* 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.
*/
#include <cstdint>
#include "cann_ops_blas_common.h"
#include "kernel_operator.h"
#include "simt_api/asc_simt.h"
#include "syr2_tiling_data.h"
#include "common/helper/kernel_constant.h"
static constexpr uint32_t UB_HALF_FLOATS = 8192;
template <bool UPLO_IS_UPPER>
__simt_callee__ inline void GetColRange(uint32_t row, uint32_t n, uint32_t& colStart, uint32_t& colEnd)
{
if constexpr (UPLO_IS_UPPER) {
colStart = row;
colEnd = n;
} else {
colStart = 0;
colEnd = row + 1;
}
}
template <bool UPLO_IS_UPPER>
__simt_vf__ __aicore__ LAUNCH_BOUND(SIMT_MAX_THREAD_NUM) inline void Ssyr2Gm(
uint32_t n, uint32_t lda, float alpha, int64_t incx, int64_t incy, __gm__ const float* xGm, __gm__ const float* yGm,
__gm__ float* aGm)
{
for (uint32_t row = blockIdx.x * blockDim.x + threadIdx.x; row < n; row += gridDim.x * blockDim.x) {
float xRow = (incx >= 0) ? xGm[row * incx] : xGm[(n - 1 - row) * (-incx)];
float yRow = (incy >= 0) ? yGm[row * incy] : yGm[(n - 1 - row) * (-incy)];
float axRow = alpha * xRow;
float ayRow = alpha * yRow;
uint32_t colStart, colEnd;
GetColRange<UPLO_IS_UPPER>(row, n, colStart, colEnd);
for (uint32_t col = colStart; col < colEnd; ++col) {
float xCol = (incx >= 0) ? xGm[col * incx] : xGm[(n - 1 - col) * (-incx)];
float yCol = (incy >= 0) ? yGm[col * incy] : yGm[(n - 1 - col) * (-incy)];
uint64_t aIdx = static_cast<uint64_t>(col) * lda + row;
aGm[aIdx] = aGm[aIdx] + axRow * yCol + ayRow * xCol;
}
}
}
__aicore__ inline void FallbackToGm(
const Syr2TilingData& tiling, __gm__ float* __restrict xGm, __gm__ float* __restrict yGm,
__gm__ float* __restrict aGm)
{
if (tiling.uplo == ACLBLAS_UPPER) {
asc_vf_call<Ssyr2Gm<true>>(
dim3{tiling.numThreads, 1, 1}, tiling.n, tiling.lda, tiling.alpha, tiling.incx, tiling.incy, xGm, yGm, aGm);
} else {
asc_vf_call<Ssyr2Gm<false>>(
dim3{tiling.numThreads, 1, 1}, tiling.n, tiling.lda, tiling.alpha, tiling.incx, tiling.incy, xGm, yGm, aGm);
}
}
template <bool UPLO_IS_UPPER>
__simt_vf__ __aicore__ LAUNCH_BOUND(SIMT_MAX_THREAD_NUM) inline void Ssyr2Ub(
uint32_t n, uint32_t lda, float alpha, __gm__ const float* xGm, __gm__ const float* yGm, __gm__ float* aGm,
uint32_t rowStart, uint32_t rowEnd, uint32_t xBase, uint32_t xLen, uint32_t yBase, uint32_t yLen)
{
__ubuf__ float xUb[UB_HALF_FLOATS];
__ubuf__ float yUb[UB_HALF_FLOATS];
for (uint32_t i = threadIdx.x; i < xLen; i += blockDim.x)
xUb[i] = xGm[xBase + i];
for (uint32_t i = threadIdx.x; i < yLen; i += blockDim.x)
yUb[i] = yGm[yBase + i];
asc_syncthreads();
for (uint32_t row = rowStart + threadIdx.x; row < rowEnd; row += blockDim.x) {
float xRow = xUb[row - xBase];
float yRow = yUb[row - yBase];
float axRow = alpha * xRow;
float ayRow = alpha * yRow;
uint32_t colStart, colEnd;
GetColRange<UPLO_IS_UPPER>(row, n, colStart, colEnd);
for (uint32_t col = colStart; col < colEnd; ++col) {
float xCol = xUb[col - xBase];
float yCol = yUb[col - yBase];
uint64_t aIdx = static_cast<uint64_t>(col) * lda + row;
aGm[aIdx] = aGm[aIdx] + axRow * yCol + ayRow * xCol;
}
}
}
__global__ __aicore__ void syr2_kernel(GM_ADDR x, GM_ADDR y, GM_ADDR A, const Syr2TilingData tiling)
{
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIV_ONLY);
auto* __restrict xGm = reinterpret_cast<__gm__ float* __restrict>(x);
auto* __restrict yGm = reinterpret_cast<__gm__ float* __restrict>(y);
auto* __restrict aGm = reinterpret_cast<__gm__ float* __restrict>(A);
if (tiling.incx != 1 || tiling.incy != 1 || tiling.n < 32) {
FallbackToGm(tiling, xGm, yGm, aGm);
return;
}
int32_t blkIdx = AscendC::GetBlockIdx();
uint32_t rowsPerBlk = tiling.rowsPerBlock;
uint32_t rowStart = static_cast<uint32_t>(blkIdx) * rowsPerBlk;
uint32_t rowEnd = (rowStart + rowsPerBlk < tiling.n) ? (rowStart + rowsPerBlk) : tiling.n;
if (rowStart >= rowEnd) {
return;
}
uint32_t xBase = (tiling.uplo == ACLBLAS_UPPER) ? rowStart : 0;
uint32_t xLen = (tiling.uplo == ACLBLAS_UPPER) ? (tiling.n - rowStart) : rowEnd;
if (xLen > UB_HALF_FLOATS) {
FallbackToGm(tiling, xGm, yGm, aGm);
return;
}
if (tiling.uplo == ACLBLAS_UPPER) {
asc_vf_call<Ssyr2Ub<true>>(
dim3{tiling.numThreads, 1, 1}, tiling.n, tiling.lda, tiling.alpha, xGm, yGm, aGm, rowStart, rowEnd, xBase,
xLen, xBase, xLen);
} else {
asc_vf_call<Ssyr2Ub<false>>(
dim3{tiling.numThreads, 1, 1}, tiling.n, tiling.lda, tiling.alpha, xGm, yGm, aGm, rowStart, rowEnd, xBase,
xLen, xBase, xLen);
}
}
void syr2_kernel_do(GM_ADDR x, GM_ADDR y, GM_ADDR A, const Syr2TilingData& tiling, uint32_t numBlocks, void* stream)
{
syr2_kernel<<<numBlocks, nullptr, stream>>>(x, y, A, tiling);
}