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

[CK] Fix misc issues in CK examples (#2890)

Browse Source 
* [CK] Fix misc CK issues

* revert fp8 change, it causes CI fail.

* resubmit fp8 change
This commit is contained in:
linqunAMD
2025-09-25 02:28:20 +08:00
committed by GitHub
parent 8fe3838c65
commit f076f207ce
10 changed files with 74 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
View File

@@ -21,8 +21,7 @@ template <typename GridwiseGemm,
typename FloatC,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3,
typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
typename CGridDesc_M_N,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
@@ -32,17 +31,16 @@ __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_xdlops_skip_b_lds_v1(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
kernel_gemm_xdlops_skip_b_lds_v1(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
const CGridDesc_M_N c_grid_desc_m_n,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx11__) || \
defined(__gfx12__)
@@ -50,6 +48,12 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
{
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
auto b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3 =
GridwiseGemm::MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(b_grid_desc_k0_n_k1);
GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
p_b_grid,
p_c_grid,
@@ -67,8 +71,8 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
ignore = p_b_grid;
ignore = p_c_grid;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3;
ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
ignore = b_grid_desc_k0_n_k1;
ignore = c_grid_desc_m_n;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
@@ -375,20 +379,19 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_b_lds_v1
return cblockid_to_m0_n0_block_cluster_adaptor;
}
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
using DefaultBlock2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
using BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 =
decltype(MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(BGridDesc_K0_N_K1{}));
template <bool HasMainK0BlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
template <bool HasMainK0BlockLoop,
typename BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3,
typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
typename Block2CTileMap = DefaultBlock2CTileMap>
__device__ static void
Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
void* __restrict__ p_shared,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3& b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2& c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,