ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-13 10:37:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/feature/multiple-d-gemms' into 64-implement-device_gemm_multiply_multiply_instance-for-rdna4

Browse Source
This commit is contained in:
Zoltan Lakatos
2025-06-18 14:46:38 +00:00
parent cd0172bec5 055bc0227c
commit 40ce862cdf
29 changed files with 1678 additions and 290 deletions
Download Patch File Download Diff File Expand all files Collapse all files

211
library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -11,11 +11,13 @@
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#if defined(CK_ENABLE_FP16)
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#if defined(CK_USE_XDL)
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
@@ -67,8 +69,64 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn
PassThrough,
PassThrough,
AddAddFastGelu>>>&);
#endif // CK_USE_XDL
// GEMM + Add + Add + FastGelu
#if defined(CK_USE_WMMA)
void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Row,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
AddAddFastGelu>>>&);
void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
AddAddFastGelu>>>&);
void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
Row,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
AddAddFastGelu>>>&);
void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
Col,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
AddAddFastGelu>>>&);
#endif // CK_USE_WMMA
// GEMM + Add + FastGelu
// DeviceGemmMultipleDSplitK specialization
template <typename ALayout,
typename BLayout,
typename D0Layout,
@@ -79,18 +137,101 @@ template <typename ALayout,
typename D0DataType,
typename D1DataType,
typename EDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddAddFastGelu>>
struct DeviceOperationInstanceFactory<DeviceGemmMultipleDSplitK<ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
PassThrough,
PassThrough,
AddAddFastGelu>>
{
using DeviceOp = DeviceGemmMultipleDSplitK<ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
PassThrough,
PassThrough,
AddAddFastGelu>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if defined(CK_USE_XDL)
// No XDL instances for DeviceGemmMultipleDSplitK with AddFastGelu at the moment
#endif // CK_USE_XDL
#if defined(CK_USE_WMMA)
constexpr bool IsAllDRowLayout = is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row>;
constexpr bool IsAllDFloat16 =
is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t>;
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t> && IsAllDRowLayout && IsAllDFloat16)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
op_ptrs);
}
}
#endif // CK_USE_WMMA
return op_ptrs;
}
};
// GEMM + Add + Add + FastGelu
// DeviceGemmMultipleD specialization
template <typename ALayout,
typename BLayout,
typename D0Layout,
typename D1Layout,
typename ELayout,
typename ADataType,
typename BDataType,
typename D0DataType,
typename D1DataType,
typename EDataType>
struct DeviceOperationInstanceFactory<
DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
PassThrough,
PassThrough,
AddAddFastGelu>>
{
using DeviceOp = DeviceGemmMultipleD<ALayout,
BLayout,
@@ -100,47 +241,70 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddAddFastGelu>;
PassThrough,
PassThrough,
AddAddFastGelu>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if defined(CK_USE_XDL)
constexpr bool IsAllDRowLayout = is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row>;
constexpr bool IsAllDFloat16 =
is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t>;
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
is_same_v<EDataType, half_t>)
is_same_v<EDataType, half_t> && IsAllDRowLayout && IsAllDFloat16)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_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<D1Layout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_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<D1Layout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_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<D1Layout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
op_ptrs);
}
}
#endif // CK_USE_XDL
#if defined(CK_USE_WMMA)
// Reuse DeviceGemmMultipleDSplitK instances
using Wrapper = DeviceGemmMultipleDSplitKWrapper<ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
PassThrough,
PassThrough,
AddAddFastGelu>;
auto new_op_ptrs =
DeviceOperationInstanceFactory<typename Wrapper::DeviceOp>::GetInstances();
for(auto& op_ptr : new_op_ptrs)
{
op_ptrs.emplace_back(std::make_unique<Wrapper>(std::move(op_ptr)));
}
#endif // CK_USE_WMMA
return op_ptrs;
}
@@ -150,3 +314,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif // CK_ENABLE_FP16

113
library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
View File

@@ -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,
@@ -121,18 +160,17 @@ template <typename ALayout,
typename BDataType,
typename D0DataType,
typename EDataType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmMultipleDSplitK<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
AddFastGelu>>
struct DeviceOperationInstanceFactory<DeviceGemmMultipleDSplitK<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
AddFastGelu>>
{
using DeviceOp = DeviceGemmMultipleDSplitK<ALayout,
BLayout,
@@ -159,16 +197,32 @@ 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
@@ -186,18 +240,17 @@ template <typename ALayout,
typename BDataType,
typename D0DataType,
typename EDataType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
AddFastGelu>>
struct DeviceOperationInstanceFactory<DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
AddFastGelu>>
{
using DeviceOp = DeviceGemmMultipleD<ALayout,
BLayout,

176
library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -16,6 +16,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
#if defined(CK_USE_XDL)
void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
@@ -67,6 +68,132 @@ void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
PassThrough,
PassThrough,
FastGelu>>>&);
#endif // CK_USE_XDL
#if defined(CK_USE_WMMA)
void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>&);
void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>&);
void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>&);
void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>&);
#endif // CK_USE_WMMA
// GEMM + Add + FastGelu
// DeviceGemmMultipleDSplitK specialization
template <typename ALayout,
typename BLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename EDataType>
struct DeviceOperationInstanceFactory<DeviceGemmMultipleDSplitK<ALayout,
BLayout,
Empty_Tuple,
ELayout,
ADataType,
BDataType,
Empty_Tuple,
EDataType,
PassThrough,
PassThrough,
FastGelu>>
{
using DeviceOp = DeviceGemmMultipleDSplitK<ALayout,
BLayout,
Empty_Tuple,
ELayout,
ADataType,
BDataType,
Empty_Tuple,
EDataType,
PassThrough,
PassThrough,
FastGelu>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if defined(CK_USE_XDL)
// No XDL instances for DeviceGemmMultipleDSplitK with AddFastGelu at the moment
#endif // CK_USE_XDL
#if defined(CK_USE_WMMA)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
}
}
#endif // CK_USE_WMMA
return op_ptrs;
}
};
// GEMM + FastGelu
template <typename ALayout,
@@ -75,17 +202,17 @@ template <typename ALayout,
typename ADataType,
typename BDataType,
typename EDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
BLayout,
Empty_Tuple,
ELayout,
ADataType,
BDataType,
Empty_Tuple,
EDataType,
PassThrough,
PassThrough,
FastGelu>>
struct DeviceOperationInstanceFactory<DeviceGemmMultipleD<ALayout,
BLayout,
Empty_Tuple,
ELayout,
ADataType,
BDataType,
Empty_Tuple,
EDataType,
PassThrough,
PassThrough,
FastGelu>>
{
using DeviceOp = DeviceGemmMultipleD<ALayout,
BLayout,
@@ -103,6 +230,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if defined(CK_USE_XDL)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t>)
{
@@ -127,6 +255,28 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
}
}
#endif // CK_USE_XDL
#if defined(CK_USE_WMMA)
// Reuse DeviceGemmMultipleDSplitK instances
using Wrapper = DeviceGemmMultipleDSplitKWrapper<ALayout,
BLayout,
Empty_Tuple,
ELayout,
ADataType,
BDataType,
Empty_Tuple,
EDataType,
PassThrough,
PassThrough,
FastGelu>;
auto new_op_ptrs =
DeviceOperationInstanceFactory<typename Wrapper::DeviceOp>::GetInstances();
for(auto& op_ptr : new_op_ptrs)
{
op_ptrs.emplace_back(std::make_unique<Wrapper>(std::move(op_ptr)));
}
#endif // CK_USE_WMMA
return op_ptrs;
}
@@ -136,4 +286,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
#endif // CK_ENABLE_FP16

7
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
View File

@@ -1,5 +1,10 @@
# ONLY XDL_KERNELS
# ONLY XDL_AND_WMMA_KERNELS
add_instance_library(device_gemm_add_add_fastgelu_instance
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp

76
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp Normal file
View File

@@ -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, d1)
// elementwise(c, d0, d1) = fastgelu(c + d0 + d1)
// output: e[m, n]
// input: a[m, k], b[n, k], d0[m, n], d1[m, n]
template <GemmSpecialization GemmSpec>
using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
Row,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
AddAddFastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance<GemmDefault>{});
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

78
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp Normal file
View File

@@ -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, d1)
// elementwise(c, d0, d1) = fastgelu(c + d0 + d1)
// output: e[m, n]
// input: a[m, k], b[n, k], d0[m, n], d1[m, n]
template <GemmSpecialization GemmSpec>
using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
Col,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
AddAddFastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances<GemmDefault>{});
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

81
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp Normal file
View File

@@ -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, d1)
// elementwise(c, d0, d1) = fastgelu(c + d0 + d1)
// output: e[m, n]
// input: a[m, k], b[n, k], d0[m, n], d1[m, n]
template <GemmSpecialization GemmSpec>
using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Row,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
AddAddFastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances<GemmDefault>{});
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

84
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,84 @@
// 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, d1)
// elementwise(c, d0, d1) = fastgelu(c + d0 + d1)
// output: e[m, n]
// input: a[m, k], b[n, k], d0[m, n], d1[m, n]
template <GemmSpecialization GemmSpec>
using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, 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_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
AddAddFastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances<GemmDefault>{});
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

5
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
View File

@@ -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

76
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp Normal file
View File

@@ -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

78
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp Normal file
View File

@@ -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

81
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -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

83
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp Normal file
View File

@@ -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

73
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
View File

@@ -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

7
library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
View File

@@ -1,5 +1,10 @@
# ONLY XDL_KERNELS
# ONLY XDL_AND_WMMA_KERNELS
add_instance_library(device_gemm_fastgelu_instance
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp

76
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp Normal file
View File

@@ -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)
// elementwise(c) = fastgelu(c)
// output: e[m, n]
// input: a[m, k], b[n, k]
template <GemmSpecialization GemmSpec>
using device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance<GemmDefault>{});
add_device_operation_instances(
instances,
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

78
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp Normal file
View File

@@ -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)
// elementwise(c) = fastgelu(c)
// output: e[m, n]
// input: a[m, k], b[n, k]
template <GemmSpecialization GemmSpec>
using device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Col,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances<GemmDefault>{});
add_device_operation_instances(
instances,
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

81
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp Normal file
View File

@@ -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)
// elementwise(c) = fastgelu(c)
// output: e[m, n]
// input: a[m, k], b[n, k]
template <GemmSpecialization GemmSpec>
using device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances<GemmDefault>{});
add_device_operation_instances(
instances,
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

83
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp Normal file
View File

@@ -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)
// elementwise(c) = fastgelu(c)
// output: e[m, n]
// input: a[m, k], b[n, k]
template <GemmSpecialization GemmSpec>
using device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, 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_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances<GemmDefault>{});
add_device_operation_instances(
instances,
device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

12
profiler/src/CMakeLists.txt
View File

@@ -44,10 +44,8 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_OPS profile_batched_gemm_gemm.cpp)
list(APPEND PROFILER_OPS profile_batched_gemm_add_relu_gemm_add.cpp)
list(APPEND PROFILER_OPS profile_gemm_add.cpp)
list(APPEND PROFILER_OPS profile_gemm_add_add_fastgelu.cpp)
list(APPEND PROFILER_OPS profile_grouped_gemm.cpp)
list(APPEND PROFILER_OPS profile_gemm_streamk.cpp)
list(APPEND PROFILER_OPS profile_gemm_fastgelu.cpp)
list(APPEND PROFILER_OPS profile_gemm_add_relu.cpp)
list(APPEND PROFILER_OPS profile_gemm_add_silu.cpp)
list(APPEND PROFILER_OPS profile_gemm_add_relu_add_layernorm.cpp)
@@ -92,8 +90,10 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1[1
list(APPEND PROFILER_OPS profile_grouped_conv_fwd.cpp)
list(APPEND PROFILER_OPS profile_grouped_conv_bwd_data.cpp)
list(APPEND PROFILER_OPS profile_grouped_conv_bwd_weight.cpp)
list(APPEND PROFILER_OPS profile_gemm_add_fastgelu.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_OPS profile_gemm_add_fastgelu.cpp)
list(APPEND PROFILER_OPS profile_gemm_fastgelu.cpp)
list(APPEND PROFILER_OPS profile_gemm_add_add_fastgelu.cpp)
endif()
endif()
@@ -150,8 +150,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_INSTANCES device_gemm_add_instance)
list(APPEND DEVICE_INSTANCES device_gemm_add_add_fastgelu_instance)
# list(APPEND DEVICE_INSTANCES device_gemm_fastgelu_instance)
list(APPEND DEVICE_INSTANCES device_batched_gemm_gemm_instance)
list(APPEND DEVICE_INSTANCES device_batched_gemm_add_relu_gemm_add_instance)
list(APPEND DEVICE_INSTANCES device_grouped_gemm_instance)
@@ -207,8 +205,10 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1[1
list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_data_instance)
list(APPEND DEVICE_INSTANCES device_grouped_conv2d_fwd_instance)
list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_weight_instance)
list(APPEND DEVICE_INSTANCES device_gemm_add_fastgelu_instance)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_INSTANCES device_gemm_add_fastgelu_instance)
list(APPEND DEVICE_INSTANCES device_gemm_fastgelu_instance)
list(APPEND DEVICE_INSTANCES device_gemm_add_add_fastgelu_instance)
endif()
endif()

52
test/gemm_add/CMakeLists.txt
View File

@@ -1,27 +1,39 @@
# 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)
# endif()
# Implements test instances for MultipleD with xdl and wmma support.
# add_gtest_executable(test_gemm_add_relu_xdl test_gemm_add_relu_xdl.cpp)
# if(result EQUAL 0)
# target_link_libraries(test_gemm_add_relu_xdl PRIVATE utility device_gemm_add_instance device_gemm_add_relu_instance)
# endif()
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)
endif()
# add_gtest_executable(test_gemm_add_silu_xdl test_gemm_add_silu_xdl.cpp)
# if(result EQUAL 0)
# target_link_libraries(test_gemm_add_silu_xdl PRIVATE utility device_gemm_add_instance device_gemm_add_silu_instance)
# endif()
add_gtest_executable(test_gemm_add_relu_xdl test_gemm_add_relu_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_gemm_add_relu_xdl PRIVATE utility device_gemm_add_instance device_gemm_add_relu_instance)
endif()
# add_gtest_executable(test_gemm_add_fastgelu_xdl test_gemm_add_fastgelu_xdl.cpp)
# if(result EQUAL 0)
# target_link_libraries(test_gemm_add_fastgelu_xdl PRIVATE utility device_gemm_add_instance device_gemm_add_fastgelu_instance)
# endif()
add_gtest_executable(test_gemm_add_silu_xdl test_gemm_add_silu_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_gemm_add_silu_xdl PRIVATE utility device_gemm_add_instance device_gemm_add_silu_instance)
endif()
# add_gtest_executable(test_gemm_add_fastgelu_wmma test_gemm_add_fastgelu_wmma.cpp)
# if(result EQUAL 0)
# target_link_libraries(test_gemm_add_fastgelu_wmma PRIVATE utility device_gemm_add_fastgelu_instance)
# endif()
add_gtest_executable(test_gemm_add_fastgelu_xdl test_gemm_add_fastgelu_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_gemm_add_fastgelu_xdl PRIVATE utility device_gemm_add_instance device_gemm_add_fastgelu_instance)
endif()
add_gtest_executable(test_gemm_fastgelu_wmma test_gemm_fastgelu_wmma.cpp)
if(result EQUAL 0)
target_link_libraries(test_gemm_fastgelu_wmma PRIVATE utility device_gemm_fastgelu_instance)
endif()
add_gtest_executable(test_gemm_add_fastgelu_wmma test_gemm_add_fastgelu_wmma.cpp)
if(result EQUAL 0)
target_link_libraries(test_gemm_add_fastgelu_wmma PRIVATE utility device_gemm_add_fastgelu_instance)
endif()
add_gtest_executable(test_gemm_add_add_fastgelu_wmma test_gemm_add_add_fastgelu_wmma.cpp)
if(result EQUAL 0)
target_link_libraries(test_gemm_add_add_fastgelu_wmma PRIVATE utility device_gemm_add_add_fastgelu_instance)
endif()
add_gtest_executable(test_gemm_multiply_multiply_wmma test_gemm_multiply_multiply_wmma.cpp)
if(result EQUAL 0)

38
test/gemm_add/test_gemm_add_add_fastgelu_wmma.cpp Normal file
View File

@@ -0,0 +1,38 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_add_fastgelu_impl.hpp"
#include "test_gemm_common.hpp"
template <typename Tuple>
class TestGemmAddAddFastgelu : public TestGemmD0D1Common<Tuple>
{
using ProfileCall = typename TestGemmD0D1Common<Tuple>::ProfileCall;
public:
ProfileCall GetImpl() override
{
return ck::profiler::profile_gemm_add_add_fastgelu_impl<
typename TestGemmD0D1Common<Tuple>::ADataType,
typename TestGemmD0D1Common<Tuple>::BDataType,
typename TestGemmD0D1Common<Tuple>::AccDataType,
typename TestGemmD0D1Common<Tuple>::D0DataType,
typename TestGemmD0D1Common<Tuple>::D1DataType,
typename TestGemmD0D1Common<Tuple>::EDataType,
typename TestGemmD0D1Common<Tuple>::ALayout,
typename TestGemmD0D1Common<Tuple>::BLayout,
typename TestGemmD0D1Common<Tuple>::D0Layout,
typename TestGemmD0D1Common<Tuple>::D1Layout,
typename TestGemmD0D1Common<Tuple>::ELayout>;
}
};
using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, F16, F16, F16, Row, Row, Row, Row, Row>,
std::tuple<F16, F16, F32, F16, F16, F16, Row, Col, Row, Row, Row>,
std::tuple<F16, F16, F32, F16, F16, F16, Col, Row, Row, Row, Row>,
std::tuple<F16, F16, F32, F16, F16, F16, Col, Col, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAddAddFastgelu, KernelTypes);
TYPED_TEST(TestGemmAddAddFastgelu, Test_BF16FP16) { this->Run(); }

41
test/gemm_add/test_gemm_add_fastgelu_wmma.cpp
View File

@@ -9,32 +9,27 @@
template <typename Tuple>
class TestGemmAddFastgelu : public TestGemmD0Common<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
using ProfileCall = typename TestGemmD0Common<Tuple>::ProfileCall;
constexpr static auto ProfileGemmAddFastgeluImpl =
ck::profiler::profile_gemm_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddFastgeluImpl) GetImpl() override { return ProfileGemmAddFastgeluImpl; }
ProfileCall GetImpl() override
{
return ck::profiler::profile_gemm_add_fastgelu_impl<
typename TestGemmD0Common<Tuple>::ADataType,
typename TestGemmD0Common<Tuple>::BDataType,
typename TestGemmD0Common<Tuple>::AccDataType,
typename TestGemmD0Common<Tuple>::D0DataType,
typename TestGemmD0Common<Tuple>::EDataType,
typename TestGemmD0Common<Tuple>::ALayout,
typename TestGemmD0Common<Tuple>::BLayout,
typename TestGemmD0Common<Tuple>::D0Layout,
typename TestGemmD0Common<Tuple>::ELayout>;
}
};
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(); }

36
test/gemm_add/test_gemm_add_fastgelu_xdl.cpp
View File

@@ -9,29 +9,21 @@
template <typename Tuple>
class TestGemmAddFastgelu : public TestGemmD0Common<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
using ProfileCall = typename TestGemmD0Common<Tuple>::ProfileCall;
constexpr static auto ProfileGemmAddFastgeluImpl =
ck::profiler::profile_gemm_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddFastgeluImpl) GetImpl() override { return ProfileGemmAddFastgeluImpl; }
ProfileCall GetImpl() override
{
return ck::profiler::profile_gemm_add_fastgelu_impl<
typename TestGemmD0Common<Tuple>::ADataType,
typename TestGemmD0Common<Tuple>::BDataType,
typename TestGemmD0Common<Tuple>::AccDataType,
typename TestGemmD0Common<Tuple>::D0DataType,
typename TestGemmD0Common<Tuple>::EDataType,
typename TestGemmD0Common<Tuple>::ALayout,
typename TestGemmD0Common<Tuple>::BLayout,
typename TestGemmD0Common<Tuple>::D0Layout,
typename TestGemmD0Common<Tuple>::ELayout>;
}
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,

36
test/gemm_add/test_gemm_add_relu_xdl.cpp
View File

@@ -9,29 +9,21 @@
template <typename Tuple>
class TestGemmAddRelu : public TestGemmD0Common<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
using ProfileCall = typename TestGemmD0Common<Tuple>::ProfileCall;
constexpr static auto ProfileGemmAddReluImpl =
ck::profiler::profile_gemm_add_relu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddReluImpl) GetImpl() override { return ProfileGemmAddReluImpl; }
ProfileCall GetImpl() override
{
return ck::profiler::profile_gemm_add_relu_impl<
typename TestGemmD0Common<Tuple>::ADataType,
typename TestGemmD0Common<Tuple>::BDataType,
typename TestGemmD0Common<Tuple>::AccDataType,
typename TestGemmD0Common<Tuple>::D0DataType,
typename TestGemmD0Common<Tuple>::EDataType,
typename TestGemmD0Common<Tuple>::ALayout,
typename TestGemmD0Common<Tuple>::BLayout,
typename TestGemmD0Common<Tuple>::D0Layout,
typename TestGemmD0Common<Tuple>::ELayout>;
}
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,

36
test/gemm_add/test_gemm_add_silu_xdl.cpp
View File

@@ -9,29 +9,21 @@
template <typename Tuple>
class TestGemmAddSilu : public TestGemmD0Common<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
using ProfileCall = typename TestGemmD0Common<Tuple>::ProfileCall;
constexpr static auto ProfileGemmAddSiluImpl =
ck::profiler::profile_gemm_add_silu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddSiluImpl) GetImpl() override { return ProfileGemmAddSiluImpl; }
ProfileCall GetImpl() override
{
return ck::profiler::profile_gemm_add_silu_impl<
typename TestGemmD0Common<Tuple>::ADataType,
typename TestGemmD0Common<Tuple>::BDataType,
typename TestGemmD0Common<Tuple>::AccDataType,
typename TestGemmD0Common<Tuple>::D0DataType,
typename TestGemmD0Common<Tuple>::EDataType,
typename TestGemmD0Common<Tuple>::ALayout,
typename TestGemmD0Common<Tuple>::BLayout,
typename TestGemmD0Common<Tuple>::D0Layout,
typename TestGemmD0Common<Tuple>::ELayout>;
}
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,

35
test/gemm_add/test_gemm_add_xdl.hpp
View File

@@ -9,28 +9,21 @@
template <typename Tuple>
class TestGemmAdd : public TestGemmD0Common<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
using ProfileCall = typename TestGemmD0Common<Tuple>::ProfileCall;
constexpr static auto ProfileGemmAddImpl = ck::profiler::profile_gemm_add_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddImpl) GetImpl() override { return ProfileGemmAddImpl; }
ProfileCall GetImpl() override
{
return ck::profiler::profile_gemm_add_impl<
typename TestGemmD0Common<Tuple>::ADataType,
typename TestGemmD0Common<Tuple>::BDataType,
typename TestGemmD0Common<Tuple>::AccDataType,
typename TestGemmD0Common<Tuple>::D0DataType,
typename TestGemmD0Common<Tuple>::EDataType,
typename TestGemmD0Common<Tuple>::ALayout,
typename TestGemmD0Common<Tuple>::BLayout,
typename TestGemmD0Common<Tuple>::D0Layout,
typename TestGemmD0Common<Tuple>::ELayout>;
}
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,

103
test/gemm_add/test_gemm_common.hpp
View File

@@ -3,7 +3,6 @@
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_impl.hpp"
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -14,6 +13,47 @@ using F16 = ck::half_t;
using F32 = float;
using I32 = int32_t;
template <typename Tuple>
class TestGemmCommon : public ::testing::Test
{
protected:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using EDataType = std::tuple_element_t<3, Tuple>;
using ALayout = std::tuple_element_t<4, Tuple>;
using BLayout = std::tuple_element_t<5, Tuple>;
using ELayout = std::tuple_element_t<6, Tuple>;
using ProfileCall = bool(*const)(int, int, bool, bool, int, int, int, int, int, int);
virtual ProfileCall GetImpl() = 0;
void Run()
{
std::vector<std::vector<ck::index_t>> lengths = {
{16, 32, 64}, {2048, 4096, 8192}, {2048, 1024, 16}};
bool all_success = true;
for(auto length : lengths)
{
int M = length[0];
int N = length[1];
int K = length[2];
int StrideA = ck::is_same_v<ALayout, Row> ? K : M;
int StrideB = ck::is_same_v<BLayout, Row> ? N : K;
int StrideE = ck::is_same_v<ELayout, Row> ? N : M;
all_success =
all_success &
GetImpl()(1, 1, false, false, M, N, K, StrideA, StrideB, StrideE);
}
EXPECT_TRUE(all_success);
}
};
template <typename Tuple>
class TestGemmD0Common : public ::testing::Test
{
@@ -28,17 +68,9 @@ class TestGemmD0Common : public ::testing::Test
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddImpl = ck::profiler::profile_gemm_add_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
using ProfileCall = bool(*const)(int, int, bool, bool, int, int, int, int, int, int, int);
virtual decltype(ProfileGemmAddImpl) GetImpl() = 0;
virtual ProfileCall GetImpl() = 0;
void Run()
{
@@ -59,7 +91,54 @@ class TestGemmD0Common : public ::testing::Test
all_success =
all_success &
GetImpl()(true, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideE);
GetImpl()(1, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideE);
}
EXPECT_TRUE(all_success);
}
};
template <typename Tuple>
class TestGemmD0D1Common : public ::testing::Test
{
protected:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using D1DataType = std::tuple_element_t<4, Tuple>;
using EDataType = std::tuple_element_t<5, Tuple>;
using ALayout = std::tuple_element_t<6, Tuple>;
using BLayout = std::tuple_element_t<7, Tuple>;
using D0Layout = std::tuple_element_t<8, Tuple>;
using D1Layout = std::tuple_element_t<9, Tuple>;
using ELayout = std::tuple_element_t<10, Tuple>;
using ProfileCall = bool(*const)(int, int, bool, bool, int, int, int, int, int, int, int, int);
virtual ProfileCall GetImpl() = 0;
void Run()
{
std::vector<std::vector<ck::index_t>> lengths = {
{16, 32, 64}, {2048, 4096, 8192}, {2048, 1024, 16}};
bool all_success = true;
for(auto length : lengths)
{
int M = length[0];
int N = length[1];
int K = length[2];
int StrideA = ck::is_same_v<ALayout, Row> ? K : M;
int StrideB = ck::is_same_v<BLayout, Row> ? N : K;
int StrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
int StrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
int StrideE = ck::is_same_v<ELayout, Row> ? N : M;
all_success =
all_success &
GetImpl()(1, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE);
}
EXPECT_TRUE(all_success);

32
test/gemm_add/test_gemm_fastgelu_wmma.cpp Normal file
View File

@@ -0,0 +1,32 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_fastgelu_impl.hpp"
#include "test_gemm_common.hpp"
template <typename Tuple>
class TestGemmFastgelu : public TestGemmCommon<Tuple>
{
using ProfileCall = typename TestGemmCommon<Tuple>::ProfileCall;
ProfileCall GetImpl() override
{
return ck::profiler::profile_gemm_fastgelu_impl<typename TestGemmCommon<Tuple>::ADataType,
typename TestGemmCommon<Tuple>::BDataType,
typename TestGemmCommon<Tuple>::AccDataType,
typename TestGemmCommon<Tuple>::EDataType,
typename TestGemmCommon<Tuple>::ALayout,
typename TestGemmCommon<Tuple>::BLayout,
typename TestGemmCommon<Tuple>::ELayout>;
}
};
using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, F16, Row, Row, Row>,
std::tuple<F16, F16, F32, F16, Row, Col, Row>,
std::tuple<F16, F16, F32, F16, Col, Row, Row>,
std::tuple<F16, F16, F32, F16, Col, Col, Row>>;
TYPED_TEST_SUITE(TestGemmFastgelu, KernelTypes);
TYPED_TEST(TestGemmFastgelu, Test_BF16FP16) { this->Run(); }