mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
Wmma support for gemm_multiply_multiply_wp (#3278)
* Initial implementation with splitK support * Add gfx11 support * Fix compilation error * Add instances * Add irregular instances * Fix GetBuffer arguments * Minor changes * Address review comments * Fix compilation errors * Fix copyright header
This commit is contained in:
Enrico Degregori
@@ -31,6 +31,7 @@ using TGemmMulMulF8F8F16Instances =
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>;
|
||||
|
||||
#ifdef CK_USE_XDL
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p1(
|
||||
TGemmMulMulF8F8F16Instances& instances);
|
||||
|
||||
@@ -86,6 +87,21 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16
|
||||
TGemmMulMulF8F8F16Instances& instances);
|
||||
#endif
|
||||
|
||||
#ifdef CK_USE_WMMA
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p1(
|
||||
TGemmMulMulF8F8F16Instances& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p2(
|
||||
TGemmMulMulF8F8F16Instances& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p3(
|
||||
TGemmMulMulF8F8F16Instances& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p4(
|
||||
TGemmMulMulF8F8F16Instances& instances);
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
|
||||
using TGemmMulMulF8F8BF16Instances =
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
@@ -100,6 +116,7 @@ using TGemmMulMulF8F8BF16Instances =
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>;
|
||||
|
||||
#ifdef CK_USE_XDL
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances_p1(
|
||||
TGemmMulMulF8F8BF16Instances& instances);
|
||||
|
||||
@@ -153,7 +170,21 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma1
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p6(
|
||||
TGemmMulMulF8F8BF16Instances& instances);
|
||||
#endif
|
||||
|
||||
#ifdef CK_USE_WMMA
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p1(
|
||||
TGemmMulMulF8F8BF16Instances& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p2(
|
||||
TGemmMulMulF8F8BF16Instances& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p3(
|
||||
TGemmMulMulF8F8BF16Instances& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p4(
|
||||
TGemmMulMulF8F8BF16Instances& instances);
|
||||
#endif
|
||||
#endif
|
||||
|
||||
template <typename ADataType,
|
||||
@@ -200,6 +231,7 @@ 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_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p1(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p2(
|
||||
@@ -237,6 +269,17 @@ struct DeviceOperationInstanceFactory<
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instances_v2(
|
||||
op_ptrs);
|
||||
#endif
|
||||
#ifdef CK_USE_WMMA
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p1(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p2(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p3(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p4(
|
||||
op_ptrs);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
#endif
|
||||
@@ -248,6 +291,7 @@ 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_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p1(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p2(
|
||||
@@ -285,6 +329,17 @@ struct DeviceOperationInstanceFactory<
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instances_v2(
|
||||
op_ptrs);
|
||||
#endif
|
||||
#ifdef CK_USE_WMMA
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p1(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p2(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p3(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p4(
|
||||
op_ptrs);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
@@ -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_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES)
|
||||
|
||||
list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
|
||||
@@ -42,6 +42,15 @@ list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
|
||||
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p4.cpp
|
||||
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p5.cpp
|
||||
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p6.cpp
|
||||
|
||||
device_gemm_multiply_multiply_wp_wmma_f8_f8_bf16_mk_wmma_mn_default_instance_p1.cpp
|
||||
device_gemm_multiply_multiply_wp_wmma_f8_f8_bf16_mk_wmma_mn_default_instance_p2.cpp
|
||||
device_gemm_multiply_multiply_wp_wmma_f8_f8_bf16_mk_wmma_mn_default_instance_p3.cpp
|
||||
device_gemm_multiply_multiply_wp_wmma_f8_f8_bf16_mk_wmma_mn_default_instance_p4.cpp
|
||||
device_gemm_multiply_multiply_wp_wmma_f8_f8_f16_mk_wmma_mn_default_instance_p1.cpp
|
||||
device_gemm_multiply_multiply_wp_wmma_f8_f8_f16_mk_wmma_mn_default_instance_p2.cpp
|
||||
device_gemm_multiply_multiply_wp_wmma_f8_f8_f16_mk_wmma_mn_default_instance_p3.cpp
|
||||
device_gemm_multiply_multiply_wp_wmma_f8_f8_f16_mk_wmma_mn_default_instance_p4.cpp
|
||||
)
|
||||
|
||||
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
|
||||
|
||||
@@ -0,0 +1,105 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3_b_preshuffle.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
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;
|
||||
using MultiplyMultiply = element_wise::MultiplyMultiply;
|
||||
|
||||
static constexpr auto GemmDefault = GemmSpecialization::Default;
|
||||
|
||||
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
|
||||
|
||||
static constexpr auto v1 = BlockGemmPipelineVersion::v1;
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p1 =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//###########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| 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| CDEBlockTransferClusterLengths| CDEBlockTransfer| BlkGemmPipeSched| Block-wiseGemm|
|
||||
//###########################################| | | | | Type| 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| _MBlock_MPerBlock| ScalarPerVector| | Pipeline|
|
||||
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| | Verision|
|
||||
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p2 =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//###########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| 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| CDEBlockTransferClusterLengths| CDEBlockTransfer| BlkGemmPipeSched| Block-wiseGemm|
|
||||
//###########################################| | | | | Type| 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| _MBlock_MPerBlock| ScalarPerVector| | Pipeline|
|
||||
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| | Verision|
|
||||
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p3 =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//###########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| 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| CDEBlockTransferClusterLengths| CDEBlockTransfer| BlkGemmPipeSched| Block-wiseGemm|
|
||||
//###########################################| | | | | Type| 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| _MBlock_MPerBlock| ScalarPerVector| | Pipeline|
|
||||
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| | Verision|
|
||||
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p4 =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//###########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| 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| CDEBlockTransferClusterLengths| CDEBlockTransfer| BlkGemmPipeSched| Block-wiseGemm|
|
||||
//###########################################| | | | | Type| 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| _MBlock_MPerBlock| ScalarPerVector| | Pipeline|
|
||||
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| | Verision|
|
||||
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,33 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "device_gemm_multiply_multiply_wp_wmma_f8_f8_bf16_mk_wmma_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p1(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p1<
|
||||
GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,33 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "device_gemm_multiply_multiply_wp_wmma_f8_f8_bf16_mk_wmma_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p2(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p2<
|
||||
GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,33 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "device_gemm_multiply_multiply_wp_wmma_f8_f8_bf16_mk_wmma_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p3(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p3<
|
||||
GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,33 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "device_gemm_multiply_multiply_wp_wmma_f8_f8_bf16_mk_wmma_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_default_instances_p4(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_bf16_mk_wmma_mn_instances_p4<
|
||||
GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,105 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3_b_preshuffle.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
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;
|
||||
using MultiplyMultiply = element_wise::MultiplyMultiply;
|
||||
|
||||
static constexpr auto GemmDefault = GemmSpecialization::Default;
|
||||
|
||||
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
|
||||
|
||||
static constexpr auto v1 = BlockGemmPipelineVersion::v1;
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p1 =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//###########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| 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| CDEBlockTransferClusterLengths| CDEBlockTransfer| BlkGemmPipeSched| Block-wiseGemm|
|
||||
//###########################################| | | | | Type| 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| _MBlock_MPerBlock| ScalarPerVector| | Pipeline|
|
||||
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| | Verision|
|
||||
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p2 =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//###########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| 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| CDEBlockTransferClusterLengths| CDEBlockTransfer| BlkGemmPipeSched| Block-wiseGemm|
|
||||
//###########################################| | | | | Type| 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| _MBlock_MPerBlock| ScalarPerVector| | Pipeline|
|
||||
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| | Verision|
|
||||
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p3 =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//###########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| 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| CDEBlockTransferClusterLengths| CDEBlockTransfer| BlkGemmPipeSched| Block-wiseGemm|
|
||||
//###########################################| | | | | Type| 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| _MBlock_MPerBlock| ScalarPerVector| | Pipeline|
|
||||
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| | Verision|
|
||||
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p4 =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//###########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| 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| CDEBlockTransferClusterLengths| CDEBlockTransfer| BlkGemmPipeSched| Block-wiseGemm|
|
||||
//###########################################| | | | | Type| 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| _MBlock_MPerBlock| ScalarPerVector| | Pipeline|
|
||||
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| | Verision|
|
||||
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>,
|
||||
DeviceGemmMultiD_Wmma_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, v1, F8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,33 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "device_gemm_multiply_multiply_wp_wmma_f8_f8_f16_mk_wmma_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p1(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p1<
|
||||
GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,33 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "device_gemm_multiply_multiply_wp_wmma_f8_f8_f16_mk_wmma_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p2(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p2<
|
||||
GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,33 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "device_gemm_multiply_multiply_wp_wmma_f8_f8_f16_mk_wmma_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p3(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p3<
|
||||
GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,33 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include "device_gemm_multiply_multiply_wp_wmma_f8_f8_f16_mk_wmma_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_default_instances_p4(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_multiply_multiply_weight_preshuffle_wmma_f8_f8_f16_mk_wmma_mn_instances_p4<
|
||||
GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user