* Copyright (c) 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.
*/
#ifndef CSROT_KERNEL_H
#define CSROT_KERNEL_H
#include "kernel_operator.h"
#include "common/helper/kernel_utils.h"
#include "common/iterator/iterator.h"
#include "common/compute/simd.h"
using namespace AscendC;
__global__ __aicore__ void csrot_kernel(GM_ADDR gm_x, GM_ADDR gm_y,
GM_ADDR gm_workspace, GM_ADDR tiling_para_gm)
{
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIV_ONLY);
AscendC::SetAtomicNone();
AscendC::SetMaskNorm();
const int32_t elementCount = *reinterpret_cast<__gm__ int32_t *>(tiling_para_gm);
float cosValue = *reinterpret_cast<__gm__ float *>(tiling_para_gm + 4);
float sinValue = *reinterpret_cast<__gm__ float *>(tiling_para_gm + 8);
float negaSinValue = 0 - sinValue;
const int16_t aiv_id = AscendC::GetBlockIdx();
int16_t aiv_num = AscendC::GetBlockNum();
if (aiv_num == 0) {
aiv_num = 1;
}
int32_t elementPerCore = elementCount / aiv_num;
const int32_t elementRemain = elementCount % aiv_num;
int32_t startIndex = aiv_id * elementPerCore;
if (aiv_id < elementRemain) {
startIndex += aiv_id;
elementPerCore += 1;
} else {
startIndex += elementRemain;
}
AscendC::GlobalTensor<float> gm_x_tensor;
AscendC::GlobalTensor<float> gm_y_tensor;
gm_x_tensor.SetGlobalBuffer(reinterpret_cast<__gm__ float *>(gm_x));
gm_y_tensor.SetGlobalBuffer(reinterpret_cast<__gm__ float *>(gm_y));
AsdopsBuffer<ArchType::ASCEND_V220> buf;
AscendC::LocalTensor<float> buf0_x = buf.GetBuffer<BufferType::ASCEND_UB, float>(0);
AscendC::LocalTensor<float> buf0_y = buf.GetBuffer<BufferType::ASCEND_UB, float>(24 * 1024);
AscendC::LocalTensor<float> buf0_temp1 = buf.GetBuffer<BufferType::ASCEND_UB, float>(48 * 1024);
AscendC::LocalTensor<float> buf0_temp2 = buf.GetBuffer<BufferType::ASCEND_UB, float>(72 * 1024);
AscendC::LocalTensor<float> buf1_x = buf.GetBuffer<BufferType::ASCEND_UB, float>(96 * 1024);
AscendC::LocalTensor<float> buf1_y = buf.GetBuffer<BufferType::ASCEND_UB, float>(120 * 1024);
AscendC::LocalTensor<float> buf1_temp1 = buf.GetBuffer<BufferType::ASCEND_UB, float>(144 * 1024);
AscendC::LocalTensor<float> buf1_temp2 = buf.GetBuffer<BufferType::ASCEND_UB, float>(168 * 1024);
const int32_t maxElementSingleCount = 24 * 1024 / sizeof(float);
int32_t loop_count = elementPerCore / maxElementSingleCount;
const int32_t loop_remain = elementPerCore % maxElementSingleCount;
if (loop_remain > 0) {
loop_count += 1;
}
bool flag = 0;
AscendC::GlobalTensor<float> current_gm_x = gm_x_tensor[startIndex];
AscendC::GlobalTensor<float> current_gm_y = gm_y_tensor[startIndex];
int32_t currentElementSingleCount = maxElementSingleCount;
SET_FLAG(MTE3, MTE2, EVENT_ID0);
SET_FLAG(MTE3, MTE2, EVENT_ID1);
for (int32_t loop_idx = 0; loop_idx < loop_count; loop_idx++) {
currentElementSingleCount = (loop_idx == loop_count - 1 && loop_remain > 0) ? loop_remain :
maxElementSingleCount;
int32_t currentRepeatCount = (currentElementSingleCount + 63) / 64;
auto buf_x = flag ? buf0_x : buf1_x;
auto buf_y = flag ? buf0_y : buf1_y;
auto buf_temp1 = flag ? buf0_temp1 : buf1_temp1;
auto buf_temp2 = flag ? buf0_temp2 : buf1_temp2;
auto event_id = flag ? EVENT_ID0 : EVENT_ID1;
WAIT_FLAG(MTE3, MTE2, event_id);
gm_to_ub_align<ArchType::ASCEND_V220, float>(
buf_x,
current_gm_x,
0,
1,
currentElementSingleCount * sizeof(float),
0,
0,
0,
0
);
SET_FLAG(MTE2, V, EVENT_ID0);
gm_to_ub_align<ArchType::ASCEND_V220, float>(
buf_y,
current_gm_y,
0,
1,
currentElementSingleCount * sizeof(float),
0,
0,
0,
0
);
SET_FLAG(MTE2, V, EVENT_ID1);
WAIT_FLAG(MTE2, V, EVENT_ID0);
muls_v<ArchType::ASCEND_V220, float>(
buf_temp1,
buf_x,
cosValue,
currentRepeatCount,
1,
1,
8,
8
);
muls_v<ArchType::ASCEND_V220, float>(\
buf_temp2,
buf_x,
negaSinValue,
currentRepeatCount,
1,
1,
8,
8
);
PIPE_BARRIER(V);
WAIT_FLAG(MTE2, V, EVENT_ID1);
AscendC::Axpy(buf_temp1, buf_y, sinValue, currentRepeatCount * 64);
AscendC::Axpy(buf_temp2, buf_y, cosValue, currentRepeatCount * 64);
SET_FLAG(V, MTE3, event_id);
WAIT_FLAG(V, MTE3, event_id);
ub_to_gm_align<ArchType::ASCEND_V220, float>(
current_gm_x,
buf_temp1,
0,
1,
currentElementSingleCount * sizeof(float),
0,
0,
0,
0
);
ub_to_gm_align<ArchType::ASCEND_V220, float>(
current_gm_y,
buf_temp2,
0,
1,
currentElementSingleCount * sizeof(float),
0,
0,
0,
0
);
SET_FLAG(MTE3, MTE2, event_id);
current_gm_x = current_gm_x[maxElementSingleCount];
current_gm_y = current_gm_y[maxElementSingleCount];
flag = 1 - flag;
}
WAIT_FLAG(MTE3, MTE2, EVENT_ID0);
WAIT_FLAG(MTE3, MTE2, EVENT_ID1);
}
void csrot_kernel_do(GM_ADDR x, GM_ADDR y, GM_ADDR workSpace, GM_ADDR tilingGm,
uint32_t numBlocks, void *stream)
{
csrot_kernel<<<numBlocks, nullptr, stream>>>(x, y, workSpace, tilingGm);
}
#endif