ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-11 01:27:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

add type-conversion to AccDataType and then to CDataType to exactly mimic GPU's behavior

Browse Source
This commit is contained in:
Kevin Abraham
2025-09-12 15:35:24 +00:00
committed by Kevin Abraham
parent 7ed033a1b2
commit 81953c81d6
1 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
View File

@@ -105,12 +105,15 @@ struct ReferenceBatchedGemm : public device::BaseOperator
{
for(int batchIdx = 0;batchIdx < arg.k_batch_;batchIdx++)
{
v_c = ck::type_convert<CDataType>(ck::type_convert<AccDataType>(v_c) + ck::type_convert<AccDataType>(partialSums[batchIdx]));
// mimic the way fp operations would be done on GPU for k-batching
v_c = ck::type_convert<AccDataType>(ck::type_convert<CDataType>(
ck::type_convert<AccDataType>(v_c) +
ck::type_convert<AccDataType>(partialSums[batchIdx])));
}
}
else
{
v_c = partialSums[0];
v_c = ck::type_convert<AccDataType>(partialSums[0]);
}
arg.c_g_m_n_(g, m, n) = ck::type_convert<CDataType>(v_c);