ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix for Add the API to load SGPR (#2913)

Browse Source 
* Revert "Revert "[CK-Tile] Add the API to load SGPR  (#2878)" (#2904)"

This reverts commit f161b5b738.

* Fix: sgpr minor issue

* cyclic dependency resolved

* clang formatted

* removing unused variable

* clang formatted

---------

Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
This commit is contained in:
Khushbu Agarwal
2025-09-25 10:32:42 -07:00
committed by GitHub
parent 64e61b8647
commit b56e5d1d79
41 changed files with 224 additions and 173 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
include/ck_tile/ops/gemm_quant/kernel/gemm_quant_kernel.hpp
View File

@@ -270,34 +270,34 @@ struct QuantGemmKernel
const std::size_t k_id = blockIdx.z)
{
constexpr auto K1 = TilePartitioner::BlockGemmShape::WarpTile::at(I2);
const index_t K_t = __builtin_amdgcn_readfirstlane(kargs.k_batch * K1);
const index_t KRead = __builtin_amdgcn_readfirstlane((kargs.K + K_t - 1) / K_t * K1);
const index_t K_t = amd_wave_read_first_lane(kargs.k_batch * K1);
const index_t KRead = amd_wave_read_first_lane((kargs.K + K_t - 1) / K_t * K1);
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
a_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead);
a_k_split_offset = amd_wave_read_first_lane(k_id * KRead);
}
else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
a_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead * kargs.stride_A);
a_k_split_offset = amd_wave_read_first_lane(k_id * KRead * kargs.stride_A);
}
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
{
b_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead * kargs.stride_B);
b_k_split_offset = amd_wave_read_first_lane(k_id * KRead * kargs.stride_B);
}
else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
{
b_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead);
b_k_split_offset = amd_wave_read_first_lane(k_id * KRead);
}
if(k_id < static_cast<uint32_t>(kargs.k_batch - 1))
{
splitted_k = __builtin_amdgcn_readfirstlane(KRead);
splitted_k = amd_wave_read_first_lane(KRead);
}
else
{
splitted_k = __builtin_amdgcn_readfirstlane(kargs.K - KRead * (kargs.k_batch - 1));
splitted_k = amd_wave_read_first_lane(kargs.K - KRead * (kargs.k_batch - 1));
}
}
@@ -918,8 +918,8 @@ struct QuantGemmKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop = __builtin_amdgcn_readfirstlane(
TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
const index_t num_loop =
amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
// Run GEMM cooperatively by whole workgroup.
const auto& a_block_window = gemm_tile_windows.at(I0);
@@ -981,10 +981,10 @@ struct QuantGemmKernel
CK_TILE_DEVICE void operator()(QuantGemmKernelArgs kargs) const
{
const auto blockId = __builtin_amdgcn_readfirstlane(blockIdx.x);
const auto blockId = amd_wave_read_first_lane(blockIdx.x);
const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(blockId);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const SplitKBatchOffset splitk_batch_offset(kargs);
// options

4
include/ck_tile/ops/gemm_quant/kernel/grouped_gemm_quant_kernel.hpp
View File

@@ -305,8 +305,8 @@ struct QuantGroupedGemmKernel
{
const auto [iM, iN] = block_idx_2d;
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const typename Base::SplitKBatchOffset splitk_batch_offset(kargs, block_idx_z);