mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-26 16:04:58 +00:00
External Interface (#304)
* add client example * clean * clean * reorg * clean up profiler * reorg * clea * fix profiler * function for getinstances * update client example * update client example * update client example * update * update example * update Jenkins file * update cmake * update Jenkins
This commit is contained in:
Chao Liu
@@ -364,13 +364,8 @@ HostTensorDescriptor::HostTensorDescriptor(const std::vector<X>& lens,
|
||||
{
|
||||
}
|
||||
|
||||
void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream& os = std::cout);
|
||||
|
||||
#if 1
|
||||
// FIXME: remove
|
||||
void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst);
|
||||
#endif
|
||||
|
||||
template <typename T>
|
||||
float check_error(const Tensor<T>& ref, const Tensor<T>& result)
|
||||
{
|
||||
@@ -416,3 +411,4 @@ float check_error(const Tensor<T>& ref, const Tensor<T>& result)
|
||||
|
||||
return linf_error;
|
||||
}
|
||||
#endif
|
||||
|
||||
@@ -62,20 +62,20 @@ struct ReferenceBatchedGemm : public device::BaseOperator
|
||||
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
float v_a;
|
||||
float v_b;
|
||||
ADataType v_a;
|
||||
BDataType v_b;
|
||||
|
||||
arg.a_element_op_(v_a, static_cast<const float>(arg.a_g_m_k_(g, m, k)));
|
||||
arg.b_element_op_(v_b, static_cast<const float>(arg.b_g_k_n_(g, k, n)));
|
||||
arg.a_element_op_(v_a, arg.a_g_m_k_(g, m, k));
|
||||
arg.b_element_op_(v_b, arg.b_g_k_n_(g, k, n));
|
||||
|
||||
v_acc += v_a * v_b;
|
||||
v_acc += ck::type_convert<float>(v_a) * ck::type_convert<float>(v_b);
|
||||
}
|
||||
|
||||
float v_c;
|
||||
|
||||
arg.c_element_op_(v_c, v_acc);
|
||||
|
||||
arg.c_g_m_n_(g, m, n) = v_c;
|
||||
arg.c_g_m_n_(g, m, n) = ck::type_convert<CDataType>(v_c);
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f_gmk_gkn_gmn,
|
||||
|
||||
@@ -63,20 +63,21 @@ struct ReferenceGemm : public device::BaseOperator
|
||||
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
AccDataType v_a;
|
||||
AccDataType v_b;
|
||||
ADataType v_a;
|
||||
BDataType v_b;
|
||||
|
||||
arg.a_element_op_(v_a, static_cast<const AccDataType>(arg.a_m_k_(m, k)));
|
||||
arg.b_element_op_(v_b, static_cast<const AccDataType>(arg.b_k_n_(k, n)));
|
||||
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
|
||||
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
|
||||
|
||||
v_acc += v_a * v_b;
|
||||
v_acc +=
|
||||
ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
|
||||
}
|
||||
|
||||
AccDataType v_c;
|
||||
|
||||
arg.c_element_op_(v_c, v_acc);
|
||||
|
||||
arg.c_m_n_(m, n) = v_c;
|
||||
arg.c_m_n_(m, n) = ck::type_convert<CDataType>(v_c);
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(
|
||||
|
||||
@@ -0,0 +1,203 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace device_batched_gemm_instance {
|
||||
|
||||
using DeviceBatchedGemmNoOpPtr = ck::tensor_operation::device::DeviceBatchedGemmPtr<
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough>;
|
||||
|
||||
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(
|
||||
std::vector<DeviceBatchedGemmNoOpPtr>&);
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
auto get_device_batched_gemm_instances()
|
||||
{
|
||||
std::vector<DeviceBatchedGemmNoOpPtr> op_ptrs;
|
||||
|
||||
if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
|
||||
is_same<CDataType, float>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
|
||||
is_same<CDataType, half_t>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same<ADataType, bhalf_t>::value && is_same<BDataType, bhalf_t>::value &&
|
||||
is_same<CDataType, bhalf_t>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
|
||||
is_same<CDataType, int8_t>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_batched_gemm_instance::
|
||||
add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
return op_ptrs;
|
||||
}
|
||||
|
||||
} // namespace device_batched_gemm_instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,93 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace device_gemm_instance {
|
||||
|
||||
using DeviceGemmAddAddFastGeluPtr = ck::tensor_operation::device::DeviceGemmMultipleDPtr<
|
||||
2,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::AddAddFastGelu>;
|
||||
|
||||
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<DeviceGemmAddAddFastGeluPtr>&);
|
||||
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
|
||||
std::vector<DeviceGemmAddAddFastGeluPtr>&);
|
||||
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
|
||||
std::vector<DeviceGemmAddAddFastGeluPtr>&);
|
||||
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
|
||||
std::vector<DeviceGemmAddAddFastGeluPtr>&);
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename AccDataType,
|
||||
typename D0DataType,
|
||||
typename D1DataType,
|
||||
typename EDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename D0Layout,
|
||||
typename D1Layout,
|
||||
typename ELayout>
|
||||
auto get_device_gemm_add_add_fastgelu_instances()
|
||||
{
|
||||
std::vector<DeviceGemmAddAddFastGeluPtr> op_ptrs;
|
||||
|
||||
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, tensor_layout::gemm::RowMajor> &&
|
||||
is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
|
||||
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
|
||||
is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
|
||||
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
|
||||
is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
|
||||
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
|
||||
is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
|
||||
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
return op_ptrs;
|
||||
}
|
||||
|
||||
} // namespace device_gemm_instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,286 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace device_gemm_instance {
|
||||
|
||||
using DeviceGemmNoOpPtr =
|
||||
ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough>;
|
||||
|
||||
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
|
||||
std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
|
||||
std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
|
||||
std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
|
||||
std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
|
||||
std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
auto get_device_gemm_instances()
|
||||
{
|
||||
std::vector<DeviceGemmNoOpPtr> op_ptrs;
|
||||
|
||||
if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
|
||||
is_same<CDataType, float>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
|
||||
is_same<CDataType, half_t>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
|
||||
is_same<BDataType, ck::bhalf_t>::value &&
|
||||
is_same<CDataType, ck::bhalf_t>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
|
||||
is_same<CDataType, int8_t>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(op_ptrs);
|
||||
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
return op_ptrs;
|
||||
}
|
||||
|
||||
} // namespace device_gemm_instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,124 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace device_gemm_instance {
|
||||
|
||||
using DeviceGemmSplitKNoOpPtr = ck::tensor_operation::device::DeviceGemmSplitKPtr<
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough>;
|
||||
|
||||
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(
|
||||
std::vector<DeviceGemmSplitKNoOpPtr>&);
|
||||
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
|
||||
std::vector<DeviceGemmSplitKNoOpPtr>&);
|
||||
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(
|
||||
std::vector<DeviceGemmSplitKNoOpPtr>&);
|
||||
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(
|
||||
std::vector<DeviceGemmSplitKNoOpPtr>&);
|
||||
|
||||
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<DeviceGemmSplitKNoOpPtr>&);
|
||||
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
|
||||
std::vector<DeviceGemmSplitKNoOpPtr>&);
|
||||
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(
|
||||
std::vector<DeviceGemmSplitKNoOpPtr>&);
|
||||
void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(
|
||||
std::vector<DeviceGemmSplitKNoOpPtr>&);
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
auto get_device_gemm_splitk_instances()
|
||||
{
|
||||
std::vector<DeviceGemmSplitKNoOpPtr> op_ptrs;
|
||||
|
||||
if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
|
||||
is_same<CDataType, float>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
|
||||
is_same<CDataType, half_t>::value)
|
||||
{
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
||||
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
ck::tensor_operation::device::device_gemm_instance::
|
||||
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
return op_ptrs;
|
||||
}
|
||||
|
||||
} // namespace device_gemm_instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user