mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-07 08:15:04 +00:00
added add_fastgelu instances
This commit is contained in:
Zoltan Lakatos
@@ -97,8 +97,47 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_in
|
||||
#endif // CK_USE_XDL
|
||||
|
||||
#if defined(CK_USE_WMMA)
|
||||
void add_device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances(
|
||||
void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
|
||||
Row,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
F16_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
AddFastGelu>>>&);
|
||||
|
||||
void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
|
||||
Col,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
F16_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
AddFastGelu>>>&);
|
||||
|
||||
void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
|
||||
Row,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
F16_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
AddFastGelu>>>&);
|
||||
|
||||
void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
|
||||
Col,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
@@ -159,16 +198,32 @@ struct DeviceOperationInstanceFactory<
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
|
||||
is_same_v<D0DataType, half_t> && is_same_v<EDataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances(
|
||||
add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
// TODO: Add other types and layouts
|
||||
|
||||
|
||||
#endif // CK_ENABLE_FP16
|
||||
#endif // CK_USE_WMMA
|
||||
|
||||
|
||||
@@ -1,6 +1,9 @@
|
||||
# ONLY XDL_AND_WMMA_KERNELS
|
||||
add_instance_library(device_gemm_add_fastgelu_instance
|
||||
device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
|
||||
|
||||
device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
|
||||
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
|
||||
|
||||
@@ -0,0 +1,76 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/utility/sequence.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
|
||||
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
|
||||
|
||||
static constexpr auto V1 = BlockGemmPipelineVersion::v1;
|
||||
static constexpr auto V3 = BlockGemmPipelineVersion::v3;
|
||||
|
||||
// e = elementwise((a * b), d0)
|
||||
// elementwise(c, d0) = fastgelu(c + d0)
|
||||
// output: e[m, n]
|
||||
// input: a[m, k], b[n, k], d0[m, n]
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm|
|
||||
//##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer|
|
||||
//##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | |
|
||||
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | |
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 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, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
|
||||
Row,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
F16_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
AddFastGelu>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance<GemmDefault>{});
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance<GemmMNKPadding>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,78 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/utility/sequence.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
|
||||
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
|
||||
|
||||
static constexpr auto V1 = BlockGemmPipelineVersion::v1;
|
||||
static constexpr auto V3 = BlockGemmPipelineVersion::v3;
|
||||
|
||||
// e = elementwise((a * b), d0)
|
||||
// elementwise(c, d0) = fastgelu(c + d0)
|
||||
// output: e[m, n]
|
||||
// input: a[m, k], b[n, k], d0[m, n]
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm|
|
||||
//##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer|
|
||||
//##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | |
|
||||
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | |
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 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>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 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, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
|
||||
Col,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
F16_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
AddFastGelu>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances<GemmDefault>{});
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances<GemmMNKPadding>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,81 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/utility/sequence.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
|
||||
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
|
||||
|
||||
static constexpr auto V1 = BlockGemmPipelineVersion::v1;
|
||||
static constexpr auto V3 = BlockGemmPipelineVersion::v3;
|
||||
|
||||
// e = elementwise((a * b), d0)
|
||||
// elementwise(c, d0) = fastgelu(c + d0)
|
||||
// output: e[m, n]
|
||||
// input: a[m, k], b[n, k], d0[m, n]
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm|
|
||||
//##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer|
|
||||
//##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | |
|
||||
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | |
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Intrawave, V3>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
|
||||
Row,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
F16_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
AddFastGelu>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances<GemmDefault>{});
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances<GemmMNKPadding>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,83 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/utility/sequence.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
|
||||
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
|
||||
|
||||
static constexpr auto V1 = BlockGemmPipelineVersion::v1;
|
||||
static constexpr auto V3 = BlockGemmPipelineVersion::v3;
|
||||
|
||||
// e = elementwise((a * b), d0)
|
||||
// elementwise(c, d0) = fastgelu(c + d0)
|
||||
// output: e[m, n]
|
||||
// input: a[m, k], b[n, k], d0[m, n]
|
||||
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm|
|
||||
//##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer|
|
||||
//##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | |
|
||||
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | |
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 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>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 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, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 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>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 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, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>,
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
|
||||
Col,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
F16_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
AddFastGelu>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances<GemmDefault>{});
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances<GemmMNKPadding>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -1,73 +0,0 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/utility/sequence.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
|
||||
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
|
||||
|
||||
// e = elementwise((a * b), d0)
|
||||
// elementwise(c, d0) = fastgelu(c + d0)
|
||||
// output: e[m, n]
|
||||
// input: a[m, k], b[n, k], d0[m, n]
|
||||
|
||||
using device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_generic_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm|
|
||||
//##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer|
|
||||
//##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | |
|
||||
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | |
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, BlockGemmPipelineVersion::v1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm|
|
||||
//##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer|
|
||||
//##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | |
|
||||
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | |
|
||||
DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmDefault, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, BlockGemmPipelineVersion::v1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
|
||||
Col,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
F16_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
AddFastGelu>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_generic_instances{});
|
||||
add_device_operation_instances(
|
||||
instances, device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -47,7 +47,7 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_insta
|
||||
>;
|
||||
|
||||
void add_device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
|
||||
Row,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
|
||||
@@ -1,3 +1,5 @@
|
||||
# Implements test instances for MultipleD with xdl and wmma support.
|
||||
|
||||
add_gtest_executable(test_gemm_add_xdl test_gemm_add_xdl.cpp)
|
||||
if(result EQUAL 0)
|
||||
target_link_libraries(test_gemm_add_xdl PRIVATE utility device_gemm_add_instance)
|
||||
|
||||
@@ -34,7 +34,10 @@ class TestGemmAddFastgelu : public TestGemmD0Common<Tuple>
|
||||
decltype(ProfileGemmAddFastgeluImpl) GetImpl() override { return ProfileGemmAddFastgeluImpl; }
|
||||
};
|
||||
|
||||
using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, F16, F16, Row, Col, Row, Row>>;
|
||||
using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, F16, F16, Row, Row, Row, Row>,
|
||||
std::tuple<F16, F16, F32, F16, F16, Row, Col, Row, Row>,
|
||||
std::tuple<F16, F16, F32, F16, F16, Col, Row, Row, Row>,
|
||||
std::tuple<F16, F16, F32, F16, F16, Col, Col, Row, Row>>;
|
||||
|
||||
TYPED_TEST_SUITE(TestGemmAddFastgelu, KernelTypes);
|
||||
TYPED_TEST(TestGemmAddFastgelu, Test_BF16FP16) { this->Run(); }
|
||||
|
||||
Reference in New Issue
Block a user