Ascend C Sinh算子开发——香橙派310B

/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/bin/msopgen gen -i SinhCustom.json -c ai_core-Ascend310B -lan cpp -out Sinh

"ASCEND_CANN_PACKAGE_PATH": {
    "type": "PATH",
    "value": "/usr/local/Ascend/ascend-toolkit/latest"
}

#include "register/tilingdata_base.h"

namespace optiling {
BEGIN_TILING_DATA_DEF(SinhCustomTilingData)
  
  TILING_DATA_FIELD_DEF(uint32_t, totalLength);
  TILING_DATA_FIELD_DEF(uint32_t, tileNum);
END_TILING_DATA_DEF;

REGISTER_TILING_DATA_CLASS(SinhCustom, SinhCustomTilingData)
}

#include "sinh_custom_tiling.h"
#include "register/op_def_registry.h"


namespace optiling {
const uint32_t BLOCK_DIM = 8;
const uint32_t TILE_NUM = 8;
static ge::graphStatus TilingFunc(gert::TilingContext* context)
{
    SinhCustomTilingData tiling;
    uint32_t totalLength = context->GetInputTensor(0)->GetShapeSize();
    context->SetBlockDim(BLOCK_DIM);
    tiling.set_totalLength(totalLength);
    tiling.set_tileNum(TILE_NUM);
    tiling.SaveToBuffer(context->GetRawTilingData()->GetData(), context->GetRawTilingData()->GetCapacity());
    context->GetRawTilingData()->SetDataSize(tiling.GetDataSize());
    size_t *currentWorkspace = context->GetWorkspaceSizes(1);
    currentWorkspace[0] = 0;
    return ge::GRAPH_SUCCESS;
}
}

#include "kernel_operator.h"
using namespace AscendC;
constexpr int32_t BUFFER_NUM = 2;

class KernelSinh {
public:
    __aicore__ inline KernelSinh() {}
    __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, uint32_t totalLength, uint32_t tileNum)
    {
        ASSERT(GetBlockNum() != 0 && "block dim can not be zero!");
        this->blockLength = totalLength / GetBlockNum();
        this->tileNum = tileNum;
        ASSERT(tileNum != 0 && "tile num can not be zero!");
        this->tileLength = this->blockLength / tileNum / BUFFER_NUM;

        xGm.SetGlobalBuffer((__gm__ half*)x + this->blockLength * GetBlockIdx(), this->blockLength);
        yGm.SetGlobalBuffer((__gm__ half*)y + this->blockLength * GetBlockIdx(), this->blockLength);
        pipe.InitBuffer(inQueueX, BUFFER_NUM, this->tileLength * sizeof(half));
        pipe.InitBuffer(outQueueY, BUFFER_NUM, this->tileLength * sizeof(half));
    }
    __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)
    {
        LocalTensor<half> xLocal = inQueueX.AllocTensor<half>();
        DataCopy(xLocal, xGm[progress * tileLength], tileLength);
        inQueueX.EnQue(xLocal);
    }
    __aicore__ inline void Compute(int32_t progress)
    {
        LocalTensor<half> xLocal = inQueueX.DeQue<half>();
        LocalTensor<half> yLocal = outQueueY.AllocTensor<half>();
        
        Exp(xLocal, xLocal, this->tileLength);   // x = e^{x}
        Reciprocal(yLocal, xLocal, this->tileLength);  // y = e^{-x}
        Sub(yLocal, xLocal, yLocal, this->tileLength);  // x = x - y = e^{x} - e^{-x};
        half scalar = 0.5;
        Muls(yLocal, yLocal, (half)0.5, this->tileLength);  // y = x * 0.5 = (e^{x} - e^{-x}) * 0.5;
        
        outQueueY.EnQue<half>(yLocal);
        inQueueX.FreeTensor(xLocal);

    }
    __aicore__ inline void CopyOut(int32_t progress)
    {
        LocalTensor<half> yLocal = outQueueY.DeQue<half>();
        DataCopy(yGm[progress * tileLength], yLocal, tileLength);
        outQueueY.FreeTensor(yLocal);
    }

private:
    TPipe pipe;
    TQue<QuePosition::VECIN, BUFFER_NUM> inQueueX;
    TQue<QuePosition::VECOUT, BUFFER_NUM> outQueueY;
    GlobalTensor<half> xGm;
    GlobalTensor<half> yGm;
    
    uint32_t blockLength;
    uint32_t tileNum;
    uint32_t tileLength;
    
};

extern "C" __global__ __aicore__ void sinh_custom(GM_ADDR x, GM_ADDR y, GM_ADDR workspace, GM_ADDR tiling) {
    GET_TILING_DATA(tiling_data, tiling);
    KernelSinh op;
    
    op.Init(x, y, tiling_data.totalLength, tiling_data.tileNum);
    op.Process();

}

# cd进入算子工程目录内
# 编译
# chmod +x ./build.sh # 若没有执行权限，可以使用
./build.sh

# 部署
cd ./build_out
./custom_opp_ubuntu_aarch64.run

#!/usr/bin/python3
# -*- coding:utf-8 -*-
# Copyright 2022-2023 Huawei Technologies Co., Ltd
import numpy as np

def gen_golden_data_simple():
    input_x = np.random.uniform(1, 10, [8, 2048]).astype(np.float16)
    golden = np.sinh(input_x)

    input_x.tofile("./input/input_x.bin")
    golden.tofile("./output/golden.bin")

if __name__ == "__main__":
    gen_golden_data_simple()

OperatorDesc CreateOpDesc()
{
    // define operator
    std::vector<int64_t> shape { 8, 2048 };
    aclDataType dataType = ACL_FLOAT16;
    aclFormat format = ACL_FORMAT_ND;
    OperatorDesc opDesc;
    opDesc.AddInputTensorDesc(dataType, shape.size(), shape.data(), format);
    opDesc.AddOutputTensorDesc(dataType, shape.size(), shape.data(), format);
    return opDesc;
}

bool SetInputData(OpRunner &runner)
{
    size_t fileSize = 0;
    ReadFile("../input/input_x.bin", fileSize, runner.GetInputBuffer<void>(0), runner.GetInputSize(0));
    INFO_LOG("Set input success");
    return true;
}

bool ProcessOutputData(OpRunner &runner)
{
    WriteFile("../output/output_z.bin", runner.GetOutputBuffer<void>(0), runner.GetOutputSize(0));
    INFO_LOG("Write output success");
    return true;
}

/**
* @file op_runner.cpp
*
* Copyright (C) 2020. Huawei Technologies Co., Ltd. All rights reserved.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
*/
#include "op_runner.h"
#include "aclnn_sinh_custom.h"
#include <limits>
#include <cassert>
#include "acl/acl_op_compiler.h"
#include "common.h"

bool OpRunner::RunOp()
{
    for (size_t i = 0; i < numInputs_; ++i) {
        auto size = GetInputSize(i);
        aclrtMemcpyKind kind = ACL_MEMCPY_HOST_TO_DEVICE;
        if (g_isDevice) {
            kind = ACL_MEMCPY_DEVICE_TO_DEVICE;
        }
        if (aclrtMemcpy(devInputs_[i], size, hostInputs_[i], size, kind) != ACL_SUCCESS) {
            ERROR_LOG("Copy input[%zu] failed", i);
            return false;
        }
        INFO_LOG("Copy input[%zu] success", i);
    }

    aclrtStream stream = nullptr;
    if (aclrtCreateStream(&stream) != ACL_SUCCESS) {
        ERROR_LOG("Create stream failed");
        return false;
    }
    INFO_LOG("Create stream success");

    size_t workspaceSize = 0;
    aclOpExecutor *handle = nullptr;
    //添加计算workspace大小并申请内存代码
    auto ret = aclnnSinhCustomGetWorkspaceSize(inputTensor_[0], outputTensor_[0],
                                              &workspaceSize, &handle);
    if (ret != ACL_SUCCESS) {
        (void)aclrtDestroyStream(stream);
        ERROR_LOG("Get Operator Workspace failed. error code is %d", static_cast<int32_t>(ret));
        return false;
    }
   
    void *workspace = nullptr;
    if (workspaceSize != 0) {
        if (aclrtMalloc(&workspace, workspaceSize, ACL_MEM_MALLOC_NORMAL_ONLY) != ACL_SUCCESS) {
            ERROR_LOG("Malloc device memory failed");
        }
    }
    //添加执行算子代码
    if (aclnnSinhCustom(workspace, workspaceSize, handle, stream) != ACL_SUCCESS) {
        (void)aclrtDestroyStream(stream);
        ERROR_LOG("Execute Operator failed. error code is %d", static_cast<int32_t>(ret));
        return false;
    }

    ret = aclrtSynchronizeStreamWithTimeout(stream, 5000);
    if (ret != SUCCESS) {
        ERROR_LOG("Synchronize stream failed. error code is %d", static_cast<int32_t>(ret));
        (void)aclrtDestroyStream(stream);
        return false;
    }
    INFO_LOG("Synchronize stream success");

    for (size_t i = 0; i < numOutputs_; ++i) {
        auto size = GetOutputSize(i);
        aclrtMemcpyKind kind = ACL_MEMCPY_DEVICE_TO_HOST;
        if (g_isDevice) {
            kind = ACL_MEMCPY_DEVICE_TO_DEVICE;
        }
        if (aclrtMemcpy(hostOutputs_[i], size, devOutputs_[i], size, kind) != ACL_SUCCESS) {
            INFO_LOG("Copy output[%zu] success", i);
            (void)aclrtDestroyStream(stream);
            return false;
        }
        INFO_LOG("Copy output[%zu] success", i);
    }

    (void)aclrtDestroyStream(stream);
    return true;
}

bash run.sh

[
    {
        "op": "SinhCustom",
        "language":"cpp",
        "input_desc": [
            {
                "name": "x",
                "param_type": "required",
                "format": [
                    "ND"
                ],
                "type": [
                    "fp16"
                ]
            }
        ],
        "output_desc": [
            {
                "name": "y",
                "param_type": "required",
                "format": [
                    "ND"
                ],
                "type": [
                    "fp16"
                ]
            }
        ]
    }
]