Skip to content

GitLab

"Src/YOLOV9.cpp" did not exist on "a63911bb575bea84b31c8cc7fc8dcc750e80e470"
Commit c5138aa1 authored by Artur Wojcik's avatar Artur Wojcik
Browse files

Merge branch 'develop' into uif2-initial

parents 7830272f 82f3a835
Show whitespace changes
Inline Side-by-side
Showing with 444 additions and 306 deletions
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
View file @ c5138aa1
add_instance_library(device_grouped_conv3d_fwd_instance
set(GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp
)
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
list(APPEND GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp)
endif()
add_instance_library(device_grouped_conv3d_fwd_instance ${GROUPED_CONV3D_FWD})
library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt
View file @ c5138aa1
set(GROUPED_GEMM_FIXED_NK_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp)
endif()
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp)
endif()
if((DTYPES MATCHES "int8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp)
endif()
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp)
add_instance_library(device_grouped_gemm_fixed_nk_instance ${GROUPED_GEMM_FIXED_NK_INSTANCES})
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp
View file @ c5138aa1
......@@ -11,7 +11,7 @@ namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_5_3_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 5, 3>>>&
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 5, 3>>>&
instances)
{
add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp
View file @ c5138aa1
......@@ -11,7 +11,7 @@ namespace instance {
using Swish = ck::tensor_operation::element_wise::Swish;
void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F32, F16, Swish, 5, 3>>>&
std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F16, F32, Swish, 5, 3>>>&
instances)
{
add_device_operation_instances(
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
View file @ c5138aa1
......@@ -11,7 +11,7 @@ namespace instance {
using Swish = ck::tensor_operation::element_wise::Swish;
void add_device_normalization_rank_5_3_swish_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Swish, 5, 3>>>&
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Swish, 5, 3>>>&
instances)
{
add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp
View file @ c5138aa1
......@@ -11,7 +11,7 @@ namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_2_1_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 2, 1>>>&
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 2, 1>>>&
instances)
{
add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp
View file @ c5138aa1
......@@ -11,7 +11,7 @@ namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_4_3_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 4, 3>>>&
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 4, 3>>>&
instances)
{
add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
View file @ c5138aa1
......@@ -22,25 +22,25 @@ template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_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>
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
// clang-format on
>;
......@@ -48,150 +48,150 @@ 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>
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_generic_instance = std::tuple<
// clang-format off
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f32_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
// 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>
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f32_generic_instance = std::tuple<
// clang-format off
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_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>
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
// 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>
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple<
// clang-format off
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
// clang-format on
>;
......
profiler/README.md
View file @ c5138aa1
......@@ -22,7 +22,7 @@ c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
```
## Profile 2d forward convolution kernels
## Profile 2D forward convolution kernels
```bash
#arg1: tensor operation (conv=Convolution)
#arg2: data type (0=fp32, 1=fp16)
......@@ -115,7 +115,7 @@ Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
# Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
......@@ -158,7 +158,7 @@ GB/s: 127.947
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
# Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
......@@ -201,7 +201,7 @@ Note: This kernel use atomic add, this will cause output buffer to be accumulate
# arg7: time kernel (0: no, 1: yes)
# arg8: operation type (0: ImageToColumn, 1: ColumnToImage)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
# Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
......
profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
View file @ c5138aa1
......@@ -80,6 +80,8 @@ bool profile_elementwise_layernorm_impl(int do_verification,
Tensor<BetaDataType> beta(gammaBetaLength);
Tensor<YDataType> y(length);
Tensor<YDataType> host_y(length);
Tensor<AccDataType> host_save_mean({M});
Tensor<AccDataType> host_save_inv_std({M});
switch(init_method)
{
......@@ -152,13 +154,22 @@ bool profile_elementwise_layernorm_impl(int do_verification,
BetaDataType,
YDataType,
AccDataType,
AccDataType,
PassThrough,
Rank,
NumReduceDim>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, {M, N}, {1}, 1e-4);
auto ref_argument = ref.MakeArgument(x,
gamma,
beta,
host_y,
host_save_mean,
host_save_inv_std,
PassThrough{},
{M, N},
{1},
1e-4);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
......
profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
View file @ c5138aa1
......@@ -66,12 +66,15 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
BetaDataType,
HDataType,
AccDataType,
AccDataType,
HElementOp,
2,
1>;
Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> save_mean({M});
Tensor<AccDataType> save_inv_std({M});
auto ref_gemm = ReferenceGemm{};
auto ref_gemm_invoker = ref_gemm.MakeInvoker();
......@@ -97,7 +100,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
auto ref_layernorm_argument = ref_layernorm.MakeArgument(
e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
e_m_n, gamma_n, beta_n, h_m_n, save_mean, save_inv_std, h_element_op, {M, N}, {1}, epsilon);
ref_layernorm_invoker.Run(ref_layernorm_argument);
}
......
profiler/include/profiler/profile_groupnorm_impl.hpp
View file @ c5138aa1
......@@ -21,8 +21,10 @@ namespace profiler {
template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename AccDataType,
typename YDataType>
typename ComputeDataType,
typename YDataType,
typename SaveMeanInvStdDataType,
bool SaveMeanInvStd>
bool profile_groupnorm_impl(int do_verification,
int init_method,
bool do_log,
......@@ -34,6 +36,7 @@ bool profile_groupnorm_impl(int do_verification,
if(length.size() != 5)
return false;
index_t N = length[0];
index_t G = length[3];
index_t C = length[4];
......@@ -45,7 +48,14 @@ bool profile_groupnorm_impl(int do_verification,
Tensor<GammaDataType> gamma(gammaBetaLength);
Tensor<BetaDataType> beta(gammaBetaLength);
Tensor<YDataType> y(length);
Tensor<SaveMeanInvStdDataType> save_mean({N, G});
Tensor<SaveMeanInvStdDataType> save_inv_std({N, G});
Tensor<YDataType> host_y(length);
Tensor<SaveMeanInvStdDataType> host_save_mean({N, G});
Tensor<SaveMeanInvStdDataType> host_save_inv_std({N, G});
std::vector<index_t> strideSaveMeanInvStd = {1};
switch(init_method)
{
......@@ -69,6 +79,9 @@ bool profile_groupnorm_impl(int do_verification,
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
DeviceMem save_mean_dev(sizeof(SaveMeanInvStdDataType) * save_mean.mDesc.GetElementSpaceSize());
DeviceMem save_inv_std_dev(sizeof(SaveMeanInvStdDataType) *
save_inv_std.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
......@@ -78,8 +91,8 @@ bool profile_groupnorm_impl(int do_verification,
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
5,
3>;
......@@ -97,29 +110,56 @@ bool profile_groupnorm_impl(int do_verification,
if(do_verification)
{
using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
using ReferenceInstance =
ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
SaveMeanInvStdDataType,
ComputeDataType,
PassThrough>;
ReferenceInstance ref;
auto ref_argument = ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, 1e-6);
auto ref_argument = ref.MakeArgument(
x, gamma, beta, host_y, host_save_mean, host_save_inv_std, PassThrough{}, length, 1e-6);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(
auto f_get_argument = [&](auto& inst_ptr) {
if constexpr(SaveMeanInvStd)
return inst_ptr->MakeArgumentPointer(
length,
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
gammaBetaStride,
gammaBetaStride,
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_inv_std.mDesc.GetStrides().begin(),
save_inv_std.mDesc.GetStrides().end()},
reduce_dim,
1e-6,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
save_mean_dev.GetDeviceBuffer(),
save_inv_std_dev.GetDeviceBuffer(),
PassThrough{});
else
return inst_ptr->MakeArgumentPointer(
length,
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
gammaBetaStride,
gammaBetaStride,
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_inv_std.mDesc.GetStrides().begin(),
save_inv_std.mDesc.GetStrides().end()},
reduce_dim,
1e-6,
x_dev.GetDeviceBuffer(),
......@@ -129,6 +169,11 @@ bool profile_groupnorm_impl(int do_verification,
nullptr,
nullptr,
PassThrough{});
};
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = f_get_argument(inst_ptr);
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
......@@ -152,6 +197,10 @@ bool profile_groupnorm_impl(int do_verification,
beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
y.mDesc.GetElementSize() * sizeof(YDataType);
if constexpr(SaveMeanInvStd)
num_bytes += save_mean.mDesc.GetElementSpaceSize() * sizeof(SaveMeanInvStdDataType) +
save_inv_std.mDesc.GetElementSpaceSize() * sizeof(SaveMeanInvStdDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
......@@ -168,9 +217,22 @@ bool profile_groupnorm_impl(int do_verification,
if(do_verification)
{
y_dev.FromDevice(y.mData.data());
bool pass = ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
if constexpr(SaveMeanInvStd)
{
save_mean_dev.FromDevice(save_mean.mData.data());
pass &= ck::utils::check_err(
save_mean.mData, host_save_mean.mData, "Error: Incorrect results", 1e-3, 1e-3);
save_inv_std_dev.FromDevice(save_inv_std.mData.data());
pass &= ck::utils::check_err(save_inv_std.mData,
host_save_inv_std.mData,
"Error: Incorrect results",
1e-3,
1e-3);
}
if(do_log)
{
LogRangeAsType<float>(std::cout << "x : ", x.mData, ",") << std::endl;
......
profiler/include/profiler/profile_layernorm_impl.hpp
View file @ c5138aa1
......@@ -21,6 +21,8 @@ template <typename XDataType,
typename BetaDataType,
typename ComputeDataType,
typename YDataType,
typename SaveMeanInvStdDataType,
bool SaveMeanInvStd,
index_t Rank>
bool profile_layernorm_impl(int do_verification,
int init_method,
......@@ -43,13 +45,19 @@ bool profile_layernorm_impl(int do_verification,
Tensor<GammaDataType> gamma(reduce_length);
Tensor<BetaDataType> beta(reduce_length);
Tensor<YDataType> y(length);
Tensor<SaveMeanInvStdDataType> save_mean({length[0]});
Tensor<SaveMeanInvStdDataType> save_inv_std({length[0]});
Tensor<YDataType> host_y(length);
Tensor<SaveMeanInvStdDataType> host_save_mean({length[0]});
Tensor<SaveMeanInvStdDataType> host_save_inv_std({length[0]});
std::vector<index_t> strideXY =
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()};
std::vector<index_t> strideGammaBeta = strideXY;
strideGammaBeta[0] = 0;
std::vector<index_t> strideSaveMeanInvStd = {1};
switch(init_method)
{
case 0:
......@@ -75,6 +83,9 @@ bool profile_layernorm_impl(int do_verification,
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
DeviceMem save_mean_dev(sizeof(SaveMeanInvStdDataType) * save_mean.mDesc.GetElementSpaceSize());
DeviceMem save_inv_std_dev(sizeof(SaveMeanInvStdDataType) *
save_inv_std.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
......@@ -86,8 +97,8 @@ bool profile_layernorm_impl(int do_verification,
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim>;
......@@ -105,31 +116,60 @@ bool profile_layernorm_impl(int do_verification,
if(do_verification)
{
using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
ComputeDataType,
PassThrough,
Rank,
NumReduceDim>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, reduce_dim, 1e-4);
auto ref_argument = ref.MakeArgument(x,
gamma,
beta,
host_y,
host_save_mean,
host_save_inv_std,
PassThrough{},
length,
reduce_dim,
1e-4);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
auto f_get_argument = [&](auto& inst_ptr) {
if constexpr(SaveMeanInvStd)
return inst_ptr->MakeArgumentPointer(length,
strideXY,
strideGammaBeta,
strideGammaBeta,
strideXY,
strideSaveMeanInvStd,
strideSaveMeanInvStd,
reduce_dim,
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
save_mean_dev.GetDeviceBuffer(),
save_inv_std_dev.GetDeviceBuffer(),
PassThrough{});
else
return inst_ptr->MakeArgumentPointer(length,
strideXY,
strideGammaBeta,
strideGammaBeta,
strideXY,
strideSaveMeanInvStd,
strideSaveMeanInvStd,
reduce_dim,
1e-4,
x_dev.GetDeviceBuffer(),
......@@ -139,6 +179,11 @@ bool profile_layernorm_impl(int do_verification,
nullptr,
nullptr,
PassThrough{});
};
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = f_get_argument(inst_ptr);
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
......@@ -168,6 +213,10 @@ bool profile_layernorm_impl(int do_verification,
beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
y.mDesc.GetElementSize() * sizeof(YDataType);
if constexpr(SaveMeanInvStd)
num_bytes += save_mean.mDesc.GetElementSpaceSize() * sizeof(SaveMeanInvStdDataType) +
save_inv_std.mDesc.GetElementSpaceSize() * sizeof(SaveMeanInvStdDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
......@@ -184,10 +233,23 @@ bool profile_layernorm_impl(int do_verification,
if(do_verification)
{
y_dev.FromDevice(y.mData.data());
bool pass =
ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
if constexpr(SaveMeanInvStd)
{
save_mean_dev.FromDevice(save_mean.mData.data());
pass &= ck::utils::check_err(
save_mean.mData, host_save_mean.mData, "Error: Incorrect results", 1e-3, 1e-3);
save_inv_std_dev.FromDevice(save_inv_std.mData.data());
pass &= ck::utils::check_err(save_inv_std.mData,
host_save_inv_std.mData,
"Error: Incorrect results",
1e-3,
1e-3);
}
if(do_log)
{
LogRangeAsType<float>(std::cout << "x : ", x.mData, ",") << std::endl;
......
profiler/src/CMakeLists.txt
View file @ c5138aa1
......@@ -25,8 +25,6 @@ set(PROFILER_SOURCES
profile_batchnorm_fwd.cpp
profile_batchnorm_bwd.cpp
profile_batchnorm_infer.cpp
profile_contraction_bilinear.cpp
profile_contraction_scale.cpp
profile_grouped_conv_bwd_data.cpp
profile_conv_tensor_rearrange.cpp
)
......@@ -46,6 +44,11 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
endif()
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
endif()
set(PROFILER_EXECUTABLE ckProfiler)
add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
......@@ -76,8 +79,6 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instan
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
......@@ -85,9 +86,18 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_d
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
endif()
if(DL_KERNELS)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
......
profiler/src/profile_grouped_conv_bwd_weight.cpp
View file @ c5138aa1
......@@ -86,12 +86,8 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
#ifdef CK_ENABLE_BF8
using BF8 = ck::bf8_t;
#endif
using namespace ck::tensor_layout::convolution;
......@@ -141,59 +137,59 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
{
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
......@@ -204,22 +200,22 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, int8_t{}, int8_t{}, int8_t{}, int8_t{}, int8_t{});
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F16_F16_F16_BF8_F8)
if(data_type == ConvDataType::F16_F16_F16_BF8_F8)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{}, BF8{}, F8{});
}
......
profiler/src/profile_groupnorm.cpp
View file @ c5138aa1
......@@ -93,12 +93,12 @@ int profile_groupnorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32>(
ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32, F32, false>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16>(
ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16, F32, false>(
do_verification, init_method, do_log, time_kernel, length);
}
else
......
profiler/src/profile_layernorm.cpp
View file @ c5138aa1
......@@ -82,12 +82,12 @@ int profile_layernorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, rank>(
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F32, false, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, rank>(
ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, F32, false, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else
......
test/contraction/CMakeLists.txt
View file @ c5138aa1
add_gtest_executable(test_contraction test_contraction.cpp)
target_link_libraries(test_contraction PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
if((DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64") OR NOT DEFINED DTYPES)
add_gtest_executable(test_contraction test_contraction.cpp)
target_link_libraries(test_contraction PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
add_gtest_executable(test_contraction_interface test_contraction_interface.cpp)
target_link_libraries(test_contraction_interface PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
set(target 1)
endif()
endif()
endforeach()
test/conv_tensor_rearrange/CMakeLists.txt
View file @ c5138aa1
add_gtest_executable(test_conv_tensor_rearrange test_conv_tensor_rearrange.cpp)
target_link_libraries(test_conv_tensor_rearrange PRIVATE utility device_image_to_column_instance device_column_to_image_instance)
add_gtest_executable(test_conv_tensor_rearrange_interface test_conv_tensor_rearrange_interface.cpp)
target_link_libraries(test_conv_tensor_rearrange_interface PRIVATE utility)
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