* Copyright (c) 2025 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 eye.h
* \brief eye
*/
#ifndef ASCENDC_EYE_H_
#define ASCENDC_EYE_H_
#include "kernel_operator.h"
#include "simt_api/asc_simt.h"
namespace Eye {
using namespace AscendC;
constexpr uint32_t USED_THREAD = 512;
template <typename T, typename U>
__simt_vf__ __aicore__ LAUNCH_BOUND(USED_THREAD) inline void EyeSimt(
__gm__ T* y, U allAxis, U perBatchAxis, uint32_t magic, uint32_t shift,
int64_t numRows, int64_t numColumns);
template <typename T, typename U> class EyeKernel {
public:
__aicore__ inline EyeKernel(const EyeForAscendCTilingData& tilingData, TPipe& pipe)
: td_(tilingData), pipe_(pipe){};
__aicore__ inline void Init(GM_ADDR y);
__aicore__ inline void Process();
__aicore__ inline void DumpGmZero();
private:
GlobalTensor<T> y_;
TQue<QuePosition::VECOUT, 1> zeroBuf_;
TPipe &pipe_;
U allAxis_{ 0 };
U perBatchAxis_{ 0 };
U blockIdx_{ 0 };
U blockNum_{ 0 };
uint32_t shift_{ 0 };
uint32_t magic_{ 0 };
int64_t loopNum_{ 0 };
int64_t tailLoopLength_{ 0 };
int64_t curCoreData_{ 0 };
const EyeForAscendCTilingData &td_;
};
template <typename T, typename U>
__aicore__ inline void EyeKernel<T, U>::Init(GM_ADDR y)
{
#if defined(__CCE_KT_TEST__)
perBatchAxis_ = std::min(td_.numRows, td_.numColumns);
#else
perBatchAxis_ = min(td_.numRows, td_.numColumns);
#endif
allAxis_ = static_cast<U>(td_.batch) * perBatchAxis_;
y_.SetGlobalBuffer((__gm__ T *)(y));
pipe_.InitBuffer(zeroBuf_, 1, td_.loopLength * sizeof(T));
blockIdx_ = GetBlockIdx();
blockNum_ = GetBlockNum();
curCoreData_ = blockIdx_ != (td_.usedCoreNum - 1) ? td_.normBlockData : td_.tailBlockData;
loopNum_ = curCoreData_ / td_.loopLength;
tailLoopLength_ = curCoreData_ - loopNum_ * td_.loopLength;
if constexpr (sizeof(U) == sizeof(uint32_t)) {
GetUintDivMagicAndShift(magic_, shift_, perBatchAxis_);
}
}
template <typename T, typename U>
__simt_vf__ __aicore__ LAUNCH_BOUND(USED_THREAD) inline void EyeSimt(
__gm__ T* y, U allAxis, U perBatchAxis, uint32_t magic, uint32_t shift, int64_t numRows, int64_t numColumns)
{
for (U i = blockIdx.x * blockDim.x + threadIdx.x; i < allAxis; i += gridDim.x * blockDim.x) {
if constexpr (sizeof(U) == sizeof(uint32_t)) {
uint32_t t1 = __umulhi(i, magic);
t1 = t1 + i;
uint32_t batchIdx = t1 >> shift;
uint32_t rowIdx = i - batchIdx * perBatchAxis;
y[batchIdx * static_cast<U>(numRows) * static_cast<U>(numColumns) +
rowIdx * static_cast<U>(numColumns) + rowIdx] = static_cast<T>(1);
} else {
U batchIdx = i / perBatchAxis;
U rowIdx = i - batchIdx * perBatchAxis;
y[batchIdx * static_cast<U>(numRows) * static_cast<U>(numColumns) +
rowIdx * static_cast<U>(numColumns) + rowIdx] = static_cast<T>(1);
}
}
}
template <typename T, typename U> __aicore__ inline void EyeKernel<T, U>::DumpGmZero()
{
LocalTensor<T> tmpLocal = zeroBuf_.AllocTensor<T>();
Duplicate(tmpLocal, static_cast<T>(0), td_.loopLength);
zeroBuf_.EnQue<T>(tmpLocal);
LocalTensor<T> srcLocal = zeroBuf_.DeQue<T>();
DataCopyExtParams copyParams = { static_cast<uint16_t>(1), static_cast<uint32_t>(td_.loopLength * sizeof(T)),
static_cast<uint32_t>(0), static_cast<uint32_t>(0), static_cast<uint32_t>(0) };
for (int64_t i = 0; i < loopNum_; i++) {
DataCopyPad(y_[blockIdx_ * td_.normBlockData + i * td_.loopLength], srcLocal, copyParams);
}
if (tailLoopLength_ > 0) {
copyParams.blockLen = tailLoopLength_ * sizeof(T);
DataCopyPad(y_[blockIdx_ * td_.normBlockData + loopNum_ * td_.loopLength], srcLocal, copyParams);
}
zeroBuf_.FreeTensor(srcLocal);
}
template <typename T, typename U> __aicore__ inline void EyeKernel<T, U>::Process()
{
if (blockIdx_ < td_.usedCoreNum) {
DumpGmZero();
}
SyncAll();
asc_vf_call<EyeSimt<T, U>>(dim3(USED_THREAD), (__gm__ T*)(y_.GetPhyAddr()),
allAxis_, perBatchAxis_, magic_, shift_, td_.numRows, td_.numColumns);
}
}
#endif