ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Do not use warpSize as compile time constant as it is removed (#2320)

Browse Source 
* Do not use warpSize as compile time constant as it is removed

* Update tile_image_to_column_shape.hpp

update warpSize usage.

* clean-up all use of warpSize, make sure code builds

* fix

---------

Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: illsilin <Illia.Silin@amd.com>
Co-authored-by: Bartlomiej Kocot <barkocot@amd.com>
This commit is contained in:
Satyanvesh Dittakavi
2025-06-18 00:24:30 +05:30
committed by GitHub
parent 3af66e99ab
commit 4c57157d50
31 changed files with 213 additions and 206 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
include/ck/tensor_operation/gpu/block/blockwise_gemm_mx_pipeline_xdlops_base.hpp
View File

@@ -45,7 +45,7 @@ struct BlockwiseGemmXdlops_mx_pipeline_base
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
// Hardcode to 64, as HIP-provided "warpSize" would return 32 on RDNA GPUs.
// Hardcode to 64, as HIP-provided "WarpSize" would return 32 on RDNA GPUs.
static constexpr index_t WaveSize = 64;
static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);

2
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp
View File

@@ -40,7 +40,7 @@ struct BlockwiseGemmXdlops_pipeline_base
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
// Hardcode to 64, as HIP-provided "warpSize" would return 32 on RDNA GPUs.
// Hardcode to 64, as HIP-provided "WarpSize" would return 32 on RDNA GPUs.
static constexpr index_t WaveSize = 64;
static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);

4
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2.hpp
View File

@@ -141,7 +141,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
using Base::BMmaKStride;
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
(4 * WarpSize / BlockSize) >= 1 ? 4 * WarpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
@@ -631,7 +631,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
static constexpr index_t KRepeat = KPerThread / KPerInnerLoop;
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
(4 * WarpSize / BlockSize) >= 1 ? 4 * WarpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);

2
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp
View File

@@ -143,7 +143,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
using Base::BMmaKStride;
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
(4 * WarpSize / BlockSize) >= 1 ? 4 * WarpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);

4
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_b_scale.hpp
View File

@@ -141,7 +141,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_b_scale<BlockGemmPipelineScheduler::Intra
using Base::BMmaKStride;
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
(4 * WarpSize / BlockSize) >= 1 ? 4 * WarpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
@@ -632,7 +632,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_b_scale<BlockGemmPipelineScheduler::Inter
static constexpr index_t KRepeat = KPerThread / KPerInnerLoop;
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
(4 * WarpSize / BlockSize) >= 1 ? 4 * WarpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);