diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/CMakeLists.txt b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..e31108ae16cf7c722a9d758e17b90610ad940db7 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/CMakeLists.txt @@ -0,0 +1,49 @@ +cmake_minimum_required(VERSION 3.16) +project(Ascend_c) + +set(RUN_MODE "npu" CACHE STRING "cpu/sim/npu") +set(SOC_VERSION "Ascend310P3" CACHE STRING "system on chip type") +set(ASCEND_CANN_PACKAGE_PATH "/usr/local/Ascend/ascend-toolkit/latest" + CACHE STRING "ASCEND CANN package installation directory" +) +if(NOT CMAKE_BUILD_TYPE) + set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "Build type Release/Debug (default Debug)" FORCE) +endif() +if(CMAKE_INSTALL_PREFIX STREQUAL /usr/local) + set(CMAKE_INSTALL_PREFIX "${CMAKE_CURRENT_LIST_DIR}/out" CACHE STRING "path for install()" FORCE) +endif() + +# ${KERNEL_FILES} are used to compile library, push files written by ascendc in ${KERNEL_FILES}. +# ref to cmake/npu.cmake ascendc_library, cmake/cpu.cmake add_library +file(GLOB KERNEL_FILES + ${CMAKE_CURRENT_SOURCE_DIR}/add_custom_v1.cpp + ${CMAKE_CURRENT_SOURCE_DIR}/add_custom_v2.cpp + ${CMAKE_CURRENT_SOURCE_DIR}/add_custom_v3.cpp + ${CMAKE_CURRENT_SOURCE_DIR}/add_custom_v4.cpp +) + +if("${RUN_MODE}" STREQUAL "cpu") + include(cmake/cpu_lib.cmake) +elseif("${RUN_MODE}" STREQUAL "sim" OR "${RUN_MODE}" STREQUAL "npu") + include(cmake/npu_lib.cmake) +else() + message("invalid RUN_MODE: ${RUN_MODE}") +endif() +add_executable(ascendc_kernels_bbit ${CMAKE_CURRENT_SOURCE_DIR}/main.cpp) + +target_compile_options(ascendc_kernels_bbit PRIVATE + $:-g>> + -O2 -std=c++17 -D_GLIBCXX_USE_CXX11_ABI=0 -Wall -Werror +) + +target_link_libraries(ascendc_kernels_bbit PRIVATE + $,$>:host_intf_pub>> + $:ascendcl>> + ascendc_kernels_${RUN_MODE} +) + +install(TARGETS ascendc_kernels_bbit + LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR} + ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR} + RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR} +) diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/README.md b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/README.md new file mode 100644 index 0000000000000000000000000000000000000000..bf4099816e4b9d8a2a8ba52d67c30417ab12dd88 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/README.md @@ -0,0 +1,96 @@ +## 目录结构介绍 + +``` +├── KernelLaunch +│ ├── cmake // 编译工程文件 +│ ├── scripts +│ │ ├── gen_data.py // 输入数据和真值数据生成脚本 +│ │ └── verify_result.py // 验证输出数据和真值数据是否一致的验证脚本 +│ ├── add_custom_tiling.h // tiling结构体 +│ ├── add_custom_v1.cpp // 算子kernel实现1：未优化前实现 +│ ├── add_custom_v2.cpp // 算子kernel实现2：基于实现1，实现double buffer +│ ├── add_custom_v3.cpp // 算子kernel实现3：优化double buffer实现，简化判断逻辑，并使用LocalMemAllocator简化代码 +│ ├── add_custom_v4.cpp // 算子kernel实现4：基于add_custom_v3，修改地址分配逻辑，消除bank冲突 +│ ├── CMakeLists.txt // 编译工程文件 +│ ├── data_utils.h // 数据读入写出函数 +│ ├── main.cpp // 主函数，调用算子的应用程序，含CPU域及NPU域调用 +│ └── run.sh // 编译运行算子的脚本 +``` + +## 代码实现介绍 + +本样例中实现的是固定shape为72*4096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。 + + 实现1：请参考[add_custom_v1.cpp](./add_custom_v1.cpp)，使用静态Tensor编程方法，进行add算子的编程。 + + 实现2：请参考[add_custom_v2.cpp](./add_custom_v2.cpp)，优化性能，使用double buffer进行流水排布。 + + 实现3：请参考[add_custom_v3.cpp](./add_custom_v3.cpp)，优化add_custom_v2中反向同步，替换为MTE2等待MTE3执行结束。减少分支判断的同时，算子性能因为double buffer的原因不受影响。另外使用LocalMemAllocator进行线性内存分配，Bank冲突不敏感场景可以使用这种方式简化分配。 + + 实现4：请参考[add_custom_v4.cpp](./add_custom_v4.cpp)，基于add_custom_v3的实现，优化地址分配消除Bank冲突。 + +- 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ACLRT_LAUNCH_KERNEL内核调用宏来完成。 + + 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 + 以命令行方式下载样例代码，master分支为例。 + + ```bash + cd ${git_clone_path}/samples/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch + ``` +- 配置环境变量 + + 请根据当前环境上CANN开发套件包的[安装方式](https://hiascend.com/document/redirect/CannCommunityInstSoftware)，选择对应配置环境变量的命令。 + + - 默认路径，root用户安装CANN软件包 + ```bash + export ASCEND_INSTALL_PATH=/usr/local/Ascend/ascend-toolkit/latest + ``` + - 默认路径，非root用户安装CANN软件包 + ```bash + export ASCEND_INSTALL_PATH=$HOME/Ascend/ascend-toolkit/latest + ``` + - 指定路径install_path，安装CANN软件包 + ```bash + export ASCEND_INSTALL_PATH=${install_path}/ascend-toolkit/latest + ``` +- 样例执行 + + ```bash + bash run.sh -r [RUN_MODE] -v [SOC_VERSION] + ``` + + - RUN_MODE：编译方式，可选择CPU调试，NPU仿真，NPU上板。支持参数为[cpu /sim / npu] + - SOC_VERSION：昇腾AI处理器型号，如果无法确定具体的[SOC_VERSION]，则在安装昇腾AI处理器的服务器执行npu-smi info命令进行查询，在查询到的“Name”前增加Ascend信息，例如“Name”对应取值为xxxyy，实际配置的[SOC_VERSION]值为Ascendxxxyy。支持以下产品型号： + - Atlas A2训练系列产品/Atlas 800I A2推理产品 + + 示例如下，Ascendxxxyy请替换为实际的AI处理器型号。 + + ```bash + bash run.sh -r cpu -v Ascendxxxyy + ``` + +## 更新说明 + + +| 时间 | 更新事项 | +| ---------- | ------------ | +| 2025/09/06 | 新增本readme | diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_tiling.h b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_tiling.h new file mode 100644 index 0000000000000000000000000000000000000000..278a6e336f07522ffd4c66ce1fb0640494607629 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_tiling.h @@ -0,0 +1,17 @@ +/** + * @file add_custom_tiling.h + * + * Copyright (C) 2025. 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. + */ +#ifndef ADD_CUSTOM_TILING_H +#define ADD_CUSTOM_TILING_H +#include + +struct AddCustomTilingData { + uint32_t singleCoreLength; +}; +#endif \ No newline at end of file diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v1.cpp b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v1.cpp new file mode 100644 index 0000000000000000000000000000000000000000..831ae4c3c6a9623281a6140810d8d3abb54974d8 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v1.cpp @@ -0,0 +1,88 @@ +/** + * @file add_custom_v1.cpp + * + * Copyright (C) 2025. 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 "add_custom_tiling.h" +#include "kernel_operator.h" + +using AscendC::TPosition; +namespace { +constexpr uint32_t TILE_LENGTH = 4096; +} + +class KernelAddV1 { +public: + __aicore__ inline KernelAddV1() = default; + __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z, uint32_t singleCoreLength) + { + xGm.SetGlobalBuffer((__gm__ float *)x + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + yGm.SetGlobalBuffer((__gm__ float *)y + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + zGm.SetGlobalBuffer((__gm__ float *)z + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + loopCount = singleCoreLength / TILE_LENGTH; + } + __aicore__ inline void Process() + { + AscendC::LocalTensor xLocal(AscendC::TPosition::VECCALC, xAddr, TILE_LENGTH); + AscendC::LocalTensor yLocal(AscendC::TPosition::VECCALC, yAddr, TILE_LENGTH); + AscendC::LocalTensor zLocal(AscendC::TPosition::VECCALC, zAddr, TILE_LENGTH); + + // one buffer + for (uint32_t i = 0; i < loopCount; i++) { + // dependency of PIPE_V & PIPE_MTE2 caused by xLocal/yLocal between 2 sequential loops + if (i != 0) { + AscendC::WaitFlag(EVENT_ID0); + } + AscendC::DataCopy(xLocal, xGm[i * TILE_LENGTH], TILE_LENGTH); + AscendC::DataCopy(yLocal, yGm[i * TILE_LENGTH], TILE_LENGTH); + // dependency of PIPE_MTE2 & PIPE_V caused by xLocal/yLocal in one single loop + AscendC::SetFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID0); + if (i != 0) { + // dependency of PIPE_MTE3 & PIPE_V caused by zLocal between 2 sequential loops + AscendC::WaitFlag(EVENT_ID0); + } + AscendC::Add(zLocal, xLocal, yLocal, TILE_LENGTH); + if (i != (loopCount - 1)) { + // dependency of PIPE_V & PIPE_MTE2 caused by xLocal/yLocal between 2 sequential loops + AscendC::SetFlag(EVENT_ID0); + } + // dependency of PIPE_V & PIPE_MTE3 caused by zLocal in one single loop + AscendC::SetFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID0); + AscendC::DataCopy(zGm[i * TILE_LENGTH], zLocal, TILE_LENGTH); + if (i != (loopCount - 1)) { + // dependency of PIPE_MTE3 & PIPE_V caused by zLocal between 2 sequential loops + AscendC::SetFlag(EVENT_ID0); + } + } + } + +private: + static constexpr uint32_t xAddr = 0; + static constexpr uint32_t yAddr = TILE_LENGTH * sizeof(float); + static constexpr uint32_t zAddr = TILE_LENGTH * sizeof(float) * 2; + AscendC::GlobalTensor xGm; + AscendC::GlobalTensor yGm; + AscendC::GlobalTensor zGm; + uint32_t loopCount; +}; + +extern "C" __global__ __aicore__ void add_custom_v1(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling) +{ + AscendC::InitSocState(); + KernelAddV1 op; + op.Init(x, y, z, ((__gm__ AddCustomTilingData *)tiling)->singleCoreLength); + op.Process(); +} + +#ifndef ASCENDC_CPU_DEBUG +void add_custom_do_v1(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z, uint8_t *tiling) +{ + add_custom_v1<<>>(x, y, z, tiling); +} +#endif diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v2.cpp b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v2.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2c4525d553c2ceb5b37f3f2044df2418c1c5a26b --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v2.cpp @@ -0,0 +1,145 @@ +/** + * @file add_custom_v2.cpp + * + * Copyright (C) 2025. 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 "add_custom_tiling.h" +#include "kernel_operator.h" + +using AscendC::TPosition; +namespace { +constexpr uint32_t TILE_LENGTH = 4096; +} + +class KernelAddV2 { +public: + __aicore__ inline KernelAddV2() = default; + __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z, uint32_t singleCoreLength) + { + xGm.SetGlobalBuffer((__gm__ float *)x + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + yGm.SetGlobalBuffer((__gm__ float *)y + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + zGm.SetGlobalBuffer((__gm__ float *)z + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + loopCount = singleCoreLength / TILE_LENGTH; + } + __aicore__ inline void Process() + { + // ping + AscendC::LocalTensor xLocalPing(AscendC::TPosition::VECCALC, xAddrPing, TILE_LENGTH); + AscendC::LocalTensor yLocalPing(AscendC::TPosition::VECCALC, yAddrPing, TILE_LENGTH); + AscendC::LocalTensor zLocalPing(AscendC::TPosition::VECCALC, zAddrPing, TILE_LENGTH); + // pong + AscendC::LocalTensor xLocalPong(AscendC::TPosition::VECCALC, xAddrPong, TILE_LENGTH); + AscendC::LocalTensor yLocalPong(AscendC::TPosition::VECCALC, yAddrPong, TILE_LENGTH); + AscendC::LocalTensor zLocalPong(AscendC::TPosition::VECCALC, zAddrPong, TILE_LENGTH); + + // double buffer + for (uint32_t i = 0; i < loopCount / 2; i++) { + // ping part + // dependency of PIPE_V & PIPE_MTE2 caused by xLocalPing/yLocalPing between 2 sequential loops + if (i != 0) { + AscendC::WaitFlag(EVENT_ID0); + } + AscendC::DataCopy(xLocalPing, xGm[2 * i * TILE_LENGTH], TILE_LENGTH); + AscendC::DataCopy(yLocalPing, yGm[2 * i * TILE_LENGTH], TILE_LENGTH); + // dependency of PIPE_MTE2 & PIPE_V caused by xLocalPing/yLocalPing in one single loop + AscendC::SetFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID0); + if (i != 0) { + // dependency of PIPE_MTE3 & PIPE_V caused by zLocalPing between 2 sequential loops + AscendC::WaitFlag(EVENT_ID0); + } + AscendC::Add(zLocalPing, xLocalPing, yLocalPing, TILE_LENGTH); + if (i != (loopCount / 2 - 1)) { + // dependency of PIPE_V & PIPE_MTE2 caused by xLocalPing/yLocalPing between 2 sequential loops + AscendC::SetFlag(EVENT_ID0); + } + // dependency of PIPE_V & PIPE_MTE3 caused by zLocalPing in one single loop + AscendC::SetFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID0); + AscendC::DataCopy(zGm[2 * i * TILE_LENGTH], zLocalPing, TILE_LENGTH); + if (i != (loopCount / 2 - 1)) { + // dependency of PIPE_MTE3 & PIPE_V caused by zLocalPing between 2 sequential loops + AscendC::SetFlag(EVENT_ID0); + } + + // pong part + // dependency of PIPE_V & PIPE_MTE2 caused by xLocalPong/yLocalPong between 2 sequential loops + if (i != 0) { + AscendC::WaitFlag(EVENT_ID1); + } + AscendC::DataCopy(xLocalPong, xGm[(2 * i + 1) * TILE_LENGTH], TILE_LENGTH); + AscendC::DataCopy(yLocalPong, yGm[(2 * i + 1) * TILE_LENGTH], TILE_LENGTH); + // dependency of PIPE_MTE2 & PIPE_V caused by xLocalPong/yLocalPong in one single loop + AscendC::SetFlag(EVENT_ID1); + AscendC::WaitFlag(EVENT_ID1); + if (i != 0) { + // dependency of PIPE_MTE3 & PIPE_V caused by zLocalPong between 2 sequential loops + AscendC::WaitFlag(EVENT_ID1); + } + AscendC::Add(zLocalPong, xLocalPong, yLocalPong, TILE_LENGTH); + if (i != (loopCount / 2 - 1)) { + // dependency of PIPE_V & PIPE_MTE2 caused by xLocalPong/yLocalPong between 2 sequential loops + AscendC::SetFlag(EVENT_ID1); + } + // dependency of PIPE_V & PIPE_MTE3 caused by zLocalPong in one single loop + AscendC::SetFlag(EVENT_ID1); + AscendC::WaitFlag(EVENT_ID1); + AscendC::DataCopy(zGm[(2 * i + 1) * TILE_LENGTH], zLocalPong, TILE_LENGTH); + if (i != (loopCount / 2 - 1)) { + // dependency of PIPE_MTE3 & PIPE_V caused by zLocalPong between 2 sequential loops + AscendC::SetFlag(EVENT_ID1); + } + } + + // tail block + if (loopCount % 2 != 0) { + // dependency of PIPE_V & PIPE_MTE2 caused by xLocalPing/yLocalPing with the previous for loop + AscendC::SetFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID0); + AscendC::DataCopy(xLocalPing, xGm[(loopCount - 1) * TILE_LENGTH], TILE_LENGTH); + AscendC::DataCopy(yLocalPing, yGm[(loopCount - 1) * TILE_LENGTH], TILE_LENGTH); + // dependency of PIPE_MTE2 & PIPE_V caused by xLocalPing/yLocalPing in one loop + AscendC::SetFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID0); + // dependency of PIPE_MTE3 & PIPE_V caused by zLocalPing with the previous for loop + AscendC::SetFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID0); + AscendC::Add(zLocalPing, xLocalPing, yLocalPing, TILE_LENGTH); + // dependency of PIPE_V & PIPE_MTE3 caused by zLocalPing in one loop + AscendC::SetFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID0); + AscendC::DataCopy(zGm[(loopCount - 1) * TILE_LENGTH], zLocalPing, TILE_LENGTH); + } + } + +private: + static constexpr uint32_t xAddrPing = 0; + static constexpr uint32_t yAddrPing = TILE_LENGTH * sizeof(float); + static constexpr uint32_t zAddrPing = TILE_LENGTH * sizeof(float) * 2; + static constexpr uint32_t xAddrPong = TILE_LENGTH * sizeof(float) * 3; + static constexpr uint32_t yAddrPong = TILE_LENGTH * sizeof(float) * 4; + static constexpr uint32_t zAddrPong = TILE_LENGTH * sizeof(float) * 5; + AscendC::GlobalTensor xGm; + AscendC::GlobalTensor yGm; + AscendC::GlobalTensor zGm; + uint32_t loopCount; +}; + +extern "C" __global__ __aicore__ void add_custom_v2(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling) +{ + AscendC::InitSocState(); + KernelAddV2 op; + op.Init(x, y, z, ((__gm__ AddCustomTilingData *)tiling)->singleCoreLength); + op.Process(); +} + +#ifndef ASCENDC_CPU_DEBUG +void add_custom_do_v2(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z, uint8_t *tiling) +{ + add_custom_v2<<>>(x, y, z, tiling); +} +#endif diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v3.cpp b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v3.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d424b54f1336823fbdbcb5f6629a4a37f5541b96 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v3.cpp @@ -0,0 +1,90 @@ +/** + * @file add_custom_v3.cpp + * + * Copyright (C) 2025. 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 "add_custom_tiling.h" +#include "kernel_operator.h" + +using AscendC::TPosition; +namespace { +constexpr uint32_t TILE_LENGTH = 4096; +} + +class KernelAddV3 { +public: + __aicore__ inline KernelAddV3() = default; + __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z, uint32_t singleCoreLength) + { + xGm.SetGlobalBuffer((__gm__ float *)x + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + yGm.SetGlobalBuffer((__gm__ float *)y + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + zGm.SetGlobalBuffer((__gm__ float *)z + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + loopCount = singleCoreLength / TILE_LENGTH; + } + + __aicore__ inline void Process() + { + // use local memory allocator to simplify memor allocation + AscendC::LocalMemAllocator ubAllocator; + // ping + AscendC::LocalTensor xLocalPing = ubAllocator.Alloc(); + AscendC::LocalTensor yLocalPing = ubAllocator.Alloc(); + AscendC::LocalTensor zLocalPing = ubAllocator.Alloc(); + // pong + AscendC::LocalTensor xLocalPong = ubAllocator.Alloc(); + AscendC::LocalTensor yLocalPong = ubAllocator.Alloc(); + AscendC::LocalTensor zLocalPong = ubAllocator.Alloc(); + + // double buffer + AscendC::SetFlag(EVENT_ID0); + AscendC::SetFlag(EVENT_ID1); + for (uint32_t i = 0; i < loopCount; i++) { + int32_t eventID = (i % 2 == 0 ? EVENT_ID0 : EVENT_ID1); + AscendC::LocalTensor &xLocal = (i % 2 == 0 ? xLocalPing : xLocalPong); + AscendC::LocalTensor &yLocal = (i % 2 == 0 ? yLocalPing : yLocalPong); + AscendC::LocalTensor &zLocal = (i % 2 == 0 ? zLocalPing : zLocalPong); + // dependency of PIPE_MTE3 & PIPE_MTE2 caused by xLocal/yLocal between 2 sequential loops + AscendC::WaitFlag(eventID); + AscendC::DataCopy(xLocal, xGm[i * TILE_LENGTH], TILE_LENGTH); + AscendC::DataCopy(yLocal, yGm[i * TILE_LENGTH], TILE_LENGTH); + + // dependency of PIPE_MTE2 & PIPE_V caused by xLocal/yLocal in one single loop + AscendC::SetFlag(eventID); + AscendC::WaitFlag(eventID); + AscendC::Add(zLocal, xLocal, yLocal, TILE_LENGTH); + // dependency of PIPE_V & PIPE_MTE3 caused by zLocal in one single loop + AscendC::SetFlag(eventID); + AscendC::WaitFlag(eventID); + AscendC::DataCopy(zGm[i * TILE_LENGTH], zLocal, TILE_LENGTH); + // dependency of PIPE_MTE3 & PIPE_MTE2 caused by zLocal between 2 sequential loops + AscendC::SetFlag(eventID); + } + AscendC::WaitFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID1); + } + +private: + AscendC::GlobalTensor xGm; + AscendC::GlobalTensor yGm; + AscendC::GlobalTensor zGm; + uint32_t loopCount; +}; + +extern "C" __global__ __aicore__ void add_custom_v3(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling) +{ + AscendC::InitSocState(); + KernelAddV3 op; + op.Init(x, y, z, ((__gm__ AddCustomTilingData *)tiling)->singleCoreLength); + op.Process(); +} + +#ifndef ASCENDC_CPU_DEBUG +void add_custom_do_v3(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z, uint8_t *tiling) +{ + add_custom_v3<<>>(x, y, z, tiling); +} +#endif diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v4.cpp b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v4.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ad45b3b44fb055c33a89398a6f49d82e3820281d --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/add_custom_v4.cpp @@ -0,0 +1,99 @@ +/** + * @file add_custom_v4.cpp + * + * Copyright (C) 2025. 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 "add_custom_tiling.h" +#include "kernel_operator.h" + +using AscendC::TPosition; +namespace { +constexpr uint32_t TILE_LENGTH = 4096; +constexpr uint32_t DST_START_ADDRESS = 0x20000; +} + +class KernelAddV4 { +public: + __aicore__ inline KernelAddV4() = default; + __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z, uint32_t singleCoreLength) + { + xGm.SetGlobalBuffer((__gm__ float *)x + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + yGm.SetGlobalBuffer((__gm__ float *)y + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + zGm.SetGlobalBuffer((__gm__ float *)z + AscendC::GetBlockIdx() * singleCoreLength, singleCoreLength); + loopCount = singleCoreLength / TILE_LENGTH; + } + + __aicore__ inline void Process() + { + // ping + AscendC::LocalTensor xLocalPing(AscendC::TPosition::VECCALC, xAddrPing, TILE_LENGTH); + AscendC::LocalTensor yLocalPing(AscendC::TPosition::VECCALC, yAddrPing, TILE_LENGTH); + AscendC::LocalTensor zLocalPing(AscendC::TPosition::VECCALC, zAddrPing, TILE_LENGTH); + // pong + AscendC::LocalTensor xLocalPong(AscendC::TPosition::VECCALC, xAddrPong, TILE_LENGTH); + AscendC::LocalTensor yLocalPong(AscendC::TPosition::VECCALC, yAddrPong, TILE_LENGTH); + AscendC::LocalTensor zLocalPong(AscendC::TPosition::VECCALC, zAddrPong, TILE_LENGTH); + + // double buffer + AscendC::SetFlag(EVENT_ID0); + AscendC::SetFlag(EVENT_ID1); + for (uint32_t i = 0; i < loopCount; i++) { + int32_t eventID = (i % 2 == 0 ? EVENT_ID0 : EVENT_ID1); + AscendC::LocalTensor &xLocal = (i % 2 == 0 ? xLocalPing : xLocalPong); + AscendC::LocalTensor &yLocal = (i % 2 == 0 ? yLocalPing : yLocalPong); + AscendC::LocalTensor &zLocal = (i % 2 == 0 ? zLocalPing : zLocalPong); + // dependency of PIPE_MTE3 & PIPE_MTE2 caused by xLocal/yLocal between 2 sequential loops + AscendC::WaitFlag(eventID); + AscendC::DataCopy(xLocal, xGm[i * TILE_LENGTH], TILE_LENGTH); + AscendC::DataCopy(yLocal, yGm[i * TILE_LENGTH], TILE_LENGTH); + + // dependency of PIPE_MTE2 & PIPE_V caused by xLocal/yLocal in one single loop + AscendC::SetFlag(eventID); + AscendC::WaitFlag(eventID); + AscendC::Add(zLocal, xLocal, yLocal, TILE_LENGTH); + // dependency of PIPE_V & PIPE_MTE3 caused by zLocal in one single loop + AscendC::SetFlag(eventID); + AscendC::WaitFlag(eventID); + AscendC::DataCopy(zGm[i * TILE_LENGTH], zLocal, TILE_LENGTH); + // dependency of PIPE_MTE3 & PIPE_MTE2 caused by zLocal between 2 sequential loops + AscendC::SetFlag(eventID); + } + AscendC::WaitFlag(EVENT_ID0); + AscendC::WaitFlag(EVENT_ID1); + } + +private: + // according to bank conflict rule: + // rr conflict happened when 2 read requests are in the same bank group + // rw conflict happened when read and write requests are in the same bank + // so we adjust the address to avoid bank conflicts + static constexpr uint32_t xAddrPing = 0x0; + static constexpr uint32_t yAddrPing = TILE_LENGTH * sizeof(float) + 256; + static constexpr uint32_t zAddrPing = DST_START_ADDRESS; + static constexpr uint32_t xAddrPong = TILE_LENGTH * sizeof(float) * 2 + 256; + static constexpr uint32_t yAddrPong = TILE_LENGTH * sizeof(float) * 3 + 512; + static constexpr uint32_t zAddrPong = DST_START_ADDRESS + TILE_LENGTH * sizeof(float); + AscendC::GlobalTensor xGm; + AscendC::GlobalTensor yGm; + AscendC::GlobalTensor zGm; + uint32_t loopCount; +}; + +extern "C" __global__ __aicore__ void add_custom_v4(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling) +{ + AscendC::InitSocState(); + KernelAddV4 op; + op.Init(x, y, z, ((__gm__ AddCustomTilingData *)tiling)->singleCoreLength); + op.Process(); +} + +#ifndef ASCENDC_CPU_DEBUG +void add_custom_do_v4(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z, uint8_t *tiling) +{ + add_custom_v4<<>>(x, y, z, tiling); +} +#endif diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/cmake/cpu_lib.cmake b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/cmake/cpu_lib.cmake new file mode 100644 index 0000000000000000000000000000000000000000..5362c8b5a53b1f730ac6fe542ee226a42dff40ff --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/cmake/cpu_lib.cmake @@ -0,0 +1,9 @@ +if(NOT DEFINED ENV{CMAKE_PREFIX_PATH}) + set(CMAKE_PREFIX_PATH ${ASCEND_CANN_PACKAGE_PATH}/tools/tikicpulib/lib/cmake) +endif() +find_package(tikicpulib REQUIRED) + +add_library(ascendc_kernels_${RUN_MODE} SHARED ${KERNEL_FILES}) +target_link_libraries(ascendc_kernels_${RUN_MODE} PUBLIC tikicpulib::${SOC_VERSION}) +target_compile_options(ascendc_kernels_${RUN_MODE} PRIVATE -g -O0 -std=c++17) +install(TARGETS ascendc_kernels_${RUN_MODE} DESTINATION ${CMAKE_INSTALL_LIBDIR}) diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/cmake/npu_lib.cmake b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/cmake/npu_lib.cmake new file mode 100644 index 0000000000000000000000000000000000000000..f92b095d1f4f258af274b98dfcba2dccf6165b30 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/cmake/npu_lib.cmake @@ -0,0 +1,11 @@ +if(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake) + set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake) +elseif(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/tools/tikcpp/ascendc_kernel_cmake) + set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/tools/tikcpp/ascendc_kernel_cmake) +else() + message(FATAL_ERROR "ascendc_kernel_cmake does not exist ,please check whether the cann package is installed") +endif() +include(${ASCENDC_CMAKE_DIR}/ascendc.cmake) + +# ascendc_library use to add kernel file to generate ascendc library +ascendc_library(ascendc_kernels_${RUN_MODE} SHARED ${KERNEL_FILES}) diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/data_utils.h b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/data_utils.h new file mode 100644 index 0000000000000000000000000000000000000000..9d34457807dda2d0c79ce19eadc7fcbd4667ece2 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/data_utils.h @@ -0,0 +1,203 @@ +/** + * @file data_utils.h + * + * Copyright (C) 2025. 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. + */ +#ifndef DATA_UTILS_H +#define DATA_UTILS_H +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include + +#include "acl/acl.h" + +typedef enum { + DT_UNDEFINED = -1, + FLOAT = 0, + HALF = 1, + INT8_T = 2, + INT32_T = 3, + UINT8_T = 4, + INT16_T = 6, + UINT16_T = 7, + UINT32_T = 8, + INT64_T = 9, + UINT64_T = 10, + DOUBLE = 11, + BOOL = 12, + STRING = 13, + COMPLEX64 = 16, + COMPLEX128 = 17, + BF16 = 27 +} printDataType; + +#define INFO_LOG(fmt, args...) fprintf(stdout, "[INFO] " fmt "\n", ##args) +#define WARN_LOG(fmt, args...) fprintf(stdout, "[WARN] " fmt "\n", ##args) +#define ERROR_LOG(fmt, args...) fprintf(stdout, "[ERROR] " fmt "\n", ##args) +#define CHECK_ACL(x) \ + do { \ + aclError __ret = x; \ + if (__ret != ACL_ERROR_NONE) { \ + std::cerr << __FILE__ << ":" << __LINE__ << " aclError:" << __ret << std::endl; \ + } \ + } while (0); + +/** + * @brief Read data from file + * @param [in] filePath: file path + * @param [out] fileSize: file size + * @return read result + */ +bool ReadFile(const std::string &filePath, size_t &fileSize, void *buffer, size_t bufferSize) +{ + struct stat sBuf; + int fileStatus = stat(filePath.data(), &sBuf); + if (fileStatus == -1) { + ERROR_LOG("failed to get file"); + return false; + } + if (S_ISREG(sBuf.st_mode) == 0) { + ERROR_LOG("%s is not a file, please enter a file", filePath.c_str()); + return false; + } + + std::ifstream file; + file.open(filePath, std::ios::binary); + if (!file.is_open()) { + ERROR_LOG("Open file failed. path = %s", filePath.c_str()); + return false; + } + + std::filebuf *buf = file.rdbuf(); + size_t size = buf->pubseekoff(0, std::ios::end, std::ios::in); + if (size == 0) { + ERROR_LOG("file size is 0"); + file.close(); + return false; + } + if (size > bufferSize) { + ERROR_LOG("file size is larger than buffer size"); + file.close(); + return false; + } + buf->pubseekpos(0, std::ios::in); + buf->sgetn(static_cast(buffer), size); + fileSize = size; + 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 + */ +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_IWRITE); + if (fd < 0) { + ERROR_LOG("Open file failed. path = %s", filePath.c_str()); + return false; + } + + size_t writeSize = write(fd, buffer, size); + (void)close(fd); + if (writeSize != size) { + ERROR_LOG("Write file Failed."); + return false; + } + + return true; +} + +template void DoPrintData(const T *data, size_t count, size_t elementsPerRow) +{ + assert(elementsPerRow != 0); + for (size_t i = 0; i < count; ++i) { + std::cout << std::setw(10) << data[i]; + if (i % elementsPerRow == elementsPerRow - 1) { + std::cout << std::endl; + } + } +} + +void DoPrintHalfData(const aclFloat16 *data, size_t count, size_t elementsPerRow) +{ + assert(elementsPerRow != 0); + for (size_t i = 0; i < count; ++i) { + std::cout << std::setw(10) << std::setprecision(6) << aclFloat16ToFloat(data[i]); + if (i % elementsPerRow == elementsPerRow - 1) { + std::cout << std::endl; + } + } +} + +void PrintData(const void *data, size_t count, printDataType dataType, size_t elementsPerRow = 16) +{ + if (data == nullptr) { + ERROR_LOG("Print data failed. data is nullptr"); + return; + } + + switch (dataType) { + case BOOL: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case INT8_T: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case UINT8_T: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case INT16_T: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case UINT16_T: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case INT32_T: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case UINT32_T: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case INT64_T: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case UINT64_T: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case HALF: + DoPrintHalfData(reinterpret_cast(data), count, elementsPerRow); + break; + case FLOAT: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + case DOUBLE: + DoPrintData(reinterpret_cast(data), count, elementsPerRow); + break; + default: + ERROR_LOG("Unsupported type: %d", dataType); + } + std::cout << std::endl; +} +#endif // DATA_UTILS_H diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/main.cpp b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/main.cpp new file mode 100644 index 0000000000000000000000000000000000000000..caf5653e8d65d315d6221fb21f7725194cf5b50d --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/main.cpp @@ -0,0 +1,148 @@ +/** + * @file main.cpp + * + * Copyright (C) 2025. 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 "add_custom_tiling.h" +#include "data_utils.h" +#ifndef ASCENDC_CPU_DEBUG +#include "acl/acl.h" +extern void add_custom_do_v1(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z, uint8_t *tiling); +extern void add_custom_do_v2(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z, uint8_t *tiling); +extern void add_custom_do_v3(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z, uint8_t *tiling); +extern void add_custom_do_v4(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z, uint8_t *tiling); +using KernelEntry = void (*)(uint32_t, void *, uint8_t *, uint8_t *, uint8_t *, uint8_t *); +#else +#include "tikicpulib.h" +extern "C" __global__ __aicore__ void add_custom_v1(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling); +extern "C" __global__ __aicore__ void add_custom_v2(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling); +extern "C" __global__ __aicore__ void add_custom_v3(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling); +extern "C" __global__ __aicore__ void add_custom_v4(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR tiling); +using KernelEntry = void (*)(GM_ADDR, GM_ADDR, GM_ADDR, GM_ADDR); + +#endif + +struct ArgInfo { + std::string fileName; + size_t length; +}; + +#ifndef ASCENDC_CPU_DEBUG + +void KernelCall(KernelEntry kernelEntry, uint32_t blockDim, void *stream, std::vector &inputsInfo, + std::vector &outputsInfo, uint8_t *tiling) +{ + std::vector inputHost(inputsInfo.size()); + std::vector inputDevice(inputsInfo.size()); + std::vector outputHost(outputsInfo.size()); + std::vector outputDevice(outputsInfo.size()); + uint8_t *tilingDevice; + + CHECK_ACL(aclrtMalloc((void **)(&tilingDevice), sizeof(AddCustomTilingData), ACL_MEM_MALLOC_HUGE_FIRST)); + CHECK_ACL(aclrtMemcpy( + tilingDevice, sizeof(AddCustomTilingData), tiling, sizeof(AddCustomTilingData), ACL_MEMCPY_HOST_TO_DEVICE)); + + for (uint32_t i = 0; i < inputsInfo.size(); i++) { + CHECK_ACL(aclrtMallocHost((void **)(&inputHost[i]), inputsInfo[i].length)); + CHECK_ACL(aclrtMalloc((void **)(&inputDevice[i]), inputsInfo[i].length, ACL_MEM_MALLOC_HUGE_FIRST)); + ReadFile(inputsInfo[i].fileName, inputsInfo[i].length, inputHost[i], inputsInfo[i].length); + CHECK_ACL(aclrtMemcpy( + inputDevice[i], inputsInfo[i].length, inputHost[i], inputsInfo[i].length, ACL_MEMCPY_HOST_TO_DEVICE)); + } + + for (uint32_t i = 0; i < outputsInfo.size(); i++) { + CHECK_ACL(aclrtMallocHost((void **)(&outputHost[i]), outputsInfo[i].length)); + CHECK_ACL(aclrtMalloc((void **)(&outputDevice[i]), outputsInfo[i].length, ACL_MEM_MALLOC_HUGE_FIRST)); + } + + kernelEntry(blockDim, stream, inputDevice[0], inputDevice[1], outputDevice[0], tilingDevice); + CHECK_ACL(aclrtSynchronizeStream(stream)); + + CHECK_ACL(aclrtFree(tilingDevice)); + for (uint32_t i = 0; i < outputsInfo.size(); i++) { + CHECK_ACL(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); + CHECK_ACL(aclrtFree(outputDevice[i])); + CHECK_ACL(aclrtFreeHost(outputHost[i])); + } + + for (uint32_t i = 0; i < inputsInfo.size(); i++) { + CHECK_ACL(aclrtFree(inputDevice[i])); + CHECK_ACL(aclrtFreeHost(inputHost[i])); + } +} + +#else + +#define KernelCall(kernelEntry, blockDim, inputsInfo, outputsInfo, tiling) \ + do { \ + std::vector input(inputsInfo.size()); \ + std::vector output(outputsInfo.size()); \ + \ + for (uint32_t i = 0; i < inputsInfo.size(); i++) { \ + input[i] = (uint8_t *)AscendC::GmAlloc(inputsInfo[i].length); \ + ReadFile(inputsInfo[i].fileName, inputsInfo[i].length, input[i], inputsInfo[i].length); \ + } \ + \ + for (uint32_t i = 0; i < outputsInfo.size(); i++) { \ + output[i] = (uint8_t *)AscendC::GmAlloc(outputsInfo[i].length); \ + } \ + \ + AscendC::SetKernelMode(KernelMode::AIV_MODE); \ + ICPU_RUN_KF(kernelEntry, blockDim, input[0], input[1], output[0], tiling); \ + for (uint32_t i = 0; i < inputsInfo.size(); i++) { \ + AscendC::GmFree((void *)input[i]); \ + } \ + \ + for (uint32_t i = 0; i < outputsInfo.size(); i++) { \ + WriteFile(outputsInfo[i].fileName, output[i], outputsInfo[i].length); \ + AscendC::GmFree((void *)output[i]); \ + } \ + } while (0) + +#endif + +int32_t main(int32_t argc, char *argv[]) +{ + uint32_t blockDim = 8; + // set data length, in this case we use 8 cores and length of each core is 4096 * 9 + uint32_t dataLen = 4096 * 9 * blockDim; + size_t inputByteSize = dataLen * sizeof(float); + size_t outputByteSize = dataLen * sizeof(float); + AddCustomTilingData tiling; + tiling.singleCoreLength = dataLen / blockDim; + + std::vector inputsInfo = {{"./input/input_x.bin", inputByteSize}, {"./input/input_y.bin", inputByteSize}}; + std::vector outputsV1Info = {{"./output/output_z_v1.bin", outputByteSize}}; + std::vector outputsV2Info = {{"./output/output_z_v2.bin", outputByteSize}}; + std::vector outputsV3Info = {{"./output/output_z_v3.bin", outputByteSize}}; + std::vector outputsV4Info = {{"./output/output_z_v4.bin", outputByteSize}}; + +#ifndef ASCENDC_CPU_DEBUG + CHECK_ACL(aclInit(nullptr)); + int32_t deviceId = 0; + CHECK_ACL(aclrtSetDevice(deviceId)); + aclrtStream stream = nullptr; + CHECK_ACL(aclrtCreateStream(&stream)); + + KernelCall(add_custom_do_v1, blockDim, stream, inputsInfo, outputsV1Info, (uint8_t *)&tiling); + KernelCall(add_custom_do_v2, blockDim, stream, inputsInfo, outputsV2Info, (uint8_t *)&tiling); + KernelCall(add_custom_do_v3, blockDim, stream, inputsInfo, outputsV3Info, (uint8_t *)&tiling); + KernelCall(add_custom_do_v4, blockDim, stream, inputsInfo, outputsV4Info, (uint8_t *)&tiling); + + CHECK_ACL(aclrtDestroyStream(stream)); + CHECK_ACL(aclrtResetDevice(deviceId)); + CHECK_ACL(aclFinalize()); +#else + KernelCall(add_custom_v1, blockDim, inputsInfo, outputsV1Info, (uint8_t *)&tiling); + KernelCall(add_custom_v2, blockDim, inputsInfo, outputsV2Info, (uint8_t *)&tiling); + KernelCall(add_custom_v3, blockDim, inputsInfo, outputsV3Info, (uint8_t *)&tiling); + KernelCall(add_custom_v4, blockDim, inputsInfo, outputsV4Info, (uint8_t *)&tiling); +#endif + return 0; +} diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/run.sh b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/run.sh new file mode 100644 index 0000000000000000000000000000000000000000..6c691801e909f79a5b61fa41e6d632d6ec308191 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/run.sh @@ -0,0 +1,121 @@ +#!/bin/bash +CURRENT_DIR=$( + cd $(dirname ${BASH_SOURCE:-$0}) + pwd +) + +BUILD_TYPE="Debug" +INSTALL_PREFIX="${CURRENT_DIR}/out" + +SHORT=r:,v:,i:,b:,p:, +LONG=run-mode:,soc-version:,install-path:,build-type:,install-prefix:, +OPTS=$(getopt -a --options $SHORT --longoptions $LONG -- "$@") +eval set -- "$OPTS" + +while :; do + case "$1" in + -r | --run-mode) + RUN_MODE="$2" + shift 2 + ;; + -v | --soc-version) + SOC_VERSION="$2" + shift 2 + ;; + -i | --install-path) + ASCEND_INSTALL_PATH="$2" + shift 2 + ;; + -b | --build-type) + BUILD_TYPE="$2" + shift 2 + ;; + -p | --install-prefix) + INSTALL_PREFIX="$2" + shift 2 + ;; + --) + shift + break + ;; + *) + echo "[ERROR] Unexpected option: $1" + break + ;; + esac +done + +RUN_MODE_LIST="cpu sim npu" +if [[ " $RUN_MODE_LIST " != *" $RUN_MODE "* ]]; then + echo "ERROR: RUN_MODE error, This sample only support specify cpu, sim or npu!" + exit -1 +fi + +VERSION_LIST="Ascend910B1 Ascend910B2 Ascend910B3 Ascend910B4" +if [[ " $VERSION_LIST " != *" $SOC_VERSION "* ]]; then + echo "ERROR: SOC_VERSION should be in [$VERSION_LIST]" + exit -1 +fi + +if [ -n "$ASCEND_INSTALL_PATH" ]; then + _ASCEND_INSTALL_PATH=$ASCEND_INSTALL_PATH +elif [ -n "$ASCEND_HOME_PATH" ]; then + _ASCEND_INSTALL_PATH=$ASCEND_HOME_PATH +else + if [ -d "$HOME/Ascend/ascend-toolkit/latest" ]; then + _ASCEND_INSTALL_PATH=$HOME/Ascend/ascend-toolkit/latest + else + _ASCEND_INSTALL_PATH=/usr/local/Ascend/ascend-toolkit/latest + fi +fi + +export ASCEND_TOOLKIT_HOME=${_ASCEND_INSTALL_PATH} +export ASCEND_HOME_PATH=${_ASCEND_INSTALL_PATH} +source ${_ASCEND_INSTALL_PATH}/bin/setenv.bash +if [ "${RUN_MODE}" = "sim" ]; then + # in case of running op in simulator, use stub .so instead + export LD_LIBRARY_PATH=${_ASCEND_INSTALL_PATH}/tools/simulator/${SOC_VERSION}/lib:$LD_LIBRARY_PATH +elif [ "${RUN_MODE}" = "cpu" ]; then + export LD_LIBRARY_PATH=${_ASCEND_INSTALL_PATH}/tools/tikicpulib/lib:${_ASCEND_INSTALL_PATH}/tools/tikicpulib/lib/${SOC_VERSION}:${_ASCEND_INSTALL_PATH}/tools/simulator/${SOC_VERSION}/lib:$LD_LIBRARY_PATH +fi + +set -e +rm -rf build out +mkdir -p build +cmake -B build \ + -DRUN_MODE=${RUN_MODE} \ + -DSOC_VERSION=${SOC_VERSION} \ + -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \ + -DCMAKE_INSTALL_PREFIX=${INSTALL_PREFIX} \ + -DASCEND_CANN_PACKAGE_PATH=${_ASCEND_INSTALL_PATH} +cmake --build build -j +cmake --install build + +rm -f ascendc_kernels_bbit +cp ./out/bin/ascendc_kernels_bbit ./ +rm -rf input output +mkdir -p input output +python3 scripts/gen_data.py +( + export LD_LIBRARY_PATH=$(pwd)/out/lib:$(pwd)/out/lib64:${_ASCEND_INSTALL_PATH}/lib64:$LD_LIBRARY_PATH + if [[ "$RUN_WITH_TOOLCHAIN" -eq 1 ]]; then + if [ "${RUN_MODE}" = "npu" ]; then + msprof op --launch-count=4 --output=./prof ./ascendc_kernels_bbit + elif [ "${RUN_MODE}" = "sim" ]; then + msprof op simulator --launch-count=4 --output=./prof ./ascendc_kernels_bbit + elif [ "${RUN_MODE}" = "cpu" ]; then + ./ascendc_kernels_bbit + fi + else + ./ascendc_kernels_bbit + fi +) +# tidy folder by delete log files +if [ "${RUN_MODE}" = "sim" ]; then + rm -f *.log *.dump *.vcd *.toml *_log +fi +md5sum output/*.bin +python3 scripts/verify_result.py output/output_z_v1.bin output/golden.bin +python3 scripts/verify_result.py output/output_z_v2.bin output/golden.bin +python3 scripts/verify_result.py output/output_z_v3.bin output/golden.bin +python3 scripts/verify_result.py output/output_z_v4.bin output/golden.bin diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/scripts/gen_data.py b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/scripts/gen_data.py new file mode 100644 index 0000000000000000000000000000000000000000..b8f7ccb5b94e8dcbf285c6000e0418a2ee7f22e8 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/scripts/gen_data.py @@ -0,0 +1,25 @@ +#!/usr/bin/python3 +# coding=utf-8 +# +# Copyright (C) 2025. 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. +# =============================================================================== + +import numpy as np + + +def gen_golden_data_simple(): + input_x = np.random.uniform(1, 100, [8 * 9, 4096]).astype(np.float32) + input_y = np.random.uniform(1, 100, [8 * 9, 4096]).astype(np.float32) + golden = (input_x + input_y).astype(np.float32) + + input_x.tofile("./input/input_x.bin") + input_y.tofile("./input/input_y.bin") + golden.tofile("./output/golden.bin") + + +if __name__ == "__main__": + gen_golden_data_simple() diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/scripts/verify_result.py b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/scripts/verify_result.py new file mode 100644 index 0000000000000000000000000000000000000000..a5019f30fdf1e34188f6f777e5ef3e4aad3491c2 --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/KernelLaunch/scripts/verify_result.py @@ -0,0 +1,53 @@ +#!/usr/bin/python3 +# coding=utf-8 +# +# Copyright (C) 2025. 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. +# =============================================================================== + +import sys +import numpy as np + +# for float32 +relative_tol = 1e-4 +absolute_tol = 1e-5 +error_tol = 1e-4 + + +def verify_result(output, golden): + output = np.fromfile(output, dtype=np.float32).reshape(-1) + golden = np.fromfile(golden, dtype=np.float32).reshape(-1) + different_element_results = np.isclose(output, + golden, + rtol=relative_tol, + atol=absolute_tol, + equal_nan=True) + different_element_indexes = np.where(different_element_results == False)[0] + for index in range(len(different_element_indexes)): + real_index = different_element_indexes[index] + golden_data = golden[real_index] + output_data = output[real_index] + print( + "data index: %06d, expected: %-.9f, actual: %-.9f, rdiff: %-.6f" % + (real_index, golden_data, output_data, + abs(output_data - golden_data) / golden_data)) + if index == 100: + break + error_ratio = float(different_element_indexes.size) / golden.size + print("error ratio: %.4f, tolerance: %.4f" % (error_ratio, error_tol)) + return error_ratio <= error_tol + + +if __name__ == '__main__': + try: + res = verify_result(sys.argv[1], sys.argv[2]) + if not res: + raise ValueError("[ERROR] result error") + else: + print("test pass") + except Exception as e: + print(e) + sys.exit(1) diff --git a/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/README.md b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/README.md new file mode 100644 index 0000000000000000000000000000000000000000..b6a82a0b1039e68c804b7662f607ea2b8522b01a --- /dev/null +++ b/operator/ascendc/0_introduction/23_static_tensor_programming_kernellaunch/README.md @@ -0,0 +1,69 @@ +## 概述 + +本样例介绍基于静态Tensor方式编程的场景下Add算子的实现方法，并提供核函数直调方法。 + +## 目录结构介绍 + +``` +├── 23_static_tensor_programming_kernellaunch // 使用核函数直调的方式调用Add自定义算子 +│ └── KernelLaunch // Kernel Launch方式调用核函数样例 +``` + +## 算子描述 + +算子实现的是固定shape为72×4096的Add算子。 + +Add的计算公式为： + +```python +z = x + y +``` + +- x：输入，形状为\[72, 4096]，数据类型为float； +- y：输入，形状为\[72, 4096]，数据类型为float； +- z：输出，形状为\[72, 4096]，数据类型为float； + +## 算子规格描述 + + + + + + + + + + + + +
算子类型(OpType)Add
算子输入nameshapedata typeformat
x72 * 4096floatND
y72 * 4096floatND
算子输出y72 * 4096floatND
核函数名add_custom_v1 / add_custom_v2 / add_custom_v3 / add_custom_v4
+ +## 支持的产品型号 + +本样例支持如下产品型号： + +- Atlas A2训练系列产品/Atlas 800I A2推理产品 + +## 编译运行样例算子 + +针对自定义算子工程，编译运行包含如下步骤： + +- 编译自定义算子工程； +- 调用执行自定义算子； + +详细操作如下所示。 + +### 1. 获取源码包 + +编译运行此样例前，请参考[准备：获取样例代码](../README.md#codeready)获取源码包。 + +### 2. 编译运行样例工程 + +- [KernelLaunch样例运行](./KernelLaunch/README.md) + +## 更新说明 + + +| 时间 | 更新事项 | +| ---------- | ---------------- | +| 2025/09/06 | 新增直调方式样例 | diff --git a/operator/ascendc/0_introduction/README.md b/operator/ascendc/0_introduction/README.md index 2f95f076d558c2eb0a59fca318a10fb2b1910e26..44a722d77609bdc2bc4063125b904e61b561d38a 100644 --- a/operator/ascendc/0_introduction/README.md +++ b/operator/ascendc/0_introduction/README.md @@ -37,6 +37,7 @@ | [20_mmad_kernellaunch](./20_mmad_kernellaunch) | 基于Ascend C基础API的Matmul自定义Cube算子及KernelLaunch调用样例 | Atlas 推理系列产品AI Core
Atlas A2训练系列产品/Atlas 800I A2推理产品 | | [21_vectoradd_kernellaunch](./21_vectoradd_kernellaunch) | 基于Ascend C的Add多场景自定义Vector算子的KernelLaunch调用样例 | Atlas A2训练系列产品/Atlas 800I A2推理产品 | [22_baremix_kernellaunch](./22_baremix_kernellaunch) | 通过更底层的编码方式，实现MatmulLeayrelu融合算子的样例 | Atlas A2训练系列产品/Atlas 800I A2推理产品 +| [23_static_tensor_programming_kernellaunch](./23_static_tensor_programming_kernellaunch) | 通过静态Tensor编程方式，实现Add算子的样例 | Atlas A2训练系列产品/Atlas 800I A2推理产品 ## 获取样例代码