未验证 提交 087af6a6 编写于 作者: X Xin Pan 提交者: GitHub

Merge pull request #15131 from panyx0718/clean

hide temp tensor allocation
......@@ -377,6 +377,30 @@ class ExecutionContext {
return op_.Outputs(name);
}
template <typename T, typename DevContext>
Tensor AllocateTmpTensor(const framework::DDim& dim,
const DevContext& dev_ctx) const {
auto tmp_allocation_ptr = platform::DeviceTemporaryAllocator::Instance()
.Get<DevContext>(dev_ctx)
.Allocate(product(dim) * sizeof(T));
auto& deleter = tmp_allocation_ptr.get_deleter();
auto* allocation_ptr = tmp_allocation_ptr.release();
auto shared_allocation = std::shared_ptr<memory::allocation::Allocation>(
allocation_ptr, deleter);
PADDLE_ENFORCE(
dynamic_cast<platform::TemporaryAllocation*>(allocation_ptr) != nullptr,
"The AllocationPtr must be TemporaryAllocation.");
PADDLE_ENFORCE_EQ(allocation_ptr->size(),
framework::product(dim) * sizeof(T));
paddle::framework::Tensor temp_tensor(
framework::ToDataType(std::type_index(typeid(T))));
temp_tensor.Resize(dim);
temp_tensor.ResetHolder(std::move(shared_allocation));
return temp_tensor;
}
private:
const OperatorBase& op_;
const Scope& scope_;
......
......@@ -151,27 +151,5 @@ void TensorToVector(const Tensor& src, std::vector<T>* dst) {
memory::Copy(dst_place, dst_ptr, boost::get<platform::CPUPlace>(src.place()),
src_ptr, size);
}
template <typename T>
paddle::framework::Tensor GetTensor(
memory::allocation::AllocationPtr temp_allocation_ptr,
const framework::DDim& dim) {
auto& deleter = temp_allocation_ptr.get_deleter();
auto* allocation_ptr = temp_allocation_ptr.release();
auto shared_allocation =
std::shared_ptr<memory::allocation::Allocation>(allocation_ptr, deleter);
PADDLE_ENFORCE(
dynamic_cast<platform::TemporaryAllocation*>(allocation_ptr) != nullptr,
"The AllocationPtr must be TemporaryAllocation.");
PADDLE_ENFORCE_EQ(allocation_ptr->size(),
framework::product(dim) * sizeof(T));
paddle::framework::Tensor temp_tensor(
framework::ToDataType(std::type_index(typeid(T))));
temp_tensor.Resize(dim);
temp_tensor.ResetHolder(std::move(shared_allocation));
return temp_tensor;
}
} // namespace framework
} // namespace paddle
......@@ -18,7 +18,6 @@ limitations under the License. */
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/depthwise_conv.h"
#include "paddle/fluid/operators/math/im2col.h"
......@@ -158,10 +157,7 @@ class GemmConvKernel : public framework::OpKernel<T> {
// to call the matrix multiplication interface.
Tensor col_matrix;
if (is_expand) {
auto tmp_allocation_ptr =
platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
framework::product(col_shape) * sizeof(T));
col = framework::GetTensor<T>(std::move(tmp_allocation_ptr), col_shape);
col = context.AllocateTmpTensor<T, DeviceContext>(col_shape, dev_ctx);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
......@@ -293,10 +289,7 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
// to call the matrix multiplication interface.
Tensor col_matrix;
if (is_expand) {
auto tmp_allocation_ptr =
platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
framework::product(col_shape) * sizeof(T));
col = framework::GetTensor<T>(std::move(tmp_allocation_ptr), col_shape);
col = context.AllocateTmpTensor<T, DeviceContext>(col_shape, dev_ctx);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
......
......@@ -100,7 +100,7 @@ ENDIF()
nv_library(cuda_device_guard SRCS cuda_device_guard.cc DEPS gpu_info)
if(WITH_GPU)
nv_test(temporal_allocator_test SRCS temporary_allocator_test.cc DEPS temp_allocator tensor)
nv_test(temporal_allocator_test SRCS temporary_allocator_test.cc DEPS temp_allocator tensor operator)
else()
cc_test(temporal_allocator_test SRCS temporary_allocator_test.cc DEPS temp_allocator tensor)
cc_test(temporal_allocator_test SRCS temporary_allocator_test.cc DEPS temp_allocator tensor operator)
endif()
......@@ -14,12 +14,27 @@
#include "paddle/fluid/platform/temporary_allocator.h"
#include <gtest/gtest.h>
#include <string>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/tensor_util.h"
DECLARE_double(limit_of_temporary_allocation);
namespace paddle {
namespace platform {
class DummyOp : public framework::OperatorBase {
public:
DummyOp(const std::string& type, const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
protected:
void RunImpl(const framework::Scope& scope,
const platform::Place& place) const override {}
};
TEST(temporary_allocator, temporary_allocator) {
platform::CPUPlace cpu_place;
TemporaryAllocator alloc(cpu_place);
......@@ -68,96 +83,92 @@ TEST(temporary_allocator, add_callback) {
}
TEST(temporary_allocator, create_tensor_with_allocationptr) {
platform::CPUPlace cpu_place;
TemporaryAllocator cpu_alloc(cpu_place);
framework::VariableNameMap dummy_vars;
framework::AttributeMap dummy_attrs;
DummyOp op("dummy", dummy_vars, dummy_vars, dummy_attrs);
framework::Scope scope;
framework::VariableValueMap vars;
framework::RuntimeContext run_ctx(vars, vars);
size_t memory_size = 300;
{
size_t memory_size = 200;
auto allocation = cpu_alloc.Allocate(memory_size);
void* address = allocation->ptr();
platform::CPUPlace cpu_place;
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx =
static_cast<platform::CPUDeviceContext*>(pool.Get(cpu_place));
framework::ExecutionContext ctx(op, scope, *dev_ctx, run_ctx);
int numel = memory_size / sizeof(float);
framework::Tensor tensor = framework::GetTensor<float>(
std::move(allocation), framework::make_ddim({numel}));
PADDLE_ENFORCE_EQ(address, tensor.data<float>());
framework::Tensor tensor =
ctx.AllocateTmpTensor<float, platform::CPUDeviceContext>(
framework::make_ddim({numel}), *dev_ctx);
PADDLE_ENFORCE_EQ(tensor.numel(), numel);
}
#ifdef PADDLE_WITH_CUDA
platform::CUDAPlace gpu_place(0);
TemporaryAllocator gpu_alloc(gpu_place);
{
size_t memory_size = 300;
auto allocation = gpu_alloc.Allocate(memory_size);
void* address = allocation->ptr();
platform::CUDAPlace gpu_place(0);
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx =
static_cast<platform::CUDADeviceContext*>(pool.Get(gpu_place));
framework::ExecutionContext ctx(op, scope, *dev_ctx, run_ctx);
int numel = memory_size / sizeof(float);
framework::Tensor tensor = framework::GetTensor<float>(
std::move(allocation), framework::make_ddim({numel}));
PADDLE_ENFORCE_EQ(address, tensor.data<float>());
framework::Tensor tensor =
ctx.AllocateTmpTensor<float, platform::CUDADeviceContext>(
framework::make_ddim({numel}), *dev_ctx);
PADDLE_ENFORCE_EQ(tensor.numel(), numel);
}
// The allocation is not holded now, it should be placed to
// TemporaryAllocationQueue.
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
gpu_alloc.Release([]() {});
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
#endif
}
TEST(temporary_allocator, create_tensor_with_allocationptr2) {
platform::CPUPlace cpu_place;
TemporaryAllocator cpu_alloc(cpu_place);
framework::VariableNameMap dummy_vars;
framework::AttributeMap dummy_attrs;
DummyOp op("dummy", dummy_vars, dummy_vars, dummy_attrs);
framework::Scope scope;
framework::VariableValueMap vars;
framework::RuntimeContext run_ctx(vars, vars);
size_t memory_size = 400;
{
size_t memory_size = 400;
platform::CPUPlace cpu_place;
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx =
static_cast<platform::CPUDeviceContext*>(pool.Get(cpu_place));
framework::ExecutionContext ctx(op, scope, *dev_ctx, run_ctx);
int numel = memory_size / sizeof(float);
framework::Tensor out_side_tensor;
void* address;
{
auto allocation = cpu_alloc.Allocate(memory_size);
address = allocation->ptr();
framework::Tensor tensor = framework::GetTensor<float>(
std::move(allocation), framework::make_ddim({numel}));
PADDLE_ENFORCE_EQ(address, tensor.data<float>());
framework::Tensor tensor =
ctx.AllocateTmpTensor<float, platform::CPUDeviceContext>(
framework::make_ddim({numel}), *dev_ctx);
PADDLE_ENFORCE_EQ(tensor.numel(), numel);
out_side_tensor.ShareDataWith(tensor);
}
PADDLE_ENFORCE_EQ(address, out_side_tensor.data<float>());
PADDLE_ENFORCE_EQ(out_side_tensor.numel(), numel);
}
#ifdef PADDLE_WITH_CUDA
platform::CUDAPlace gpu_place(0);
TemporaryAllocator gpu_alloc(gpu_place);
{
void* address;
platform::CUDAPlace gpu_place(0);
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx =
static_cast<platform::CUDADeviceContext*>(pool.Get(gpu_place));
framework::ExecutionContext ctx(op, scope, *dev_ctx, run_ctx);
size_t memory_size = 500;
int numel = memory_size / sizeof(float);
framework::Tensor out_side_tensor;
{
auto allocation = gpu_alloc.Allocate(memory_size);
address = allocation->ptr();
framework::Tensor tensor = framework::GetTensor<float>(
std::move(allocation), framework::make_ddim({numel}));
PADDLE_ENFORCE_EQ(address, tensor.data<float>());
framework::Tensor tensor =
ctx.AllocateTmpTensor<float, platform::CUDADeviceContext>(
framework::make_ddim({numel}), *dev_ctx);
PADDLE_ENFORCE_EQ(tensor.numel(), numel);
out_side_tensor.ShareDataWith(tensor);
}
PADDLE_ENFORCE_EQ(address, out_side_tensor.data<float>());
PADDLE_ENFORCE_EQ(out_side_tensor.numel(), numel);
// The allocation is holded by out_side_tensor.
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
gpu_alloc.Release([]() {});
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
}
// The allocation is not holded now, it should be placed to
// TemporaryAllocationQueue.
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
gpu_alloc.Release([]() {});
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
#endif
}
......
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