devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Enable efficient Tensor reduction for doubles

Browse Source
This commit is contained in:
Igor Babuschkin
2016-07-01 19:08:26 +01:00
parent d161b8f03a
commit eeb0d880ee
2 changed files with 64 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
unsupported/test/cxx11_tensor_reduction_cuda.cu
View File

@@ -16,7 +16,7 @@
#include <unsupported/Eigen/CXX11/Tensor>
template<int DataLayout>
template<typename Type, int DataLayout>
static void test_full_reductions() {
Eigen::CudaStreamDevice stream;
@@ -25,24 +25,24 @@ static void test_full_reductions() {
const int num_rows = internal::random<int>(1024, 5*1024);
const int num_cols = internal::random<int>(1024, 5*1024);
Tensor<float, 2, DataLayout> in(num_rows, num_cols);
Tensor<Type, 2, DataLayout> in(num_rows, num_cols);
in.setRandom();
Tensor<float, 0, DataLayout> full_redux;
Tensor<Type, 0, DataLayout> full_redux;
full_redux = in.sum();
std::size_t in_bytes = in.size() * sizeof(float);
std::size_t out_bytes = full_redux.size() * sizeof(float);
float* gpu_in_ptr = static_cast<float*>(gpu_device.allocate(in_bytes));
float* gpu_out_ptr = static_cast<float*>(gpu_device.allocate(out_bytes));
std::size_t in_bytes = in.size() * sizeof(Type);
std::size_t out_bytes = full_redux.size() * sizeof(Type);
Type* gpu_in_ptr = static_cast<Type*>(gpu_device.allocate(in_bytes));
Type* gpu_out_ptr = static_cast<Type*>(gpu_device.allocate(out_bytes));
gpu_device.memcpyHostToDevice(gpu_in_ptr, in.data(), in_bytes);
TensorMap<Tensor<float, 2, DataLayout> > in_gpu(gpu_in_ptr, num_rows, num_cols);
TensorMap<Tensor<float, 0, DataLayout> > out_gpu(gpu_out_ptr);
TensorMap<Tensor<Type, 2, DataLayout> > in_gpu(gpu_in_ptr, num_rows, num_cols);
TensorMap<Tensor<Type, 0, DataLayout> > out_gpu(gpu_out_ptr);
out_gpu.device(gpu_device) = in_gpu.sum();
Tensor<float, 0, DataLayout> full_redux_gpu;
Tensor<Type, 0, DataLayout> full_redux_gpu;
gpu_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_ptr, out_bytes);
gpu_device.synchronize();
@@ -54,6 +54,8 @@ static void test_full_reductions() {
}
void test_cxx11_tensor_reduction_cuda() {
CALL_SUBTEST_1(test_full_reductions<ColMajor>());
CALL_SUBTEST_2(test_full_reductions<RowMajor>());
CALL_SUBTEST_1((test_full_reductions<float, ColMajor>()));
CALL_SUBTEST_1((test_full_reductions<double, ColMajor>()));
CALL_SUBTEST_2((test_full_reductions<float, RowMajor>()));
CALL_SUBTEST_2((test_full_reductions<double, RowMajor>()));
}