// Copyright (c) 2026 Huawei Technologies Co., Ltd
// All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <torch/extension.h>
#include "torch_npu/csrc/core/npu/NPUStream.h"
#include "kernel_operator.h"
constexpr uint32_t BUFFER_NUM = 2; //tensor num for each queue
class KernelTrig {
public:
__aicore__ inline KernelTrig() {}
// Initialize the global memory and buffer queues
__aicore__ inline void Init(GM_ADDR x, GM_ADDR out_sin, GM_ADDR out_cos, GM_ADDR out_tan, uint32_t totalLength)
{
this->blockLength = totalLength / AscendC::GetBlockNum();
this->tileNum = 8;
this->tileLength = this->blockLength / this->tileNum / BUFFER_NUM;
xGm.SetGlobalBuffer((__gm__ float *)x + this->blockLength * AscendC::GetBlockIdx(), this->blockLength);
sinGm.SetGlobalBuffer((__gm__ float *)out_sin + this->blockLength * AscendC::GetBlockIdx(), this->blockLength);
cosGm.SetGlobalBuffer((__gm__ float *)out_cos + this->blockLength * AscendC::GetBlockIdx(), this->blockLength);
tanGm.SetGlobalBuffer((__gm__ float *)out_tan + this->blockLength * AscendC::GetBlockIdx(), this->blockLength);
pipe.InitBuffer(inQueueX, BUFFER_NUM, this->tileLength * sizeof(float));
pipe.InitBuffer(outQueueSin, BUFFER_NUM, this->tileLength * sizeof(float));
pipe.InitBuffer(outQueueCos, BUFFER_NUM, this->tileLength * sizeof(float));
pipe.InitBuffer(outQueueTan, BUFFER_NUM, this->tileLength * sizeof(float));
}
__aicore__ inline void Process()
{
int32_t loopCount = this->tileNum * BUFFER_NUM;
for (int32_t i = 0; i < loopCount; i++) {
CopyIn(i);
Compute(i);
CopyOut(i);
}
}
private:
__aicore__ inline void CopyIn(int32_t progress)
{
AscendC::LocalTensor<float> xLocal = inQueueX.AllocTensor<float>();
AscendC::DataCopy(xLocal, xGm[progress * this->tileLength], this->tileLength);
inQueueX.EnQue(xLocal);
}
__aicore__ inline void Compute(int32_t progress)
{
AscendC::LocalTensor<float> xLocal = inQueueX.DeQue<float>();
AscendC::LocalTensor<float> sinLocal = outQueueSin.AllocTensor<float>();
AscendC::LocalTensor<float> cosLocal = outQueueCos.AllocTensor<float>();
AscendC::LocalTensor<float> tanLocal = outQueueTan.AllocTensor<float>();
AscendC::Sin(sinLocal, xLocal, this->tileLength);
AscendC::Cos(cosLocal, xLocal, this->tileLength);
AscendC::Tan(tanLocal, xLocal, this->tileLength);
outQueueSin.EnQue<float>(sinLocal);
outQueueCos.EnQue<float>(cosLocal);
outQueueTan.EnQue<float>(tanLocal);
inQueueX.FreeTensor(xLocal);
}
__aicore__ inline void CopyOut(int32_t progress)
{
// Copy the sin, cos, and tan values from local memory to global memory (inplace modification)
AscendC::LocalTensor<float> sinLocal = outQueueSin.DeQue<float>();
AscendC::LocalTensor<float> cosLocal = outQueueCos.DeQue<float>();
AscendC::LocalTensor<float> tanLocal = outQueueTan.DeQue<float>();
AscendC::DataCopy(sinGm[progress * this->tileLength], sinLocal, this->tileLength);
AscendC::DataCopy(cosGm[progress * this->tileLength], cosLocal, this->tileLength);
AscendC::DataCopy(tanGm[progress * this->tileLength], tanLocal, this->tileLength);
outQueueSin.FreeTensor(sinLocal);
outQueueCos.FreeTensor(cosLocal);
outQueueTan.FreeTensor(tanLocal);
}
private:
AscendC::TPipe pipe;
AscendC::TQue<AscendC::TPosition::VECIN, BUFFER_NUM> inQueueX;
AscendC::TQue<AscendC::TPosition::VECOUT, BUFFER_NUM> outQueueSin, outQueueCos, outQueueTan;
AscendC::GlobalTensor<float> xGm;
AscendC::GlobalTensor<float> sinGm;
AscendC::GlobalTensor<float> cosGm;
AscendC::GlobalTensor<float> tanGm;
uint32_t blockLength;
uint32_t tileNum;
uint32_t tileLength;
};
__global__ __vector__ void trig_inplace_custom(GM_ADDR x, GM_ADDR out_sin, GM_ADDR out_cos, GM_ADDR out_tan,
uint32_t totalLength)
{
KernelTrig op;
op.Init(x, out_sin, out_cos, out_tan, totalLength);
op.Process();
}
namespace cpp_extension_acs {
at::Tensor ascendc_trig(const at::Tensor &x, const at::Tensor &out_sin, const at::Tensor &out_cos)
{
auto acl_stream = c10_npu::getCurrentNPUStream().stream(true);
at::Tensor out_tan = at::empty_like(x);
uint32_t blockDim = 8;
uint32_t totalLength = 1;
for (uint32_t size : x.sizes()) {
totalLength *= size;
}
// Launch the custom kernel using <<<>>>
trig_inplace_custom<<<blockDim, nullptr, acl_stream>>>(
(uint8_t *)(x.mutable_data_ptr()), (uint8_t *)(out_sin.mutable_data_ptr()),
(uint8_t *)(out_cos.mutable_data_ptr()), (uint8_t *)(out_tan.mutable_data_ptr()), totalLength);
return out_tan;
}
} // namespace cpp_extension_acs
at::Tensor trig_impl_meta(const at::Tensor& x, const at::Tensor& out_sin, const at::Tensor& out_cos)
{
return at::empty_like(x);
}
// Define a new operator
TORCH_LIBRARY_FRAGMENT(cpp_extension_acs, m)
{
m.def("ascendc_trig(Tensor x, Tensor(a!) out_sin, Tensor(b!) out_cos) -> Tensor");
}
// Register implementation for the "PrivateUse1" backend
TORCH_LIBRARY_IMPL(cpp_extension_acs, PrivateUse1, m)
{
m.impl("ascendc_trig", TORCH_FN(cpp_extension_acs::ascendc_trig));
}
// Define a simple model using the custom operation
TORCH_LIBRARY_IMPL(cpp_extension_acs, Meta, m)
{
m.impl("ascendc_trig", &trig_impl_meta);
}