ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-18 20:09:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

experimenting

Browse Source 
[ROCm/composable_kernel commit: 766b0a9eaf]
This commit is contained in:
Chao Liu
2019-03-24 12:09:57 -05:00
parent 6f1651f8a7
commit cd883e7581
33 changed files with 1886 additions and 1822 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
driver/device_direct_convolution_1.hpp
View File

@@ -10,7 +10,7 @@ void device_direct_convolution_1(InDesc,
const Tensor<T>& wei,
OutDesc,
Tensor<T>& out,
unsigned nrepeat)
index_t nrepeat)
{
std::size_t data_sz = sizeof(T);
DeviceMem in_device_buf(data_sz * in.mDesc.GetElementSpace());
@@ -34,28 +34,28 @@ void device_direct_convolution_1(InDesc,
#if 1
// 3x3, 34x34
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 16;
constexpr unsigned CPerBlock = 2;
constexpr unsigned HoPerBlock = 4;
constexpr unsigned WoPerBlock = 32;
constexpr index_t NPerBlock = 2;
constexpr index_t KPerBlock = 16;
constexpr index_t CPerBlock = 2;
constexpr index_t HoPerBlock = 4;
constexpr index_t WoPerBlock = 32;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2;
constexpr index_t NPerThread = 2;
constexpr index_t KPerThread = 4;
constexpr index_t CPerThread = 2;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr unsigned BlockSize = 128;
constexpr index_t BlockSize = 128;
#endif
constexpr unsigned GridSize =
constexpr index_t GridSize =
(out_desc.GetLength(I0) / NPerBlock) * (out_desc.GetLength(I1) / KPerBlock) *
(out_desc.GetLength(I2) / HoPerBlock) * (out_desc.GetLength(I3) / WoPerBlock);
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
for(unsigned i = 0; i < nrepeat; ++i)
for(index_t i = 0; i < nrepeat; ++i)
{
float time = launch_kernel(gridwise_direct_convolution_1<T,
InDesc,