mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
Add grouped gemm instances for RDNA4 (#3237)
* wip: grouped_gemm implementation based on wmma kernel + example for fp16 * chore: clean up grouped_gem_wmma_splitk_fp16 example * chore: add cmake options to fully disable XDL or WMMA kernels * feat: add tests for grouped gemma wmma instances for f16 and bf16 (all layouts) * chore: add grouped gemm wmma bf16 example * refactor: reuse more code between instance factory functions * chore: turn test failure if not all batch sizes are supported into a warning * chore: made failing of test on unsupported instances conditional to not break old tests * chore: add log message to failure case where AK1/BK1/KBatch is too high for K value * fix: issue with new overloads of GridwiseGemm_wmma_cshuffle_v3::Run() * fix: stray comma after parameter list * fix: compilation issues on RDNA3 and tests failing due to unsupported problems still being ran * chore: update copyright in header comments * nit: minor feebdack * refactor: unified XDL / wma tests * fix: properly disable FP8 instances when ONLY targeting gfx11 * refactor: add v3 suffix to grouped_gemm device struct name * fix: small typos in example code * fix: fully exclude xdl/wmma instances when using the corresponding cmake flags * chore: remove unused destructor and added pipeline support checks to remove unnecessary paths * fix: make sure to not add instance library to group if library was skipped * fix: make sure xdl grouped gemm doesnt fail the new test * fix: explicitly exclude test if no xdl/wmma support, as pattern matching fails in this case * fix: examples not working since dependent types and functions were moved to ck namespace in develop * fix: tests failing when compiling for just gfx11 due to trying to run unsupported instances * chore: replace/add copyright headers with new format
This commit is contained in:
Erwin Terpstra
@@ -15,6 +15,142 @@ namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
#if defined(CK_USE_WMMA)
|
||||
#if defined(CK_ENABLE_FP16)
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f16_f16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f16_f16_km_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f16_f16_km_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif // CK_ENABLE_FP16
|
||||
#if defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8) && defined(__gfx12__)
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F8,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F8,
|
||||
F16,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
#if defined(CK_ENABLE_BF16)
|
||||
void add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif // CK_ENABLE_BF16
|
||||
#endif // CK_USE_WMMA
|
||||
|
||||
#if defined(CK_USE_XDL)
|
||||
#if defined(CK_ENABLE_FP16)
|
||||
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
|
||||
@@ -409,6 +545,81 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
#if defined(CK_USE_WMMA)
|
||||
#if defined(CK_ENABLE_FP16)
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
|
||||
is_same_v<EDataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_f16_f16_f16_km_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_f16_f16_f16_km_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif // CK_ENABLE_FP16
|
||||
#if defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8) && defined(__gfx12__)
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
|
||||
is_same_v<EDataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
|
||||
is_same_v<EDataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
#if defined(CK_ENABLE_BF16)
|
||||
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
|
||||
is_same_v<EDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif // CK_ENABLE_BF16
|
||||
#endif // CK_USE_WMMA
|
||||
|
||||
#if defined(CK_USE_XDL)
|
||||
#if defined(CK_ENABLE_FP16)
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
|
||||
|
||||
@@ -0,0 +1,205 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#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_grouped_gemm_wmma_splitk_cshuffle_v3.hpp"
|
||||
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/utility/loop_scheduler.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F8 = ck::f8_t;
|
||||
using F16 = ck::half_t;
|
||||
using BF16 = ck::bhalf_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using AccDataType = F32;
|
||||
using DsDataType = Empty_Tuple;
|
||||
|
||||
using DsLayout = Empty_Tuple;
|
||||
using ELayout = Row;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = PassThrough;
|
||||
|
||||
static constexpr auto PipelineV1 = BlockGemmPipelineVersion::v1;
|
||||
static constexpr auto PipelineV3 = BlockGemmPipelineVersion::v3;
|
||||
static constexpr auto IntrawaveScheduler = BlockGemmPipelineScheduler::Intrawave;
|
||||
static constexpr auto InterwaveScheduler = BlockGemmPipelineScheduler::Interwave;
|
||||
static constexpr auto GemmMNKPadding = device::GemmSpecialization::MNKPadding;
|
||||
static constexpr auto GemmDefault = device::GemmSpecialization::Default;
|
||||
|
||||
// Instances for 2 byte datatypes in CRR layout with ADataType = BDataType = EDataType
|
||||
template <typename T,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
enable_if_t<sizeof(T) == 2, bool> = false>
|
||||
using device_grouped_gemm_wmma_universal_km_kn_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| 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| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//##############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Col, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Col, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Col, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang`-format on
|
||||
>;
|
||||
|
||||
// Instances for 2 byte datatypes in CCR layout with ADataType = BDataType = EDataType
|
||||
template <typename T,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
enable_if_t<sizeof(T) == 2, bool> = false>
|
||||
using device_grouped_gemm_wmma_universal_km_nk_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//##############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| 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| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//##############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Col, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Col, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Col, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
// Instances for 2 byte datatypes in RRR layout with ADataType = BDataType = EDataType
|
||||
template <typename T,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
enable_if_t<sizeof(T) == 2, bool> = false>
|
||||
using device_grouped_gemm_wmma_universal_mk_kn_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| 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| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//##############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 2, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
// Instances for 2 byte datatypes in RCR layout with ADataType = BDataType = EDataType
|
||||
template <typename T,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
enable_if_t<sizeof(T) == 2, bool> = false>
|
||||
using device_grouped_gemm_wmma_universal_mk_nk_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| 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| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//##############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
// Helper function to add a list of layout instances with specific A/B/E datatypes for all supported
|
||||
// padding/scheduler/pipeline version combinations
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
template <device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer>
|
||||
typename LayoutInstances,
|
||||
typename ADataType, // NOTE: type parameters as last so that they can be inferred from the
|
||||
typename BDataType, // vector argument
|
||||
typename EDataType>
|
||||
void add_device_grouped_gemm_wmma_universal_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
LayoutInstances<GemmDefault, IntrawaveScheduler, PipelineV1>{});
|
||||
add_device_operation_instances(instances,
|
||||
LayoutInstances<GemmDefault, InterwaveScheduler, PipelineV1>{});
|
||||
add_device_operation_instances(instances,
|
||||
LayoutInstances<GemmDefault, IntrawaveScheduler, PipelineV3>{});
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<GemmMNKPadding, IntrawaveScheduler, PipelineV1>{});
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<GemmMNKPadding, InterwaveScheduler, PipelineV1>{});
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<GemmMNKPadding, IntrawaveScheduler, PipelineV3>{});
|
||||
}
|
||||
|
||||
// Helper function to add a list of layout instances for instances with matching A/B/E data types
|
||||
// for all supported padding/scheduler/pipeline version combinations
|
||||
template <typename T,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
template <typename T2,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer>
|
||||
typename LayoutInstances>
|
||||
void add_device_grouped_gemm_wmma_universal_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
T,
|
||||
T,
|
||||
DsDataType,
|
||||
T,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<T, GemmDefault, IntrawaveScheduler, PipelineV1>{});
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<T, GemmDefault, InterwaveScheduler, PipelineV1>{});
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<T, GemmDefault, IntrawaveScheduler, PipelineV3>{});
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<T, GemmMNKPadding, IntrawaveScheduler, PipelineV1>{});
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<T, GemmMNKPadding, InterwaveScheduler, PipelineV1>{});
|
||||
add_device_operation_instances(
|
||||
instances, LayoutInstances<T, GemmMNKPadding, IntrawaveScheduler, PipelineV3>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -57,7 +57,7 @@ function(add_instance_library INSTANCE_NAME)
|
||||
list(REMOVE_ITEM ARGN "${source}")
|
||||
endif()
|
||||
# Do not build XDL instances if gfx9 targets are not on the target list
|
||||
if(NOT INST_TARGETS MATCHES "gfx9" AND NOT INST_TARGETS MATCHES "gfx11" AND NOT INST_TARGETS MATCHES "gfx12" AND source_name MATCHES "_xdl")
|
||||
if(((NOT INST_TARGETS MATCHES "gfx9" AND NOT INST_TARGETS MATCHES "gfx11" AND NOT INST_TARGETS MATCHES "gfx12") OR FORCE_DISABLE_XDL) AND source_name MATCHES "_xdl")
|
||||
message(DEBUG "removing xdl instance ${source} ")
|
||||
list(REMOVE_ITEM ARGN "${source}")
|
||||
endif()
|
||||
@@ -67,7 +67,7 @@ function(add_instance_library INSTANCE_NAME)
|
||||
list(REMOVE_ITEM ARGN "${source}")
|
||||
endif()
|
||||
# Do not build WMMA instances if gfx11 targets are not on the target list
|
||||
if(NOT INST_TARGETS MATCHES "gfx11" AND NOT INST_TARGETS MATCHES "gfx12" AND source_name MATCHES "_wmma")
|
||||
if(((NOT INST_TARGETS MATCHES "gfx11" AND NOT INST_TARGETS MATCHES "gfx12") OR FORCE_DISABLE_WMMA) AND source_name MATCHES "_wmma")
|
||||
message(DEBUG "removing wmma instance ${source} ")
|
||||
list(REMOVE_ITEM ARGN "${source}")
|
||||
endif()
|
||||
@@ -88,7 +88,7 @@ function(add_instance_library INSTANCE_NAME)
|
||||
endif()
|
||||
endif()
|
||||
# Do not build WMMA gemm_universal_f8 for any targets except gfx12+
|
||||
if(NOT INST_TARGETS MATCHES "gfx12" AND source_name MATCHES "gemm_wmma_universal" AND source_name MATCHES "_f8_")
|
||||
if((NOT INST_TARGETS MATCHES "gfx12" OR FORCE_DISABLE_WMMA) AND source_name MATCHES "gemm_wmma_universal" AND source_name MATCHES "_f8_")
|
||||
message(DEBUG "removing gemm_universal_f8 instance ${source} ")
|
||||
list(REMOVE_ITEM ARGN "${source}")
|
||||
endif()
|
||||
@@ -274,7 +274,7 @@ FOREACH(subdir_path ${dir_list})
|
||||
message(DEBUG "Found only dl instances, but DL_KERNELS is not set. Skipping.")
|
||||
set(add_inst 0)
|
||||
endif()
|
||||
if(("${cmake_instance}" MATCHES "ONLY XDL_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx9|gfx11|gfx12"))
|
||||
if(("${cmake_instance}" MATCHES "ONLY XDL_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx9|gfx11|gfx12" OR FORCE_DISABLE_XDL))
|
||||
message(DEBUG "Found only xdl instances, but gfx9 is not on the targets list. Skipping.")
|
||||
set(add_inst 0)
|
||||
endif()
|
||||
@@ -282,7 +282,7 @@ FOREACH(subdir_path ${dir_list})
|
||||
message(DEBUG "Found only MX instances, but gfx950 is not on the targets list. Skipping.")
|
||||
set(add_inst 0)
|
||||
endif()
|
||||
if(("${cmake_instance}" MATCHES "ONLY WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12"))
|
||||
if(("${cmake_instance}" MATCHES "ONLY WMMA_KERNELS") AND (((NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12")) OR FORCE_DISABLE_WMMA))
|
||||
message(DEBUG "Found only wmma instances, but gfx11 is not on the targets list. Skipping.")
|
||||
set(add_inst 0)
|
||||
endif()
|
||||
@@ -290,7 +290,7 @@ FOREACH(subdir_path ${dir_list})
|
||||
message(DEBUG "Found only xdl and dl instances, but gfx9 is not on the targets listand DL_KERNELS is not set. Skipping.")
|
||||
set(add_inst 0)
|
||||
endif()
|
||||
if(("${cmake_instance}" MATCHES "ONLY XDL_AND_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx9|gfx11|gfx12"))
|
||||
if(("${cmake_instance}" MATCHES "ONLY XDL_AND_WMMA_KERNELS") AND ((NOT INST_TARGETS MATCHES "gfx9|gfx11|gfx12") OR (FORCE_DISABLE_XDL AND FORCE_DISABLE_WMMA)))
|
||||
message(DEBUG "Found only xdl and wmma instances, but gfx11 and gfx9 are not on the targets list. Skipping.")
|
||||
set(add_inst 0)
|
||||
endif()
|
||||
@@ -333,20 +333,22 @@ FOREACH(subdir_path ${dir_list})
|
||||
if((add_inst EQUAL 1))
|
||||
get_filename_component(target_dir ${subdir_path} NAME)
|
||||
add_subdirectory(${target_dir})
|
||||
if("${cmake_instance}" MATCHES "gemm")
|
||||
list(APPEND CK_DEVICE_GEMM_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
elseif("${cmake_instance}" MATCHES "conv")
|
||||
list(APPEND CK_DEVICE_CONV_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
elseif("${cmake_instance}" MATCHES "mha")
|
||||
list(APPEND CK_DEVICE_MHA_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
elseif("${cmake_instance}" MATCHES "contr")
|
||||
list(APPEND CK_DEVICE_CONTRACTION_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
elseif("${cmake_instance}" MATCHES "reduce")
|
||||
list(APPEND CK_DEVICE_REDUCTION_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
else()
|
||||
list(APPEND CK_DEVICE_OTHER_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
endif()
|
||||
message(DEBUG "add_instance_directory ${subdir_path}")
|
||||
if (TARGET device_${target_dir}_instance)
|
||||
if("${cmake_instance}" MATCHES "gemm")
|
||||
list(APPEND CK_DEVICE_GEMM_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
elseif("${cmake_instance}" MATCHES "conv")
|
||||
list(APPEND CK_DEVICE_CONV_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
elseif("${cmake_instance}" MATCHES "mha")
|
||||
list(APPEND CK_DEVICE_MHA_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
elseif("${cmake_instance}" MATCHES "contr")
|
||||
list(APPEND CK_DEVICE_CONTRACTION_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
elseif("${cmake_instance}" MATCHES "reduce")
|
||||
list(APPEND CK_DEVICE_REDUCTION_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
else()
|
||||
list(APPEND CK_DEVICE_OTHER_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
|
||||
endif()
|
||||
message(DEBUG "add_instance_directory ${subdir_path}")
|
||||
endif()
|
||||
else()
|
||||
message(DEBUG "skip_instance_directory ${subdir_path}")
|
||||
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
|
||||
add_instance_library(device_grouped_gemm_instance
|
||||
device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
|
||||
@@ -36,4 +36,17 @@ add_instance_library(device_grouped_gemm_instance
|
||||
device_grouped_gemm_multiple_d_splitk_xdl_two_stage_bf16_bf16_bf16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_multiple_d_splitk_xdl_two_stage_bf16_i8_bf16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_multiple_d_splitk_xdl_two_stage_bf16_i8_bf16_mk_nk_mn_instance.cpp
|
||||
|
||||
device_grouped_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_instance.cpp
|
||||
|
||||
device_grouped_gemm_wmma_universal_f16_f16_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_universal_f16_f16_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_universal_f16_f16_f16_km_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_universal_f16_f16_f16_km_nk_mn_instance.cpp
|
||||
|
||||
device_grouped_gemm_wmma_universal_bf16_bf16_bf16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_universal_bf16_bf16_bf16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_instance.cpp
|
||||
)
|
||||
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
|
||||
Row,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
BF16,
|
||||
BF16,
|
||||
DsDataType,
|
||||
BF16,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
BF16,
|
||||
Col,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_universal_km_kn_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
|
||||
Col,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
BF16,
|
||||
BF16,
|
||||
DsDataType,
|
||||
BF16,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
BF16,
|
||||
Col,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_universal_km_nk_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
DsDataType,
|
||||
BF16,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
BF16,
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_universal_mk_kn_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_bf16_bf16_bf16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Col,
|
||||
DsLayout,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
DsDataType,
|
||||
BF16,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
BF16,
|
||||
Row,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_universal_mk_nk_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f16_f16_km_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
DsDataType,
|
||||
F16,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
F16,
|
||||
Col,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_universal_km_kn_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f16_f16_km_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
|
||||
Col,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
F16,
|
||||
F16,
|
||||
DsDataType,
|
||||
F16,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
F16,
|
||||
Col,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_universal_km_nk_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,38 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
DsDataType,
|
||||
F16,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
F16,
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_universal_mk_kn_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,38 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f16_f16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Col,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
F16,
|
||||
F16,
|
||||
DsDataType,
|
||||
F16,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
F16,
|
||||
Row,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_universal_mk_nk_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,57 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F8;
|
||||
using EDataType = F16;
|
||||
|
||||
template <device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer>
|
||||
using device_grouped_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| 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| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//##############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, ADataType, BDataType, AccDataType, EDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, ADataType, BDataType, AccDataType, EDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 2, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, ADataType, BDataType, AccDataType, EDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,57 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_splitk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using ADataType = F8;
|
||||
using BDataType = F16;
|
||||
using EDataType = F16;
|
||||
|
||||
template <device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer>
|
||||
using device_grouped_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| 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| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//##############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, ADataType, BDataType, AccDataType, EDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, ADataType, BDataType, AccDataType, EDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 2, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_CShuffleV3< Row, Row, DsLayout, ELayout, ADataType, BDataType, AccDataType, EDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
|
||||
add_device_grouped_gemm_wmma_universal_instances<
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user