已合并
update v9.1.0-beta2 release note #4104
gao_dafa创建于 7月7日
update v9.1.0-beta2 release note #4104
已合并
共 4 个文件变更+108-32
| @@ -1,5 +1,37 @@ | |||
| 1 | # CHANGELOG | 1 | # CHANGELOG |
| 2 | 2 | ||
| 3 | +## v9.1.0-beta.2 | ||
| 4 | +发布日期：2026/05/31 | ||
| 5 | +| CANN版本 | 版本源码标签 | | ||
| 6 | +|--|--| | ||
| 7 | +| [CANN 9.1.0-beta.2](https://www.hiascend.com/developer/download/community/result?module=cann&cann=9.1.0-beta.2) | [v9.1.0-beta.2](https://gitcode.com/cann/asc-devkit/tags/v9.1.0-beta.2) | | ||
| 8 | + | ||
| 9 | +### 🚀 关键特性 | ||
| 10 | +- AscendC框架基础API支持NPU Check（[PR#1557](https://gitcode.com/cann/asc-devkit/pull/1557) [PR#1467](https://gitcode.com/cann/asc-devkit/pull/1467)），增强算子运行时校验能力。 | ||
| 11 | +- SIMD VF内支持printf和reg dump打印（[PR#1605](https://gitcode.com/cann/asc-devkit/pull/1605)），提供调试打印和寄存器数据dump能力。 | ||
| 12 | +- A5支持L1 Tensor数据的DumpTensor（[PR#2175](https://gitcode.com/cann/asc-devkit/pull/2175)），扩展L1层数据调试支持。 | ||
| 13 | +- 编译工程CMakeModule支持CMAKE<LANG>编译选项（[PR#2055](https://gitcode.com/cann/asc-devkit/pull/2055)）；新增optype_collector工具，支持检查optype重名（[PR#285](https://gitcode.com/cann/asc-tools/pull/285)）。 | ||
| 14 | +- 基础API支持ctrl中的功能行为（饱和溢出管理）（[PR#2077](https://gitcode.com/cann/asc-devkit/pull/2077)）。 | ||
| 15 | +- SIMT编程新增ld/st接口（[PR#2058](https://gitcode.com/cann/asc-devkit/pull/2058)）和AddrSpace类接口（[PR#1597](https://gitcode.com/cann/asc-devkit/pull/1597)），丰富SIMT内存访问编程能力。 | ||
| 16 | + | ||
| 17 | +### 🎯 样例更新 | ||
| 18 | +- 最佳实践样例开发：matmul+gelu融合、datacopy优化、bank冲突优化、group_matmul量化组矩阵乘、simt&simd高性能编程（[PR#1814](https://gitcode.com/cann/asc-devkit/pull/1814) [PR#2137](https://gitcode.com/cann/asc-devkit/pull/2137) [PR#2141](https://gitcode.com/cann/asc-devkit/pull/2141) [PR#2166](https://gitcode.com/cann/asc-devkit/pull/2166) [PR#2363](https://gitcode.com/cann/asc-devkit/pull/2363)）。 | ||
| 19 | +- Ascend 950新特性补充及兼容性样例整改：loopmode数据搬运、interleave矢量计算、datacopy_gm2l1、loadmx（Load2DMX）、mmad_mx、data_copy_pad等（[PR#2336](https://gitcode.com/cann/asc-devkit/pull/2336) [PR#1899](https://gitcode.com/cann/asc-devkit/pull/1899) [PR#2124](https://gitcode.com/cann/asc-devkit/pull/2124)）。 | ||
| 20 | +- RegBase新增基础样例：基础算术、数据类型转换、归约、比较、索引等样例（[PR#1459](https://gitcode.com/cann/asc-devkit/pull/1459) [PR#1575](https://gitcode.com/cann/asc-devkit/pull/1575) [PR#2024](https://gitcode.com/cann/asc-devkit/pull/2024)）。 | ||
| 21 | +- 新增SIMD VF print样例和dump样例（[PR#2558](https://gitcode.com/cann/asc-devkit/pull/2558)）。 | ||
| 22 | +- SIMT新增DCache访问优化样例（[PR#2453](https://gitcode.com/cann/asc-devkit/pull/2453)）、基于transpose的仿存合并和bank冲突样例（[PR#1753](https://gitcode.com/cann/asc-devkit/pull/1753)）、最佳实践样例：通过类型对齐提升搬运效率（[PR#2297](https://gitcode.com/cann/asc-devkit/pull/2297)）。 | ||
| 23 | +- SIMT新增功能特性样例：pytorch注册自定义算子（[PR#2769](https://gitcode.com/cann/asc-devkit/pull/2769)）、编译相关样例（动态、静态、分离编译等）（[PR#2356](https://gitcode.com/cann/asc-devkit/pull/2356)）、profiling样例（[PR#1989](https://gitcode.com/cann/asc-devkit/pull/1989)）、内存屏障特性样例（[PR#1923](https://gitcode.com/cann/asc-devkit/pull/1923)）、Warp类特性样例（[PR#2876](https://gitcode.com/cann/asc-devkit/pull/2876)）、simulator样例（[PR#2692](https://gitcode.com/cann/asc-devkit/pull/2692)）、kernel log样例（[PR#2131](https://gitcode.com/cann/asc-devkit/pull/2131)）。 | ||
| 24 | +- SIMT入门样例修改为gather（[PR#2405](https://gitcode.com/cann/asc-devkit/pull/2405)）。 | ||
| 25 | +- 新增Tensor API入门及最佳实践样例：Matmul入门、数据搬入搬出、搬出随路量化、MX FP4最佳实践（[PR#2553](https://gitcode.com/cann/asc-devkit/pull/2553)）。 | ||
| 26 | + | ||
| 27 | +### 📖 资料文档 | ||
| 28 | +- 新增矩阵计算概述和计算分形介绍的文档（[PR#2533](https://gitcode.com/cann/asc-devkit/pull/2533)）。 | ||
| 29 | +- 优化矢量计算API文档，补充指令约束等（[PR#2676](https://gitcode.com/cann/asc-devkit/pull/2676)）。 | ||
| 30 | +- 搭建VitePress文档站点，提供AscendC资料预览功能（[PR#2547](https://gitcode.com/cann/asc-devkit/pull/2547)）。 | ||
| 31 | +- 增加SIMD与SIMT混合编程性能优化概述（[PR#2736](https://gitcode.com/cann/asc-devkit/pull/2736)）。 | ||
| 32 | + | ||
| 33 | +--- | ||
| 34 | + | ||
| 3 | ## v9.1.0-beta.1 | 35 | ## v9.1.0-beta.1 |
| 4 | 发布日期：2026/04/30 | 36 | 发布日期：2026/04/30 |
| 5 | | CANN版本 | 版本源码标签 | | 37 | | CANN版本 | 版本源码标签 | |
| @@ -1,5 +1,37 @@ | |||
| 1 | # CHANGELOG | 1 | # CHANGELOG |
| 2 | 2 | ||
| 3 | +## v9.1.0-beta.2 | ||
| 4 | +Release Date: 2026/05/31 | ||
| 5 | +| CANN Version | Version Source Tag | | ||
| 6 | +|--|--| | ||
| 7 | +| [CANN 9.1.0-beta.2](https://www.hiascend.com/developer/download/community/result?module=cann&cann=9.1.0-beta.2) | [v9.1.0-beta.2](https://gitcode.com/cann/asc-devkit/tags/v9.1.0-beta.2) | | ||
| 8 | + | ||
| 9 | +### 🚀 Key Features | ||
| 10 | +- The AscendC framework basic API supports NPU Check ([PR#1557](https://gitcode.com/cann/asc-devkit/pull/1557) [PR#1467](https://gitcode.com/cann/asc-devkit/pull/1467)), enhancing operator runtime verification capability. | ||
| 11 | +- SIMD VF supports printf and reg dump printing ([PR#1605](https://gitcode.com/cann/asc-devkit/pull/1605)), providing debug printing and register data dump capability. | ||
| 12 | +- A5 supports DumpTensor for L1 Tensor data ([PR#2175](https://gitcode.com/cann/asc-devkit/pull/2175)), extending L1 layer data debugging support. | ||
| 13 | +- The compilation project CMakeModule supports CMAKE<LANG> compilation options ([PR#2055](https://gitcode.com/cann/asc-devkit/pull/2055)); adds the optype_collector tool, supporting optype duplicate name checking ([PR#285](https://gitcode.com/cann/asc-tools/pull/285)). | ||
| 14 | +- The basic API supports functional behavior in ctrl (saturation overflow management) ([PR#2077](https://gitcode.com/cann/asc-devkit/pull/2077)). | ||
| 15 | +- SIMT programming adds ld/st interfaces ([PR#2058](https://gitcode.com/cann/asc-devkit/pull/2058)) and AddrSpace class interfaces ([PR#1597](https://gitcode.com/cann/asc-devkit/pull/1597)), enriching SIMT memory access programming capability. | ||
| 16 | + | ||
| 17 | +### 🎯 Sample Updates | ||
| 18 | +- Best practice sample development: matmul+gelu fusion, datacopy optimization, bank conflict optimization, group_matmul quantized group matrix multiplication, and simt&simd high-performance programming ([PR#1814](https://gitcode.com/cann/asc-devkit/pull/1814) [PR#2137](https://gitcode.com/cann/asc-devkit/pull/2137) [PR#2141](https://gitcode.com/cann/asc-devkit/pull/2141) [PR#2166](https://gitcode.com/cann/asc-devkit/pull/2166) [PR#2363](https://gitcode.com/cann/asc-devkit/pull/2363)). | ||
| 19 | +- Ascend 950 new feature supplements and compatibility sample rectification: loopmode data movement, interleave vector computation, datacopy_gm2l1, loadmx (Load2DMX), mmad_mx, data_copy_pad, and so on ([PR#2336](https://gitcode.com/cann/asc-devkit/pull/2336) [PR#1899](https://gitcode.com/cann/asc-devkit/pull/1899) [PR#2124](https://gitcode.com/cann/asc-devkit/pull/2124)). | ||
| 20 | +- RegBase adds basic samples: basic arithmetic, data type conversion, reduction, comparison, indexing, and other samples ([PR#1459](https://gitcode.com/cann/asc-devkit/pull/1459) [PR#1575](https://gitcode.com/cann/asc-devkit/pull/1575) [PR#2024](https://gitcode.com/cann/asc-devkit/pull/2024)). | ||
| 21 | +- Added SIMD VF print samples and dump samples ([PR#2558](https://gitcode.com/cann/asc-devkit/pull/2558)). | ||
| 22 | +- SIMT adds DCache access optimization samples ([PR#2453](https://gitcode.com/cann/asc-devkit/pull/2453)), transpose-based memory coalescing and bank conflict samples ([PR#1753](https://gitcode.com/cann/asc-devkit/pull/1753)), and best practice samples for improving data movement efficiency through type alignment ([PR#2297](https://gitcode.com/cann/asc-devkit/pull/2297)). | ||
| 23 | +- SIMT adds functional feature samples: PyTorch custom operator registration ([PR#2769](https://gitcode.com/cann/asc-devkit/pull/2769)), compilation-related samples (dynamic, static, separate compilation, and so on) ([PR#2356](https://gitcode.com/cann/asc-devkit/pull/2356)), profiling samples ([PR#1989](https://gitcode.com/cann/asc-devkit/pull/1989)), memory barrier feature samples ([PR#1923](https://gitcode.com/cann/asc-devkit/pull/1923)), Warp class feature samples ([PR#2876](https://gitcode.com/cann/asc-devkit/pull/2876)), simulator samples ([PR#2692](https://gitcode.com/cann/asc-devkit/pull/2692)), and kernel log samples ([PR#2131](https://gitcode.com/cann/asc-devkit/pull/2131)). | ||
| 24 | +- The SIMT getting-started sample is modified to gather ([PR#2405](https://gitcode.com/cann/asc-devkit/pull/2405)). | ||
| 25 | +- Added Tensor API getting-started and best practice samples: Matmul getting started, data movement in/out, data movement out with on-path quantization, and MX FP4 best practices ([PR#2553](https://gitcode.com/cann/asc-devkit/pull/2553)). | ||
| 26 | + | ||
| 27 | +### 📖 Documentation | ||
| 28 | +- Added matrix computation overview and computation fractal introduction documentation ([PR#2533](https://gitcode.com/cann/asc-devkit/pull/2533)). | ||
| 29 | +- Optimized vector computation API documentation, supplementing instruction constraints, and so on ([PR#2676](https://gitcode.com/cann/asc-devkit/pull/2676)). | ||
| 30 | +- Set up a VitePress documentation site, providing AscendC documentation preview functionality ([PR#2547](https://gitcode.com/cann/asc-devkit/pull/2547)). | ||
| 31 | +- Added SIMD and SIMT hybrid programming performance optimization overview ([PR#2736](https://gitcode.com/cann/asc-devkit/pull/2736)). | ||
| 32 | + | ||
| 33 | +--- | ||
| 34 | + | ||
| 3 | ## v9.1.0-beta.1 | 35 | ## v9.1.0-beta.1 |
| 4 | Release Date: 2026/04/30 | 36 | Release Date: 2026/04/30 |
| 5 | | CANN Version | Version Source Tag | | 37 | | CANN Version | Version Source Tag | |
| @@ -17,24 +17,30 @@ | |||
| 17 | </div> | 17 | </div> |
| 18 | 18 | ||
| 19 | ## 🔥Latest News | 19 | ## 🔥Latest News |
| 20 | -[2026/03] v9.0.0-beta.2 版本关键特性 | 20 | +[2026/05] v9.1.0-beta.2 版本关键特性 |
| 21 | ### 🚀 关键特性 | 21 | ### 🚀 关键特性 |
| 22 | -- Ascend 950PR支持SIMD编程模式，提供200+ [API接口](./impl/basic_api/dav_3510)跨代兼容能力，可实现Atlas A2系列产品和Atlas A3系列产品算子平滑迁移。 | 22 | +- Ascend C框架基础API支持NPU Check（[PR#1557](https://gitcode.com/cann/asc-devkit/pull/1557) [PR#1467](https://gitcode.com/cann/asc-devkit/pull/1467)），增强算子运行时校验能力。 |
| 23 | -- Ascend 950PR新增基于Reg的编程方式，提供Reg数据搬运、基础算术、规约计算、同步控制等90+ [Reg编程接口](./impl/basic_api/reg_compute/dav_3510)。 | 23 | +- SIMD VF内支持printf和reg dump打印（[PR#1605](https://gitcode.com/cann/asc-devkit/pull/1605)），提供调试打印和寄存器数据dump能力。 |
| 24 | -- Atlas A2系列产品、Atlas A3系列产品、Ascend 950PR支持[语言扩展层纯C接口](./include/c_api)，支持数组式内存分配与指针型计算接口，提供原生纯 C 编程体验。 | 24 | +- A5支持L1 Tensor数据的DumpTensor（[PR#2175](https://gitcode.com/cann/asc-devkit/pull/2175)），扩展L1层数据调试支持。 |
| 25 | -- Ascend 950PR支持SIMD与SIMT混合编程，提供约700个[SIMT API接口](./include/simt_api)，包含warp、atomic、基本数学计算、类型转换等基础接口。 | 25 | +- 编译工程CMakeModule支持CMAKE<LANG>编译选项（[PR#2055](https://gitcode.com/cann/asc-devkit/pull/2055)）；新增optype_collector工具，支持检查optype重名（[PR#285](https://gitcode.com/cann/asc-tools/pull/285)）。 |
| 26 | -- Ascend 950PR支持通信高阶API的CCU通信接口，提供基于CCU的[Allreduce，Allgather，Reducescatter，AlltoAll等主流通信原语](./impl/adv_api/detail/hccl/impl/platform_v310)；Matmul高阶API新增支持[MXFP4/8低比特数据类型的矩阵运算](./impl/adv_api/detail/matmul/mx_matmul_impl.h)，实现内存占用减半、算力吞吐倍增。 | 26 | +- 基础API支持ctrl中的功能行为（饱和溢出管理）（[PR#2077](https://gitcode.com/cann/asc-devkit/pull/2077)）。 |
| 27 | -- Ascend 950PR新增及兼容支持样例共计约260个，包含[SIMT样例](./examples/03_simt_api)、[SIMD样例](./examples/01_simd_cpp_api)（框架类、基础API、高阶API、最佳实践等），并按照编程模型和样例类别对[样例目录结构进行调整](https://gitcode.com/cann/asc-devkit/pull/1223)，提升样例目录结构的易读性。 | 27 | +- SIMT编程新增ld/st接口（[PR#2058](https://gitcode.com/cann/asc-devkit/pull/2058)）和AddrSpace类接口（[PR#1597](https://gitcode.com/cann/asc-devkit/pull/1597)），丰富SIMT内存访问编程能力。 |
| 28 | -- 融合编译与<<<>>>调用方式支持[CPU模式](https://gitcode.com/cann/asc-tools/pull/138)以及[SIM仿真模式](https://gitcode.com/cann/asc-devkit/blob/master/cmake/asc/asc_modules/CMakeASCInformation.cmake)。 | 28 | +### 🎯 样例更新 |
| 29 | +- 最佳实践样例开发：matmul+gelu融合、datacopy优化、bank冲突优化、group_matmul量化组矩阵乘、simt&simd高性能编程（[PR#1814](https://gitcode.com/cann/asc-devkit/pull/1814) [PR#2137](https://gitcode.com/cann/asc-devkit/pull/2137) [PR#2141](https://gitcode.com/cann/asc-devkit/pull/2141) [PR#2166](https://gitcode.com/cann/asc-devkit/pull/2166) [PR#2363](https://gitcode.com/cann/asc-devkit/pull/2363)）。 | ||
| 30 | +- Ascend 950新特性补充及兼容性样例整改：loopmode数据搬运、interleave矢量计算、datacopy_gm2l1、loadmx（Load2DMX）、mmad_mx、data_copy_pad等（[PR#2336](https://gitcode.com/cann/asc-devkit/pull/2336) [PR#1899](https://gitcode.com/cann/asc-devkit/pull/1899) [PR#2124](https://gitcode.com/cann/asc-devkit/pull/2124)）。 | ||
| 31 | +- RegBase新增基础样例：基础算术、数据类型转换、归约、比较、索引等样例（[PR#1459](https://gitcode.com/cann/asc-devkit/pull/1459) [PR#1575](https://gitcode.com/cann/asc-devkit/pull/1575) [PR#2024](https://gitcode.com/cann/asc-devkit/pull/2024)）。 | ||
| 32 | +- 新增SIMD VF print样例和dump样例（[PR#2558](https://gitcode.com/cann/asc-devkit/pull/2558)）。 | ||
| 33 | +- SIMT新增DCache访问优化样例（[PR#2453](https://gitcode.com/cann/asc-devkit/pull/2453)）、基于transpose的仿存合并和bank冲突样例（[PR#1753](https://gitcode.com/cann/asc-devkit/pull/1753)）、最佳实践样例：通过类型对齐提升搬运效率（[PR#2297](https://gitcode.com/cann/asc-devkit/pull/2297)）。 | ||
| 34 | +- SIMT新增功能特性样例：pytorch注册自定义算子（[PR#2769](https://gitcode.com/cann/asc-devkit/pull/2769)）、编译相关样例（动态、静态、分离编译等）（[PR#2356](https://gitcode.com/cann/asc-devkit/pull/2356)）、profiling样例（[PR#1989](https://gitcode.com/cann/asc-devkit/pull/1989)）、内存屏障特性样例（[PR#1923](https://gitcode.com/cann/asc-devkit/pull/1923)）、Warp类特性样例（[PR#2876](https://gitcode.com/cann/asc-devkit/pull/2876)）、simulator样例（[PR#2692](https://gitcode.com/cann/asc-devkit/pull/2692)）、kernel log样例（[PR#2131](https://gitcode.com/cann/asc-devkit/pull/2131)）。 | ||
| 35 | +- SIMT入门样例修改为gather（[PR#2405](https://gitcode.com/cann/asc-devkit/pull/2405)）。 | ||
| 36 | +- 新增Tensor API入门及最佳实践样例：Matmul入门、数据搬入搬出、搬出随路量化、MX FP4最佳实践（[PR#2553](https://gitcode.com/cann/asc-devkit/pull/2553)）。 | ||
| 29 | ### 📖 资料文档 | 37 | ### 📖 资料文档 |
| 30 | -- 新增90+ [Reg编程接口API](https://gitcode.com/cann/asc-devkit/blob/master/docs/api/SIMD-API/基础API/Reg矢量计算/Reg矢量计算.md)资料，Reg矢量计算API是面向RegBase架构开发的API，用户可通过该API直接对芯片中涉及Vector计算的寄存器进行操作，实现更大的灵活性和更好的性能。 | 38 | +- 新增矩阵计算概述和计算分形介绍的文档（[PR#2533](https://gitcode.com/cann/asc-devkit/pull/2533)）。 |
| 31 | -- 新增SIMT[快速入门](docs/guide/入门教程/快速入门/基于SIMT编程/Gather算子快速入门.md)、[编程模型](docs/guide/编程指南/编程模型/AI-Core-SIMT编程/抽象硬件架构.md)和[算子实现](docs/guide/算子实践参考/SIMT算子实现/基础知识.md)介绍。 | 39 | +- 优化矢量计算API文档，补充指令约束等（[PR#2676](https://gitcode.com/cann/asc-devkit/pull/2676)）。 |
| 32 | -- 新增SIMD与SIMT[混合编程模型](docs/guide/编程指南/高级编程/高级AI-Core编程模型/SIMD与SIMT混合编程/概述.md)、[算子实现](docs/guide/算子实践参考/SIMD与SIMT混合算子实现/基础知识.md)、[性能优化](docs/guide/算子实践参考/SIMD与SIMT混合算子性能优化/内存访问/使用Unified-Buffer提升内存访问效率.md)介绍。 | 40 | +- 搭建VitePress文档站点，提供AscendC资料预览功能（[PR#2547](https://gitcode.com/cann/asc-devkit/pull/2547)）。 |
| 33 | -- 新增[SIMT API](https://gitcode.com/cann/asc-devkit/blob/master/docs/api/SIMT-API/SIMT-API.md)资料章节。 | 41 | +- 增加SIMD与SIMT混合编程性能优化概述（[PR#2736](https://gitcode.com/cann/asc-devkit/pull/2736)）。 |
| 34 | -- 新增[兼容性迁移指南](docs/guide/跨代迁移兼容性指南/Ascend-C-API兼容策略.md)（220x架构版本迁移到351x架构版本）。 | ||
| 35 | -- 昇腾社区中，Ascend C算子开发新增[可视化专区](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/900beta2/opdevg/Ascendcopdevg/atlas_ascendc_map_10_0017.html)，通过视频呈现Cube和Vector算子的执行过程。 | ||
| 36 | 42 | ||
| 37 | -有关所有历史版本及更新的详细信息，请参阅[CHANGELOG.md](./CHANGELOG.md) | 43 | +有关所有历史版本及更新的详细信息，请参阅[CHANGELOG.md](./CHANGELOG.md)。 |
| 38 | 44 | ||
| 39 | ## 🚀概述 | 45 | ## 🚀概述 |
| 40 | 46 | ||
| @@ -62,7 +68,7 @@ Ascend C提供三类接口，均可实现底层的完备编程能力： | |||
| 62 | 68 | ||
| 63 | | API层级 | 目标用户 | 主要用途 | | 69 | | API层级 | 目标用户 | 主要用途 | |
| 64 | |----------|----------|----------| | 70 | |----------|----------|----------| |
| 65 | -| **算子模板库 (CATLASS/ATVOSS等)** | 算法开发人员 | 基于典型算子实现进行自定义扩展，满足特定场景高性能需求 | | 71 | +| **算子模板库 (CATLASS/ATVOSS等)** | 算法开发人员 | 基于典型算子实现进行自定义扩展，满足特定场景高性能需求 | |
| 66 | | **高阶API** |算法开发人员 | 复用通用单核算法，快速完成算法验证 | | 72 | | **高阶API** |算法开发人员 | 复用通用单核算法，快速完成算法验证 | |
| 67 | 73 | ||
| 68 | 74 | ||
| @@ -17,24 +17,30 @@ | |||
| 17 | </div> | 17 | </div> |
| 18 | 18 | ||
| 19 | ## 🔥Latest News | 19 | ## 🔥Latest News |
| 20 | -[2026/03] Key features of v9.0.0-beta.2 | 20 | +[2026/05] Key features of v9.1.0-beta.2 |
| 21 | ### 🚀 Key Features | 21 | ### 🚀 Key Features |
| 22 | -- Ascend 950PR supports SIMD programming mode and provides over 200 [API interfaces](./impl/basic_api/dav_3510) with cross-generation compatibility capabilities, enabling smooth operator migration between Atlas A2 series products and Atlas A3 series products. | 22 | +- The Ascend C framework basic API supports NPU Check ([PR#1557](https://gitcode.com/cann/asc-devkit/pull/1557) [PR#1467](https://gitcode.com/cann/asc-devkit/pull/1467)), enhancing operator runtime verification capability. |
| 23 | -- Ascend 950PR adds Reg-based programming and provides over 90 [Reg programming interfaces](./impl/basic_api/reg_compute/dav_3510), including Reg datamove, basic arithmetic, reduction calculation, and synchronization control. | 23 | +- SIMD VF supports printf and reg dump printing ([PR#1605](https://gitcode.com/cann/asc-devkit/pull/1605)), providing debug printing and register data dump capability. |
| 24 | -- Atlas A2 series products, Atlas A3 series products, and Ascend 950PR support [language extension layer pure C interfaces](./include/c_api), supporting array-style memory allocation and pointer-based calculation interfaces, providing a native pure C programming experience. | 24 | +- A5 supports DumpTensor for L1 Tensor data ([PR#2175](https://gitcode.com/cann/asc-devkit/pull/2175)), extending L1 layer data debugging support. |
| 25 | -- Ascend 950PR supports SIMD and SIMT hybrid programming and provides approximately 700 [SIMT API interfaces](./include/simt_api), including warp, atomic, basic mathematical calculation, type conversion, and other fundamental interfaces. | 25 | +- The compilation project CMakeModule supports CMAKE<LANG> compilation options ([PR#2055](https://gitcode.com/cann/asc-devkit/pull/2055)); adds the optype_collector tool, supporting optype duplicate name checking ([PR#285](https://gitcode.com/cann/asc-tools/pull/285)). |
| 26 | -- Ascend 950PR supports CCU communication interfaces for communication high-level APIs and provides mainstream communication primitives based on CCU, including [Allreduce, Allgather, Reducescatter, AlltoAll, and others](./impl/adv_api/detail/hccl/impl/platform_v310). Matmul high-level API adds support for [MXFP4/8 low-bit data type matrix operations](./impl/adv_api/detail/matmul/mx_matmul_impl.h), achieving half memory usage and doubled computing throughput. | 26 | +- The basic API supports functional behavior in ctrl (saturation overflow management) ([PR#2077](https://gitcode.com/cann/asc-devkit/pull/2077)). |
| 27 | -- Ascend 950PR adds and supports approximately 260 sample cases in total, including [SIMT samples](./examples/03_simt_api), [SIMD samples](./examples/01_simd_cpp_api) (framework, basic API, high-level API, best practices, and others), and adjusts the [sample directory structure](https://gitcode.com/cann/asc-devkit/pull/1223) according to programming model and sample category to improve readability. | 27 | +- SIMT programming adds ld/st interfaces ([PR#2058](https://gitcode.com/cann/asc-devkit/pull/2058)) and AddrSpace class interfaces ([PR#1597](https://gitcode.com/cann/asc-devkit/pull/1597)), enriching SIMT memory access programming capability. |
| 28 | -- Fusion compilation and <<<>>> invocation support [CPU mode](https://gitcode.com/cann/asc-tools/pull/138) and [SIM simulation mode](https://gitcode.com/cann/asc-devkit/blob/master/cmake/asc/asc_modules/CMakeASCInformation.cmake). | 28 | +### 🎯 Sample Updates |
| 29 | +- Best practice sample development: matmul+gelu fusion, datacopy optimization, bank conflict optimization, group_matmul quantized group matrix multiplication, and simt&simd high-performance programming ([PR#1814](https://gitcode.com/cann/asc-devkit/pull/1814) [PR#2137](https://gitcode.com/cann/asc-devkit/pull/2137) [PR#2141](https://gitcode.com/cann/asc-devkit/pull/2141) [PR#2166](https://gitcode.com/cann/asc-devkit/pull/2166) [PR#2363](https://gitcode.com/cann/asc-devkit/pull/2363)). | ||
| 30 | +- Ascend 950 new feature supplements and compatibility sample rectification: loopmode data movement, interleave vector computation, datacopy_gm2l1, loadmx (Load2DMX), mmad_mx, data_copy_pad, and so on ([PR#2336](https://gitcode.com/cann/asc-devkit/pull/2336) [PR#1899](https://gitcode.com/cann/asc-devkit/pull/1899) [PR#2124](https://gitcode.com/cann/asc-devkit/pull/2124)). | ||
| 31 | +- RegBase adds basic samples: basic arithmetic, data type conversion, reduction, comparison, indexing, and other samples ([PR#1459](https://gitcode.com/cann/asc-devkit/pull/1459) [PR#1575](https://gitcode.com/cann/asc-devkit/pull/1575) [PR#2024](https://gitcode.com/cann/asc-devkit/pull/2024)). | ||
| 32 | +- Adds SIMD VF print samples and dump samples ([PR#2558](https://gitcode.com/cann/asc-devkit/pull/2558)). | ||
| 33 | +- SIMT adds DCache access optimization samples ([PR#2453](https://gitcode.com/cann/asc-devkit/pull/2453)), transpose-based memory coalescing and bank conflict samples ([PR#1753](https://gitcode.com/cann/asc-devkit/pull/1753)), and best practice samples for improving data movement efficiency through type alignment ([PR#2297](https://gitcode.com/cann/asc-devkit/pull/2297)). | ||
| 34 | +- SIMT adds functional feature samples: PyTorch custom operator registration ([PR#2769](https://gitcode.com/cann/asc-devkit/pull/2769)), compilation-related samples (dynamic, static, separate compilation, and so on) ([PR#2356](https://gitcode.com/cann/asc-devkit/pull/2356)), profiling samples ([PR#1989](https://gitcode.com/cann/asc-devkit/pull/1989)), memory barrier feature samples ([PR#1923](https://gitcode.com/cann/asc-devkit/pull/1923)), Warp class feature samples ([PR#2876](https://gitcode.com/cann/asc-devkit/pull/2876)), simulator samples ([PR#2692](https://gitcode.com/cann/asc-devkit/pull/2692)), and kernel log samples ([PR#2131](https://gitcode.com/cann/asc-devkit/pull/2131)). | ||
| 35 | +- The SIMT getting-started sample is modified to gather ([PR#2405](https://gitcode.com/cann/asc-devkit/pull/2405)). | ||
| 36 | +- Adds Tensor API getting-started and best practice samples: Matmul getting started, data movement in/out, data movement out with on-path quantization, and MX FP4 best practices ([PR#2553](https://gitcode.com/cann/asc-devkit/pull/2553)). | ||
| 29 | ### 📖 Documentation | 37 | ### 📖 Documentation |
| 30 | -- Adds over 90 [Reg programming interface API](https://gitcode.com/cann/asc-devkit/blob/master/docs/api/SIMD-API/基础API/Reg矢量计算/Reg矢量计算.md) documentation. Reg vector calculation API is developed for the RegBase architecture. Users can directly operate registers involved in Vector calculation on the chip through this API to achieve greater flexibility and better performance. | 38 | +- Adds matrix computation overview and computation fractal introduction documentation ([PR#2533](https://gitcode.com/cann/asc-devkit/pull/2533)). |
| 31 | -- Adds SIMT [quick start](docs/guide/入门教程/快速入门/基于SIMT编程/Gather算子快速入门.md), [programming model](docs/guide/编程指南/编程模型/AI-Core-SIMT编程/抽象硬件架构.md), and [operator implementation](docs/guide/算子实践参考/SIMT算子实现/基础知识.md) introduction. | 39 | +- Optimizes vector computation API documentation, supplementing instruction constraints, and so on ([PR#2676](https://gitcode.com/cann/asc-devkit/pull/2676)). |
| 32 | -- Adds SIMD and SIMT [hybrid programming model](docs/guide/编程指南/高级编程/高级AI-Core编程模型/SIMD与SIMT混合编程/概述.md), [operator implementation](docs/guide/算子实践参考/SIMD与SIMT混合算子实现/基础知识.md), and [performance optimization](docs/guide/算子实践参考/SIMD与SIMT混合算子性能优化/内存访问/使用Unified-Buffer提升内存访问效率.md) introduction. | 40 | +- Sets up a VitePress documentation site, providing AscendC documentation preview functionality ([PR#2547](https://gitcode.com/cann/asc-devkit/pull/2547)). |
| 33 | -- Adds [SIMT API](https://gitcode.com/cann/asc-devkit/blob/master/docs/api/SIMT-API/SIMT-API.md) documentation chapter. | 41 | +- Adds SIMD and SIMT hybrid programming performance optimization overview ([PR#2736](https://gitcode.com/cann/asc-devkit/pull/2736)). |
| 34 | -- Adds [compatibility migration guide](docs/guide/跨代迁移兼容性指南/Ascend-C-API兼容策略.md) (migration from 220x architecture version to 351x architecture version). | ||
| 35 | -- In the Ascend community, Ascend C operator development adds a [visualization section](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/900beta2/opdevg/Ascendcopdevg/atlas_ascendc_map_10_0017.html), presenting the execution process of Cube and Vector operators through videos. | ||
| 36 | 42 | ||
| 37 | -For detailed information about all historical versions and updates, see [CHANGELOG.md](./CHANGELOG_en.md) | 43 | +For detailed information about all historical versions and updates, see [CHANGELOG.md](./CHANGELOG_en.md). |
| 38 | 44 | ||
| 39 | ## 🚀 Overview | 45 | ## 🚀 Overview |
| 40 | 46 | ||
| @@ -62,7 +68,7 @@ Additionally, Ascend C provides high-level APIs and operator template libraries | |||
| 62 | 68 | ||
| 63 | | API Level | Target Users | Primary Use | | 69 | | API Level | Target Users | Primary Use | |
| 64 | |----------|----------|----------| | 70 | |----------|----------|----------| |
| 65 | -| **Operator Template Library (CATLASS/ATVOSS and others)** | Algorithm developers | Perform custom extensions based on typical operator implementations to meet high-performance requirements in specific scenarios | | 71 | +| **Operator Template Library (CATLASS/ATVOSS and others)** | Algorithm developers | Perform custom extensions based on typical operator implementations to meet high-performance requirements in specific scenarios | |
| 66 | | **High-level API** |Algorithm developers | Reuse general single-core algorithms to quickly complete algorithm verification | | 72 | | **High-level API** |Algorithm developers | Reuse general single-core algorithms to quickly complete algorithm verification | |
| 67 | 73 | ||
| 68 | 74 | ||