ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 05:01:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Remove "basic" and universal GEMM tests, and incorporate their test cases into the GEMM pipeline tests (#3094)

Browse Source 
* Add missing copyright statements

* Use ck_tile::host_tensor_descriptor instead of a custom lambda

* Refactor use of check_data_type in test classes

* Use TEST_SUITE_NAME with TYPED_TEST_SUITE

* Remove an unused namespace

* Make dim3 const

* Add BF8 x BF8 tests for CompV3 in test_gemm_pipeline_kernel_types.hpp

* Add F8 x BF8 tests for CompV3 in test_gemm_pipeline_kernel_types.hpp

* Add BF16 x I4 tests for CompV3 in test_gemm_pipeline_kernel_types.hpp

* Add BF16 x BF16 tests for CompV3 in test_gemm_pipeline_kernel_types.hpp

* Add BF8 x I4 tests for CompV3 in test_gemm_pipeline_kernel_types.hpp

* Add F8 x I4 tests for CompV3 in test_gemm_pipeline_kernel_types.hpp

* Add F16 x I4 tests for CompV3 in test_gemm_pipeline_kernel_types.hpp

* Skip failing tests of F16 x I4 for CompV3 with K == 2 * K_Tile

* Add missing precision type combinations to CompV4 from CompV3

* Move the INT8 tests around for consistency with KernelTypesCompV3Wmma

* Add missing precision type combinations to CompV3Wmma from CompV3

* Remove the basic and universal tests and their dependencies

* On __gfx950__, avoid using transposed loading of A with datatype pk_int4_t of B

* Use ADataType and BDataType instead of ComputeDataType for WarpGemm

* Explicitly set some return types to void

* Use more general typenames in InterleavedPKTypeLoader

* Add load_interleaved_pk_type.hpp to common.hpp

* Use std::is_same_v in load_int4_tile

* Add handling of LoadTranspose to load_int4_tile

* Factor out common code in several places using load_int4_tile

* Add support for pk_int4_t using load_int4_tile

* Fix formatting
This commit is contained in:
SamiAario-AMD
2025-11-13 21:01:27 +02:00
committed by GitHub
parent 7d57bc169f
commit f2cfc6b94e
38 changed files with 352 additions and 1888 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp
View File

@@ -26,8 +26,21 @@ struct GemmPipelineAgBgCrImplBase
static constexpr index_t NPerBlock = BlockGemmShape::kN;
static constexpr index_t KPerBlock = BlockGemmShape::kK;
#if defined(__gfx950__)
static constexpr bool is_a_load_tr = std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>;
static constexpr bool is_b_load_tr = std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>;
// The combination of pk_int4_t and transposed loading causes numerical errors.
// Therefore do not use transposed loading in this case.
static constexpr bool is_a_load_tr = []() {
if constexpr(std::is_same_v<BDataType, pk_int4_t>)
return false;
else
return std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>;
}();
static constexpr bool is_b_load_tr = []() {
if constexpr(std::is_same_v<BDataType, pk_int4_t>)
return false;
else
return std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>;
}();
#else
static constexpr bool is_a_load_tr = false;
static constexpr bool is_b_load_tr = false;

38
include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp
View File

@@ -33,12 +33,27 @@ template <typename Derived>
struct UniversalGemmBasePolicy
{
#if defined(__gfx950__)
// The combination of pk_int4_t and transposed loading causes numerical errors.
// Therefore do not use transposed loading in this case.
template <typename Problem>
static constexpr bool is_a_load_tr =
std::is_same_v<remove_cvref_t<typename Problem::ALayout>, tensor_layout::gemm::ColumnMajor>;
static constexpr bool is_a_load_tr = []() {
using BDataType = remove_cvref_t<typename Problem::BDataType>;
if constexpr(std::is_same_v<BDataType, pk_int4_t>)
return false;
else
return std::is_same_v<remove_cvref_t<typename Problem::ALayout>,
tensor_layout::gemm::ColumnMajor>;
}();
template <typename Problem>
static constexpr bool is_b_load_tr =
std::is_same_v<remove_cvref_t<typename Problem::BLayout>, tensor_layout::gemm::RowMajor>;
static constexpr bool is_b_load_tr = []() {
using BDataType = remove_cvref_t<typename Problem::BDataType>;
if constexpr(std::is_same_v<BDataType, pk_int4_t>)
return false;
else
return std::is_same_v<remove_cvref_t<typename Problem::BLayout>,
tensor_layout::gemm::RowMajor>;
}();
#else
template <typename Problem>
static constexpr bool is_a_load_tr = false;
@@ -707,8 +722,15 @@ struct UniversalGemmPipelineAgBgCrPolicy
: vector_size * 4 == thread_elements ? WGAttrNumAccessEnum::Quad
: WGAttrNumAccessEnum::Invalid;
using WarpGemm = WarpGemmDispatcher<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
using ADataType = remove_cvref_t<typename Problem::ADataType>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
using ATypeToUse =
std::conditional_t<std::is_same_v<ADataType, pk_int4_t>, BDataType, ADataType>;
using BTypeToUse =
std::conditional_t<std::is_same_v<BDataType, pk_int4_t>, ADataType, BDataType>;
using WarpGemm = WarpGemmDispatcher<ATypeToUse,
BTypeToUse,
typename Problem::CDataType,
WarpTile::at(I0),
WarpTile::at(I1),
@@ -718,8 +740,8 @@ struct UniversalGemmPipelineAgBgCrPolicy
Problem::UseStructuredSparsity,
wg_attr_num_access>;
using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1CustomPolicy<typename Problem::ADataType,
typename Problem::BDataType,
using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1CustomPolicy<ATypeToUse,
BTypeToUse,
typename Problem::CDataType,
BlockWarps,
WarpGemm>;