cann-workbench 0.0.0-stage → 0.0.2-alpha

package/skills/asc-tools/tool-ops-tune/ops-family/references/elementwise/templates/add_custom/op_host/add_custom.asc

@@ -0,0 +1,149 @@
+/**
+ * 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.
+ */
+
+// ============================================================================
+// Ascend C Kernel 直调样例 - Add 算子（Host + Main 入口）
+// ============================================================================
+//
+// 文件结构说明：
+//   op_kernel/add_custom_tiling.h   - Tiling 常量 + 结构体（纯 C/C++，kernel 和 host 共用）
+//   op_kernel/add_custom_kernel.asc - Kernel 实现（KernelAdd 类 + add_custom_kernel 核函数入口）
+//   op_host/add_custom.asc          - 本文件：Host 端代码 + main 入口（引用 op_kernel/ 下的文件）
+//
+// [MODIFY] 创建新算子时的修改点（搜索 [MODIFY] 标记）:
+// 1. KernelCall 中的函数名和参数
+// 2. Tiling 计算逻辑
+// 3. 输入/输出数量
+// 4. CMakeLists.txt 中的目标名
+// ============================================================================
+
+#include <cstdint>
+#include <iostream>
+#include <vector>
+#include "acl/acl.h"
+#include "../op_kernel/add_custom_tiling.h"
+#include "data_utils.h"
+
+// 引用拆出的纯 kernel 文件（包含 kernel_operator.h）
+#include "../op_kernel/add_custom_kernel.asc"
+
+// ============================================================================
+// Host 端代码 - 运行在 CPU 上
+// ============================================================================
+
+struct ArgInfo {
+    std::string fileName;
+    size_t length;
+};
+
+// [PATTERN] KernelCall 封装 - 处理完整的核函数调用流程
+// [MODIFY] 修改函数名和参数
+void KernelCall(uint32_t blockNum, aclrtStream stream,
+    std::vector<ArgInfo> &inputsInfo, std::vector<ArgInfo> &outputsInfo,
+    uint8_t *tilingData)
+{
+    std::vector<uint8_t *> inputHost(inputsInfo.size());
+    std::vector<uint8_t *> inputDevice(inputsInfo.size());
+    std::vector<uint8_t *> outputHost(outputsInfo.size());
+    std::vector<uint8_t *> outputDevice(outputsInfo.size());
+    uint8_t *tilingDevice = nullptr;
+
+    aclrtMalloc((void **)&tilingDevice, sizeof(AddTilingData), ACL_MEM_MALLOC_HUGE_FIRST);
+    aclrtMemcpy(tilingDevice, sizeof(AddTilingData), tilingData,
+        sizeof(AddTilingData), ACL_MEMCPY_HOST_TO_DEVICE);
+
+    for (size_t i = 0; i < inputsInfo.size(); i++) {
+        aclrtMallocHost((void **)&inputHost[i], inputsInfo[i].length);
+        aclrtMalloc((void **)&inputDevice[i], inputsInfo[i].length, ACL_MEM_MALLOC_HUGE_FIRST);
+        ReadFile(inputsInfo[i].fileName, inputsInfo[i].length, inputHost[i], inputsInfo[i].length);
+        aclrtMemcpy(inputDevice[i], inputsInfo[i].length, inputHost[i],
+            inputsInfo[i].length, ACL_MEMCPY_HOST_TO_DEVICE);
+    }
+
+    for (size_t i = 0; i < outputsInfo.size(); i++) {
+        aclrtMallocHost((void **)&outputHost[i], outputsInfo[i].length);
+        aclrtMalloc((void **)&outputDevice[i], outputsInfo[i].length, ACL_MEM_MALLOC_HUGE_FIRST);
+    }
+
+    // [MODIFY] kernel 入口函数名
+    add_custom_kernel<<<blockNum, nullptr, stream>>>(
+        inputDevice[0], inputDevice[1],
+        outputDevice[0], tilingDevice);
+    aclrtSynchronizeStream(stream);
+
+    aclrtFree(tilingDevice);
+    for (size_t i = 0; i < outputsInfo.size(); i++) {
+        aclrtMemcpy(outputHost[i], outputsInfo[i].length, outputDevice[i],
+            outputsInfo[i].length, ACL_MEMCPY_DEVICE_TO_HOST);
+        WriteFile(outputsInfo[i].fileName, outputHost[i], outputsInfo[i].length);
+        aclrtFree(outputDevice[i]);
+        aclrtFreeHost(outputHost[i]);
+    }
+
+    for (size_t i = 0; i < inputsInfo.size(); i++) {
+        aclrtFree(inputDevice[i]);
+        aclrtFreeHost(inputHost[i]);
+    }
+}
+
+// [PATTERN] ACL 标准初始化流程
+int32_t main(int32_t argc, char *argv[])
+{
+    aclInit(nullptr);
+    int32_t deviceId = 0;
+    auto ret = aclrtSetDevice(deviceId);
+    if (ret != ACL_SUCCESS) {
+        printf("aclrtSetDevice failed, ret=%d\n", ret);
+        return ret;
+    }
+
+    // [MODIFY] 获取设备最大核数
+    int64_t availableCoreNum = 0;
+    ret = aclrtGetDeviceInfo(deviceId, ACL_DEV_ATTR_VECTOR_CORE_NUM, &availableCoreNum);
+    if (ret != ACL_SUCCESS) {
+        printf("aclrtGetDeviceInfo failed, ret=%d\n", ret);
+        aclrtResetDevice(deviceId);
+        return ret;
+    }
+
+    // [MODIFY] 根据你的算子调整这些参数
+    uint32_t totalLength = 4096 * 8;
+    size_t inputByteSize = totalLength * sizeof(float);
+    size_t outputByteSize = totalLength * sizeof(float);
+
+    // [MODIFY] 设置 Tiling 数据
+    AddTilingData tiling;
+    tiling.totalLength = totalLength;
+    uint64_t totalTiles = (totalLength + TILE_LENGTH - 1) / TILE_LENGTH;
+    uint64_t tilesPerCore = (totalTiles + availableCoreNum - 1) / availableCoreNum;
+    uint32_t blockNum = tiling.blockNum = (totalTiles + tilesPerCore - 1) / tilesPerCore;
+    tiling.numPerCore = tilesPerCore * TILE_LENGTH;
+    tiling.tailNumLastCore = totalLength - tiling.numPerCore * (blockNum - 1);
+
+    // [MODIFY] 根据你的输入/输出文件调整
+    std::vector<ArgInfo> inputsInfo = {
+        {"./input/input_x.bin", inputByteSize},
+        {"./input/input_y.bin", inputByteSize}
+    };
+    std::vector<ArgInfo> outputsInfo = {
+        {"./output/output.bin", outputByteSize}
+    };
+
+    aclrtStream stream = nullptr;
+    aclrtCreateStream(&stream);
+
+    KernelCall(blockNum, stream, inputsInfo, outputsInfo, (uint8_t *)&tiling);
+
+    aclrtDestroyStream(stream);
+    aclrtResetDevice(deviceId);
+    aclFinalize();
+
+    return 0;
+}

package/skills/asc-tools/tool-ops-tune/ops-family/references/elementwise/templates/add_custom/op_host/data_utils.h

@@ -0,0 +1,97 @@
+/**
+ * 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 DATA_UTILS_H
+#define DATA_UTILS_H
+
+#include <fcntl.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#include <fstream>
+
+#define ERROR_LOG(fmt, args...) fprintf(stdout, "[ERROR]  " fmt "\n", ##args)
+
+/**
+ * @brief Read data from file
+ * @param [in] filePath: file path
+ * @param [in] bufferSize: expected file size
+ * @param [out] buffer: buffer to store data
+ * @param [in] bufferLen: buffer length
+ * @return read result
+ */
+inline bool ReadFile(const std::string& filePath, size_t bufferSize, void* buffer, size_t bufferLen)
+{
+    if (buffer == nullptr) {
+        ERROR_LOG("buffer is nullptr");
+        return false;
+    }
+    if (bufferSize > bufferLen) {
+        ERROR_LOG("buffer size is larger than buffer length");
+        return false;
+    }
+
+    std::ifstream file(filePath, std::ios::binary);
+    if (!file.is_open()) {
+        ERROR_LOG("Open file failed. path = %s", filePath.c_str());
+        return false;
+    }
+
+    file.seekg(0, std::ios::end);
+    size_t fileSize = file.tellg();
+    file.seekg(0, std::ios::beg);
+
+    if (fileSize != bufferSize) {
+        ERROR_LOG("file size %zu != expected size %zu", fileSize, bufferSize);
+        file.close();
+        return false;
+    }
+
+    file.read(static_cast<char*>(buffer), bufferSize);
+    if (!file) {
+        ERROR_LOG("Read file failed");
+        file.close();
+        return false;
+    }
+
+    file.close();
+    return true;
+}
+
+/**
+ * @brief Write data to file
+ * @param [in] filePath: file path
+ * @param [in] buffer: data to write to file
+ * @param [in] size: size to write
+ * @return write result
+ */
+inline bool WriteFile(const std::string& filePath, const void* buffer, size_t size)
+{
+    if (buffer == nullptr) {
+        ERROR_LOG("Write file failed. buffer is nullptr");
+        return false;
+    }
+
+    int fd = open(filePath.c_str(), O_RDWR | O_CREAT | O_TRUNC, S_IRUSR | S_IWUSR);
+    if (fd < 0) {
+        ERROR_LOG("Open file failed. path = %s", filePath.c_str());
+        return false;
+    }
+
+    ssize_t writeSize = write(fd, buffer, size);
+    close(fd);
+    if (static_cast<size_t>(writeSize) != size) {
+        ERROR_LOG("Write file Failed.");
+        return false;
+    }
+
+    return true;
+}
+
+#endif

package/skills/asc-tools/tool-ops-tune/ops-family/references/elementwise/templates/add_custom/op_kernel/add_custom_kernel.asc

@@ -0,0 +1,136 @@
+/**
+ * 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.
+ */
+
+/* Generated by Ascend C project template */
+
+// ============================================================================
+// Ascend C Kernel 实现 - Add 算子（纯 kernel，不含 host 代码）
+// ============================================================================
+//
+// [MODIFY] 创建新算子时的修改点（搜索 [MODIFY] 标记）:
+// 1. #include "add_custom_tiling.h" → "#include "<your_op>_tiling.h"
+// 2. 类名: KernelAdd → Kernel<YourOp>
+// 3. Kernel函数名: add_custom → <your_op>（保持 _kernel 后缀）
+// 4. 计算逻辑: AscendC::Add → 你的算子逻辑
+// 5. 输入/输出数量: 根据你的算子调整
+// ============================================================================
+
+#include "kernel_operator.h"
+#include "add_custom_tiling.h"
+
+// [MODIFY] Kernel 类 - 封装 NPU 上的计算逻辑
+class KernelAdd {
+public:
+    __aicore__ inline KernelAdd(AscendC::TPipe* pipe)
+    {
+        pipe_ = pipe;
+    };
+
+    // [MODIFY] 根据你的算子修改 Init 参数和 Global Tensor
+    __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z,
+                                const __gm__ AddTilingData* tiling)
+    {
+        this->tiling = tiling;
+        uint32_t blockIdx = AscendC::GetBlockIdx();
+        if (blockIdx < tiling->blockNum - 1) {
+            total = tiling->numPerCore;
+            tileNum = total / TILE_LENGTH;
+            tailTileElementNum = TILE_LENGTH;
+        } else {
+            total = tiling->tailNumLastCore;
+            tileNum = (total + (TILE_LENGTH - 1))/ TILE_LENGTH;
+            tailTileElementNum = total - TILE_LENGTH * (tileNum - 1);
+        }
+
+        // [MODIFY] 根据你的输入/输出数量调整 Global Tensor
+        xGm.SetGlobalBuffer((__gm__ float *)x + blockIdx * tiling->numPerCore, total);
+        yGm.SetGlobalBuffer((__gm__ float *)y + blockIdx * tiling->numPerCore, total);
+        zGm.SetGlobalBuffer((__gm__ float *)z + blockIdx * tiling->numPerCore, total);
+
+        // [MODIFY] 根据你的算子计算逻辑，调整UB Buffer初始化
+        pipe_->InitBuffer(inQueueX, DOUBLE_BUFFER, TILE_LENGTH * sizeof(float));
+        pipe_->InitBuffer(inQueueY, DOUBLE_BUFFER, TILE_LENGTH * sizeof(float));
+        pipe_->InitBuffer(outQueueZ, DOUBLE_BUFFER, TILE_LENGTH * sizeof(float));
+    }
+
+    __aicore__ inline void Process()
+    {
+        for (uint32_t i = 0; i < tileNum; i++) {
+            uint32_t count = (i == tileNum - 1) ? tailTileElementNum : TILE_LENGTH;
+            CopyIn(count);
+            Compute(count);
+            CopyOut(count);
+        }
+    }
+
+    __aicore__ inline void CopyIn(uint32_t count)
+    {
+        // [MODIFY] 根据你的输入数量调整
+        AscendC::LocalTensor<float> xLocal = inQueueX.AllocTensor<float>();
+        AscendC::LocalTensor<float> yLocal = inQueueY.AllocTensor<float>();
+
+        AscendC::DataCopyPad(xLocal, xGm,
+            {1, static_cast<uint16_t>(count * sizeof(float)), 0, 0},
+            {false, 0, 0, 0});
+        AscendC::DataCopyPad(yLocal, yGm,
+            {1, static_cast<uint16_t>(count * sizeof(float)), 0, 0},
+            {false, 0, 0, 0});
+
+        inQueueX.EnQue(xLocal);
+        inQueueY.EnQue(yLocal);
+    }
+
+    __aicore__ inline void Compute(uint32_t count)
+    {
+        AscendC::LocalTensor<float> xLocal = inQueueX.DeQue<float>();
+        AscendC::LocalTensor<float> yLocal = inQueueY.DeQue<float>();
+        AscendC::LocalTensor<float> zLocal = outQueueZ.AllocTensor<float>();
+
+        // [MODIFY] 替换为你的计算逻辑
+        AscendC::Add(zLocal, xLocal, yLocal, count);
+
+        outQueueZ.EnQue<float>(zLocal);
+        inQueueX.FreeTensor(xLocal);
+        inQueueY.FreeTensor(yLocal);
+    }
+
+    __aicore__ inline void CopyOut(uint32_t count)
+    {
+        AscendC::LocalTensor<float> zLocal = outQueueZ.DeQue<float>();
+        AscendC::DataCopyPad(zGm, zLocal,
+            {1, static_cast<uint16_t>(count * sizeof(float)), 0, 0});
+        outQueueZ.FreeTensor(zLocal);
+    }
+
+private:
+    AscendC::TPipe* pipe_;
+    const __gm__ AddTilingData* tiling;
+
+    // [MODIFY] 根据你的输入/输出数量调整
+    AscendC::GlobalTensor<float> xGm;
+    AscendC::GlobalTensor<float> yGm;
+    AscendC::GlobalTensor<float> zGm;
+
+    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueueX, inQueueY;
+    AscendC::TQue<AscendC::TPosition::VECOUT, 1> outQueueZ;
+
+    uint32_t total;
+    uint32_t tileNum;
+    uint32_t tailTileElementNum;
+};
+
+// [MODIFY] 核函数入口 - add_custom → <your_op>（保持 _kernel 后缀）
+extern "C" __global__ __vector__ void add_custom_kernel(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling)
+{
+    AscendC::TPipe pipe;
+    KernelAdd op(&pipe);
+    op.Init(x, y, z, (__gm__ AddTilingData*)tiling);
+    op.Process();
+}

package/skills/asc-tools/tool-ops-tune/ops-family/references/elementwise/templates/add_custom/op_kernel/add_custom_tiling.h

@@ -0,0 +1,31 @@
+/**
+ * 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.
+ */
+
+/* Generated by Ascend C project template */
+
+// [MODIFY] Tiling 常量和结构体 - kernel 和 host 共用
+// 此文件只含纯 C/C++ 语法，不含 __aicore__、__gm__ 等 ASC 关键字
+
+#pragma once
+
+#include <cstdint>
+
+// [CONFIG] 单次搬运/处理的数据量（元素数），影响 UB 内存占用
+// 常用值: float32 用 4096/8192，float16 用 8192/16384
+constexpr uint32_t TILE_LENGTH = 4096;
+constexpr uint32_t DOUBLE_BUFFER = 2;
+
+// [MODIFY] Tiling 数据结构 - 向核函数传递的运行时参数
+struct AddTilingData {
+    uint32_t blockNum;
+    uint64_t totalLength;
+    uint64_t numPerCore;
+    uint64_t tailNumLastCore;
+};

package/skills/asc-tools/tool-ops-tune/ops-family/references/elementwise/templates/add_custom/run.sh

@@ -0,0 +1,87 @@
+#!/bin/bash
+# ----------------------------------------------------------------------------------------------------------
+# 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.
+# ----------------------------------------------------------------------------------------------------------
+
+# ============================================================================
+# Ascend C Kernel 直调运行脚本
+# 编译 → 生成测试数据 → 运行 → 精度验证
+#
+# 用法：
+#   bash run.sh              # 完整流程（含编译）
+#   bash run.sh --skip-build # 跳过编译，复用已有产物（代码审查阶段使用）
+#   bash run.sh --torch      # 跳过可执行文件通路，只跑 PyTorch 通路
+#
+# 退出码：
+#   0  全部步骤成功
+#   1  某步骤失败（stderr 有具体说明）
+# ============================================================================
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+cd "${SCRIPT_DIR}"
+
+# [MODIFY] 替换为实际的算子二进制名称
+OP_NAME="add"
+
+SKIP_BUILD=0
+TORCH_ONLY=0
+for arg in "$@"; do
+    case "$arg" in
+        --skip-build) SKIP_BUILD=1 ;;
+        --torch)      TORCH_ONLY=1 ;;
+    esac
+done
+
+# 统一错误输出函数
+die() { echo "ERROR: $*" >&2; exit 1; }
+
+echo "=== [1/4] 设置 CANN 环境 ==="
+[ -n "${ASCEND_HOME_PATH:-}" ] || die "ASCEND_HOME_PATH 未设置，请先配置 CANN 环境"
+source "${ASCEND_HOME_PATH}/set_env.sh" || die "set_env.sh 执行失败"
+
+if [ "${TORCH_ONLY}" -eq 1 ]; then
+    echo "=== PyTorch 通路验证 ==="
+    cd build
+    python3 ../scripts/test_torch.py || die "PyTorch 通路验证失败"
+    echo "=== 完成 ==="
+    exit 0
+fi
+
+if [ "${SKIP_BUILD}" -eq 1 ]; then
+    [ -f "build/${OP_NAME}" ] || die "--skip-build 指定但 build/${OP_NAME} 不存在，请先完整编译"
+    echo "=== [2/4] 跳过编译（复用已有产物）==="
+else
+    echo "=== [2/4] 编译 ==="
+    mkdir -p build
+    cd build
+    cmake .. || die "cmake 配置失败"
+    make -j4  || die "make 编译失败"
+    cd ..
+fi
+
+echo "=== [3/4] 生成测试数据 ==="
+cd build
+python3 ../scripts/gen_data.py || die "gen_data.py 执行失败"
+
+echo "=== [4/4] 运行 Kernel ==="
+rm -f output/output.bin
+"./${OP_NAME}" || die "Kernel 运行失败（exit code $?）"
+[ -f output/output.bin ] || die "Kernel 运行后 output.bin 不存在（静默失败）"
+
+echo "=== 精度验证 ==="
+python3 ../scripts/verify_result.py output/output.bin output/golden.bin \
+    || die "精度验证失败（verify_result.py 返回非零）"
+
+echo "=== PyTorch 通路验证 ==="
+python3 ../scripts/test_torch.py \
+    || die "PyTorch 通路验证失败（test_torch.py 返回非零）"
+
+echo "=== 完成 ==="
+exit 0

package/skills/asc-tools/tool-ops-tune/ops-family/references/elementwise/templates/add_custom/scripts/gen_data.py

@@ -0,0 +1,70 @@
+# ----------------------------------------------------------------------------------------------------------
+# 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.
+# ----------------------------------------------------------------------------------------------------------
+
+
+# ============================================================================
+# 测试数据生成脚本
+# ============================================================================
+#
+# [MODIFY] 创建新算子时修改:
+# 1. total_length: 总元素数（需与 add.asc 中的保持一致）
+# 2. dtype: 数据类型（需与 add.asc 中的保持一致）
+# 3. 输入数据: 根据你的算子生成对应的输入
+# 4. golden 计算: 修改 golden.py 中的 compute_golden()
+# ============================================================================
+
+import argparse
+import os
+
+import numpy as np
+
+from golden import compute_golden
+
+BIN_EXT = "." + "bin"  # 数据文件扩展名
+parser = argparse.ArgumentParser(description="add_custom 测试数据生成")
+parser.add_argument(
+    "--input-x",
+    default=os.path.join("input", "input_x" + BIN_EXT),
+    help="x 输入文件路径",
+)
+parser.add_argument(
+    "--input-y",
+    default=os.path.join("input", "input_y" + BIN_EXT),
+    help="y 输入文件路径",
+)
+parser.add_argument(
+    "--golden",
+    default=os.path.join("output", "golden" + BIN_EXT),
+    help="golden 输出路径",
+)
+args = parser.parse_args()
+
+os.makedirs("input", exist_ok=True)
+os.makedirs("output", exist_ok=True)
+
+# [MODIFY] 数据规格参数
+total_length = 4096 * 8  # 总元素数，需与 add.asc 中 main 函数保持一致
+dtype = np.float32  # 数据类型，需与 add.asc 中模板参数保持一致
+
+# [MODIFY] 生成输入数据（根据你的算子调整）
+x = np.random.randn(total_length).astype(dtype)
+y = np.random.randn(total_length).astype(dtype)
+
+x.tofile(args.input_x)
+y.tofile(args.input_y)
+
+# [MODIFY] 计算 golden（逻辑定义在 golden.py 中，修改那里即可）
+golden = compute_golden(x, y)
+golden.tofile(args.golden)
+
+print(f"Generated test data: {total_length} elements, dtype={dtype}")
+print(f"  {args.input_x}: {x.shape}, {x.dtype}")
+print(f"  {args.input_y}: {y.shape}, {y.dtype}")
+print(f"  {args.golden}: {golden.shape}, {golden.dtype}")

package/skills/asc-tools/tool-ops-tune/ops-family/references/elementwise/templates/add_custom/scripts/golden.py

@@ -0,0 +1,34 @@
+# ----------------------------------------------------------------------------------------------------------
+# 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.
+# ----------------------------------------------------------------------------------------------------------
+
+# ============================================================================
+# Golden 计算（双通路共用）
+# ============================================================================
+#
+# [MODIFY] 创建新算子时修改:
+# 1. compute_golden(): 替换为你的算子的正确计算逻辑
+#
+# 本文件被 gen_data.py 和 test_torch.py 共同引用，
+# 修改此处即同时更新两条通路的 golden 基准。
+# ============================================================================
+
+
+# [MODIFY] 替换为你的算子的正确计算逻辑
+def compute_golden(x1, x2):
+    """计算算子的参考输出。
+
+    Args:
+        x1: numpy array 或 torch.Tensor
+        x2: numpy array 或 torch.Tensor
+
+    Returns:
+        与输入同类型的参考输出
+    """
+    return x1 + x2