num) {
platform::SetDeviceId(dst_place.device);
platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice);
}
#endif // PADDLE_ONLY_CPU #endif // PADDLE_ONLY_CPU
} // namespace memory } // namespace memory
......
...@@ -56,7 +56,7 @@ class CosSimKernel : public framework::OpKernel { ...@@ -56,7 +56,7 @@ class CosSimKernel : public framework::OpKernel {
x_norm.device(place) = x.square().sum(row_along).sqrt(); x_norm.device(place) = x.square().sum(row_along).sqrt();
y_norm.device(place) = y.square().sum(row_along).sqrt(); y_norm.device(place) = y.square().sum(row_along).sqrt();
if (rows_x == rows_y) { if (rows_x == rows_y) {
auto xy = (x * y).sum(Eigen::array<int, 1>({1})); auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
z.device(place) = xy / x_norm / y_norm; z.device(place) = xy / x_norm / y_norm;
} else { } else {
Eigen::DSizes<int, 2> bcast(rows_x, 1); Eigen::DSizes<int, 2> bcast(rows_x, 1);
...@@ -134,7 +134,7 @@ class CosSimGradKernel : public framework::OpKernel { ...@@ -134,7 +134,7 @@ class CosSimGradKernel : public framework::OpKernel {
out_grad_y->mutable_data<T>(context.GetPlace()); out_grad_y->mutable_data<T>(context.GetPlace());
auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1); auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast; auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({0})); dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({{0}}));
} }
} }
} }
......
...@@ -24,3 +24,4 @@ cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator ...@@ -24,3 +24,4 @@ cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info) nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)
nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda) nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place)
/* 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
#ifndef __NVCC__
#error device_ptr_cast must be include by .cu file
#endif
#include <thrust/device_ptr.h>
namespace paddle {
namespace platform {
namespace details {
template <typename T, bool is_ptr>
struct DevicePtrCast;
template <typename T>
struct DevicePtrCast<T, true> {
using ELEM = typename std::remove_pointer<T>::type;
using RTYPE = thrust::device_ptr<ELEM>;
inline thrust::device_ptr<ELEM> operator()(ELEM* ele) const {
return thrust::device_pointer_cast(ele);
}
};
template <typename T>
struct DevicePtrCast<T, false> {
using RTYPE = T;
inline RTYPE operator()(RTYPE it) const { return it; }
};
// Cast T to thrust::device_ptr if T is a pointer.
// Otherwise, e.g., T is a iterator, return T itself.
template <typename T>
auto DevPtrCast(T t) ->
typename DevicePtrCast<T, std::is_pointer<T>::value>::RTYPE {
DevicePtrCast<T, std::is_pointer<T>::value> cast;
return cast(t);
}
} // namespace details
} // namespace platform
} // namespace paddle
/* 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/platform/enforce.h"
#include "paddle/platform/hostdevice.h"
#include "paddle/platform/place.h"
#include <algorithm>
#include <type_traits>
#ifdef __NVCC__
#include <thrust/transform.h>
#include "paddle/platform/details/device_ptr_cast.h"
#endif
namespace paddle {
namespace platform {
// Transform on host or device. It provides the same API in std library.
template <typename Place, typename InputIter, typename OutputIter,
typename UnaryOperation>
void Transform(Place place, InputIter first, InputIter last, OutputIter result,
UnaryOperation op) {
if (is_cpu_place(place)) {
std::transform(first, last, result, op);
} else {
#ifdef __NVCC__
using namespace details;
thrust::transform(DevPtrCast(first), DevPtrCast(last), DevPtrCast(result),
op);
#else
PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
#endif
}
}
template <typename Place, typename InputIter1, typename InputIter2,
typename OutputIter, typename BinaryOperation>
void Transform(Place place, InputIter1 first1, InputIter1 last1,
InputIter2 first2, OutputIter result, BinaryOperation op) {
if (is_cpu_place(place)) {
std::transform(first1, last1, first2, result, op);
} else {
#ifdef __NVCC__
using namespace details;
thrust::transform(DevPtrCast(first1), DevPtrCast(last1), DevPtrCast(first2),
DevPtrCast(result), op);
#else
PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
#endif
}
};
} // namespace platform
} // namespace paddle
/* 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>
#include "paddle/memory/memcpy.h"
#include "paddle/memory/memory.h"
#include "paddle/platform/transform.h"
template <typename T>
class Scale {
public:
explicit Scale(const T& scale) : scale_(scale) {}
HOSTDEVICE T operator()(const T& a) const { return a * scale_; }
private:
T scale_;
};
template <typename T>
class Multiply {
public:
HOSTDEVICE T operator()(const T& a, const T& b) const { return a * b; }
};
TEST(Transform, CPUUnary) {
using namespace paddle::platform;
float buf[4] = {0.1, 0.2, 0.3, 0.4};
Transform(CPUPlace(), buf, buf + 4, buf, Scale<float>(10));
for (int i = 0; i < 4; ++i) {
ASSERT_NEAR(buf[i], static_cast<float>(i + 1), 1e-5);
}
}
TEST(Transform, GPUUnary) {
using namespace paddle::platform;
using namespace paddle::memory;
GPUPlace gpu0(0);
float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4};
float* gpu_buf = static_cast<float*>(Alloc(gpu0, sizeof(float) * 4));
Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf));
Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf));
Free(gpu0, gpu_buf);
for (int i = 0; i < 4; ++i) {
ASSERT_NEAR(cpu_buf[i], static_cast<float>(i + 1), 1e-5);
}
}
TEST(Transform, CPUBinary) {
using namespace paddle::platform;
using namespace paddle::memory;
int buf[4] = {1, 2, 3, 4};
Transform(CPUPlace(), buf, buf + 4, buf, buf, Multiply<int>());
for (int i = 0; i < 4; ++i) {
ASSERT_EQ((i + 1) * (i + 1), buf[i]);
}
}
TEST(Transform, GPUBinary) {
using namespace paddle::platform;
using namespace paddle::memory;
int buf[4] = {1, 2, 3, 4};
GPUPlace gpu0(0);
int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf));
Free(gpu0, gpu_buf);
for (int i = 0; i < 4; ++i) {
ASSERT_EQ((i + 1) * (i + 1), buf[i]);
}
}
\ No newline at end of file
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