ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-20 21:09:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix K padding calculation for grouped conv data (#876)

Browse Source 
* Fix K padding calculation for grouped conv data

* Restore previous padd for 1x1 specialization

[ROCm/composable_kernel commit: c981f6d033]
This commit is contained in:
Bartłomiej Kocot
2023-09-05 17:07:41 +02:00
committed by GitHub
parent 47958ebd07
commit fcafba0fd4
3 changed files with 21 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
View File

@@ -280,6 +280,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
BK1,
MPerBlock,
NPerBlock,
KPerBlock,
DoPadGemmM,
DoPadGemmN>{};

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
View File

@@ -268,6 +268,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
static_assert(KPerBlock % AK1Value == 0 && KPerBlock % BK1Value == 0,
"KPerBlock must be divisible by AK1Value and BK1Value!");
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto N = b_grid_desc_n_k.GetLength(I0);