* Copyright (c) 2025-2026 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file col2im_simt.h
* \brief col2im_simt
*/
#ifndef COL2IM_SIMT_H
#define COL2IM_SIMT_H
#include "kernel_operator.h"
#include "simt_api/asc_simt.h"
#include "col2im_tiling_data.h"
namespace Col2imOps {
using namespace AscendC;
const uint32_t VF_MAX_THREAD_NUM = 1024;
template <typename ACC_T, typename D_T>
class Col2imSimt {
public:
__aicore__ inline Col2imSimt()
{}
__aicore__ inline void Init(
GM_ADDR gradIn, GM_ADDR gradOut, GM_ADDR workspace,
const Col2imRegBaseTilingData* __restrict tilingData);
__aicore__ inline void Process();
private:
GlobalTensor<D_T> inputXGm_;
GlobalTensor<D_T> yGm_;
uint32_t blockId_ = GetBlockIdx();
const Col2imRegBaseTilingData* tiling_;
};
template <typename ACC_T, typename D_T>
__aicore__ inline void Col2imSimt<ACC_T, D_T>::Init(
GM_ADDR gradIn, GM_ADDR gradOut, GM_ADDR workspace, const Col2imRegBaseTilingData* __restrict tilingData)
{
inputXGm_.SetGlobalBuffer((__gm__ D_T*)(gradOut));
yGm_.SetGlobalBuffer((__gm__ D_T*)(gradIn));
tiling_ = tilingData;
}
template <typename ACC_T, typename D_T>
__simt_vf__ LAUNCH_BOUND(VF_MAX_THREAD_NUM) inline void Col2imSimtCompute(
__gm__ D_T* inputXGmAddr, __gm__ D_T* yGmAddr, const uint32_t cnt, const uint32_t gradInH, const uint32_t gradInW,
const uint32_t kernelSizeH, const uint32_t kernelSizeW, const uint32_t padHeight, const uint32_t padWidth,
const uint32_t strideHeight, const uint32_t strideWidth, const uint32_t dilationHeight, const uint32_t dilationWidth,
const uint32_t heightGradOut, const uint32_t widthGradOut, uint32_t shiftGW_, uint32_t mGW_, uint32_t shiftGWH_,
uint32_t mGWH_, uint32_t shiftSW_, uint32_t mSW_, uint32_t shiftSH_, uint32_t mSH_, uint32_t shiftDH_,
uint32_t mDH_, uint32_t shiftDW_, uint32_t mDW_)
{
for(uint32_t idx = threadIdx.x + blockIdx.x * blockDim.x; idx < cnt; idx += gridDim.x * blockDim.x) {
ACC_T valv = static_cast<ACC_T>(0);
const uint32_t wIm = idx % gradInW + padWidth;
const uint32_t hIm = Simt::UintDiv(idx, mGW_, shiftGW_) % gradInH + padHeight;
const uint32_t cIm = Simt::UintDiv(idx, mGWH_, shiftGWH_);
uint32_t kernelExtentW = (kernelSizeW - 1) * dilationWidth + 1;
uint32_t kernelExtentH = (kernelSizeH - 1) * dilationHeight + 1;
uint32_t wGradOutStart = 0;
if (wIm >= kernelExtentW) {
wGradOutStart = Simt::UintDiv((wIm - kernelExtentW), mSW_, shiftSW_) + 1;
}
uint32_t wGradOutEnd = Simt::UintDiv(wIm, mSW_, shiftSW_) + 1;
if (wGradOutEnd > widthGradOut) {
wGradOutEnd = widthGradOut;
}
uint32_t hGradOutStart = 0;
if (hIm >= kernelExtentH) {
hGradOutStart = Simt::UintDiv((hIm - kernelExtentH), mSH_, shiftSH_) + 1;
}
uint32_t hGradOutEnd = Simt::UintDiv(hIm, mSH_, shiftSH_) + 1;
if (hGradOutEnd > heightGradOut) {
hGradOutEnd = heightGradOut;
}
for (uint32_t hGradOut = hGradOutStart; hGradOut < hGradOutEnd; hGradOut += 1) {
for (uint32_t wGradOut = wGradOutStart; wGradOut < wGradOutEnd; wGradOut += 1) {
uint32_t hK = (hIm - hGradOut * strideHeight);
uint32_t wK = (wIm - wGradOut * strideWidth);
if (hK - Simt::UintDiv(hK, mDH_, shiftDH_) * dilationHeight == 0 &&
wK - Simt::UintDiv(wK, mDW_, shiftDW_) * dilationWidth == 0) {
hK = Simt::UintDiv(hK, mDH_, shiftDH_);
wK = Simt::UintDiv(wK, mDW_, shiftDW_);
uint32_t gradOutIdx =
(((cIm * kernelSizeH + hK) * kernelSizeW + wK) * heightGradOut +
hGradOut) *
widthGradOut +
wGradOut;
valv += inputXGmAddr[gradOutIdx];
}
}
}
yGmAddr[idx] = static_cast<D_T>(valv);
}
}
template <typename ACC_T, typename D_T>
__aicore__ inline void Col2imSimt<ACC_T, D_T>::Process()
{
uint32_t shiftGW_, mGW_, shiftGWH_, mGWH_, shiftSW_, mSW_, shiftSH_, mSH_, shiftDH_, mDH_, shiftDW_, mDW_;
uint32_t gradInW = static_cast<uint32_t>(tiling_->outputSizeW);
uint32_t gradInWH = static_cast<uint32_t>(tiling_->outputSizeW * tiling_->outputSizeH);
uint32_t strideWidth = static_cast<uint32_t>(tiling_->strideW);
uint32_t strideHeight = static_cast<uint32_t>(tiling_->strideH);
uint32_t dilationHeight = static_cast<uint32_t>(tiling_->dilationH);
uint32_t dilationWidth = static_cast<uint32_t>(tiling_->dilationW);
GetUintDivMagicAndShift(mGW_, shiftGW_, gradInW);
GetUintDivMagicAndShift(mGWH_, shiftGWH_, gradInWH);
GetUintDivMagicAndShift(mSW_, shiftSW_, strideWidth);
GetUintDivMagicAndShift(mSH_, shiftSH_, strideHeight);
GetUintDivMagicAndShift(mDH_, shiftDH_, dilationHeight);
GetUintDivMagicAndShift(mDW_, shiftDW_, dilationWidth);
asc_vf_call<Col2imSimtCompute<ACC_T, D_T>>(
dim3{VF_MAX_THREAD_NUM, 1, 1}, (__gm__ D_T*)(inputXGm_.GetPhyAddr()), (__gm__ D_T*)(yGm_.GetPhyAddr()),
static_cast<uint32_t>(tiling_->totalLength), static_cast<uint32_t>(tiling_->outputSizeH),
static_cast<uint32_t>(tiling_->outputSizeW), static_cast<uint32_t>(tiling_->kernelSizeH),
static_cast<uint32_t>(tiling_->kernelSizeW), static_cast<uint32_t>(tiling_->paddingH),
static_cast<uint32_t>(tiling_->paddingW), static_cast<uint32_t>(tiling_->strideH),
static_cast<uint32_t>(tiling_->strideW), static_cast<uint32_t>(tiling_->dilationH),
static_cast<uint32_t>(tiling_->dilationW), static_cast<uint32_t>(tiling_->colH),
static_cast<uint32_t>(tiling_->colW), shiftGW_, mGW_, shiftGWH_, mGWH_,
shiftSW_, mSW_, shiftSH_, mSH_, shiftDH_, mDH_, shiftDW_, mDW_);
}
}
#endif
