mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-27 08:25:46 +00:00
fixed vector load siz for fp4
This commit is contained in:
@@ -316,10 +316,10 @@ struct MXGemmPipelineAgBgCrCompAsync : public BaseMXGemmPipelineAgBgCrCompAsync<
|
||||
number<BsLayout::size()>{});
|
||||
|
||||
/// Check tile window traits for vector size
|
||||
using ATileDstr = remove_cvref_t<decltype(Policy::template MakeADramTileDistribution<Problem>())>;
|
||||
// using ATileDstr = remove_cvref_t<decltype(Policy::template MakeADramTileDistribution<Problem>())>;
|
||||
// static_assert(ATileDstr::LargestVec >= 16, "wrong! not implemented vector size");
|
||||
// static_assert(ATileDstr::X1 >= 16, "wrong! not implemented vector size");
|
||||
using BTileDstr = remove_cvref_t<decltype(Policy::template MakeBDramTileDistribution<Problem>())>;
|
||||
// using BTileDstr = remove_cvref_t<decltype(Policy::template MakeBDramTileDistribution<Problem>())>;
|
||||
// static_assert(BTileDstr::LargestVec >= 16, "wrong! not implemented vector size");
|
||||
// static_assert(BTileDstr::X1 >= 16, "wrong! not implemented vector size");
|
||||
using ATileType = remove_cvref_t<decltype(a_tile_windows[number<0>{}])>;
|
||||
|
||||
@@ -39,7 +39,15 @@ struct MXGemmPipelineAgBgCrCompAsyncDefaultPolicy
|
||||
// constexpr index_t vector_size_for_16_bytes = 16 / sizeof(ADataType);
|
||||
|
||||
// return vector_size_for_16_bytes;
|
||||
return 16;
|
||||
static_assert(std::is_same_v<ADataType, pk_fp4_t>, "ADataType must be pk_fp4_t or pk_fp4_raw_t");
|
||||
if constexpr(std::is_same_v<ADataType, pk_fp4_t> || std::is_same_v<ADataType, pk_fp4_raw_t>)
|
||||
{
|
||||
return 32;
|
||||
}
|
||||
else
|
||||
{
|
||||
return 16;
|
||||
}
|
||||
}
|
||||
|
||||
template <typename Problem, bool IsWave32Host = false>
|
||||
@@ -55,7 +63,15 @@ struct MXGemmPipelineAgBgCrCompAsyncDefaultPolicy
|
||||
// constexpr index_t vector_size_for_16_bytes = 16 / sizeof(BDataType);
|
||||
|
||||
// return vector_size_for_16_bytes;
|
||||
return 16;
|
||||
static_assert(std::is_same_v<BDataType, pk_fp4_t>, "BDataType must be pk_fp4_t or pk_fp4_raw_t");
|
||||
if constexpr(std::is_same_v<BDataType, pk_fp4_t> || std::is_same_v<BDataType, pk_fp4_raw_t>)
|
||||
{
|
||||
return 32;
|
||||
}
|
||||
else
|
||||
{
|
||||
return 16;
|
||||
}
|
||||
}
|
||||
|
||||
// Override DRAM tile distributions to use the constrained vector sizes
|
||||
|
||||
Reference in New Issue
Block a user