ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 17:00:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

use dedicated threadwise_copy for 1x1, perf at 80%

Browse Source
This commit is contained in:
Chao Liu
2019-04-06 18:40:54 -05:00
parent 5245a0162b
commit bd0098afb3
3 changed files with 64 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
driver/device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp
View File

@@ -192,7 +192,6 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t BlockSize = 256;
#elif 0
// 1x1, 14x14, Vega 20, disable lds_double_buffer, enable register double buffer
// 1x1, 14x14, Pascal, enable lds_double_buffer, disable register double buffer
constexpr index_t BPerBlock = 64;
constexpr index_t KPerBlock = 128;
@@ -245,6 +244,8 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t InBlockCopyDataPerRead = 4;
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 256;
#endif
@@ -295,7 +296,8 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
WeiBlockCopyThreadPerDim0,
WeiBlockCopyThreadPerDim1,
InBlockCopyDataPerRead,
WeiBlockCopyDataPerRead>{};
WeiBlockCopyDataPerRead,
OutThreadCopyDataPerWrite>{};
float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
dim3(GridSize),