Skip to content

GitLab

Unverified Commit 9f8ab221 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Merge branch 'develop' into add_int8_wmma_example_instance

parents 755ace59 b4fc4d0b
Hide whitespace changes
Inline Side-by-side
Showing with 811 additions and 285 deletions
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
View file @ 9f8ab221
......@@ -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 @ 9f8ab221
......@@ -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 @ 9f8ab221
......@@ -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 @ 9f8ab221
......@@ -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
>;
......
library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
View file @ 9f8ab221
set(DEVICE_POOL3D_FWD_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
device_max_pool3d_fwd_ndhwc_f16_instance.cpp)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
device_max_pool3d_fwd_ndhwc_bf16_instance.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
device_max_pool3d_fwd_ndhwc_f32_instance.cpp)
endif()
list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
device_max_pool3d_fwd_ndhwc_f16_instance.cpp
device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
device_max_pool3d_fwd_ndhwc_f32_instance.cpp
device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
device_max_pool3d_fwd_ndhwc_bf16_instance.cpp)
add_instance_library(device_pool3d_fwd_instance ${DEVICE_POOL3D_FWD_INSTANCES})
library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
View file @ 9f8ab221
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
set(CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp)
set(CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp)
set(CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp)
......@@ -10,17 +8,16 @@ set(GEMM_QUANT_SRC
gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
)
if(DL_KERNELS)
list(APPEND CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp)
list(APPEND CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp)
list(APPEND CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp)
list(APPEND CONV2D_BIAS_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp)
list(APPEND GEMM_QUANT_SRC
gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp)
endif()
list(APPEND CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp)
list(APPEND CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp)
list(APPEND CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp)
list(APPEND CONV2D_BIAS_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp)
list(APPEND GEMM_QUANT_SRC
gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp)
add_instance_library(device_quantization_instance
${CONV2D_PERLAYER_QUANT_SRC}
......@@ -29,4 +26,3 @@ add_instance_library(device_quantization_instance
${CONV2D_BIAS_PERCHANNEL_QUANT_SRC}
${GEMM_QUANT_SRC}
)
endif()
\ No newline at end of file
library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
View file @ 9f8ab221
set(DEVICE_SOFTMAX_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f16_f16_instance_rank3_reduce1.cpp
list(APPEND DEVICE_SOFTMAX_INSTANCES
device_softmax_f16_f16_instance_rank3_reduce1.cpp
device_softmax_f16_f16_instance_rank3_reduce2.cpp
device_softmax_f16_f16_instance_rank3_reduce3.cpp
device_softmax_f16_f16_instance_rank4_reduce1.cpp
device_softmax_f16_f16_instance_rank4_reduce2.cpp
device_softmax_f16_f16_instance_rank4_reduce3.cpp
device_softmax_f16_f16_instance_rank4_reduce4.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f32_f32_instance_rank3_reduce1.cpp
device_softmax_f16_f16_instance_rank4_reduce4.cpp
device_softmax_f32_f32_instance_rank3_reduce1.cpp
device_softmax_f32_f32_instance_rank3_reduce2.cpp
device_softmax_f32_f32_instance_rank3_reduce3.cpp
device_softmax_f32_f32_instance_rank4_reduce1.cpp
device_softmax_f32_f32_instance_rank4_reduce2.cpp
device_softmax_f32_f32_instance_rank4_reduce3.cpp
device_softmax_f32_f32_instance_rank4_reduce4.cpp)
endif()
add_instance_library(device_softmax_instance ${DEVICE_SOFTMAX_INSTANCES})
profiler/README.md
View file @ 9f8ab221
......@@ -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)
......@@ -147,7 +147,9 @@ GB/s: 127.947
# arg1: tensor operation (grouped_conv_bwd_weight: Grouped Convolution Backward Weight)
# arg2: data type (0: Input fp32, Weight fp32, Output fp32
# 1: Input fp16, Weight fp16, Output fp16
# 2: Input bf16, Weight fp32, Output bf16)
# 2: Input bf16, Weight fp32, Output bf16
# 3: Input fp16, Weight fp16, Output fp16, Gemm bf8@fp8
# 4: Input int8, Weight int8, Output int8)
# arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, N, K, Ho, Wo]
# 1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]
# 2: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K]
......@@ -156,7 +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)
......@@ -167,7 +169,7 @@ GB/s: 127.947
# SplitK
################ op datatype layout verify init log time Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx SplitK
./bin/ckProfiler grouped_conv_bwd_weight 1 0 1 1 0 1 2 32 256 256 512 3 3 28 28 1 1 1 1 1 0 0 0 1
./bin/ckProfiler grouped_conv_bwd_weight 1 1 0 1 0 1 2 32 256 256 512 3 3 28 28 1 1 1 1 1 0 0 0 1
```
......@@ -185,7 +187,7 @@ GB/s: 69.2301
```
Note: This kernel use atomic add, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.
## Profile image to column kernels
## Profile image to column/column to image kernels
```bash
# arg1: tensor operation (" OP_NAME ": " OP_DESC ")
# arg2: data type (0: Input fp32, Weight fp32, Output fp32
......@@ -197,8 +199,9 @@ Note: This kernel use atomic add, this will cause output buffer to be accumulate
# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# 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)
......@@ -207,8 +210,8 @@ Note: This kernel use atomic add, this will cause output buffer to be accumulate
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
################ op datatype layout verify init log time Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx
./bin/ckProfiler image_to_column 0 0 1 1 0 1 2 1 256 1 512 3 3 28 28 1 1 1 1 0 0 0 0
################ op datatype layout verify init log time opType Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx
./bin/ckProfiler conv_tensor_rearrange 0 0 0 1 0 1 0 2 1 256 1 512 3 3 28 28 1 1 1 1 0 0 0 0
```
......@@ -222,3 +225,4 @@ name: DeviceImageToColumn<128, 32, 64, 4>
avg_time: 3.12326
GB/s: 2042.59
```
Note: Column to image kernel adds to the output memory, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.
profiler/include/profiler/profile_image_to_column_impl.hpp → profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp
View file @ 9f8ab221
......@@ -9,9 +9,11 @@
#include <limits>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp"
#include "ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/image_to_column.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
......@@ -19,22 +21,88 @@
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp"
namespace ck {
namespace profiler {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using namespace conv_tensor_rearrange_op;
template <typename InputDataType, typename ConvTensorRearrangeOp>
Tensor<InputDataType> create_input(const HostTensorDescriptor& image_desc,
const HostTensorDescriptor& gemm_desc)
{
if constexpr(std::is_same_v<ConvTensorRearrangeOp, ImageToColumn>)
{
Tensor<InputDataType> input(image_desc);
return input;
}
else if constexpr(std::is_same_v<ConvTensorRearrangeOp, ColumnToImage>)
{
Tensor<InputDataType> input(gemm_desc);
return input;
}
else
{
throw std::runtime_error("Unsupported op!");
}
}
template <typename OutputDataType, typename ConvTensorRearrangeOp>
Tensor<OutputDataType> create_output(const HostTensorDescriptor& image_desc,
const HostTensorDescriptor& gemm_desc)
{
if constexpr(std::is_same_v<ConvTensorRearrangeOp, ImageToColumn>)
{
Tensor<OutputDataType> output(gemm_desc);
return output;
}
else if constexpr(std::is_same_v<ConvTensorRearrangeOp, ColumnToImage>)
{
Tensor<OutputDataType> output(image_desc);
return output;
}
else
{
throw std::runtime_error("Unsupported op!");
}
}
template <index_t NDimSpatial,
typename InputLayout,
typename InputDataType,
typename OutputDataType,
typename ConvTensorRearrangeOp>
static auto make_ref_op()
{
if constexpr(std::is_same_v<ConvTensorRearrangeOp, ImageToColumn>)
{
return ck::tensor_operation::host::
ReferenceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>{};
}
else if constexpr(std::is_same_v<ConvTensorRearrangeOp, ColumnToImage>)
{
return ck::tensor_operation::host::
ReferenceColumnToImage<NDimSpatial, InputLayout, InputDataType, OutputDataType>{};
}
else
{
throw std::runtime_error("Unsupported op!");
}
}
template <index_t NDimSpatial,
typename InputLayout,
typename InputDataType,
typename OutputDataType>
bool profile_image_to_column_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param)
typename OutputDataType,
typename ConvTensorRearrangeOp>
bool profile_conv_tensor_rearrange_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param)
{
const ck::index_t NDoHoWo =
conv_param.N_ *
......@@ -45,16 +113,16 @@ bool profile_image_to_column_impl(int do_verification,
ck::accumulate_n<ck::index_t>(
conv_param.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
const auto in_desc =
const auto image_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InputLayout>(
conv_param);
const auto out_desc = HostTensorDescriptor({NDoHoWo, CZYX});
const auto gemm_desc = HostTensorDescriptor({NDoHoWo, CZYX});
std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
std::array<ck::index_t, NDimSpatial + 3> input_g_n_c_wis_strides{};
std::array<ck::index_t, 2> output_m_k_strides{};
std::array<ck::index_t, NDimSpatial + 3> image_g_n_c_wis_strides{};
std::array<ck::index_t, 2> gemm_m_k_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
......@@ -65,16 +133,19 @@ bool profile_image_to_column_impl(int do_verification,
copy(conv_param.input_spatial_lengths_, input_spatial_lengths);
copy(conv_param.filter_spatial_lengths_, filter_spatial_lengths);
copy(conv_param.output_spatial_lengths_, output_spatial_lengths);
copy(in_desc.GetStrides(), input_g_n_c_wis_strides);
copy(out_desc.GetStrides(), output_m_k_strides);
copy(image_desc.GetStrides(), image_g_n_c_wis_strides);
copy(gemm_desc.GetStrides(), gemm_m_k_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads);
Tensor<InputDataType> input(in_desc);
Tensor<OutputDataType> host_output(out_desc);
Tensor<OutputDataType> device_output(out_desc);
Tensor<InputDataType> input =
create_input<InputDataType, ConvTensorRearrangeOp>(image_desc, gemm_desc);
Tensor<OutputDataType> device_output =
create_output<OutputDataType, ConvTensorRearrangeOp>(image_desc, gemm_desc);
Tensor<OutputDataType> host_output =
create_output<OutputDataType, ConvTensorRearrangeOp>(image_desc, gemm_desc);
std::cout << "input: " << input.mDesc << std::endl;
std::cout << "output: " << host_output.mDesc << std::endl;
......@@ -94,17 +165,21 @@ bool profile_image_to_column_impl(int do_verification,
// run reference op
if(do_verification)
{
auto ref_image_to_column = ck::tensor_operation::host::
ReferenceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>{};
auto ref_conv_tensor_rearrange = make_ref_op<NDimSpatial,
InputLayout,
InputDataType,
OutputDataType,
ConvTensorRearrangeOp>();
auto ref_invoker = ref_image_to_column.MakeInvoker();
auto ref_argument = ref_image_to_column.MakeArgument(input,
host_output,
conv_param.filter_spatial_lengths_,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_);
auto ref_invoker = ref_conv_tensor_rearrange.MakeInvoker();
auto ref_argument =
ref_conv_tensor_rearrange.MakeArgument(input,
host_output,
conv_param.filter_spatial_lengths_,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_);
// init host output to zero
host_output.SetZero();
......@@ -112,8 +187,11 @@ bool profile_image_to_column_impl(int do_verification,
ref_invoker.Run(ref_argument);
}
using DeviceOp = ck::tensor_operation::device::
DeviceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>;
using DeviceOp = ck::tensor_operation::device::DeviceConvTensorRearrange<NDimSpatial,
InputLayout,
InputDataType,
OutputDataType,
ConvTensorRearrangeOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......@@ -139,8 +217,8 @@ bool profile_image_to_column_impl(int do_verification,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_g_n_c_wis_strides,
output_m_k_strides,
image_g_n_c_wis_strides,
gemm_m_k_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......
profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
View file @ 9f8ab221
......@@ -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)
{
......@@ -152,14 +154,23 @@ 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_invoker = ref.MakeInvoker();
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 @ 9f8ab221
......@@ -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_gemm_impl.hpp
View file @ 9f8ab221
......@@ -223,6 +223,12 @@ int profile_gemm_impl(int do_verification,
{
std::cout << "Best Perf for datatype = int8";
}
#if defined CK_ENABLE_FP8
else if constexpr(is_same<CDataType, f8_t>::value)
{
std::cout << "Best Perf for datatype = fp8";
}
#endif
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
{
......
profiler/include/profiler/profile_gemm_splitk_impl.hpp
View file @ 9f8ab221
......@@ -30,7 +30,8 @@ template <typename ADataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
typename CLayout,
typename ComputeType = CDataType>
bool profile_gemm_splitk_impl(int do_verification,
int init_method,
bool do_log,
......@@ -103,7 +104,8 @@ bool profile_gemm_splitk_impl(int do_verification,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
CElementOp,
ComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......@@ -120,7 +122,8 @@ bool profile_gemm_splitk_impl(int do_verification,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
CElementOp,
ComputeType>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......
profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
View file @ 9f8ab221
......@@ -33,7 +33,9 @@ template <ck::index_t NDimSpatial,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename OutDataType>
typename OutDataType,
typename ComputeTypeA = InDataType,
typename ComputeTypeB = ComputeTypeA>
bool profile_grouped_conv_bwd_weight_impl(int do_verification,
int init_method,
bool do_log,
......@@ -120,7 +122,9 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
OutElementOp,
ComputeTypeA,
ComputeTypeB>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
profiler/include/profiler/profile_groupnorm_impl.hpp
View file @ 9f8ab221
......@@ -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>;
......@@ -97,38 +110,70 @@ bool profile_groupnorm_impl(int do_verification,
if(do_verification)
{
using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
PassThrough>;
using ReferenceInstance =
ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
ComputeDataType,
PassThrough>;
ReferenceInstance ref;
auto ref_argument = ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, 1e-6);
auto ref_invoker = ref.MakeInvoker();
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;
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(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{});
};
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = 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()},
reduce_dim,
1e-6,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{});
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 @ 9f8ab221
......@@ -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>;
......@@ -105,40 +116,74 @@ bool profile_layernorm_impl(int do_verification,
if(do_verification)
{
using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
PassThrough,
Rank,
NumReduceDim>;
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_invoker = ref.MakeInvoker();
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;
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(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{});
};
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
strideXY,
strideGammaBeta,
strideGammaBeta,
strideXY,
reduce_dim,
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{});
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 @ 9f8ab221
......@@ -25,10 +25,8 @@ 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_image_to_column.cpp
profile_conv_tensor_rearrange.cpp
)
if(DL_KERNELS)
list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.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} ${PROFILER_SOURCES})
......@@ -76,17 +79,25 @@ target_link_libraries(${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)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
target_link_libraries(${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_image_to_column.cpp → profiler/src/profile_conv_tensor_rearrange.cpp
View file @ 9f8ab221
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_image_to_column_impl.hpp"
#include "profiler/profile_conv_tensor_rearrange_impl.hpp"
#include "profiler_operation_registry.hpp"
namespace {
enum struct RearrangeOp
{
ImageToColumn, // 0
ColumnToImage, // 1
};
enum struct ConvLayout
{
NHWC, // 0
......@@ -24,8 +30,8 @@ enum struct DataType
INT8_INT8, // 3
};
#define OP_NAME "image_to_column"
#define OP_DESC "Image To Column"
#define OP_NAME "conv_tensor_rearrange"
#define OP_DESC "Conv Tensor Rearrange"
static void print_helper_msg()
{
......@@ -41,16 +47,17 @@ static void print_helper_msg()
<< "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg8: operation type (0: ImageToColumn, 1: ColumnToImage)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
// clang-format on
}
} // namespace
int profile_image_to_column(int argc, char* argv[])
int profile_conv_tensor_rearrange(int argc, char* argv[])
{
// 8 for control, 1 for num_dim_spatial
if(argc < 9)
// 9 for control, 1 for num_dim_spatial
if(argc < 10)
{
print_helper_msg();
return 1;
......@@ -62,16 +69,17 @@ int profile_image_to_column(int argc, char* argv[])
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int num_dim_spatial = std::stoi(argv[8]);
const auto rearrange_op = static_cast<RearrangeOp>(std::stoi(argv[8]));
const int num_dim_spatial = std::stoi(argv[9]);
// 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
// 9 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
if(argc != 9 + 1 + 4 + 6 * num_dim_spatial)
{
print_helper_msg();
return 1;
}
const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 10, argv);
using F32 = float;
using F16 = ck::half_t;
......@@ -79,12 +87,17 @@ int profile_image_to_column(int argc, char* argv[])
using INT8 = int8_t;
using namespace ck::tensor_layout::convolution;
using namespace ck::conv_tensor_rearrange_op;
constexpr auto I1 = ck::Number<1>{};
constexpr auto I2 = ck::Number<2>{};
constexpr auto I3 = ck::Number<3>{};
auto profile = [&](auto num_dim_spatial_tmp, auto in_layout, auto in_type, auto out_type) {
auto profile = [&](auto num_dim_spatial_tmp,
auto in_layout,
auto in_type,
auto out_type,
auto rearrange_op_type) {
constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
using InLayout = decltype(in_layout);
......@@ -92,78 +105,147 @@ int profile_image_to_column(int argc, char* argv[])
using InDataType = decltype(in_type);
using OutDataType = decltype(out_type);
using Op = decltype(rearrange_op_type);
bool pass = ck::profiler::
profile_image_to_column_impl<NDimSpatial, InLayout, InDataType, OutDataType>(
profile_conv_tensor_rearrange_impl<NDimSpatial, InLayout, InDataType, OutDataType, Op>(
do_verification, init_method, do_log, time_kernel, params);
return pass ? 0 : 1;
};
// NHWC
if(layout == ConvLayout::NHWC)
// Image To Column
if(rearrange_op == RearrangeOp::ImageToColumn)
{
if(num_dim_spatial == 1)
// NHWC
if(layout == ConvLayout::NHWC)
{
if(data_type == DataType::F32_F32)
{
return profile(I1, GNWC{}, F32{}, F32{});
}
else if(data_type == DataType::F16_F16)
if(num_dim_spatial == 1)
{
return profile(I1, GNWC{}, F16{}, F16{});
if(data_type == DataType::F32_F32)
{
return profile(I1, GNWC{}, F32{}, F32{}, ImageToColumn{});
}
else if(data_type == DataType::F16_F16)
{
return profile(I1, GNWC{}, F16{}, F16{}, ImageToColumn{});
}
else if(data_type == DataType::BF16_BF16)
{
return profile(I1, GNWC{}, BF16{}, BF16{}, ImageToColumn{});
}
else if(data_type == DataType::INT8_INT8)
{
return profile(I1, GNWC{}, INT8{}, INT8{}, ImageToColumn{});
}
}
else if(data_type == DataType::BF16_BF16)
else if(num_dim_spatial == 2)
{
return profile(I1, GNWC{}, BF16{}, BF16{});
if(data_type == DataType::F32_F32)
{
return profile(I2, GNHWC{}, F32{}, F32{}, ImageToColumn{});
}
else if(data_type == DataType::F16_F16)
{
return profile(I2, GNHWC{}, F16{}, F16{}, ImageToColumn{});
}
else if(data_type == DataType::BF16_BF16)
{
return profile(I2, GNHWC{}, BF16{}, BF16{}, ImageToColumn{});
}
else if(data_type == DataType::INT8_INT8)
{
return profile(I2, GNHWC{}, INT8{}, INT8{}, ImageToColumn{});
}
}
else if(data_type == DataType::INT8_INT8)
else if(num_dim_spatial == 3)
{
return profile(I1, GNWC{}, INT8{}, INT8{});
if(data_type == DataType::F32_F32)
{
return profile(I3, GNDHWC{}, F32{}, F32{}, ImageToColumn{});
}
else if(data_type == DataType::F16_F16)
{
return profile(I3, GNDHWC{}, F16{}, F16{}, ImageToColumn{});
}
else if(data_type == DataType::BF16_BF16)
{
return profile(I3, GNDHWC{}, BF16{}, BF16{}, ImageToColumn{});
}
else if(data_type == DataType::INT8_INT8)
{
return profile(I3, GNDHWC{}, INT8{}, INT8{}, ImageToColumn{});
}
}
}
else if(num_dim_spatial == 2)
{
if(data_type == DataType::F32_F32)
{
return profile(I2, GNHWC{}, F32{}, F32{});
}
else if(data_type == DataType::F16_F16)
{
return profile(I2, GNHWC{}, F16{}, F16{});
}
else if(data_type == DataType::BF16_BF16)
{
return profile(I2, GNHWC{}, BF16{}, BF16{});
}
else if(data_type == DataType::INT8_INT8)
{
return profile(I2, GNHWC{}, INT8{}, INT8{});
}
}
else if(num_dim_spatial == 3)
}
else if(rearrange_op == RearrangeOp::ColumnToImage)
{
// NHWC
if(layout == ConvLayout::NHWC)
{
if(data_type == DataType::F32_F32)
if(num_dim_spatial == 1)
{
return profile(I3, GNDHWC{}, F32{}, F32{});
if(data_type == DataType::F32_F32)
{
return profile(I1, GNWC{}, F32{}, F32{}, ColumnToImage{});
}
else if(data_type == DataType::F16_F16)
{
return profile(I1, GNWC{}, F16{}, F16{}, ColumnToImage{});
}
else if(data_type == DataType::BF16_BF16)
{
return profile(I1, GNWC{}, BF16{}, BF16{}, ColumnToImage{});
}
else if(data_type == DataType::INT8_INT8)
{
return profile(I1, GNWC{}, INT8{}, INT8{}, ColumnToImage{});
}
}
else if(data_type == DataType::F16_F16)
else if(num_dim_spatial == 2)
{
return profile(I3, GNDHWC{}, F16{}, F16{});
if(data_type == DataType::F32_F32)
{
return profile(I2, GNHWC{}, F32{}, F32{}, ColumnToImage{});
}
else if(data_type == DataType::F16_F16)
{
return profile(I2, GNHWC{}, F16{}, F16{}, ColumnToImage{});
}
else if(data_type == DataType::BF16_BF16)
{
return profile(I2, GNHWC{}, BF16{}, BF16{}, ColumnToImage{});
}
else if(data_type == DataType::INT8_INT8)
{
return profile(I2, GNHWC{}, INT8{}, INT8{}, ColumnToImage{});
}
}
else if(data_type == DataType::BF16_BF16)
else if(num_dim_spatial == 3)
{
return profile(I3, GNDHWC{}, BF16{}, BF16{});
}
else if(data_type == DataType::INT8_INT8)
{
return profile(I3, GNDHWC{}, INT8{}, INT8{});
if(data_type == DataType::F32_F32)
{
return profile(I3, GNDHWC{}, F32{}, F32{}, ColumnToImage{});
}
else if(data_type == DataType::F16_F16)
{
return profile(I3, GNDHWC{}, F16{}, F16{}, ColumnToImage{});
}
else if(data_type == DataType::BF16_BF16)
{
return profile(I3, GNDHWC{}, BF16{}, BF16{}, ColumnToImage{});
}
else if(data_type == DataType::INT8_INT8)
{
return profile(I3, GNDHWC{}, INT8{}, INT8{}, ColumnToImage{});
}
}
}
}
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_image_to_column);
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_tensor_rearrange);
profiler/src/profile_gemm.cpp
View file @ 9f8ab221
......@@ -23,6 +23,7 @@ enum struct GemmDataType
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F8_F8, // 4
};
#define OP_NAME "gemm"
......@@ -31,7 +32,7 @@ enum struct GemmDataType
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: fp8)\n"
<< "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
<< " 1: A[m, k] * B[n, k] = C[m, n];\n"
<< " 2: A[k, m] * B[k, n] = C[m, n];\n"
......@@ -76,6 +77,9 @@ int profile_gemm(int argc, char* argv[])
using INT8 = int8_t;
using INT32 = int32_t;
#endif
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -194,6 +198,24 @@ int profile_gemm(int argc, char* argv[])
{
return profile(Col{}, Col{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
}
#endif
#ifdef CK_ENABLE_FP8
else if(data_type == GemmDataType::F8_F8_F8 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(Row{}, Row{}, Row{}, F8{}, F8{}, F32{}, F8{});
}
else if(data_type == GemmDataType::F8_F8_F8 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(Row{}, Col{}, Row{}, F8{}, F8{}, F32{}, F8{});
}
else if(data_type == GemmDataType::F8_F8_F8 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(Col{}, Row{}, Row{}, F8{}, F8{}, F32{}, F8{});
}
else if(data_type == GemmDataType::F8_F8_F8 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(Col{}, Col{}, Row{}, F8{}, F8{}, F32{}, F8{});
}
#endif
else
{
......
profiler/src/profile_gemm_splitk.cpp
View file @ 9f8ab221
......@@ -25,6 +25,7 @@ enum struct GemmDataType
INT8_INT8_INT8, // 3
F8_F16_F16, // 4
F16_F8_F16, // 5
F16_F16_F16_F8, // 6
};
#define OP_NAME "gemm_splitk"
......@@ -35,7 +36,8 @@ int profile_gemm_splitk(int argc, char* argv[])
if(argc != 15)
{
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16, "
"comp f8)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
......@@ -80,7 +82,8 @@ int profile_gemm_splitk(int argc, char* argv[])
auto c_type,
auto a_layout,
auto b_layout,
auto c_layout) {
auto c_layout,
auto compute_type) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
......@@ -90,6 +93,8 @@ int profile_gemm_splitk(int argc, char* argv[])
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
using ComputeType = decltype(compute_type);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
......@@ -100,7 +105,8 @@ int profile_gemm_splitk(int argc, char* argv[])
CDataType,
ALayout,
BLayout,
CLayout>(
CLayout,
ComputeType>(
do_verification,
init_method,
do_log,
......@@ -118,68 +124,84 @@ int profile_gemm_splitk(int argc, char* argv[])
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{}, F32{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{}, F32{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{}, F32{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{}, F32{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{}, F16{});
}
#if defined CK_ENABLE_FP8
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{});
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{});
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Row{}, Row{});
return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Row{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Col{}, Row{});
return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Col{}, Row{}, F16{});
}
else if(data_type == GemmDataType::F16_F16_F16_F8 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{}, F8{});
}
else if(data_type == GemmDataType::F16_F16_F16_F8 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}, F8{});
}
else if(data_type == GemmDataType::F16_F16_F16_F8 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{}, F8{});
}
else if(data_type == GemmDataType::F16_F16_F16_F8 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{}, F8{});
}
#endif
else
......
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