| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
补充TransposedBatchedMatmul Co-authored-by: init__zhb__<zhanghaobo6@huawei.com> # message auto-generated for no-merge-commit merge: !882 merge br.bmm.transposed into master 补充TransposedBatchedMatmul Created-by: init__zhb__ Commit-by: init__zhb__ Merged-by: cann-robot Description: ## 描述 - 新增 TransposedBatchedMatmul（A可以形如(M, B, K)或(K, B, M), B可以形如(N, B, K)或者(K, B, N)） - 将strided_batched_matmul_tla 和 strided_batched_matmul_transpose_tla 中的bmm的golden计算移到公共组件部分 其他： - 移除experimental目录下与78_matrix_inverse相关的测试件tests/optest/tests/test_78_matrix_inverse.py ## 关联的Issue <!-- 如果这个PR是为了解决特定的Issue，请在这里提供Issue链接。--> ## 原因 <!--说明此次改动的目的、解决的问题等，应与类型标签匹配 --> ## 测试 ### 单算子测试 已进行200例子泛化验证 bash $ ./output/bin/45_strided_transposed_batched_matmul_tla 24 67 283 577 0 Compare success. ... ### optest接入验证 测试通过（包含transposed以及非trans在Row/Col组合下的情况） bash $ pytest tests/test_45_strided_batched_matmul_tla.py ... collected 8 items tests/test_45_strided_batched_matmul_tla.py ........ [100%] ========================================================================= 8 passed in 33.18s ========================================================================= ### 性能测试 基本持平 batch M N K STATUS1 TASK_TIME1 STATUS2 TASK_TIME2 STATUS3 TASK_TIME3 TASK_TIME_AVG CATLASS_TASK_TIME SpeedUP 80 512 80 640 unknown 46.834999 unknown 48.091 unknown 47.761002 47.562334 47.277333 1.00602828 80 512 27 640 unknown 42.473 unknown 42.888 unknown 42.831001 42.730667 39.800667 1.073616857 ## 文档更新 <!--如果这个PR包含文档的更新，请在这里指出。例如：更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [x] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他，请描述： See merge request: cann/catlass!882 | 19 天前 | |
Add catlass_cppgen into main repo, a tool for generating catlass template code. Co-authored-by: init__zhb__<zhanghaobo6@huawei.com> # message auto-generated for no-merge-commit merge: !817 merge br.cppgen into master 补充cppgen框架 Created-by: init__zhb__ Commit-by: init__zhb__ Merged-by: cann-robot Description: ## 描述 catlass_cppgen 是一个基于 Python 的代码生成框架，用于构建和生成 CATLASS 高性能算子。该框架提供接口以定义算子参数、选择优化策略，并自动生成对应的 C++ 核函数代码。 本次补充的内容结构如下： plain ./catlass_cppgen ├── catlass_cppgen │ ├── __init__.py │ ├── catlass # CATLASS 相关特性组件 │ │ ├── __init__.py │ │ ├── arch # 代际声明 │ │ ├── evg # EVG 特性承载 │ │ ├── evg_extension.py │ │ ├── gemm │ │ ├── gemm_coord.py │ │ ├── layout │ │ └── library.py │ ├── common # 通用组件 │ │ ├── __init__.py │ │ ├── data_type.py │ │ ├── op_tensor.py │ │ ├── typing.py │ │ └── utils.py │ ├── kernel │ │ ├── __init__.py │ │ ├── gemm # GEMM 类算子特化类 │ │ ├── group_gemm # Group GEMM 类算子特化类 │ │ ├── kernel_base.py │ │ └── visitor_kernel_base.py │ ├── op # 算子Kernel基类 │ │ ├── __init__.py │ │ ├── gemm.py │ │ ├── group_gemm.py │ │ └── op.py │ └── _version.py ├── docs # API 文档 │ ├── evg_api.md │ ├── kernel_api.md │ └── optensor_api.md ├── pyproject.toml ├── README.md # 主README文档 └── uv.lock 相关的测试路径： plain ./ ├── catlass │ ├── test_arch.py │ ├── test_dispatch_policy.py │ ├── test_gemm_coord.py │ └── test_layout.py ├── common │ ├── test_data_type.py │ └── test_op_tensor.py └── op ├── assertion_helper.py ├── test_evg.py ├── test_gemm.py └── test_group_gemm.py ## 关联的Issue <!-- 如果这个PR是为了解决特定的Issue，请在这里提供Issue链接。--> ## 原因 演进与完善 ## 测试 **gen-match 测试(UT-level)** 执行tests/run_cppgen.sh测试<span style="color:green">PASS</span>。 plain ============================================ Step 3: Preparing test all testcases... ============================================ --- catlass/test_arch.py --- ... ---------------------------------------------------------------------- Ran 3 tests in 0.000s OK --- catlass/test_dispatch_policy.py --- ....... ---------------------------------------------------------------------- Ran 7 tests in 0.001s OK --- catlass/test_gemm_coord.py --- .... ---------------------------------------------------------------------- Ran 4 tests in 0.000s OK --- catlass/test_layout.py --- ......................... ---------------------------------------------------------------------- Ran 25 tests in 0.001s OK --- common/test_data_type.py --- ........................ ---------------------------------------------------------------------- Ran 24 tests in 0.001s OK --- common/test_op_tensor.py --- ....... ---------------------------------------------------------------------- Ran 7 tests in 0.001s OK --- op/test_evg.py --- ............ ---------------------------------------------------------------------- Ran 12 tests in 0.030s OK --- op/test_gemm.py --- ............. ---------------------------------------------------------------------- Ran 13 tests in 0.014s OK --- op/test_group_gemm.py --- ........ ---------------------------------------------------------------------- Ran 8 tests in 0.003s OK ============================================ Step 4: Cleaning and uninstalling catlass_cppgen... ============================================ Found existing installation: catlass-cppgen 0.1.0+20260706144530 Uninstalling catlass-cppgen-0.1.0+20260706144530: Successfully uninstalled catlass-cppgen-0.1.0+20260706144530 Found existing installation: build 1.5.0 Uninstalling build-1.5.0: Successfully uninstalled build-1.5.0 ============================================ All steps completed successfully! SUMMARY: 9 passed, 0 failed ============================================ ## 文档更新 更新下述文档： - python/catlass_cppgen/README.md - python/catlass_cppgen/docs/evg_api.md - python/catlass_cppgen/docs/kernel_api.md - python/catlass_cppgen/docs/optensor_api.md ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [x] 新特性 - [ ] 性能优化 - [x] 文档更新 - [ ] 其他，请描述： See merge request: cann/catlass!817 | 1 个月前 | |
补充TransposedBatchedMatmul Co-authored-by: init__zhb__<zhanghaobo6@huawei.com> # message auto-generated for no-merge-commit merge: !882 merge br.bmm.transposed into master 补充TransposedBatchedMatmul Created-by: init__zhb__ Commit-by: init__zhb__ Merged-by: cann-robot Description: ## 描述 - 新增 TransposedBatchedMatmul（A可以形如(M, B, K)或(K, B, M), B可以形如(N, B, K)或者(K, B, N)） - 将strided_batched_matmul_tla 和 strided_batched_matmul_transpose_tla 中的bmm的golden计算移到公共组件部分 其他： - 移除experimental目录下与78_matrix_inverse相关的测试件tests/optest/tests/test_78_matrix_inverse.py ## 关联的Issue <!-- 如果这个PR是为了解决特定的Issue，请在这里提供Issue链接。--> ## 原因 <!--说明此次改动的目的、解决的问题等，应与类型标签匹配 --> ## 测试 ### 单算子测试 已进行200例子泛化验证 bash $ ./output/bin/45_strided_transposed_batched_matmul_tla 24 67 283 577 0 Compare success. ... ### optest接入验证 测试通过（包含transposed以及非trans在Row/Col组合下的情况） bash $ pytest tests/test_45_strided_batched_matmul_tla.py ... collected 8 items tests/test_45_strided_batched_matmul_tla.py ........ [100%] ========================================================================= 8 passed in 33.18s ========================================================================= ### 性能测试 基本持平 batch M N K STATUS1 TASK_TIME1 STATUS2 TASK_TIME2 STATUS3 TASK_TIME3 TASK_TIME_AVG CATLASS_TASK_TIME SpeedUP 80 512 80 640 unknown 46.834999 unknown 48.091 unknown 47.761002 47.562334 47.277333 1.00602828 80 512 27 640 unknown 42.473 unknown 42.888 unknown 42.831001 42.730667 39.800667 1.073616857 ## 文档更新 <!--如果这个PR包含文档的更新，请在这里指出。例如：更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [x] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他，请描述： See merge request: cann/catlass!882 | 19 天前 |