Skip to content

GitLab

Commit 48ae4051 authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge remote-tracking branch 'origin/develop' into aosewski/gemm_tile_loop

parents 6b1738f3 d52ec016
Hide whitespace changes
Inline Side-by-side
Showing with 402 additions and 148 deletions
library/include/ck/library/tensor_operation_instance/gpu/pool2d_fwd.hpp deleted 100644 → 0
View file @ 6b1738f3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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_pool_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto InOutRank = 4;
static constexpr auto WindowRank = 2;
static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
#ifdef __fp16__
// FP16
void add_device_pool2d_fwd_nhwc_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, false>>>&);
void add_device_pool2d_fwd_nhwc_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, AvgOp, false>>>&);
// FP16 - return index
void add_device_pool2d_fwd_nhwc_index_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, true>>>&);
#endif
#ifdef __fp32__
// FP32
void add_device_pool2d_fwd_nhwc_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, false>>>&);
void add_device_pool2d_fwd_nhwc_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, AvgOp, false>>>&);
// FP32 - return index
void add_device_pool2d_fwd_nhwc_index_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, true>>>&);
#endif
template <typename InDataType,
typename OutDataType,
typename IndexDataType,
ck::ReduceTensorOp ReduceOpId,
bool OutputIndex>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>>
{
using DeviceOp = DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef __fp16__
if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
is_same_v<IndexDataType, I32>)
{
if constexpr(OutputIndex && ReduceOpId == MaxOp)
{
add_device_pool2d_fwd_nhwc_index_f16_instances(op_ptrs);
}
else
{
add_device_pool2d_fwd_nhwc_f16_instances(op_ptrs);
}
}
#endif
#ifdef __fp32__
if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
is_same_v<IndexDataType, I32>)
{
if constexpr(OutputIndex && ReduceOpId == MaxOp)
{
add_device_pool2d_fwd_nhwc_index_f32_instances(op_ptrs);
}
else
{
add_device_pool2d_fwd_nhwc_f32_instances(op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp
View file @ 48ae4051
...@@ -25,36 +25,38 @@ static constexpr auto AvgOp = ck::ReduceTensorOp::AVG; ...@@ -25,36 +25,38 @@ static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
#ifdef __fp16__ #ifdef __fp16__
// FP16 // FP16
void add_device_pool3d_fwd_ndhwc_f16_instances( void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector< std::vector<std::unique_ptr<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, false>>>&); DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NDHWC, NDHWC, MaxOp, false>>>&);
void add_device_pool3d_fwd_ndhwc_f16_instances( void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector< std::vector<std::unique_ptr<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, AvgOp, false>>>&); DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NDHWC, NDHWC, AvgOp, false>>>&);
// FP16 - return index // FP16 - return index
void add_device_pool3d_fwd_ndhwc_index_f16_instances( void add_device_pool3d_fwd_ndhwc_index_f16_instances(
std::vector< std::vector<std::unique_ptr<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, true>>>&); DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NDHWC, NDHWC, MaxOp, true>>>&);
#endif #endif
#ifdef __fp32__ #ifdef __fp32__
// FP32 // FP32
void add_device_pool3d_fwd_ndhwc_f32_instances( void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector< std::vector<std::unique_ptr<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, false>>>&); DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NDHWC, NDHWC, MaxOp, false>>>&);
void add_device_pool3d_fwd_ndhwc_f32_instances( void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector< std::vector<std::unique_ptr<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, AvgOp, false>>>&); DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NDHWC, NDHWC, AvgOp, false>>>&);
// FP32 - return index // FP32 - return index
void add_device_pool3d_fwd_ndhwc_index_f32_instances( void add_device_pool3d_fwd_ndhwc_index_f32_instances(
std::vector< std::vector<std::unique_ptr<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, true>>>&); DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NDHWC, NDHWC, MaxOp, true>>>&);
#endif #endif
template <typename InDataType, template <typename InDataType,
typename OutDataType, typename OutDataType,
typename IndexDataType, typename IndexDataType,
typename InLayout,
typename OutLayout,
ck::ReduceTensorOp ReduceOpId, ck::ReduceTensorOp ReduceOpId,
bool OutputIndex> bool OutputIndex>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank, struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank,
...@@ -62,6 +64,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFw ...@@ -62,6 +64,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFw
InDataType, InDataType,
OutDataType, OutDataType,
IndexDataType, IndexDataType,
InLayout,
OutLayout,
ReduceOpId, ReduceOpId,
OutputIndex>> OutputIndex>>
{ {
...@@ -70,40 +74,46 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFw ...@@ -70,40 +74,46 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFw
InDataType, InDataType,
OutDataType, OutDataType,
IndexDataType, IndexDataType,
InLayout,
OutLayout,
ReduceOpId, ReduceOpId,
OutputIndex>; OutputIndex>;
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef __fp16__ if constexpr(is_same_v<InLayout, NDHWC> && is_same_v<OutLayout, NDHWC>)
if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
is_same_v<IndexDataType, I32>)
{ {
if constexpr(OutputIndex && ReduceOpId == MaxOp) #ifdef __fp16__
{ if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
add_device_pool3d_fwd_ndhwc_index_f16_instances(op_ptrs); is_same_v<IndexDataType, I32>)
}
else
{ {
add_device_pool3d_fwd_ndhwc_f16_instances(op_ptrs); if constexpr(OutputIndex && ReduceOpId == MaxOp)
{
add_device_pool3d_fwd_ndhwc_index_f16_instances(op_ptrs);
}
else
{
add_device_pool3d_fwd_ndhwc_f16_instances(op_ptrs);
}
} }
}
#endif #endif
#ifdef __fp32__ #ifdef __fp32__
if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> && if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
is_same_v<IndexDataType, I32>) is_same_v<IndexDataType, I32>)
{
if constexpr(OutputIndex && ReduceOpId == MaxOp)
{ {
add_device_pool3d_fwd_ndhwc_index_f32_instances(op_ptrs); if constexpr(OutputIndex && ReduceOpId == MaxOp)
{
add_device_pool3d_fwd_ndhwc_index_f32_instances(op_ptrs);
}
else
{
add_device_pool3d_fwd_ndhwc_f32_instances(op_ptrs);
}
} }
else
{
add_device_pool3d_fwd_ndhwc_f32_instances(op_ptrs);
}
}
#endif #endif
}
return op_ptrs; return op_ptrs;
} }
}; };
......
library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
View file @ 48ae4051
...@@ -31,6 +31,10 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) ...@@ -31,6 +31,10 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp) list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp) list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp) list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instance.cpp)
endif() endif()
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp) list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp) list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp)
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instance.cpp 0 → 100644
View file @ 48ae4051
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances = std::tuple<
// clang-format off
// ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 8, 1, S<1, 8>, S<1, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<4, 1>, S<1, 1, 4, 2>, S<16, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<16, 1, 2, 2>, S<1, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 2, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<4, 1>, S<1, 1, 4, 2>, S<16, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<4, 1, 8, 2>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<8, 1>, S<1, 1, 2, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 2, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 8, 1, S<2, 8>, S<4, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 16, 2, 8, 1, 1, S<1, 1>, S<8, 8>, S<1, 1, 2, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 2, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 64, 16, 2, 4, 8, 1, S<2, 8>, S<8, 1>, S<2, 1, 4, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 1, 8, 8, S<4, 8>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 1, 8, 8, S<8, 8>, S<4, 1>, S<1, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 8, 8, S<2, 8>, S<16, 1>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(instances, device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instance.cpp 0 → 100644
View file @ 48ae4051
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances = std::tuple<
// clang-format off
// ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 8, 1, S<1, 8>, S<1, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<4, 1>, S<1, 1, 4, 2>, S<16, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<16, 1, 2, 2>, S<1, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<4, 1>, S<1, 1, 4, 2>, S<16, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<4, 1, 8, 2>, S<4, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<8, 1>, S<1, 1, 2, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 2, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 8, 1, S<2, 8>, S<4, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 16, 2, 8, 1, 1, S<1, 1>, S<8, 8>, S<1, 1, 2, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 2, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 64, 16, 2, 4, 8, 1, S<2, 8>, S<8, 1>, S<2, 1, 4, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<4, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 1, 8, 8, S<4, 8>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<4, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 1, 8, 8, S<8, 8>, S<4, 1>, S<1, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 8, 8, S<2, 8>, S<16, 1>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<4, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(instances, device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instance.cpp 0 → 100644
View file @ 48ae4051
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances = std::tuple<
// clang-format off
// ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 8, 1, S<1, 8>, S<1, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0 ,3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<4, 1>, S<4, 1, 1, 2>, S<4, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<16, 1, 2, 2>, S<1, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 2, 1>, S<0, 3, 1, 2>, S<1, 1, 2, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<4, 1>, S<4, 1, 1, 2>, S<4, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 8, 2>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<8, 1>, S<2, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<2, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 8, 1, S<2, 8>, S<4, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 16, 2, 8, 1, 1, S<1, 1>, S<8, 8>, S<2, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<2, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 64, 16, 2, 4, 8, 1, S<2, 8>, S<8, 1>, S<4, 1, 2, 2>, S<4, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 1, 8, 8, S<4, 8>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 1, 8, 8, S<8, 8>, S<4, 1>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 8, 8, S<2, 8>, S<16, 1>, S<4, 1, 2, 2>, S<4, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(instances, device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instance.cpp 0 → 100644
View file @ 48ae4051
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off
// ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 8, 1, S<1, 8>, S<1, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0 ,3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<4, 1>, S<4, 1, 1, 2>, S<4, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<16, 1, 2, 2>, S<1, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<4, 1>, S<4, 1, 1, 2>, S<4, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 8, 2>, S<4, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 16, 2, 1, 8, 1, S<1, 8>, S<8, 1>, S<2, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<2, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 8, 1, S<2, 8>, S<4, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 16, 2, 8, 1, 1, S<1, 1>, S<8, 8>, S<2, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<2, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 64, 16, 2, 4, 8, 1, S<2, 8>, S<8, 1>, S<4, 1, 2, 2>, S<4, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<4, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 1, 8, 8, S<4, 8>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<4, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 1, 8, 8, S<8, 8>, S<4, 1>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlDpp8< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 8, 8, S<2, 8>, S<16, 1>, S<4, 1, 2, 2>, S<4, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<4, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(instances, device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp
View file @ 48ae4051
...@@ -17,6 +17,8 @@ void add_device_normalization_rank_5_3_f16_instances( ...@@ -17,6 +17,8 @@ void add_device_normalization_rank_5_3_f16_instances(
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f16_generic_instance<Pass, 5, 3>{}); device_normalization_f16_generic_instance<Pass, 5, 3>{});
add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 5, 3>{}); add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 5, 3>{});
add_device_operation_instances(instances,
device_normalization_splitk_f16_instances<Pass, 5, 3>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp
View file @ 48ae4051
...@@ -17,6 +17,8 @@ void add_device_normalization_rank_5_3_f32_instances( ...@@ -17,6 +17,8 @@ void add_device_normalization_rank_5_3_f32_instances(
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f32_generic_instance<Pass, 5, 3>{}); device_normalization_f32_generic_instance<Pass, 5, 3>{});
add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 5, 3>{}); add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 5, 3>{});
add_device_operation_instances(instances,
device_normalization_splitk_f32_instances<Pass, 5, 3>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp
View file @ 48ae4051
...@@ -18,6 +18,8 @@ void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances( ...@@ -18,6 +18,8 @@ void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances(
instances, device_normalization_f16_f32_f32_f16_generic_instance<Swish, 5, 3>{}); instances, device_normalization_f16_f32_f32_f16_generic_instance<Swish, 5, 3>{});
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f16_f32_f32_f16_instances<Swish, 5, 3>{}); device_normalization_f16_f32_f32_f16_instances<Swish, 5, 3>{});
add_device_operation_instances(
instances, device_normalization_splitk_f16_f32_f32_f16_instances<Swish, 5, 3>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
View file @ 48ae4051
...@@ -17,6 +17,8 @@ void add_device_normalization_rank_5_3_swish_f16_instances( ...@@ -17,6 +17,8 @@ void add_device_normalization_rank_5_3_swish_f16_instances(
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f16_generic_instance<Swish, 5, 3>{}); device_normalization_f16_generic_instance<Swish, 5, 3>{});
add_device_operation_instances(instances, device_normalization_f16_instances<Swish, 5, 3>{}); add_device_operation_instances(instances, device_normalization_f16_instances<Swish, 5, 3>{});
add_device_operation_instances(instances,
device_normalization_splitk_f16_instances<Swish, 5, 3>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp
View file @ 48ae4051
...@@ -17,6 +17,8 @@ void add_device_normalization_rank_5_3_swish_f32_instances( ...@@ -17,6 +17,8 @@ void add_device_normalization_rank_5_3_swish_f32_instances(
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f32_generic_instance<Swish, 5, 3>{}); device_normalization_f32_generic_instance<Swish, 5, 3>{});
add_device_operation_instances(instances, device_normalization_f32_instances<Swish, 5, 3>{}); add_device_operation_instances(instances, device_normalization_f32_instances<Swish, 5, 3>{});
add_device_operation_instances(instances,
device_normalization_splitk_f32_instances<Swish, 5, 3>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp
View file @ 48ae4051
...@@ -17,6 +17,8 @@ void add_device_normalization_rank_2_1_f16_instances( ...@@ -17,6 +17,8 @@ void add_device_normalization_rank_2_1_f16_instances(
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f16_generic_instance<Pass, 2, 1>{}); device_normalization_f16_generic_instance<Pass, 2, 1>{});
add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 2, 1>{}); add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 2, 1>{});
add_device_operation_instances(instances,
device_normalization_splitk_f16_instances<Pass, 2, 1>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp
View file @ 48ae4051
...@@ -17,6 +17,8 @@ void add_device_normalization_rank_2_1_f32_instances( ...@@ -17,6 +17,8 @@ void add_device_normalization_rank_2_1_f32_instances(
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f32_generic_instance<Pass, 2, 1>{}); device_normalization_f32_generic_instance<Pass, 2, 1>{});
add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 2, 1>{}); add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 2, 1>{});
add_device_operation_instances(instances,
device_normalization_splitk_f32_instances<Pass, 2, 1>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp
View file @ 48ae4051
...@@ -17,6 +17,8 @@ void add_device_normalization_rank_4_3_f16_instances( ...@@ -17,6 +17,8 @@ void add_device_normalization_rank_4_3_f16_instances(
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f16_generic_instance<Pass, 4, 3>{}); device_normalization_f16_generic_instance<Pass, 4, 3>{});
add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 4, 3>{}); add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 4, 3>{});
add_device_operation_instances(instances,
device_normalization_splitk_f16_instances<Pass, 4, 3>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp
View file @ 48ae4051
...@@ -17,6 +17,8 @@ void add_device_normalization_rank_4_3_f32_instances( ...@@ -17,6 +17,8 @@ void add_device_normalization_rank_4_3_f32_instances(
add_device_operation_instances(instances, add_device_operation_instances(instances,
device_normalization_f32_generic_instance<Pass, 4, 3>{}); device_normalization_f32_generic_instance<Pass, 4, 3>{});
add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 4, 3>{}); add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 4, 3>{});
add_device_operation_instances(instances,
device_normalization_splitk_f32_instances<Pass, 4, 3>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
View file @ 48ae4051
...@@ -5,6 +5,7 @@ ...@@ -5,6 +5,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
#include "ck/utility/data_type.hpp" #include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
...@@ -43,6 +44,32 @@ using device_normalization_f16_instances = ...@@ -43,6 +44,32 @@ using device_normalization_f16_instances =
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f16_instances =
// clang-format off
std::tuple <
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_generic_instance = std::tuple< using device_normalization_f16_generic_instance = std::tuple<
// clang-format off // clang-format off
...@@ -76,6 +103,32 @@ using device_normalization_f32_instances = std::tuple< ...@@ -76,6 +103,32 @@ using device_normalization_f32_instances = std::tuple<
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f32_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f32_generic_instance = std::tuple< using device_normalization_f32_generic_instance = std::tuple<
// clang-format off // clang-format off
...@@ -109,6 +162,32 @@ using device_normalization_f16_f32_f32_f16_instances = std::tuple< ...@@ -109,6 +162,32 @@ using device_normalization_f16_f32_f32_f16_instances = std::tuple<
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f16_f32_f32_f16_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple< using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple<
// clang-format off // clang-format off
......
library/src/tensor_operation_instance/gpu/pool_fwd/CMakeLists.txt → library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
View file @ 48ae4051
set(DEVICE_POOL_FWD_INSTANCES) set(DEVICE_POOL3D_FWD_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_POOL_FWD_INSTANCES device_avg_pool2d_fwd_nhwc_f16_instance.cpp list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
device_max_pool2d_fwd_nhwc_f16_instance.cpp
device_max_pool3d_fwd_ndhwc_f16_instance.cpp) device_max_pool3d_fwd_ndhwc_f16_instance.cpp)
endif() endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_POOL_FWD_INSTANCES device_avg_pool2d_fwd_nhwc_f32_instance.cpp list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
device_max_pool2d_fwd_nhwc_f32_instance.cpp
device_max_pool3d_fwd_ndhwc_f32_instance.cpp) device_max_pool3d_fwd_ndhwc_f32_instance.cpp)
endif() endif()
add_instance_library(device_pool_fwd_instance ${DEVICE_POOL_FWD_INSTANCES}) add_instance_library(device_pool3d_fwd_instance ${DEVICE_POOL3D_FWD_INSTANCES})
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
View file @ 48ae4051
...@@ -11,7 +11,9 @@ namespace instance { ...@@ -11,7 +11,9 @@ namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG; static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool3d_fwd_ndhwc_f16_instances( void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, false>>>& instances) std::vector<
std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{}); instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{});
......
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
View file @ 48ae4051
...@@ -11,7 +11,9 @@ namespace instance { ...@@ -11,7 +11,9 @@ namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG; static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool3d_fwd_ndhwc_f32_instances( void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, false>>>& instances) std::vector<
std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{}); instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment