ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 10:09:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Disable conv for Filter1x1Stride1Pad0 when K or C is even (#2186)

Browse Source 
[ROCm/composable_kernel commit: fa3c6811d8]
This commit is contained in:
Mateusz Ozga
2025-05-16 10:18:47 +02:00
committed by GitHub
parent 3ca1b9df66
commit ac7ecfe7aa
3 changed files with 11 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
include/ck/ck.hpp
View File

@@ -222,6 +222,9 @@
// TODO: separate index calculation into "compile-time", "global", "block", "wave", "thread"
#define CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE 0
// workaround: conv crash when K, C is even
#define CK_WORKAROUND_DISABLE_FILTER1x1STRIDE1PAD0_WHEN_K_C_IS_EVEN 1
// workaround: compiler crash when compiling recursive lambda
#define CK_WORKAROUND_SWDEV_275126 1

7
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle_v3.hpp
View File

@@ -1206,6 +1206,13 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffleV3
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
// workaround: disable when K, C is even
#if CK_WORKAROUND_DISABLE_FILTER1x1STRIDE1PAD0_WHEN_K_C_IS_EVEN
if(arg.Conv_C_ % 2 == 0 || arg.Conv_K_ % 2 == 0)
{
return false;
}
#endif
// check if it's 1x1, stride=1 pad = 0 conv
for(int i = 0; i < NDimSpatial; i++)
{

1
test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
View File

@@ -188,6 +188,7 @@ TYPED_TEST(TestGroupedConvndBwdWeight1d, Test1D)
TYPED_TEST(TestGroupedConvndBwdWeight2d, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back({2, 2, 64, 4, 4, {1, 1}, {7, 7}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 2, 64, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back({2, 2, 64, 3, 3, {1, 1}, {7, 7}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});