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

scatter optimizers

Browse Source
del_some_in_makelist
typhoonzero 7 years ago
parent 641b4c0fe6
commit 1039c1e3b7
4 changed files with 32 additions and 30 deletions
Show Stats Download Patch File Download Diff File

10
paddle/operators/adagrad_op.cu
Unescape View File

@ -79,12 +79,12 @@ struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
framework::Tensor* moment, framework::Tensor* param) {
// 1. g_m.rows = set(g.rows)
auto grad_width = grad.value().dims()[1];
math::scatter::MergeAdd<platform::CPUDeviceContext, T> merge_func;
math::scatter::MergeAdd<platform::CUDADeviceContext, T> merge_func;
auto grad_merge = merge_func(context, grad);
auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
auto& merge_rows = grad_merge.rows;
auto& merge_rows = grad_merge.rows();
// 2. m += g_m * g_m
math::scatter::Mul<platform::CPUDeviceContext, T> sqare_func;
math::scatter::Mul<platform::CUDADeviceContext, T> sqare_func;
auto grad_square = sqare_func(context, grad_merge, grad_merge);
math::SelectedRowsAddToTensor<platform::CUDADeviceContext, T> functor;
@ -95,11 +95,13 @@ struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
auto* param_data = param->data<T>();
auto* moment_data = moment->data<T>();
const int block_size = 256;
dim3 threads(block_size, 1);
dim3 grid2(1, merge_rows.size());
SparseAdagradFunctorKernel<
T, 256><<<grid2, threads, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(grad_merge_data, grad_merge->rows().data(),
.stream()>>>(grad_merge_data, grad_merge.rows().data(),
lr, param_data, moment_data, grad_width,
epsilon);
}

7
paddle/operators/math/selected_rows_functor.cc
Unescape View File

@ -233,10 +233,9 @@ template struct MergeAdd<platform::CPUDeviceContext, int64_t>;
template <typename T>
struct UpdateToTensor<platform::CPUDeviceContext, T> {
framework::Tensor operator()(const platform::CPUDeviceContext& context,
const ScatterOps& op,
const framework::SelectedRows& input1,
framework::Tensor* input2) {
void operator()(const platform::CPUDeviceContext& context,
const ScatterOps& op, const framework::SelectedRows& input1,
framework::Tensor* input2) {
auto in1_height = input1.height();
auto in2_dims = input2->dims();
PADDLE_ENFORCE_EQ(in1_height, in2_dims[0]);

38
paddle/operators/math/selected_rows_functor.cu
Unescape View File

@ -12,6 +12,8 @@ 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 <set>
#include "paddle/operators/math/math_function.h"
#include "paddle/operators/math/selected_rows_functor.h"
#include "paddle/platform/cuda_helper.h"
@ -251,8 +253,8 @@ __global__ void MergeAddKernel(const T* input, const int64_t* input_rows,
}
template <typename T>
struct MergeAdd<platform::GPUDeviceContext, T> {
framework::SelectedRows operator()(const platform::GPUDeviceContext& context,
struct MergeAdd<platform::CUDADeviceContext, T> {
framework::SelectedRows operator()(const platform::CUDADeviceContext& context,
const framework::SelectedRows& input) {
framework::SelectedRows out;
auto input_rows = input.rows();
@ -288,10 +290,10 @@ struct MergeAdd<platform::GPUDeviceContext, T> {
}
};
template struct MergeAdd<platform::GPUDeviceContext, float>;
template struct MergeAdd<platform::GPUDeviceContext, double>;
template struct MergeAdd<platform::GPUDeviceContext, int>;
template struct MergeAdd<platform::GPUDeviceContext, int64_t>;
template struct MergeAdd<platform::CUDADeviceContext, float>;
template struct MergeAdd<platform::CUDADeviceContext, double>;
template struct MergeAdd<platform::CUDADeviceContext, int>;
template struct MergeAdd<platform::CUDADeviceContext, int64_t>;
template <typename T, int block_size>
__global__ void UpdateToTensorKernel(const T* selected_rows,
@ -343,14 +345,14 @@ __global__ void UpdateToTensorKernel(const T* selected_rows,
}
template <typename T>
struct UpdateToTensor<platform::GPUDeviceContext, T> {
framework::Tensor operator()(const platform::GPUDeviceContext& context,
const ScatterOps& op,
const framework::SelectedRows& input1,
framework::Tensor* input2) {
struct UpdateToTensor<platform::CUDADeviceContext, T> {
void operator()(const platform::CUDADeviceContext& context,
const ScatterOps& op, const framework::SelectedRows& input1,
framework::Tensor* input2) {
// NOTE: Use SelectedRowsAddToTensor for better performance
// no additional MergeAdd called.
auto merged_in1 = MergeAdd()(context, input1);
MergeAdd<platform::CUDADeviceContext, T> merge_func;
auto merged_in1 = merge_func(context, input1);
auto in1_height = merged_in1.height();
auto in2_dims = input2->dims();
@ -362,14 +364,14 @@ struct UpdateToTensor<platform::GPUDeviceContext, T> {
int64_t in1_row_numel = in1_value.numel() / in1_rows.size();
PADDLE_ENFORCE_EQ(in1_row_numel, input2->numel() / in1_height);
auto* in1_data = in1_value.data<T>();
auto* input2_data = input2->data<T>();
auto* in1_data = in1_value.template data<T>();
auto* in2_data = input2->data<T>();
dim3 threads(PADDLE_CUDA_NUM_THREADS, 1);
dim3 threads(platform::PADDLE_CUDA_NUM_THREADS, 1);
dim3 grid(1, in1_rows.size());
UpdateToTensorKernel<
T, PADDLE_CUDA_NUM_THREADS><<<grid, threads, 0, context.stream()>>>(
in1_data, in1_rows.data(), op, in2_data, in1_row_numel);
UpdateToTensorKernel<T, platform::PADDLE_CUDA_NUM_THREADS><<<
grid, threads, 0, context.stream()>>>(in1_data, in1_rows.data(), op,
in2_data, in1_row_numel);
}
};
} // namespace scatter

7
paddle/operators/math/selected_rows_functor.h
Unescape View File

@ -123,10 +123,9 @@ enum class ScatterOps { ASSIGN, ADD, SUB, SUBBY, MUL, DIV, DIVBY };
// out = seleted_rows_in / tensor
template <typename DeviceContext, typename T>
struct UpdateToTensor {
framework::Tensor operator()(const DeviceContext& context,
const ScatterOps& op,
const framework::SelectedRows& input1,
framework::Tensor* input2);
void operator()(const DeviceContext& context, const ScatterOps& op,
const framework::SelectedRows& input1,
framework::Tensor* input2);
};
} // namespace scatter

Write Preview
Loading…
Cancel
Save