Skip to content

GitLab

Commit defa2071 authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge branch 'develop' into aosewski/ggemm_multi_d2

parents 28a68428 f2398f61
Hide whitespace changes
Inline Side-by-side
Showing with 633 additions and 319 deletions
library/src/tensor_operation_instance/gpu/normalization_fwd/normalization_fwd_instance_common.hpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_fwd_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_fwd_splitk_impl.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_instances =
// clang-format off
std::tuple <
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<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_splitk_f16_instances =
// clang-format off
std::tuple <
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<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
DeviceNormalizationFwdImpl<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, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationFwdImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<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, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationFwdSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<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
DeviceNormalizationFwdImpl<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, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationFwdImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdImpl<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, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationFwdSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationFwdSplitKImpl<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
DeviceNormalizationFwdImpl<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
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
View file @ defa2071
......@@ -8,18 +8,18 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul2_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul2_Clamp>>>& instances)
{
// dl
add_device_operation_instances(instances,
......@@ -52,18 +52,18 @@ void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
}
void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Relu_Mul2_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Relu_Mul2_Clamp>>>& instances)
{
// dl
add_device_operation_instances(instances,
......@@ -96,18 +96,19 @@ void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
}
void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul2_TanH_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul2_TanH_Mul_Clamp>>>&
instances)
{
// dl
add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
View file @ defa2071
......@@ -8,18 +8,18 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_dl_int8_instances<NHWGC,
......@@ -51,18 +51,18 @@ void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
}
void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Relu_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Relu_Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_dl_int8_instances<NHWGC,
......@@ -96,18 +96,19 @@ void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
}
void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul_TanH_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul_TanH_Mul_Clamp>>>&
instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_dl_int8_instances<NHWGC,
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
View file @ defa2071
......@@ -8,18 +8,18 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_dl_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul2_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul2_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_dl_int8_instances<NHWGC,
......@@ -51,18 +51,18 @@ void add_device_conv2d_dl_perchannel_quantization_int8_instances(
}
void add_device_conv2d_dl_relu_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Relu_Mul2_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Relu_Mul2_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_dl_int8_instances<NHWGC,
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
View file @ defa2071
......@@ -8,18 +8,18 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_dl_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_dl_int8_instances<NHWGC,
......@@ -51,18 +51,18 @@ void add_device_conv2d_dl_perlayer_quantization_int8_instances(
}
void add_device_conv2d_dl_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Relu_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Relu_Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_dl_int8_instances<NHWGC,
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
View file @ defa2071
......@@ -8,18 +8,18 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul2_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul2_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......@@ -51,18 +51,18 @@ void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
}
void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Relu_Mul2_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Relu_Mul2_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......@@ -94,18 +94,19 @@ void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
}
void add_device_conv2d_xdl_bias_tanh_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul2_TanH_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul2_TanH_Mul_Clamp>>>&
instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
View file @ defa2071
......@@ -8,18 +8,18 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......@@ -51,18 +51,18 @@ void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
}
void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Relu_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Relu_Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......@@ -96,18 +96,19 @@ void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
}
void add_device_conv2d_xdl_bias_tanh_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul_TanH_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
int8_t,
PassThrough,
PassThrough,
Add_Mul_TanH_Mul_Clamp>>>&
instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp
View file @ defa2071
......@@ -4,7 +4,7 @@
#pragma once
#include "conv2d_quantization_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
......@@ -26,19 +26,19 @@ using device_grouped_conv2d_xdl_int8_instances =
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 32, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 32, 64, 64, 16, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 32, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 32, 64, 64, 16, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>
>;
// clang-format on
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
View file @ defa2071
......@@ -8,18 +8,18 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul2_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul2_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......@@ -51,18 +51,18 @@ void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
}
void add_device_conv2d_xdl_relu_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Relu_Mul2_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
GK_Tuple,
NHWGK,
int8_t,
int8_t,
F32_Tuple,
int8_t,
PassThrough,
PassThrough,
Relu_Mul2_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
View file @ defa2071
......@@ -8,18 +8,18 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......@@ -51,18 +51,18 @@ void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
}
void add_device_conv2d_xdl_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Relu_Mul_Clamp>>>& instances)
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<NDimSpatial,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Relu_Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<NHWGC,
......
library/src/tensor_operation_instance/gpu/transpose/CMakeLists.txt 0 → 100644
View file @ defa2071
add_instance_library(device_transpose_instance
device_transpose_instances_3d.cpp
)
library/src/tensor_operation_instance/gpu/transpose/device_transpose_instances_3d.cpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/transpose/device_transpose_instance.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
void add_device_transpose_f16_instances(
std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 5>>>&
instances)
{
#ifdef CK_ENABLE_FP16
add_device_operation_instances(instances, device_transpose_f16_instances{});
#else
ignore = instances;
#endif
}
void add_device_transpose_f32_instances(
std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 5>>>&
instances)
{
#ifdef CK_ENABLE_FP32
add_device_operation_instances(instances, device_transpose_f32_instances{});
#else
ignore = instances;
#endif
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/README.md
View file @ defa2071
......@@ -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)
......@@ -50,21 +50,23 @@ Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s
## Profile contraction kernels
```bash
#arg1: tensor operation (contraction_bilinear=CONTRACTION+Bilinear)
#arg2: data type (0: fp32; 1: f64)\n"
#arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
#arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)
#arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)
#arg4: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1])
#arg4: verification (0: no; 1: yes)
#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 and arg9: alpha and beta
#arg10 to 15: M0, M1, N0, N1, K0, K1
#arg16 to 31: Strides for A, B, D and E (skip for default)
################ op datatype layout verify init log time alpha beta M0 M1 N0 N1 K0 K1
./bin/ckProfiler contraction_bilinear 0 1 0 0 0 1 1.0 1.0 128 128 128 128 128 128
#arg5: verification (0: no; 1: yes)
#arg6: initialization (0: no init; 1: integer value; 2: decimal value)
#arg7: print tensor value (0: no; 1: yes)
#arg8: time kernel (0: no, 1: yes)
#arg9: alpha
#arg10: beta
#arg11 to 16: M0, M1, N0, N1, K0, K1
#arg17 to 32: Strides for A, B, D and E (skip for default)
################ op datatype compute_datatype layout verify init log time alpha beta M0 M1 N0 N1 K0 K1
./bin/ckProfiler contraction_bilinear 0 0 1 0 0 0 1 1.0 1.0 128 128 128 128 128 128
```
Result (MI100)
......@@ -115,7 +117,7 @@ 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 +160,7 @@ 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)
......@@ -194,14 +196,15 @@ Note: This kernel use atomic add, this will cause output buffer to be accumulate
# 1: Input fp16, Weight fp16, Output fp16
# 2: Input bf16, Weight bf16, Output bf16
# 3: Input int8, Weight int8, Output int8)
# arg3: tensor layout (0: Input[N, Hi, Wi, C], Output[N * Ho * Wo, Y * X * C])
# arg3: tensor layout (0: Input[G, N, Hi, Wi, C], Output[G * N * Ho * Wo, Y * X * C],
# 1: Input[N, Hi, Wi, G, C], Output[N * Ho * Wo * G, Y * X * C])
# arg4: verification (0: no, 1: yes)
# 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)
......
profiler/include/profiler/profile_contraction_impl.hpp
View file @ defa2071
......@@ -31,10 +31,14 @@ namespace profiler {
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
using F32 = float;
using F64 = double;
template <typename ALayout,
typename BLayout,
typename CDELayout,
typename DataType,
typename ComputeDataType,
typename DTupleDataType,
typename CDElementOp>
int profile_contraction_impl(ck::index_t do_verification,
......@@ -45,10 +49,10 @@ int profile_contraction_impl(ck::index_t do_verification,
const std::vector<ck::index_t>& M,
const std::vector<ck::index_t>& N,
const std::vector<ck::index_t>& K,
const std::vector<ck::index_t>& StridesA,
const std::vector<ck::index_t>& StridesB,
const std::vector<ck::index_t>& StridesE,
const std::vector<ck::index_t>& StridesD)
const std::vector<ck::index_t>& StridesA, // [M0, M1, K0, K1]
const std::vector<ck::index_t>& StridesB, // [N0, N1, K0, K1]
const std::vector<ck::index_t>& StridesE, // [M0, M1, N0, N1]
const std::vector<ck::index_t>& StridesD) // [M0, M1, N0, N1]
{
bool pass = true;
......@@ -63,13 +67,13 @@ int profile_contraction_impl(ck::index_t do_verification,
};
Tensor<DataType> a_m_k(f_host_tensor_descriptor(M, K, StridesA));
Tensor<DataType> b_k_n(f_host_tensor_descriptor(K, N, StridesB));
Tensor<DataType> b_n_k(f_host_tensor_descriptor(N, K, StridesB));
Tensor<DataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StridesE));
Tensor<DataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StridesE));
Tensor<DataType> d_m_n(f_host_tensor_descriptor(M, N, StridesD));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "b_n_k: " << b_n_k.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
......@@ -78,12 +82,12 @@ int profile_contraction_impl(ck::index_t do_verification,
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
b_n_k.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<DataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
b_n_k.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
}
......@@ -91,12 +95,12 @@ int profile_contraction_impl(ck::index_t do_verification,
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
DeviceMem a_device_buf(sizeof(DataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(DataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(DataType) * b_n_k.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(DataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DataType) * d_m_n.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
b_device_buf.ToDevice(b_n_k.mData.data());
e_device_buf.SetZero();
d_device_buf.ToDevice(d_m_n.mData.data());
......@@ -118,7 +122,8 @@ int profile_contraction_impl(ck::index_t do_verification,
DataType,
AElementOp,
BElementOp,
CDElementOp>;
CDElementOp,
ComputeDataType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......@@ -126,6 +131,9 @@ int profile_contraction_impl(ck::index_t do_verification,
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
using AccDataType =
typename std::conditional<std::is_same<ComputeDataType, F64>::value, F64, F32>::type;
// Run reference op
if(do_verification)
{
......@@ -136,7 +144,8 @@ int profile_contraction_impl(ck::index_t do_verification,
DataType,
DataType,
DataType,
DataType,
AccDataType,
ComputeDataType,
AElementOp,
BElementOp>;
......@@ -146,7 +155,7 @@ int profile_contraction_impl(ck::index_t do_verification,
Tensor<DataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StridesE));
auto ref_argument =
ref_op.MakeArgument(a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op);
ref_op.MakeArgument(a_m_k, b_n_k, c_m_n_host_result, a_element_op, b_element_op);
ref_invoker.Run(ref_argument);
......@@ -272,8 +281,29 @@ int profile_contraction_impl(ck::index_t do_verification,
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
float threshold =
static_cast<DataType>(nelems_k) * std::numeric_limits<DataType>::epsilon();
// Both the kernel and the reference use `AccDataType`, so an absolute error of both
// of them is bounded by `nelems_k * std::numeric_limits<AccDataType>::epsilon()`.
// Comparing one to another can result in an absolute error as high as twice that
// value.
double threshold = 2 * nelems_k * std::numeric_limits<AccDataType>::epsilon();
// Handle the possible casting error of either AccDataType -> DataType or
// DataType -> ComputeDataType.
// TODO: Add a generic solution for calculating thresholds in CK.
if constexpr(ck::is_same_v<DataType, ck::bhalf_t> ||
ck::is_same_v<ComputeDataType, ck::bhalf_t>)
{
const double epsilon = std::pow(2, -7);
// Maximum relative casting error when rounding to zero.
threshold += epsilon * 2;
}
else if constexpr(ck::is_same_v<DataType, ck::half_t> ||
ck::is_same_v<ComputeDataType, ck::half_t>)
{
const double epsilon = std::pow(2, -10);
// Maximum relative casting error when rounding to zero.
threshold += epsilon * 2;
}
pass = pass & ck::utils::check_err(e_m_n_device_result,
e_m_n_host_result,
"Error: incorrect results!",
......@@ -283,7 +313,7 @@ int profile_contraction_impl(ck::index_t do_verification,
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_n_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", e_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", e_m_n_device_result.mData, ",")
......
profiler/include/profiler/profile_contraction_utils.hpp
View file @ defa2071
......@@ -23,8 +23,18 @@ enum struct ContractionMatrixLayout
enum struct ContractionDataType
{
F32_F32_F32_F32, // 0
F64_F64_F64_F64, // 1
F32_F32_F32_F32, // 0
F64_F64_F64_F64, // 1
F16_F16_F16_F16, // 2
BF16_BF16_BF16_BF16, // 3
};
enum struct ContractionComputeDataType
{
F32 = 0,
F64,
F16,
BF16,
};
inline void collect_index_params(char* argv[],
......
profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp
View file @ defa2071
......@@ -93,6 +93,26 @@ static auto make_ref_op()
}
}
template <typename InputLayout>
static auto create_gemm_desc(const ck::index_t G, const ck::index_t NDoHoWo, const ck::index_t CZYX)
{
using namespace ck::tensor_layout::convolution;
if constexpr(std::is_same_v<InputLayout, GNWC> || std::is_same_v<InputLayout, GNHWC> ||
std::is_same_v<InputLayout, GNDHWC>)
{
return HostTensorDescriptor({G, NDoHoWo, CZYX});
}
else if constexpr(std::is_same_v<InputLayout, NWGC> || std::is_same_v<InputLayout, NHWGC> ||
std::is_same_v<InputLayout, NDHWGC>)
{
return HostTensorDescriptor({G, NDoHoWo, CZYX}, {CZYX, CZYX * G, 1});
}
else
{
throw std::runtime_error("Unsupported layout!");
}
}
template <index_t NDimSpatial,
typename InputLayout,
typename InputDataType,
......@@ -116,13 +136,13 @@ bool profile_conv_tensor_rearrange_impl(int do_verification,
const auto image_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InputLayout>(
conv_param);
const auto gemm_desc = HostTensorDescriptor({NDoHoWo, CZYX});
const auto gemm_desc = create_gemm_desc<InputLayout>(conv_param.G_, 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> image_g_n_c_wis_strides{};
std::array<ck::index_t, 2> gemm_m_k_strides{};
std::array<ck::index_t, 3> gemm_g_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{};
......@@ -134,7 +154,7 @@ bool profile_conv_tensor_rearrange_impl(int do_verification,
copy(conv_param.filter_spatial_lengths_, filter_spatial_lengths);
copy(conv_param.output_spatial_lengths_, output_spatial_lengths);
copy(image_desc.GetStrides(), image_g_n_c_wis_strides);
copy(gemm_desc.GetStrides(), gemm_m_k_strides);
copy(gemm_desc.GetStrides(), gemm_g_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);
......@@ -212,13 +232,14 @@ bool profile_conv_tensor_rearrange_impl(int do_verification,
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InputDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutputDataType*>(out_device_buf.GetDeviceBuffer()),
conv_param.G_,
conv_param.N_,
conv_param.C_,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
image_g_n_c_wis_strides,
gemm_m_k_strides,
gemm_g_m_k_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -234,7 +255,7 @@ bool profile_conv_tensor_rearrange_impl(int do_verification,
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_btype =
NDoHoWo * CZYX * (sizeof(OutputDataType) + sizeof(InputDataType));
conv_param.G_ * NDoHoWo * CZYX * (sizeof(OutputDataType) + sizeof(InputDataType));
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
......
profiler/include/profiler/profile_gemm_impl.hpp
View file @ defa2071
......@@ -75,8 +75,8 @@ int profile_gemm_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 0.1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.01, 0.01});
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
......
profiler/include/profiler/profile_gemm_splitk_impl.hpp
View file @ defa2071
......@@ -143,8 +143,7 @@ bool profile_gemm_splitk_impl(int do_verification,
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 36, 40, 60,
64, 72, 80, 88, 96, 128, 144, 160, 176, 192, 256};
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 32, 36, 40, 64, 96, 128};
if(KBatch > 0)
{
......
profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
View file @ defa2071
......@@ -198,18 +198,18 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
}
};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
profiler/include/profiler/profile_groupnorm_impl.hpp → profiler/include/profiler/profile_groupnorm_fwd_impl.hpp
View file @ defa2071
......@@ -7,7 +7,7 @@
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
#include "ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
......@@ -88,14 +88,14 @@ bool profile_groupnorm_impl(int do_verification,
beta_dev.ToDevice(beta.mData.data());
// add device normalization instances
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
5,
3>;
using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
5,
3>;
// get device op instances
const auto instance_ptrs =
......
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