mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 02:02:46 +00:00
Add bf16 and bf16@int8 mk_nk_mn instances for grouped gemm two stage (#1228)
* added bf16 and bf16@int8 mk_nk_mn instances
* fix preprocessor guards
[ROCm/composable_kernel commit: e0f3f918f1]
This commit is contained in:
jakpiase
@@ -10,12 +10,13 @@
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
#ifdef CK_ENABLE_FP16
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
#if defined(CK_USE_XDL)
|
||||
#if defined(CK_ENABLE_FP16)
|
||||
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
@@ -120,6 +121,21 @@ void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_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);
|
||||
#endif
|
||||
|
||||
#if defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8)
|
||||
void add_device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
@@ -145,20 +161,37 @@ void add_device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instances(
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_f16_f16_f16_mk_kn_mn_instances(
|
||||
#if defined(CK_ENABLE_BF16)
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_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);
|
||||
#endif
|
||||
|
||||
#if defined(CK_ENABLE_BF16) && defined(CK_ENABLE_INT8)
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Row,
|
||||
@@ -172,6 +205,20 @@ void add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_kn_mn_inst
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
#endif // CK_USE_XDL
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename ELayout,
|
||||
@@ -205,7 +252,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
#if defined(CK_USE_XDL)
|
||||
#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>)
|
||||
@@ -270,8 +317,33 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_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_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_nk_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_multiple_d_xdl_two_stage_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_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_nk_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
#endif // CK_USE_XDL
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
@@ -280,4 +352,3 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
#endif
|
||||
|
||||
@@ -11,5 +11,8 @@ add_instance_library(device_grouped_gemm_instance
|
||||
device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp
|
||||
device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instance.cpp
|
||||
device_grouped_gemm_multiple_d_splitk_xdl_two_stage_f16_f16_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_multiple_d_splitk_xdl_two_stage_bf16_bf16_bf16_mk_kn_mn_instance.cpp
|
||||
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
|
||||
)
|
||||
|
||||
@@ -0,0 +1,99 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_multiple_d_splitk_xdl_cshuffle_two_stage.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
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;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
// Instances having AK1!=BK1 are temporarily disabled and will be re-enabled in future
|
||||
// a[m, k] * b[k, n] = e[m, n]
|
||||
using device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_kn_mn_generic_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//#################################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#################################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//#################################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//#################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 1, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_kn_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#################################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#################################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//#################################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//#################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 24, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Row, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_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)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_kn_mn_instances{});
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_kn_mn_generic_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,99 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_multiple_d_splitk_xdl_cshuffle_two_stage.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
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;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
// Instances having AK1!=BK1 are temporarily disabled and will be re-enabled in future
|
||||
// a[m, k] * b[k, n] = e[m, n]
|
||||
using device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_nk_mn_generic_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//#################################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#################################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//#################################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//#################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 1, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_nk_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#################################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#################################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//#################################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//#################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 48, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 24, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_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)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_nk_mn_instances{});
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_nk_mn_generic_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,99 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_multiple_d_splitk_xdl_cshuffle_two_stage.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using BF16 = ck::bhalf_t;
|
||||
using F32 = float;
|
||||
using I8 = int8_t;
|
||||
|
||||
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;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
// Instances having AK1!=BK1 are temporarily disabled and will be re-enabled in future
|
||||
// a[m, k] * b[k, n] = e[m, n]
|
||||
using device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_nk_mn_generic_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//#################################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#################################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//#################################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//#################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 1, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_nk_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#################################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#################################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//#################################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//#################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 48, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 24, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>,
|
||||
DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage< Row, Col, Empty_Tuple, Row, BF16, I8, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, PipelineVersion::v1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances, device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_nk_mn_instances{});
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_grouped_gemm_multiple_d_xdl_two_stage_bf16_i8_bf16_mk_nk_mn_generic_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -12,12 +12,14 @@
|
||||
enum struct GemmMatrixLayout
|
||||
{
|
||||
MK_KN_MN, // 0
|
||||
MK_NK_MN, // 1
|
||||
};
|
||||
|
||||
enum struct GemmDataType
|
||||
{
|
||||
F16_F16_F16, // 0
|
||||
BF16_INT8_BF16 // 1
|
||||
F16_F16_F16, // 0
|
||||
BF16_INT8_BF16, // 1
|
||||
BF16_BF16_BF16 // 2
|
||||
};
|
||||
|
||||
#define OP_NAME "grouped_gemm_two_stage"
|
||||
@@ -47,7 +49,7 @@ int profile_grouped_gemm_two_stage(int argc, char* argv[])
|
||||
{
|
||||
std::cout
|
||||
<< "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
|
||||
<< "arg2: data type (0: fp16; 1: bf16@int8)\n"
|
||||
<< "arg2: data type (0: fp16; 1: bf16@int8; 2: bf16)\n"
|
||||
<< "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n]);\n"
|
||||
<< "arg4: verification (0: no; 1: yes)\n"
|
||||
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"
|
||||
@@ -145,6 +147,75 @@ int profile_grouped_gemm_two_stage(int argc, char* argv[])
|
||||
n_warmup,
|
||||
n_iter);
|
||||
}
|
||||
else if(data_type == GemmDataType::BF16_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
|
||||
{
|
||||
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
|
||||
int8_t,
|
||||
ck::bhalf_t,
|
||||
float,
|
||||
ck::tensor_layout::gemm::RowMajor,
|
||||
ck::tensor_layout::gemm::ColumnMajor,
|
||||
ck::tensor_layout::gemm::RowMajor>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
Ms,
|
||||
Ns,
|
||||
Ks,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideCs,
|
||||
kbatch,
|
||||
n_warmup,
|
||||
n_iter);
|
||||
}
|
||||
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
|
||||
{
|
||||
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
|
||||
ck::bhalf_t,
|
||||
ck::bhalf_t,
|
||||
float,
|
||||
ck::tensor_layout::gemm::RowMajor,
|
||||
ck::tensor_layout::gemm::RowMajor,
|
||||
ck::tensor_layout::gemm::RowMajor>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
Ms,
|
||||
Ns,
|
||||
Ks,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideCs,
|
||||
kbatch,
|
||||
n_warmup,
|
||||
n_iter);
|
||||
}
|
||||
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
|
||||
{
|
||||
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
|
||||
ck::bhalf_t,
|
||||
ck::bhalf_t,
|
||||
float,
|
||||
ck::tensor_layout::gemm::RowMajor,
|
||||
ck::tensor_layout::gemm::ColumnMajor,
|
||||
ck::tensor_layout::gemm::RowMajor>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
Ms,
|
||||
Ns,
|
||||
Ks,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideCs,
|
||||
kbatch,
|
||||
n_warmup,
|
||||
n_iter);
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
|
||||
|
||||
Reference in New Issue
Block a user