Skip to content

sync: GitCode master @ 4cd325a422dc - #276

Merged
Crystal-wzy merged 2 commits into
hw-native-sys:mainfrom
Crystal-wzy:gitcode-sync
Aug 28, 2026
Merged

Crystal-wzy merged 2 commits into
hw-native-sys:mainfrom
Crystal-wzy:gitcode-sync

Conversation

@Crystal-wzy

Copy link
Copy Markdown
Collaborator

Automated daily sync from GitCode cann/pto-isa:master into GitHub hw-native-sys/pto-isa:main.

Carries only commits after cut point:
265344a175559f2702a7c95c87fcc5afefca60bc

Cut point source:
merged-pr-273

Source SHA:
4cd325a422dce40e367cce8b9c20795281925e8a

Sync mode:
cherry-pick

Generated by the Sync GitCode to GitHub workflow.

wangrunfa and others added 2 commits August 28, 2026 02:57
…t cases

Co-authored-by: wangrunfa<wangrunfa1@huawei.com>



# message auto-generated for no-merge-commit merge:
!1550 merge fix/a2a3-tconcat-padding-pollution into master

fix(a2a3): fix TConcat padding pollution and add int8/non-aligned test cases

Created-by: wangrunfa
Commit-by: wangrunfa
Merged-by: cann-robot
Description: ## 描述

修复 TConcatImpl isAligned 路径中 pto_copy_ubuf_to_ubuf 使用 ceil 对齐 block 长度拷贝 src1 数据时,将 src1 padding 垃圾值写入 dst padding 区域的问题。

同时新增 5 条 ST 测试用例覆盖 int8、非对齐场景,全部用例 NPU 通过。

## 关联的Issue

closes #544

https://gitcode.com/cann/pto-isa/issues/544

## 测试

- 全部 14 条 TCONCAT ST 用例 NPU 板端通过,max diff = 0
- 新增用例覆盖:int8 类型、双非对齐 (scalar copy)、dstTileH ≠ srcTileH

## 文档更新

无

## 类型标签

- [x] Bug修复
- [ ] 新特性
- [ ] 性能优化
- [ ] 文档更新
- [ ] 其他,请描述:

See merge request: cann/pto-isa!1550
Co-authored-by: zhywang<wangzhiyong74@huawei.com>



# message auto-generated for no-merge-commit merge:
!1548 merge fp4 into master

fix: correct NZ C0 size for FP4 tensors

Created-by: zhhywang
Commit-by: zhywang
Merged-by: cann-robot
Description: Created-by: zhhywang
Commit-by: zhhywang
Merged-by: cann-robot
Description: 修复 pto_tile NZ FP4 的 C0、shape 和 stride 计算,使 32 Byte C0 正确表示 64 个 FP4 逻辑元素。

## 描述
1 file changed, 27 insertions(+), 11 deletions(-)

- **范围**:修改 include/pto/common/pto_tile.hpp 中 NZ 二维 shape 和 stride 辅助类型。
- **问题**:FP4 的 C++ 存储类型大小为 1 Byte,原有 32 / sizeof(T) 将 C0 错误计算为 32 个逻辑元素。
- **修复**:新增统一的 NZ C0 计算,对 FP4 和 CPU 仿真的 twin 类型返回 64,并同步更新静态、动态 shape 和 stride 计算。
- **影响**:NZ FP4 的 C0、fractal shape 和原始 shape stride 按逻辑元素正确生成;其他数据类型保持原有计算方式。

## 关联的Issue

## 测试
- gitcode_commit.py precheck --repo .
  - 结果:PASS,所有适用的 pre-commit 检查通过。

## 文档更新
无。

## 类型标签
- [x] Bug修复
- [ ] 新特性
- [ ] 性能优化
- [ ] 文档更新
- [ ] 其他,请描述:

See merge request: cann/pto-isa!1548
@Crystal-wzy
Crystal-wzy merged commit 100dd57 into hw-native-sys:main Aug 28, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant