Merge commit '4d629cd2b0bb0b4b210881be0db398bcd382f444' into develop

This commit is contained in:
assistant-librarian[bot]
2025-11-14 02:43:22 +00:00
parent a96aded2b1
commit 897c2bd422
7 changed files with 223 additions and 76 deletions

View File

@@ -662,17 +662,21 @@ struct FlatmmKernel
const auto scale_m_view = make_naive_tensor_view<address_space_enum::global>(
kargs.scale_m_ptr.ptr,
make_tuple(
kargs.M / ScaleGranularityM,
ScaleGranularityKA == 0 ? 1 : splitk_batch_offset.splitted_k / ScaleGranularityKA),
make_tuple(kargs.M / ScaleGranularityM,
ScaleGranularityKA == 0
? 1
: splitk_batch_offset.splitted_k /
(ScaleGranularityKA != 0 ? ScaleGranularityKA : 1)),
make_tuple(scale_stride_m, 0),
number < ScaleGranularityM == 1 ? FlatmmPipeline::GetVectorSizeA() : 1 > {},
number<1>{});
const auto scale_n_view = make_naive_tensor_view<address_space_enum::global>(
kargs.scale_n_ptr.ptr,
make_tuple(
ScaleGranularityKB == 0 ? 1 : (splitk_batch_offset.splitted_k / ScaleGranularityKB),
kargs.N / ScaleGranularityN),
make_tuple(ScaleGranularityKB == 0
? 1
: (splitk_batch_offset.splitted_k /
(ScaleGranularityKB != 0 ? ScaleGranularityKB : 1)),
kargs.N / ScaleGranularityN),
make_tuple(0, scale_stride_n),
number < ScaleGranularityN == 1 ? FlatmmPipeline::GetVectorSizeB() : 1 > {},
number<1>{});