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

Implement device_gemm_universal_preshuffle_instance for RDNA4 (#3429)

Browse Source 
* add device_gemm_wmma_cshuffle_v3_b_preshuffle.hpp

* add examples

* add instances to test

* remove duplicate code between examples
This commit is contained in:
Yung-sheng Tu
2026-01-15 16:19:31 +01:00
committed by GitHub
parent e30207985a
commit 6df2d70143
20 changed files with 1229 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_preshuffle.hpp
View File

@@ -3,18 +3,19 @@
#pragma once
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_v2.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef CK_USE_XDL
#include "ck/tensor_operation/gpu/device/device_gemm_v2.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/utility/amd_ck_fp8.hpp"
#include "ck/utility/data_type.hpp"
#if defined(CK_USE_XDL) || defined(CK_USE_WMMA)
#include "gemm_universal_preshuffle.inc"
#endif
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
@@ -51,7 +52,7 @@ struct DeviceOperationInstanceFactory<
static auto GetInstances()
{
#ifdef CK_USE_XDL
#if defined(CK_USE_XDL) || defined(CK_USE_WMMA)
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if(defined(CK_ENABLE_BF16) && defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
@@ -60,6 +61,7 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
#ifdef CK_USE_XDL
add_device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma32x32_mn_instances(
op_ptrs);
add_device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma32x32_mn_compute_instances(
@@ -90,6 +92,17 @@ struct DeviceOperationInstanceFactory<
op_ptrs);
add_device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma16x16_nk_mn_comp_default_instances_part1(
op_ptrs);
#endif
#ifdef CK_USE_WMMA
add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p1(
op_ptrs);
add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p2(
op_ptrs);
add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p3(
op_ptrs);
add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p4(
op_ptrs);
#endif
}
}
#endif
@@ -100,6 +113,7 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
#ifdef CK_USE_XDL
add_device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p1(
op_ptrs);
add_device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p2(
@@ -136,10 +150,21 @@ struct DeviceOperationInstanceFactory<
op_ptrs);
add_device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p6(
op_ptrs);
#endif
#ifdef CK_USE_WMMA
add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p1(
op_ptrs);
add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p2(
op_ptrs);
add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p3(
op_ptrs);
add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p4(
op_ptrs);
#endif
}
}
#endif
#endif // CK_USE_XDL
#endif // CK_USE_XDL || CK_USE_WMMA
return op_ptrs;
}

47
library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_preshuffle.inc
View File

@@ -13,8 +13,7 @@ namespace instance {
using GemmF8F8BF16InstanceVector = std::vector<std::unique_ptr<
DeviceGemmV2BPreshuffle<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&;
using GemmF8F8F16InstanceVector = std::vector<std::unique_ptr<
DeviceGemmV2BPreshuffle<Row, Col, Row, F8, F8, F16, PassThrough, PassThrough, PassThrough>>>&;
#ifdef CK_USE_XDL
void add_device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma32x32_mn_instances(
GemmF8F8BF16InstanceVector& instances);
@@ -61,7 +60,32 @@ void add_device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma16x16_nk_mn_comp
GemmF8F8BF16InstanceVector& instances);
#endif
#ifdef CK_USE_WMMA
void add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p1(
GemmF8F8BF16InstanceVector& instances);
void add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p2(
GemmF8F8BF16InstanceVector& instances);
void add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p3(
GemmF8F8BF16InstanceVector& instances);
void add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p4(
GemmF8F8BF16InstanceVector& instances);
#endif
#endif
#if(defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8))
using GemmF8F8F16InstanceVector = std::vector<std::unique_ptr<
DeviceGemmV2BPreshuffle<Row, Col, Row, F8, F8, F16, PassThrough, PassThrough, PassThrough>>>&;
#ifdef CK_USE_XDL
void add_device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p1(
GemmF8F8F16InstanceVector& instances);
void add_device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p2(
@@ -99,6 +123,25 @@ void add_device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_
GemmF8F8F16InstanceVector& instances);
void add_device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p6(
GemmF8F8F16InstanceVector& instances);
#endif
#ifdef CK_USE_WMMA
void add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p1(
GemmF8F8F16InstanceVector& instances);
void add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p2(
GemmF8F8F16InstanceVector& instances);
void add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p3(
GemmF8F8F16InstanceVector& instances);
void add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p4(
GemmF8F8F16InstanceVector& instances);
#endif
#endif
} // namespace instance
} // namespace device

18
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/CMakeLists.txt
View File

@@ -1,7 +1,7 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
# ONLY XDL_KERNELS
# ONLY XDL_AND_WMMA_KERNELS
set(GEMM_UNIVERSAL_INSTANCES)
# F8_F8_BF16
@@ -21,6 +21,10 @@ device_gemm_xdl_universal_preshuffle_f8_f8_bf16/device_gemm_xdl_universal_preshu
device_gemm_xdl_universal_preshuffle_f8_f8_bf16/device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma_mn_p5_default_instance.cpp
device_gemm_xdl_universal_preshuffle_f8_f8_bf16/device_gemm_xdl_universal_preshuffle_f8_f8_f8_bf16_mk_mfma32x32_mn_default_instance.cpp
device_gemm_xdl_universal_preshuffle_f8_f8_bf16/device_gemm_xdl_universal_preshuffle_f8_f8_f8_bf16_mk_mfma32x32_mn_comp_instance.cpp
device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p1.cpp
device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p2.cpp
device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p3.cpp
device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p4.cpp
)
# F8_F8_F16
@@ -43,6 +47,10 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES
device_gemm_xdl_universal_preshuffle_f8_f8_f16/device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p4.cpp
device_gemm_xdl_universal_preshuffle_f8_f8_f16/device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p5.cpp
device_gemm_xdl_universal_preshuffle_f8_f8_f16/device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p6.cpp
device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p1.cpp
device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p2.cpp
device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p3.cpp
device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p4.cpp
)
# F8_F8_F16
@@ -64,6 +72,10 @@ set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_f16/devic
set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_f16/device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p4.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_f16/device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p5.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_f16/device_gemm_universal_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p6.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p4.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
# F8_F8_BF16
set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_bf16/device_gemm_xdl_universal_preshuffle_f8_f8_f8_bf16_mk_mfma32x32_mn_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
@@ -81,5 +93,9 @@ set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_bf16/devi
set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_bf16/device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma_mn_p5_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_bf16/device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma_nk_mn_comp_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_xdl_universal_preshuffle_f8_f8_bf16/device_gemm_xdl_universal_preshuffle_f8_f8_bf16_mk_mfma_nk_mn_comp_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p4.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
add_instance_library(device_gemm_universal_preshuffle_instance ${GEMM_UNIVERSAL_INSTANCES})

106
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn.hpp Normal file
View File

@@ -0,0 +1,106 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3_b_preshuffle.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/utility/amd_ck_fp8.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/utility/scheduler_enum.hpp"
#include "ck/utility/sequence.hpp"
#include <tuple>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F8 = f8_t;
using BF16 = bhalf_t;
using F32 = float;
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
template <index_t... Is>
using S = Sequence<Is...>;
using PassThrough = element_wise::PassThrough;
static constexpr auto GemmDefault = GemmSpecialization::Default;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto v1 = BlockGemmPipelineVersion::v1;
template <GemmSpecialization GemmSpec>
using device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p1 = std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemmPipeSched| BlkGemmPipelineVer| ComputeTypeA|
//#####################################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| | | |
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | |
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 128, 128, 16, 16, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 8, 1, 32>, S<4, 4, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 128, 128, 16, 16, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 256, 128, 16, 16, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 256, 128, 16, 16, 16, 16, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 16, 16, 16, 16, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p2 = std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemmPipeSched| BlkGemmPipelineVer| ComputeTypeA|
//#####################################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| | | |
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | |
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 16, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 128, 16, 16, 16, 16, 16, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 128, 16, 16, 16, 16, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p3 = std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemmPipeSched| BlkGemmPipelineVer| ComputeTypeA|
//#####################################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| | | |
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | |
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 128, 256, 16, 16, 16, 16, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 128, 256, 16, 16, 16, 16, 4, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 128, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 512, 256, 16, 16, 16, 16, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 8, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 256, 16, 16, 16, 16, 8, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p4 = std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemmPipeSched| BlkGemmPipelineVer| ComputeTypeA|
//#####################################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| | | |
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | |
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 4, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 2, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 224, 128, 16, 16, 16, 16, 1, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, Intrawave, v1, F8 >
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p1.cpp Normal file
View File

@@ -0,0 +1,33 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn.hpp"
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmV2BPreshuffle<Row,
Col,
Row,
F8,
F8,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p1<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p2.cpp Normal file
View File

@@ -0,0 +1,33 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn.hpp"
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmV2BPreshuffle<Row,
Col,
Row,
F8,
F8,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p2<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p3.cpp Normal file
View File

@@ -0,0 +1,33 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn.hpp"
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p3(
std::vector<std::unique_ptr<DeviceGemmV2BPreshuffle<Row,
Col,
Row,
F8,
F8,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p3<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_bf16/device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn_default_instance_p4.cpp Normal file
View File

@@ -0,0 +1,33 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "device_gemm_wmma_universal_preshuffle_f8_f8_bf16_mk_wmma_mn.hpp"
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p4(
std::vector<std::unique_ptr<DeviceGemmV2BPreshuffle<Row,
Col,
Row,
F8,
F8,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_universal_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p4<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

106
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn.hpp Normal file
View File

@@ -0,0 +1,106 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3_b_preshuffle.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/utility/amd_ck_fp8.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/utility/scheduler_enum.hpp"
#include "ck/utility/sequence.hpp"
#include <tuple>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F8 = f8_t;
using F16 = half_t;
using F32 = float;
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
template <index_t... Is>
using S = Sequence<Is...>;
using PassThrough = element_wise::PassThrough;
static constexpr auto GemmDefault = GemmSpecialization::Default;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto v1 = BlockGemmPipelineVersion::v1;
template <GemmSpecialization GemmSpec>
using device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p1 = std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemmPipeSched| BlkGemmPipelineVer| ComputeTypeA|
//#####################################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| | | |
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | |
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 128, 128, 16, 16, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 8, 1, 32>, S<4, 4, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 128, 128, 16, 16, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 256, 128, 16, 16, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 256, 128, 16, 16, 16, 16, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 16, 16, 16, 16, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p2 = std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemmPipeSched| BlkGemmPipelineVer| ComputeTypeA|
//#####################################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| | | |
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | |
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 16, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 128, 16, 16, 16, 16, 16, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 128, 16, 16, 16, 16, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p3 = std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemmPipeSched| BlkGemmPipelineVer| ComputeTypeA|
//#####################################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| | | |
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | |
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 128, 256, 16, 16, 16, 16, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 128, 256, 16, 16, 16, 16, 4, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 128, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 512, 256, 16, 16, 16, 16, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 8, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 256, 16, 16, 16, 16, 8, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p4 = std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemmPipeSched| BlkGemmPipelineVer| ComputeTypeA|
//#####################################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| | | |
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | |
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 4, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 2, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, Intrawave, v1, F8 >,
DeviceGemm_Wmma_CShuffleV3_BPreshuffle< Row, Col, Row, F8, F8, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 224, 128, 16, 16, 16, 16, 1, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, Intrawave, v1, F8 >
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p1.cpp Normal file
View File

@@ -0,0 +1,33 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn.hpp"
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmV2BPreshuffle<Row,
Col,
Row,
F8,
F8,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p1<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p2.cpp Normal file
View File

@@ -0,0 +1,33 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn.hpp"
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmV2BPreshuffle<Row,
Col,
Row,
F8,
F8,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p2<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p3.cpp Normal file
View File

@@ -0,0 +1,33 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn.hpp"
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p3(
std::vector<std::unique_ptr<DeviceGemmV2BPreshuffle<Row,
Col,
Row,
F8,
F8,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p3<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_universal_preshuffle/device_gemm_wmma_universal_preshuffle_f8_f8_f16/device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn_default_instance_p4.cpp Normal file
View File

@@ -0,0 +1,33 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "device_gemm_wmma_universal_preshuffle_f8_f8_f16_mk_wmma_mn.hpp"
#include <memory>
#include <vector>
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p4(
std::vector<std::unique_ptr<DeviceGemmV2BPreshuffle<Row,
Col,
Row,
F8,
F8,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_universal_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p4<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck