@cannbot-plugin/cannbot-harness 0.0.0-stage → 0.1.5

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/add_custom/scripts/verify_result.py

@@ -0,0 +1,54 @@
+# ----------------------------------------------------------------------------------------------------------
+# 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.
+# ----------------------------------------------------------------------------------------------------------
+
+# ============================================================================
+# 结果验证脚本
+# ============================================================================
+#
+# [MODIFY] 创建新算子时修改:
+# 1. dtype: 数据类型（需与 gen_data.py 和 add.asc 保持一致）
+# 2. rtol/atol: 容差参数（根据算子精度要求调整）
+# ============================================================================
+
+import numpy as np
+import sys
+
+# [MODIFY] 验证参数
+dtype = np.float32   # 数据类型，需与 gen_data.py 和 add.asc 保持一致
+rtol = 1e-5          # 相对容差
+atol = 1e-8          # 绝对容差
+
+def verify_result(output_path, golden_path):
+    output = np.fromfile(output_path, dtype=dtype)
+    golden = np.fromfile(golden_path, dtype=dtype)
+    
+    if output.shape != golden.shape:
+        print(f"Shape mismatch: output {output.shape} vs golden {golden.shape}")
+        return False
+    
+    if np.allclose(output, golden, rtol=rtol, atol=atol):
+        print(f"Verification PASSED! Shape: {output.shape}")
+        print(f"Max diff: {np.max(np.abs(output - golden))}")
+        return True
+    else:
+        diff = np.abs(output - golden)
+        print(f"Verification FAILED!")
+        print(f"Max diff: {np.max(diff)}, Mean diff: {np.mean(diff)}")
+        mismatches = np.where(diff > atol + rtol * np.abs(golden))[0]
+        print(f"Mismatch count: {len(mismatches)} / {len(golden)}")
+        return False
+
+if __name__ == "__main__":
+    if len(sys.argv) < 3:
+        print("Usage: python verify_result.py <output.bin> <golden.bin>")
+        sys.exit(1)
+    
+    success = verify_result(sys.argv[1], sys.argv[2])
+    sys.exit(0 if success else 1)

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/kernel_launch_details.md

@@ -0,0 +1,116 @@
+# Ascend C Kernel 直调进阶
+
+> 基础用法和修改指南见 `add_custom/op_host/add_custom.asc` 中的注释。本文档提供深入理解。
+
+## 内存层次与数据流
+
+```
+┌─────────────────────────────────────┐
+│        Global Memory (GM)           │  大容量，高延迟
+│        (HBM/DDR)                    │
+├─────────────────────────────────────┤
+│        Local Memory (UB)            │  小容量，低延迟
+│        (Unified Buffer)             │  用于计算
+├─────────────────────────────────────┤
+│        Registers                    │  最快，最少
+└─────────────────────────────────────┘
+```
+
+数据流：Global Memory → Local Memory (CopyIn) → Compute → Local Memory → Global Memory (CopyOut)
+
+## 性能优化
+
+### Double Buffer
+
+通过重叠内存搬运和计算提高性能：
+
+```cpp
+// 使用两个 Event ID 交替执行
+AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(EVENT_ID0);
+AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(EVENT_ID1);
+
+for (uint32_t i = 0; i < loopCount; i++) {
+    int32_t eventID = (i % 2 == 0 ? EVENT_ID0 : EVENT_ID1);
+    
+    AscendC::WaitFlag<AscendC::HardEvent::MTE3_MTE2>(eventID);
+    // CopyIn + Compute + CopyOut
+    AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(eventID);
+}
+```
+
+```
+时间轴:  |------|------|------|------|------|
+Buffer A: [MTE2] [COMP] [MTE3] [MTE2] [COMP] ...
+Buffer B:        [MTE2] [COMP] [MTE3] [MTE2] ...
+```
+
+### 同步机制
+
+| 同步方式 | 适用场景 |
+|----------|----------|
+| `PipeBarrier<PIPE_*>` | 简单场景，自动等待流水线 |
+| `SetFlag/WaitFlag` | Double Buffer，精确控制 |
+
+### 内存分配策略
+
+| 方式 | 特点 | 适用场景 |
+|------|------|----------|
+| `LocalMemAllocator` | 简单易用，线性分配 | Bank 冲突不敏感 |
+| `TQue` 手动管理 | 精确控制 Bank | 性能敏感场景 |
+
+```cpp
+// 手动分配示例（高级）
+AscendC::TPipe pipe;
+AscendC::TQue<AscendC::TPosition::VECIN, 1> queue;  // 模板 depth=1 即可
+pipe.InitBuffer(queue, 2, TILE_LENGTH * sizeof(float));  // num=2 开启 Double Buffer
+auto xLocal = queue.AllocTensor<float>();
+```
+
+## 常见问题
+
+### 如何选择 TILE_LENGTH？
+
+考虑 Local Memory 大小和数据对齐要求：
+- float32: 4096 或 8192
+- float16: 8192 或 16384
+- Double Buffer 需要额外空间
+
+### 如何支持多输入多输出？
+
+修改核函数签名和 KernelCall：
+
+```cpp
+// 核函数
+__global__ __aicore__ void kernel_custom(
+    GM_ADDR input0, GM_ADDR input1,
+    GM_ADDR output0, GM_ADDR output1,
+    GM_ADDR tiling
+);
+
+// main 函数中
+std::vector<ArgInfo> inputsInfo = {
+    {"./input/input0.bin", size0},
+    {"./input/input1.bin", size1}
+};
+std::vector<ArgInfo> outputsInfo = {
+    {"./output/output0.bin", size0},
+    {"./output/output1.bin", size1}
+};
+```
+
+### 如何处理不对齐的数据？
+
+在 Tiling 中添加边界参数：
+
+```cpp
+struct TilingData {
+    uint32_t totalLength;      // 总长度
+    uint32_t alignedLength;    // 对齐长度
+    uint32_t lastTileLength;   // 最后一个 tile 的长度
+};
+```
+
+## 参考资源
+
+- [Ascend C 编程指南](https://www.hiascend.com/document)
+- [Ascend C 示例代码](https://gitcode.com/cann/asc-devkit/tree/master/examples)

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/kirin_add_template/CMakeLists.txt

@@ -0,0 +1,19 @@
+# ----------------------------------------------------------------------------------------------------------
+# 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.
+# ----------------------------------------------------------------------------------------------------------
+
+cmake_minimum_required(VERSION 3.16)
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_CXX_STANDARD 17)
+
+set(CCE_CMAKE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Modules)
+list(APPEND CMAKE_MODULE_PATH ${CCE_CMAKE_PATH})
+project(kernel_samples LANGUAGES CCE CXX)
+
+add_subdirectory(cmake/npu)

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/kirin_add_template/README.md

@@ -0,0 +1,214 @@
+## 目录结构介绍
+``` 
+├── kirin_add_template
+│   ├── cmake                   // 编译工程文件
+│   │   ├── Modules/            // CCE编译器检测模块（通用，无需修改）
+│   │   └── npu/CMakeLists.txt  // NPU编译配置（通用，无需修改）
+│   ├── input                   // 存放脚本生成的输入数据目录
+│   ├── output                  // 存放算子运行输出数据和真值数据的目录
+│   ├── scripts
+│   │   ├── acl.json            // ACL初始化配置（默认空JSON）
+│   │   ├── gen_data.py         // 生成输入数据和golden真值（需修改）
+│   │   └── verify_result.py    // 输出与golden对比验证（需修改）
+│   ├── add_custom.cpp          // 算子核函数实现（需替换）
+│   ├── data_utils.h            // 通用文件读写/日志工具（无需修改）
+│   ├── main.cpp                // Host侧主程序（需修改）
+│   ├── CMakeLists.txt          // 顶层CMake配置（无需修改）
+│   └── run.sh                  // 一键编译+执行+验证脚本（需修改）
+``` 
+
+## 代码实现介绍
+本调用样例中实现的是固定shape为8*2048的Add算子。
+- kernel实现   
+  Add算子的数学表达式为：  
+  ```
+  z = x + y
+  ```
+  计算逻辑是：Ascend C提供的矢量计算接口的操作元素都为LocalTensor，输入数据需要先搬运进片上存储，然后使用计算接口完成两个输入参数相加，得到最终结果，再搬出到外部存储上。  
+  
+  Add算子的实现流程分为3个基本任务：CopyIn，Compute，CopyOut。CopyIn任务负责将Global Memory上的输入Tensor xGm和yGm搬运到Local Memory，分别存储在xLocal、yLocal，Compute任务负责对xLocal、yLocal执行加法操作，计算结果存储在zLocal中，CopyOut任务负责将输出数据从zLocal搬运至Global Memory上的输出Tensor zGm中。具体请参考[add_custom.cpp](./add_custom.cpp)。
+
+- 调用实现  
+  1. CPU侧运行验证主要通过ICPU_RUN_KF CPU调测宏等CPU调测库提供的接口来完成；  
+  2. NPU侧运行验证主要通过使用<<<>>>内核调用符来完成。
+
+  应用程序通过ASCENDC_CPU_DEBUG 宏区分代码逻辑运行于CPU侧还是NPU侧。
+
+## 运行样例算子
+- 打开样例目录
+
+  ```
+  cd $HOME/ops/kirin_add_template
+  ```
+- 配置环境变量
+
+  这里的$HOME需要替换为本仓根目录
+  ```
+  export ASCEND_INSTALL_PATH=$HOME/Ascend/cann-mobile/cann-8.5.0
+  ```
+
+- 样例执行
+
+  ```
+  bash run.sh -r [RUN_MODE] -v  [SOC_VERSION] 
+  ```
+  - RUN_MODE：编译方式，sim（NPU仿真）。
+  - SOC_VERSION：KirinX90 或者 Kirin9030。
+
+  示例如下。
+  ```
+  bash run.sh -r sim -v Kirin9030
+  ```   
+
+  `run.sh` 的完整执行流程为：
+  1. **解析参数**：通过 `-v`（SoC版本）、`-r`（运行模式：sim/cpu/npu）、`-i`（CANN安装路径）解析命令行参数
+  2. **确定CANN路径**：按优先级从 `ASCEND_TOOLKIT_HOME` → `ASCEND_HOME_PATH` → 用户指定路径 → 默认路径查找
+  3. **校验参数**：验证SoC版本必须是 KirinX90/Kirin9030，运行模式必须是 sim/cpu/npu，Kirin SoC不支持cpu模式
+  4. **配置仿真环境**：如果是sim模式，设置模拟器库路径（`LD_LIBRARY_PATH`）和日志目录
+  5. **编译**：清理build目录，通过CMake配置并编译，生成可执行文件
+  6. **生成测试数据**：用 `gen_data.py` 生成输入和golden数据
+  7. **执行**：运行编译出的可执行文件，在模拟器上执行算子
+  8. **验证结果**：用 `verify_result.py` 对比输出与golden数据，确认误差在容忍范围内
+  9. **清理**：删除仿真日志和vcd文件
+
+## 基于本模板开发新自定义算子
+
+以下以开发一个 `mul`（乘法）算子为例，说明如何将本模板改造为新算子工程。
+
+### 步骤1：复制模板并重命名
+
+```bash
+cp -r kirin_add_template mul_template
+cd mul_template
+rm -rf build add_sim *_cpu *_npu cceprint npuchk *.vcd sim_log input/*.bin output/*.bin
+```
+
+### 步骤2：修改 `run.sh` — FILE_NAME
+
+`run.sh:27` 中的 `FILE_NAME` 控制编译产物名称和CMake target名。改为新算子名：
+
+```bash
+FILE_NAME="mul"    # 原: FILE_NAME="add"
+```
+
+### 步骤3：替换 `add_custom.cpp` — 核函数实现
+
+这是最核心的修改，整个文件替换为新算子实现。需要修改的内容：
+
+1. **计算常量**（行19-24）：根据新算子的数据总量、核心数、分块策略调整
+   ```cpp
+   constexpr int32_t TOTAL_LENGTH = 8 * 2048;
+   constexpr int32_t USE_CORE_NUM = 1;
+   constexpr int32_t BLOCK_LENGTH = TOTAL_LENGTH / USE_CORE_NUM;
+   constexpr int32_t TILE_NUM = 8;
+   constexpr int32_t BUFFER_NUM = 2;
+   constexpr int32_t TILE_LENGTH = BLOCK_LENGTH / TILE_NUM / BUFFER_NUM;
+   ```
+
+2. **Kernel类名和逻辑**：替换类名，修改 `Init`/`Process`/`CopyIn`/`Compute`/`CopyOut`
+   - `Init`：调整 GlobalTensor 数量和类型
+   - `Compute`：将 `AscendC::Add` 替换为目标计算接口，例如 `AscendC::Mul`
+   - `CopyIn`/`CopyOut`：调整输入/输出队列数量和搬运逻辑
+
+3. **核函数入口**（行84-88）：改名为新算子名，调整参数列表
+   ```cpp
+   extern "C" __global__ __aicore__ void mul_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z)
+   {
+       KernelMul op;
+       op.Init(x, y, z);
+       op.Process();
+   }
+   ```
+
+4. **Host调用桥接函数**（行94-97）：改函数名，与核函数名一致
+   ```cpp
+   void mul_custom_do(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *x, uint8_t *y, uint8_t *z)
+   {
+       mul_custom<<<blockDim, l2ctrl, stream>>>(x, y, z);
+   }
+   ```
+
+### 步骤4：修改 `main.cpp` — Host侧调用接口
+
+1. **extern声明**（行21/24）：改为新算子名
+   ```cpp
+   extern void mul_custom_do(uint32_t blockDim, void* l2ctrl, void* stream, uint8_t* x, uint8_t* y, uint8_t* z);
+   // CPU分支:
+   extern "C" __global__ __aicore__ void mul_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z);
+   ```
+
+2. **数据大小**（行30-31）：根据新算子的输入/输出shape和数据类型调整
+   ```cpp
+   size_t inputByteSize = 8 * 2048 * sizeof(uint16_t);   // half = uint16_t
+   size_t outputByteSize = 8 * 2048 * sizeof(uint16_t);
+   ```
+
+3. **调用点**（行75）：改函数名，参数与核函数保持一致
+   ```cpp
+   mul_custom_do(blockDim, nullptr, stream, xDevice, yDevice, zDevice);
+   ```
+
+4. **输入数据读取**（行68-69）：如果输入参数数量/名称变化，需同步调整 `ReadFile` 的文件名
+
+### 步骤5：修改 `scripts/gen_data.py` — 生成输入和真值
+
+根据新算子的数学定义修改：
+
+```python
+def gen_golden_data_simple():
+    input_x = np.random.uniform(1, 100, [8, 2048]).astype(np.float16)
+    input_y = np.random.uniform(1, 100, [8, 2048]).astype(np.float16)
+    golden = (input_x * input_y).astype(np.float16)    # 改 Add 为 Mul
+
+    input_x.tofile("./input/input_x.bin")
+    input_y.tofile("./input/input_y.bin")
+    golden.tofile("./output/golden.bin")
+```
+
+需要修改的内容：
+- **Shape**：如果算子shape不同，修改 `[8, 2048]` 及数据类型（`np.float16`/`np.float32`/`np.int32` 等）
+- **输入参数数量**：增减输入文件（如3输入算子需要额外生成 `input_w.bin`）
+- **计算公式**：将 `+` 改为目标运算
+- **输出文件名**：如有多输出，需生成多个 golden 文件
+
+### 步骤6：修改 `scripts/verify_result.py` — 验证逻辑
+
+根据输出数据类型调整：
+
+```python
+# 如输出为 float32:
+output = np.fromfile(output, dtype=np.float32).reshape(-1)
+golden = np.fromfile(golden, dtype=np.float32).reshape(-1)
+
+# 调整容差（float32精度更高时可用更小容差）
+relative_tol = 1e-5
+absolute_tol = 1e-8
+error_tol = 1e-5
+```
+
+### 无需修改的文件
+
+| 文件 | 原因 |
+|------|------|
+| `data_utils.h` | 通用工具函数，与算子无关 |
+| `CMakeLists.txt`（顶层） | 只做 `add_subdirectory(cmake/npu)`，无算子名 |
+| `cmake/npu/CMakeLists.txt` | 使用 `${smoke_testcase}` 变量，由CMake参数传入 |
+| `cmake/Modules/*` | CCE编译器检测模块，通用 |
+| `scripts/acl.json` | ACL配置，默认为空 `{}` |
+
+### 常见扩展场景的修改要点
+
+| 场景 | 需修改的文件 |
+|------|-------------|
+| 改数据类型（half→float） | `add_custom.cpp`（GlobalTensor/LocalTensor类型）、`main.cpp`（sizeof）、`gen_data.py`（dtype）、`verify_result.py`（dtype+容差） |
+| 改Shape | `add_custom.cpp`（TOTAL_LENGTH等常量）、`main.cpp`（inputByteSize）、`gen_data.py`（shape） |
+| 多输入（3个以上） | `add_custom.cpp`（增加队列和GlobalTensor）、`main.cpp`（增加malloc/read/memcpy）、`gen_data.py`（增加输入文件） |
+| 多输出 | `add_custom.cpp`（增加输出队列）、`main.cpp`（增加malloc+WriteFile）、`gen_data.py`（增加golden文件）、`verify_result.py`（多输出对比） |
+| 多核并行 | `add_custom.cpp`（调整USE_CORE_NUM和GetBlockIdx逻辑）、`main.cpp`（调整blockDim） |
+
+### 关键概念速查
+
+- **核函数**：`extern "C" __global__ __aicore__` 修饰的函数，是NPU上的执行入口
+- **`_do` 桥接函数**：在核函数实现文件中用 `<<<>>>` 调用符包装核函数，供 `main.cpp` 的Host侧调用
+- **`ASCENDC_CPU_DEBUG` 宏**：区分CPU调测分支和NPU/仿真分支，Kirin SoC目前仅支持NPU/仿真分支
+- **双缓冲（BUFFER_NUM=2）**：通过 `InitBuffer(que, 2, size)` 开启双缓冲实现流水线，一个buffer搬运时另一个buffer计算

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/kirin_add_template/UPSTREAM.md

@@ -0,0 +1,35 @@
+# 上游来源
+
+本目录为外部仓库样例的本地拷贝，仅用作 `ascendc-direct-invoke-template` skill 在 Kirin 平台开发场景下的参考工程模板。
+
+| 项 | 值 |
+|----|----|
+| 上游仓库 | https://gitcode.com/cann/cann-recipes-harmony-infer |
+| 上游路径 | `ops/ascendc/AddKernelInvocation/` |
+| 拷贝时 commit | `2006a857b301893453873919f6f5c01fdccf70c1` |
+| 拷贝日期 | 2026-05-23 |
+| 上游 License | Apache License 2.0 |
+
+## 与上游的差异
+
+- **已剔除运行时产物**：`build/`、`add_sim`（编译后的可执行文件）。
+- **Kirin 平台适配修改**：
+  - `add_custom.cpp:20`：`USE_CORE_NUM` 从 8 改为 1（Kirin 单核）
+  - `add_custom.cpp:77-78`：`TQue` 队列深度参数从 `BUFFER_NUM` 改为 `1`
+  - `main.cpp:29`：`blockDim` 从 8 改为 1（与 USE_CORE_NUM 一致）
+- **新增文件**：`README.md`，替换上游原始 README，改为 Kirin 适配说明文档（含目录结构、运行方式、基于模板开发新算子的完整指引）。
+- **其余源文件未做修改**：`run.sh`、`data_utils.h`、`CMakeLists.txt`、`cmake/*`、`scripts/*`、`input/.keep`、`output/.keep`。
+
+## 同步策略
+
+本地拷贝不会自动跟随上游更新。若需要同步：
+
+1. 从上游对应 commit 重新拷贝（参考本文件「拷贝时 commit」字段升级到新版本）。
+2. 重新应用「与上游的差异」中列出的 Kirin 适配修改（USE_CORE_NUM、TQue 队列深度、blockDim）。
+3. 更新本 README.md 的 Kirin 适配说明（如有新内容需补充）。
+4. 在 CHANGELOG 记录同步动作与新 commit 哈希。
+5. 同步时再次剔除 build/、可执行文件、`*.bin`。
+
+## License
+
+上游为 Apache License 2.0。本目录文件均沿用该许可。详见 [上游 LICENSE](https://gitcode.com/cann/cann-recipes-harmony-infer/blob/master/LICENSE)。

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/kirin_add_template/add_custom.cpp

@@ -0,0 +1,99 @@
+/**
+ * ----------------------------------------------------------------------------------------------------------
+ * 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 "kernel_operator.h"
+#include "add_custom_tiling.h"
+
+class KernelAdd {
+public:
+    __aicore__ inline KernelAdd() {}
+    __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);
+        }
+
+        xGm.SetGlobalBuffer((__gm__ half *)x + blockIdx * tiling->numPerCore, total);
+        yGm.SetGlobalBuffer((__gm__ half *)y + blockIdx * tiling->numPerCore, total);
+        zGm.SetGlobalBuffer((__gm__ half *)z + blockIdx * tiling->numPerCore, total);
+
+        pipe.InitBuffer(inQueueX, DOUBLE_BUFFER, TILE_LENGTH * sizeof(half));
+        pipe.InitBuffer(inQueueY, DOUBLE_BUFFER, TILE_LENGTH * sizeof(half));
+        pipe.InitBuffer(outQueueZ, DOUBLE_BUFFER, TILE_LENGTH * sizeof(half));
+    }
+
+    __aicore__ inline void Process()
+    {
+        for (uint32_t i = 0; i < tileNum; i++) {
+            uint32_t count = (i == tileNum - 1) ? tailTileElementNum : TILE_LENGTH;
+            CopyIn(i, count);
+            Compute(count);
+            CopyOut(i, count);
+        }
+    }
+
+private:
+    __aicore__ inline void CopyIn(int32_t progress, uint32_t count)
+    {
+        AscendC::LocalTensor<half> xLocal = inQueueX.AllocTensor<half>();
+        AscendC::LocalTensor<half> yLocal = inQueueY.AllocTensor<half>();
+        AscendC::DataCopy(xLocal, xGm[progress * TILE_LENGTH], count);
+        AscendC::DataCopy(yLocal, yGm[progress * TILE_LENGTH], count);
+        inQueueX.EnQue(xLocal);
+        inQueueY.EnQue(yLocal);
+    }
+    __aicore__ inline void Compute(uint32_t count)
+    {
+        AscendC::LocalTensor<half> xLocal = inQueueX.DeQue<half>();
+        AscendC::LocalTensor<half> yLocal = inQueueY.DeQue<half>();
+        AscendC::LocalTensor<half> zLocal = outQueueZ.AllocTensor<half>();
+        AscendC::Add(zLocal, xLocal, yLocal, count);
+        outQueueZ.EnQue<half>(zLocal);
+        inQueueX.FreeTensor(xLocal);
+        inQueueY.FreeTensor(yLocal);
+    }
+    __aicore__ inline void CopyOut(int32_t progress, uint32_t count)
+    {
+        AscendC::LocalTensor<half> zLocal = outQueueZ.DeQue<half>();
+        AscendC::DataCopy(zGm[progress * TILE_LENGTH], zLocal, count);
+        outQueueZ.FreeTensor(zLocal);
+    }
+
+private:
+    AscendC::TPipe pipe;
+    const __gm__ AddTilingData* tiling;
+    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueueX, inQueueY;
+    AscendC::TQue<AscendC::TPosition::VECOUT, 1> outQueueZ;
+    AscendC::GlobalTensor<half> xGm;
+    AscendC::GlobalTensor<half> yGm;
+    AscendC::GlobalTensor<half> zGm;
+
+    uint32_t total;
+    uint32_t tileNum;
+    uint32_t tailTileElementNum;
+};
+
+extern "C" __global__ __aicore__ void add_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling)
+{
+    KernelAdd op;
+    op.Init(x, y, z, (__gm__ AddTilingData*)tiling);
+    op.Process();
+}

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/kirin_add_template/add_custom_tiling.h

@@ -0,0 +1,25 @@
+/**
+ * ----------------------------------------------------------------------------------------------------------
+ * 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 ADD_CUSTOM_TILING_H
+#define ADD_CUSTOM_TILING_H
+
+constexpr uint32_t TILE_LENGTH = 1024;
+constexpr int32_t  DOUBLE_BUFFER = 2;   
+
+struct AddTilingData {
+    uint32_t totalLength;
+    uint32_t blockNum;
+    uint32_t numPerCore;
+    uint32_t tailNumLastCore;
+};
+
+#endif // ADD_CUSTOM_TILING_H

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/kirin_add_template/cmake/Modules/CMakeCCECompiler.cmake.in

@@ -0,0 +1,15 @@
+# ----------------------------------------------------------------------------------------------------------
+# 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.
+# ----------------------------------------------------------------------------------------------------------
+
+set(CMAKE_CCE_COMPILER "@CMAKE_CCE_COMPILER@")
+set(CMAKE_CCE_COMPILER_LOADED 1)
+set(CMAKE_CCE_SOURCE_FILE_EXTENSIONS @CMAKE_CCE_SOURCE_FILE_EXTENSIONS@)
+set(CMAKE_CCE_OUTPUT_EXTENSION @CMAKE_CCE_OUTPUT_EXTENSION@)
+set(CMAKE_CCE_COMPILER_ENV_VAR "@CMAKE_CCE_COMPILER_ENV_VAR@")

package/assets/ops-direct-invoke/skills/ascendc-direct-invoke-template/references/kirin_add_template/cmake/Modules/CMakeCCEInformation.cmake

@@ -0,0 +1,51 @@
+# ----------------------------------------------------------------------------------------------------------
+# 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(CMakeCommonLanguageInclude)
+
+set(CMAKE_INCLUDE_FLAG_CCE "-I")
+
+if(UNIX)
+  set(CMAKE_CCE_OUTPUT_EXTENSION .o)
+else()
+  set(CMAKE_CCE_OUTPUT_EXTENSION .obj)
+endif()
+
+set(_INCLUDED_FILE 0)
+set(CMAKE_SHARED_LIBRARY_CCE_FLAGS -fPIC)
+set(CMAKE_SHARED_LIBRARY_CREATE_CCE_FLAGS -shared)
+set(CMAKE_STATIC_LIBRARY_CREATE_CCE_FLAGS "--cce-build-static-lib")
+set(CMAKE_LIBRARY_CREATE_CCE_FLAGS "--cce-fatobj-link")
+
+if(NOT CMAKE_CCE_COMPILE_OBJECT)
+    set(CMAKE_CCE_COMPILE_OBJECT
+      "<CMAKE_CCE_COMPILER> -xcce <DEFINES> <INCLUDES>${__IMPLICIT_INCLUDES} ${_CMAKE_CCE_BUILTIN_INCLUDE_PATH} <FLAGS> ${_CMAKE_COMPILE_AS_CCE_FLAG} ${_CMAKE_CCE_COMPILE_OPTIONS} ${_CMAKE_CCE_COMMON_COMPILE_OPTIONS} <CMAKE_SHARED_LIBRARY_CCE_FLAGS> -pthread -o <OBJECT> -c <SOURCE>")
+endif()
+
+if(NOT CMAKE_CCE_CREATE_SHARED_LIBRARY)
+  set(CMAKE_CCE_CREATE_SHARED_LIBRARY
+      "<CMAKE_CCE_COMPILER> ${CMAKE_LIBRARY_CREATE_CCE_FLAGS} <CMAKE_SHARED_LIBRARY_CCE_FLAGS> <LANGUAGE_COMPILE_FLAGS> <LINK_FLAGS> <CMAKE_SHARED_LIBRARY_CREATE_CCE_FLAGS> <SONAME_FLAG><TARGET_SONAME> -o <TARGET> <OBJECTS>")
+endif()
+
+if(NOT CMAKE_CCE_CREATE_STATIC_LIBRARY)
+  set(CMAKE_CCE_CREATE_STATIC_LIBRARY
+      "<CMAKE_CCE_COMPILER> ${CMAKE_LIBRARY_CREATE_CCE_FLAGS} <CMAKE_SHARED_LIBRARY_CCE_FLAGS> <LANGUAGE_COMPILE_FLAGS> <LINK_FLAGS> <CMAKE_STATIC_LIBRARY_CREATE_CCE_FLAGS> <SONAME_FLAG><TARGET_SONAME> -o <TARGET> <OBJECTS>")
+endif()
+
+if(NOT CMAKE_CCE_CREATE_SHARED_MODULE)
+  set(CMAKE_CCE_CREATE_SHARED_MODULE ${CMAKE_CCE_CREATE_SHARED_LIBRARY})
+endif()
+
+if(NOT CMAKE_CCE_LINK_EXECUTABLE)
+  set(CMAKE_CCE_LINK_EXECUTABLE
+    "<CMAKE_CCE_COMPILER> ${CMAKE_LIBRARY_CREATE_CCE_FLAGS} <FLAGS> <CMAKE_CCE_LINK_FLAGS> <LINK_FLAGS> <OBJECTS> -o <TARGET> <LINK_LIBRARIES>${__IMPLICIT_LINKS}")
+endif()
+
+set(CMAKE_CCE_INFORMATION_LOADED 1)