ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 21:21:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fixed Gemmv2r3 kpad (#938)

Browse Source 
* added kpad support into v2r3

* add generic instances

* fixed comments

* fixed mnk padding

* Update device_batched_gemm_xdl.hpp

* fixed kpad

---------

Co-authored-by: Jing Zhang <jizha@amd.com>
This commit is contained in:
zjing14
2023-09-26 18:40:00 -05:00
committed by GitHub
parent 94bfa50256
commit 48ba6e8a69
2 changed files with 5 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
View File

@@ -9,13 +9,13 @@ namespace ck {
struct GridwiseGemmPipeline_v2
{
__host__ __device__ static constexpr bool IsSupported(index_t num_loop)
__host__ __device__ static constexpr bool IsSupported(const index_t num_loop)
{
// TODO: improve applicability
return num_loop % 2 == 0;
}
__host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
__host__ __device__ static constexpr bool CalculateHasMainLoop(const index_t num_loop)
{
return (num_loop / 2) > 1;
}

9
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View File

@@ -175,7 +175,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
return math::integer_divide_ceil(N, NPerBlock) * NPerBlock;
}
__host__ static auto CalculateK0(index_t K) { return math::integer_divide_floor(K, K1Value); }
__host__ static auto CalculateK0(index_t K) { return math::integer_divide_ceil(K, K1Value); }
// Argument
struct Problem
@@ -369,9 +369,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
"Invalid tuning param!");
// check gridwise gemm pipeline
const index_t K0 = problem.K / K1Value;
const auto num_k_loop = K0 / K0PerBlock;
const auto num_k_loop = math::integer_divide_ceil(problem.K0, K0PerBlock);
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
return false;
@@ -1026,8 +1024,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
}
// check gridwise gemm pipeline
const index_t K0 = problem.K / K1;
const auto num_k_loop = K0 / K0PerBlock;
const auto num_k_loop = math::integer_divide_ceil(problem.K0, K0PerBlock);
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{