| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
[PR3/3] migrate transdata aicpu kernel: remaining formats + OpDef + infershape + full UT Co-authored-by: pan-tong<pantong@huawei.com> # message auto-generated for no-merge-commit merge: !4115 merge transdata_pr3 into master [PR3/3] migrate transdata aicpu kernel: remaining formats + OpDef + infershape + full UT Created-by: pan-tong Commit-by: pan-tong Merged-by: cann-robot Description: ## 描述 将canndev仓TransData算子的AICPU kernel迁移到ops-math仓，PR3/3：remaining formats + OpDef + infershape + full UT + example。 ## 关联的Issue #2231 ## 测试 PR3 of 3 (final): remaining format_transfer + OpDef registration + infershape + full UT + example. - format_transfer_fractalz_3d.{cpp,h}: NCDHW/DHWCN/NDHWC <-> FRACTAL_Z_3D - format_transfer_ndc1hwc0.{cpp,h}: NCDHW/NDHWC -> NDC1HWC0 - format_transfer_transpose.{cpp,h}: format transpose - trans_data_aicpu_def.cpp: AICPU OpDef registration (activates custom_math) - trans_data_infershape.cpp: RT2 InferShape - test_trans_data.cpp: full 23 kernel UT cases - test_trans_data_infershape.cpp: 6 infershape UT cases - test_geir_trans_data.cpp: GE IR example - docs/zh/op_list.md: mark trans_data as AI Core/AI CPU This PR activates custom_math TransData. Depends on PR1 + PR2. ## 类型标签 - [ ] Bug修复 - [x] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他，请描述：算子迁仓 See merge request: cann/ops-math!4115 | 26 天前 | |
feat: 支持八个迁移算子的常量折叠(#2551) Co-authored-by: pan-tong<pantong@huawei.com> # message auto-generated for no-merge-commit merge: !4511 merge feat/constant-folding-infershape into master feat: 支持八个迁移算子的常量折叠(#2551) Created-by: pan-tong Commit-by: pan-tong Merged-by: cann-robot Description: ## 描述 为 Ceil、ConcatV2、Expand、GreaterEqual、InvertPermutation、StridedSlice、StridedSliceV2、TransData 接入 HOSTCPU 常量折叠。 - 为尚未接线的 7 个算子在 CMake 中启用 HOSTCPU TRUE，将现有 AICPU kernel 编入 libopconstant_folding_math.so。 - 8 个算子的 AICPU kernel 使用 OPS_MATH_REGISTER_CPU_KERNELV2，兼容 device 注册和 host CPU Kernel V2 注册。 - 为 GreaterEqual 补齐 OPS_FLAG_OPEN；其余算子沿用已有 OPEN 配置。 - TransData 沿用已有 HOSTCPU 配置，并复用已接入 host object 的 format_transfer 辅助源。 - 增强 Ceil AICPU UT 的输出校验，新增 Ceil 全 Const GE 图常量折叠用例。 本 PR 不修改算子的 infershape 实现。 ## 关联的Issue https://gitcode.com/cann/ops-math/issues/2551 ## 测试 - AICPU UT：8 个算子共 156 个用例全部通过。 - HOSTCPU 构建：libopconstant_folding_math.so 构建成功，包含 8 个算子的 host 注册。 - 安装级 GE 常量折叠：8 个算子的全 Const 图输出均校验通过。 - 逐算子日志确认对应 host kernel 执行，ConstantFoldingPass 返回码均为 0。 - pre-commit：clang-format、codespell、OAT 等检查全部通过。 - GitCode 服务端 Hooks：通过。 ## 文档更新 无。该改动不改变算子接口和使用方式。 ## 类型标签 - [ ] Bug修复 - [x] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他，请描述： See merge request: cann/ops-math!4511 | 7 天前 | |
[PR1/3] migrate transdata aicpu kernel: kernel impl + format_transfer base + basic UT Co-authored-by: pan-tong<pantong@huawei.com> # message auto-generated for no-merge-commit merge: !4113 merge transdata_pr1 into master [PR1/3] migrate transdata aicpu kernel: kernel impl + format_transfer base + basic UT Created-by: pan-tong Commit-by: pan-tong Merged-by: cann-robot Description: ## 描述 将canndev仓TransData算子的AICPU kernel迁移到ops-math仓，PR1/3：kernel impl + format_transfer base + basic UT。 ## 关联的Issue #2231 ## 测试 PR1 of 3: kernel + format_transfer base (utils/register) + 3 basic UT. - cmake/symbol.cmake: add COMMAND_EXPAND_LISTS - CMakeLists.txt: target_sources for format_transfer - trans_data_aicpu.{cpp,h}: main CpuKernel implementation - format_transfer/{formats_definitions,register_format_transfer,format_transfer_utils}.{h,cpp}: base infra - test_trans_data.cpp: 3 basic UT (PARAM_INVALID paths) 修复检视意见: - P2: memset_s 返回值检查 - P3: Measure 函数 gcd(0) 修复 - P3: 冗余除零检查删除 No _aicpu_def.cpp (canndev TransData still active). ## 类型标签 - [ ] Bug修复 - [x] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他，请描述：算子迁仓 See merge request: cann/ops-math!4113 | 29 天前 | |
[PR3/3] migrate transdata aicpu kernel: remaining formats + OpDef + infershape + full UT Co-authored-by: pan-tong<pantong@huawei.com> # message auto-generated for no-merge-commit merge: !4115 merge transdata_pr3 into master [PR3/3] migrate transdata aicpu kernel: remaining formats + OpDef + infershape + full UT Created-by: pan-tong Commit-by: pan-tong Merged-by: cann-robot Description: ## 描述 将canndev仓TransData算子的AICPU kernel迁移到ops-math仓，PR3/3：remaining formats + OpDef + infershape + full UT + example。 ## 关联的Issue #2231 ## 测试 PR3 of 3 (final): remaining format_transfer + OpDef registration + infershape + full UT + example. - format_transfer_fractalz_3d.{cpp,h}: NCDHW/DHWCN/NDHWC <-> FRACTAL_Z_3D - format_transfer_ndc1hwc0.{cpp,h}: NCDHW/NDHWC -> NDC1HWC0 - format_transfer_transpose.{cpp,h}: format transpose - trans_data_aicpu_def.cpp: AICPU OpDef registration (activates custom_math) - trans_data_infershape.cpp: RT2 InferShape - test_trans_data.cpp: full 23 kernel UT cases - test_trans_data_infershape.cpp: 6 infershape UT cases - test_geir_trans_data.cpp: GE IR example - docs/zh/op_list.md: mark trans_data as AI Core/AI CPU This PR activates custom_math TransData. Depends on PR1 + PR2. ## 类型标签 - [ ] Bug修复 - [x] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他，请描述：算子迁仓 See merge request: cann/ops-math!4115 | 26 天前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 26 天前 | ||
| 7 天前 | ||
| 29 天前 | ||
| 26 天前 |