Skip to content

GitLab

Unverified Commit 7bcaf2a7 authored by Adam Osewski's avatar Adam Osewski Committed by GitHub
Browse files

Merge branch 'develop' into wavelet_model

parents e59daa22 0345963e
Show whitespace changes
Inline Side-by-side
Showing with 973 additions and 14 deletions
library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
View file @ 7bcaf2a7
......@@ -3,4 +3,8 @@ add_instance_library(device_batchnorm_instance
device_batchnorm_forward_f32_instance.cpp
device_batchnorm_forward_bf16_instance.cpp
device_batchnorm_forward_f64_instance.cpp
device_batchnorm_backward_f16_instance.cpp
device_batchnorm_backward_f32_instance.cpp
device_batchnorm_backward_bf16_instance.cpp
device_batchnorm_backward_f64_instance.cpp
)
library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp 0 → 100644
View file @ 7bcaf2a7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_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 BF16 = ck::bhalf_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// clang-format off
template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
using device_batchnorm_backward_bf16_blockwise_instances =
std::tuple <
// XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 1, 1, 1, 1, 1, 1, 1>
>;
// clang-format on
// clang-format off
template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
using device_batchnorm_backward_bf16_multiblock_instances =
std::tuple <
// XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 1, 1, 1, 1, 1, 1, 1>
>;
// clang-format on
void add_device_batchnorm_backward_rank_4_3_bf16_instances(
std::vector<std::unique_ptr<
DeviceBatchNormBwd<BF16, F32, F32, F32, BF16, F32, F32, PassThrough, 4, 3>>>& instances)
{
add_device_operation_instances(
instances, device_batchnorm_backward_bf16_blockwise_instances<4, 3, PassThrough>{});
add_device_operation_instances(
instances, device_batchnorm_backward_bf16_multiblock_instances<4, 3, PassThrough>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp 0 → 100644
View file @ 7bcaf2a7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_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;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// clang-format off
template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
using device_batchnorm_backward_f16_blockwise_instances =
std::tuple <
// XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 1, 1, 1, 1, 1, 1, 1>
>;
// clang-format on
// clang-format off
template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
using device_batchnorm_backward_f16_multiblock_instances =
std::tuple <
// XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 1, 1, 1, 1, 1, 1, 1>
>;
// clang-format on
void add_device_batchnorm_backward_rank_4_3_f16_instances(
std::vector<
std::unique_ptr<DeviceBatchNormBwd<F16, F32, F32, F32, F16, F32, F32, PassThrough, 4, 3>>>&
instances)
{
add_device_operation_instances(
instances, device_batchnorm_backward_f16_blockwise_instances<4, 3, PassThrough>{});
add_device_operation_instances(
instances, device_batchnorm_backward_f16_multiblock_instances<4, 3, PassThrough>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp 0 → 100644
View file @ 7bcaf2a7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_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 F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// clang-format off
template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
using device_batchnorm_backward_f32_blockwise_instances = std::tuple<
// XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 1, 1, 1, 1, 1, 1, 1>
>;
// clang-format on
// clang-format off
template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
using device_batchnorm_backward_f32_multiblock_instances =
std::tuple <
// XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 1, 1, 1, 1, 1, 1, 1>
>;
// clang-format on
void add_device_batchnorm_backward_rank_4_3_f32_instances(
std::vector<
std::unique_ptr<DeviceBatchNormBwd<F32, F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&
instances)
{
add_device_operation_instances(
instances, device_batchnorm_backward_f32_blockwise_instances<4, 3, PassThrough>{});
add_device_operation_instances(
instances, device_batchnorm_backward_f32_multiblock_instances<4, 3, PassThrough>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp 0 → 100644
View file @ 7bcaf2a7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_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 F64 = double;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// clang-format off
template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
using device_batchnorm_backward_f64_blockwise_instances = std::tuple<
// XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64, 4, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32, 8, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8, 32, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4, 64, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256, 2, 2, 1, 1, 1, 1, 1, 1, 1>
>;
// clang-format on
// clang-format off
template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
using device_batchnorm_backward_f64_multiblock_instances =
std::tuple <
// XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64, 4, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32, 8, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8, 32, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4, 64, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128, 2, 2, 1, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 2, 2, 2, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 1, 1, 1, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 0, 2, 2, 2, 1, 1, 1>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 1, 1, 1, 1, 2, 2, 2>,
DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256, 2, 2, 1, 1, 1, 1, 1, 1, 1>
>;
// clang-format on
void add_device_batchnorm_backward_rank_4_3_f64_instances(
std::vector<
std::unique_ptr<DeviceBatchNormBwd<F64, F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&
instances)
{
add_device_operation_instances(
instances, device_batchnorm_backward_f64_blockwise_instances<4, 3, PassThrough>{});
add_device_operation_instances(
instances, device_batchnorm_backward_f64_multiblock_instances<4, 3, PassThrough>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -26,6 +26,7 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_xdl_f16_f16_f16_km_kn_mn_instances =
......@@ -71,12 +72,36 @@ using device_gemm_xdl_f16_f16_f16_km_kn_mn_instances =
// clang-format on
>;
// irregular tile size
using device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
// clang-format off
//###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch| LoopScheduler| Pipeline|
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| | | |
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(instances, device_gemm_xdl_f16_f16_f16_km_kn_mn_instances{});
add_device_operation_instances(instances,
device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -26,6 +26,7 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_gemm_xdl_f16_f16_f16_km_nk_mn_instances =
......@@ -71,12 +72,36 @@ using device_gemm_xdl_f16_f16_f16_km_nk_mn_instances =
// clang-format on
>;
// irregular tile size
using device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
// clang-format off
//###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch| LoopScheduler| Pipeline|
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| | | |
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, 7, 1, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(instances, device_gemm_xdl_f16_f16_f16_km_nk_mn_instances{});
add_device_operation_instances(instances,
device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -26,6 +26,7 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances =
......@@ -98,12 +99,36 @@ using device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances =
// clang-format on
>;
// irregular tile size
using device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
// clang-format off
//###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch| LoopScheduler| Pipeline|
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| | | |
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(instances, device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances{});
add_device_operation_instances(instances,
device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -94,17 +94,20 @@ using device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 4, 8, 16, 16, 2, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Interwave, PipelineVersion::v1>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1, LoopScheduler::Interwave, PipelineVersion::v1>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 4, 8, 16, 16, 2, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 4, 8, 16, 16, 2, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
......
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -31,6 +31,7 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
......@@ -102,6 +103,29 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn
// clang-format on
>;
// irregular tile size
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_irregular_tile_instances =
std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Row,
......@@ -118,6 +142,9 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -31,6 +31,7 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
......@@ -102,6 +103,29 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn
// clang-format on
>;
// irregular tile size
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_irregular_tile_instances =
std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Col,
......@@ -118,6 +142,9 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -31,6 +31,7 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
......@@ -102,6 +103,29 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn
// clang-format on
>;
// irregular tile size
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_irregular_tile_instances =
std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
......@@ -118,6 +142,9 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -31,6 +31,7 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
......@@ -93,6 +94,29 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn
// clang-format on
>;
// irregular tile size
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_irregular_tile_instances =
std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Col,
......@@ -109,6 +133,9 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -16,6 +16,7 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0)
// outout: e[m, n]
......@@ -86,6 +87,29 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instanc
// clang-format on
>;
// irregular tile size
using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_irregular_tile_instances =
std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Row,
......@@ -101,6 +125,9 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_inst
{
add_device_operation_instances(
instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -16,6 +16,7 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
......@@ -86,6 +87,29 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instanc
// clang-format on
>;
// irregular tile size
using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_irregular_tile_instances =
std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Col,
......@@ -101,6 +125,9 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_inst
{
add_device_operation_instances(
instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -16,6 +16,7 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
......@@ -86,6 +87,29 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instanc
// clang-format on
>;
// irregular tile size
using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_irregular_tile_instances =
std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
......@@ -101,6 +125,9 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_inst
{
add_device_operation_instances(
instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -16,6 +16,7 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
......@@ -77,6 +78,29 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instanc
// clang-format on
>;
// irregular tile size
using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_irregular_tile_instances =
std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Col,
......@@ -92,6 +116,9 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_inst
{
add_device_operation_instances(
instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -16,6 +16,7 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b))
// outout: e[m, n]
......@@ -86,6 +87,28 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::t
// clang-format on
>;
// irregular tile size
using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Row,
......@@ -101,6 +124,9 @@ void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
{
add_device_operation_instances(
instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -16,6 +16,7 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b))
// outout: e[m, n]
......@@ -86,6 +87,28 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::t
// clang-format on
>;
// irregular tile size
using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Col,
......@@ -101,6 +124,9 @@ void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
{
add_device_operation_instances(
instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
View file @ 7bcaf2a7
......@@ -16,6 +16,7 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b))
// outout: e[m, n]
......@@ -86,6 +87,28 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::t
// clang-format on
>;
// irregular tile size
using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Interwave, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
// pipeline v2, 1 wave
,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
// clang-format on
>;
void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
......@@ -101,6 +124,9 @@ void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
{
add_device_operation_instances(
instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances{});
add_device_operation_instances(
instances,
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
}
} // namespace instance
......
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