catlass/include/tla · CANN/catlass

cann-robotComplete and refine tile-level unittests.

文件	最后提交记录	最后更新时间
numeric	Complete and refine tile-level unittests. Co-authored-by: init__zhb__<zhanghaobo6@huawei.com> # message auto-generated for no-merge-commit merge: !730 merge ut.dev into master 补充TLA Tile组件UT测试件 Created-by: init__zhb__ Commit-by: init__zhb__ Merged-by: cann-robot Description: ## 描述补充TLA Tile组件UT测试件。 \|数据通路\|被测组件\|源文件\|非TLA 特化数\|TLA/Sparse 特化数\| \|---\|---\|---\|---\|---\| \|GM→L1\|CopyGmToL1 / GMMPTD / DynamicOptimized / IntervalDataCopy\|\|\~24\|10\| \|L1→L0A\|CopyL1ToL0A \|\|10\|6\| \|L1→L0B\|CopyL1ToL0B \|\|13\|6\| \|L0C→GM \|CopyL0CToGm \|\|7\|3\| \|GM→UB\|CopyGm2Ub / CopyGm2UbAligned / CopyPerTokenScale2Ub\|\|4\|1\| \|UB→GM\|CopyUb2Gm / CopyUb2GmAligned\|\|3\|1\| \|MMAD \|TileMmad\|\|2\|2\| \|合计\|\|\|\~63\|\~27\| - 特别说明：本PR涉及对公共`MakeIntegerSequence`的修改，原因在于后续`make_index_sequence = MakeIntegerSequence<size_t, N>`和`make_int_sequence = MakeIntegerSequence<int, N>`中涉及`int`和`size_t`两种类型，而`MakeIntegerSequenceImpl`中展开量的类型限定为`size_t`，GCC/Clang stub 编译阶段严格拒绝了可能的隐式转换。 ## 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 ### UT组件通过情况 - Arch2201: `bash ... [----------] Global test environment tear-down [==========] 581 tests from 19 test suites ran. (129 ms total) [ PASSED ] 548 tests. [ SKIPPED ] 33 tests, listed below: [ SKIPPED ] CopyGmToL1/TileCopyGmToL1TestAtlasA2.RowMajorTozNTestOneRow_GMMPTD/0 [ SKIPPED ] CopyGmToL1/TileCopyGmToL1TestAtlasA2.RowMajorTozNTestOneRow_GMMPTD/2 ...` - Arch3510: `bash ... [ RUN ] Ascend950/TypedTileMmadA950Suite.DisableGemvMmad/14 [ OK ] Ascend950/TypedTileMmadA950Suite.DisableGemvMmad/14 (0 ms) [----------] 120 tests from Ascend950/TypedTileMmadA950Suite (2 ms total) [----------] Global test environment tear-down [==========] 383 tests from 20 test suites ran. (32 ms total) [ PASSED ] 383 tests.` ### 覆盖率 <span style="color:green">FULL</span> bash Overall coverage rate: lines......: 100.0% (2486 of 2486 lines) functions..: 99.3% (1480 of 1490 functions) Reading tracefile /catlass/build/coverage/coverage.info \|Lines \|Functions\|Branches Filename \|Rate Num\|Rate Num\|Rate Num ================================================================================ [/catlass/include/] catlass/coord.hpp \| 100% 17\| 100% 23\| - 0 catlass/detail/alignment.hpp \| 100% 10\| 100% 25\| - 0 catlass/epilogue/tile/copy_gm_to_ub.hpp \| 100% 68\| 100% 6\| - 0 catlass/epilogue/tile/copy_gm_to_ub_tla.hpp \| 100% 19\| 100% 2\| - 0 catlass/epilogue/tile/copy_ub_to_gm.hpp \| 100% 45\| 100% 5\| - 0 catlass/epilogue/tile/copy_ub_to_gm_tla.hpp \| 100% 10\| 100% 1\| - 0 catlass/gemm/tile/ascend950/copy_gm_to_l1.hpp \| 100% 417\| 100% 46\| - 0 catlass/gemm/tile/ascend950/copy_l0c_to_gm.hpp \| 100% 99\| 100% 10\| - 0 catlass/gemm/tile/ascend950/copy_l0c_to_ub.hpp \| 100% 44\| 100% 3\| - 0 catlass/gemm/tile/ascend950/copy_l1_to_bt.hpp \| 100% 12\| 100% 4\| - 0 catlass/gemm/tile/ascend950/copy_l1_to_l0a.hpp \| 100% 98\| 100% 16\| - 0 catlass/gemm/tile/ascend950/copy_l1_to_l0b.hpp \| 100% 162\| 100% 20\| - 0 catlass/gemm/tile/atlasa2/copy_gm_to_l1.hpp \| 100% 611\| 100% 68\| - 0 catlass/gemm/tile/atlasa2/copy_gm_to_ub.hpp \| 100% 12\| 100% 2\| - 0 catlass/gemm/tile/atlasa2/copy_l0c_to_gm.hpp \| 100% 69\| 100% 9\| - 0 catlass/gemm/tile/atlasa2/copy_l1_to_l0a.hpp \| 100% 171\| 100% 22\| - 0 catlass/gemm/tile/atlasa2/copy_l1_to_l0b.hpp \| 100% 219\| 100% 26\| - 0 catlass/gemm/tile/atlasa2/copy_ub_to_gm.hpp \| 100% 11\| 100% 2\| - 0 catlass/gemm/tile/tile_mmad.hpp \| 100% 91\| 100% 20\| - 0 catlass/layout/matrix.hpp \| 100% 139\| 100% 63\| - 0 catlass/layout/vector.hpp \| 100% 5\| 100% 3\| - 0 catlass/numeric_size.hpp \| 100% 2\| 100% 1\| - 0 tla/int_tuple.hpp \| 100% 16\| 100% 47\| - 0 tla/layout.hpp \| 100% 85\|99.6% 244\| - 0 tla/numeric/integral_constant.hpp \| 100% 4\| 100% 17\| - 0 tla/numeric/math.hpp \| 100% 4\| 100% 3\| - 0 tla/tensor.hpp \| 100% 24\|99.2% 241\| - 0 tla/tuple.hpp \| 100% 22\|98.8% 561\| - 0 ================================================================================ Total:\| 100% 2486\|99.3% 1k\| - 0 [100%] Built target coverage_collect ### 变异测试（部分） 1. src和dst typo错误混杂 Ref bugfix commit: 02d1078 Mutation: diff --- a/include/catlass/gemm/tile/ascend950/copy_l1_to_l0a.hpp +++ b/include/catlass/gemm/tile/ascend950/copy_l1_to_l0a.hpp @@ -626,7 +626,7 @@ struct CopyL1ToL0A<Arch::Ascend950, Gemm::GemmType<Element, layout::nZ, AscendC: const uint32_t L0M = layoutDst.shape(0) * layoutDst.shape(1); const uint32_t L0K = layoutDst.shape(2) * layoutDst.shape(3); const uint32_t srcOuterStrideRow = layoutSrc.stride(1); - const uint32_t dstOuterStrideCol = layoutDst.stride(3); + const uint32_t dstOuterStrideCol = layoutSrc.stride(3); AscendC::LoadData2DParamsV2 loadDataParams; loadDataParams.mStartPosition = 0; Result: <span style="color:green">KILLED</span> `bash [ FAILED ] 3 tests, listed below: [ FAILED ] CopyL1ToL0A/TileCopyL1ToL0ATestAscend950.nZTozNTestBasic/0, where GetParam() = 8-byte object <80-00 00-00 40-00 00-00> [ FAILED ] CopyL1ToL0A/TileCopyL1ToL0ATestAscend950.nZTozNTestBasic/2, where GetParam() = 8-byte object <40-00 00-00 2A-00 00-00> [ FAILED ] CopyL1ToL0A/TileCopyL1ToL0ATestAscend950.nZTozNTestBasic/3, where GetParam() = 8-byte object <7B-00 00-00 08-00 00-00> 3 FAILED TESTS` 2. 不对齐的情况 Ref bugfix commit: 0439b54e Mutation: diff --- a/include/catlass/gemm/tile/ascend950/copy_l0c_to_ub.hpp +++ b/include/catlass/gemm/tile/ascend950/copy_l0c_to_ub.hpp @@ -148,7 +148,7 @@ struct CopyL0CToUBTla< AscendC::FixpipeParamsC310<AscendC::CO2Layout::ROW_MAJOR> intriParams; // Fixpipe layout information - intriParams.nSize = RoundUp(tla::get<1>(dstTensor.originShape()), 32); // n must be multiple of 32 when split n + intriParams.nSize = tla::get<1>(dstTensor.originShape()); // n must be multiple of 32 when split n intriParams.mSize = tla::get<0>(dstTensor.originShape()); intriParams.srcStride = tla::get<1, 1>(srcTensor.stride()) / tla::get<0, 0>(srcTensor.stride()); intriParams.dstStride = tla::get<0>(dstTensor.stride()); Result: <span style="color:green">KILLED</span> `[ FAILED ] 1 test, listed below: [ FAILED ] CopyL0CToUbTlaAscend950/TileCopyL0CToUbTlaAscend950Test.L0CToRowMajorTestSplitN/1, where GetParam() = 8-byte object <41-00 00-00 7F-00 00-00> 1 FAILED TEST` 3. 常量错误 Ref bugfix commit: 81e1129e Mutation: diff --- a/include/catlass/gemm/tile/ascend950/copy_l1_to_bt.hpp +++ b/include/catlass/gemm/tile/ascend950/copy_l1_to_bt.hpp @@ -80,7 +80,7 @@ struct CopyL1ToBT<Arch::Ascend950, Catlass::Gemm::GemmType<ElementSrc, layout::V using LayoutDst = layout::VectorLayout; using LayoutSrc = layout::VectorLayout; - static constexpr uint32_t ELE_NUM_PER_C0 = BytesToBits(BYTE_PER_C0) / SizeOfBits<ElementSrc>::value; + static constexpr uint32_t ELE_NUM_PER_C2 = BytesToBits(BYTE_PER_C2) / SizeOfBits<ElementSrc>::value; CATLASS_DEVICE CopyL1ToBT(){} @@ -93,7 +93,7 @@ struct CopyL1ToBT<Arch::Ascend950, Catlass::Gemm::GemmType<ElementSrc, layout::V ){ AscendC::DataCopyParams intriParams; intriParams.blockCount = 1; - intriParams.blockLen = CeilDiv(layoutDst.shape(0), ELE_NUM_PER_C0); + intriParams.blockLen = CeilDiv(layoutDst.shape(0), ELE_NUM_PER_C2); if (sizeof(ElementSrc) == 4) { Result: <span style="color:yellow">SURVIVED</span> <span style="color:green">KILLED</span>(commit:b1d60262) bash [ FAILED ] 6 tests, listed below: [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestHalf/0, where GetParam() = 8-byte object <80-00 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestHalf/1, where GetParam() = 8-byte object <43-00 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestHalf/2, where GetParam() = 8-byte object <00-01 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestFloat/0, where GetParam() = 8-byte object <80-00 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestFloat/1, where GetParam() = 8-byte object <43-00 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestFloat/2, where GetParam() = 8-byte object <00-01 00-00 01-00 00-00> 6 FAILED TESTS ## 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 - [ ] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他，请描述： See merge request: cann/catlass!730	1 个月前
int_tuple.hpp	add a declarative post-processing framework, EVG. Co-authored-by: li_wei21<liwei949@huawei.com> Co-authored-by: yuantao<taoyuan18@huawei.com> # message auto-generated for no-merge-commit merge: !679 merge catlass_evg_0608 into master add a declarative post-processing framework, EVG. Created-by: li_wei21 Commit-by: li_wei21;yuantao Merged-by: cann-robot Description: ## 描述 Catlass开源仓新增特性EVG，增强后处理代码的易用性，同时补充对应的测试样例。 EVG设计的详细说明可以参考issue。 ## 关联的Issue https://gitcode.com/cann/catlass/issues/233 ## 测试已对每一个EVG类型的模板样例分别进行200个case测试，结果通过 ## 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!679	2 个月前
layout.hpp	add conv2d_tla Co-authored-by: chen_xu0513<chenxu210@huawei.com> # message auto-generated for no-merge-commit merge: !569 merge master into master add conv2d_tla Created-by: chen_xu0513 Commit-by: chen_xu0513 Merged-by: cann-robot Description: ## 描述补充了2d卷积的tla组件，并新增tla版本的基础2d卷积模板样例<!--在这里详细描述你的改动，包括改动的原因和所采取的方法。--> ## 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 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 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!569	3 个月前
tensor.hpp	TLA增加origin_shape，新增LocalTile、MakeTensorLike、Tensor的operator()和增加batched_matmul_tla样例 Co-authored-by: sinceseto<wangyuanfei6@huawei.com> # message auto-generated for no-merge-commit merge: !469 merge tla_modify_pr into master TLA增加originShape，新增TileView、MakeTensorLike、Tensor的operator()和增加strided_batched_matmul_tla样例 Created-by: sinceseto Commit-by: sinceseto Merged-by: cann-robot Description: ## 描述 <!--在这里详细描述你的改动，包括改动的原因和所采取的方法。--> 为了基于TLA提供kernel和block层不感知尾块逻辑，tile层感知origin_shape的新流程。为tla::layout增加originShape，为了不影响当前其它算子原来流程，支持写layout类型中originShape类型自动推导，makeLayout自动推导originShape。 MakeLayoutTile更名GetTileLayout，增加coord参数，根据coord和origin_shape自动裁剪更新GetTile（如果原来流程不规范写法涉及到coord>shape的GetTile操作会导致origin_shape被裁剪到0，导致tile层不工作，精度错误）新增TileView（tile粒度的GetTIle)、MakeTensorLike（跨内存域自动origin_shape继承的新建Tensor)、Tensor的operator()接口（slice语义）以一个strided_batched_matmul_tla样例承载新的TLA特性与流程。 ## 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 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 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 - [x] 文档更新 - [ ] 其他，请描述： See merge request: cann/catlass!469	4 个月前
tuple.hpp	add a declarative post-processing framework, EVG. Co-authored-by: li_wei21<liwei949@huawei.com> Co-authored-by: yuantao<taoyuan18@huawei.com> # message auto-generated for no-merge-commit merge: !679 merge catlass_evg_0608 into master add a declarative post-processing framework, EVG. Created-by: li_wei21 Commit-by: li_wei21;yuantao Merged-by: cann-robot Description: ## 描述 Catlass开源仓新增特性EVG，增强后处理代码的易用性，同时补充对应的测试样例。 EVG设计的详细说明可以参考issue。 ## 关联的Issue https://gitcode.com/cann/catlass/issues/233 ## 测试已对每一个EVG类型的模板样例分别进行200个case测试，结果通过 ## 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!679	2 个月前
type_traits.hpp	Complete and refine tile-level unittests. Co-authored-by: init__zhb__<zhanghaobo6@huawei.com> # message auto-generated for no-merge-commit merge: !730 merge ut.dev into master 补充TLA Tile组件UT测试件 Created-by: init__zhb__ Commit-by: init__zhb__ Merged-by: cann-robot Description: ## 描述补充TLA Tile组件UT测试件。 \|数据通路\|被测组件\|源文件\|非TLA 特化数\|TLA/Sparse 特化数\| \|---\|---\|---\|---\|---\| \|GM→L1\|CopyGmToL1 / GMMPTD / DynamicOptimized / IntervalDataCopy\|\|\~24\|10\| \|L1→L0A\|CopyL1ToL0A \|\|10\|6\| \|L1→L0B\|CopyL1ToL0B \|\|13\|6\| \|L0C→GM \|CopyL0CToGm \|\|7\|3\| \|GM→UB\|CopyGm2Ub / CopyGm2UbAligned / CopyPerTokenScale2Ub\|\|4\|1\| \|UB→GM\|CopyUb2Gm / CopyUb2GmAligned\|\|3\|1\| \|MMAD \|TileMmad\|\|2\|2\| \|合计\|\|\|\~63\|\~27\| - 特别说明：本PR涉及对公共`MakeIntegerSequence`的修改，原因在于后续`make_index_sequence = MakeIntegerSequence<size_t, N>`和`make_int_sequence = MakeIntegerSequence<int, N>`中涉及`int`和`size_t`两种类型，而`MakeIntegerSequenceImpl`中展开量的类型限定为`size_t`，GCC/Clang stub 编译阶段严格拒绝了可能的隐式转换。 ## 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 ### UT组件通过情况 - Arch2201: `bash ... [----------] Global test environment tear-down [==========] 581 tests from 19 test suites ran. (129 ms total) [ PASSED ] 548 tests. [ SKIPPED ] 33 tests, listed below: [ SKIPPED ] CopyGmToL1/TileCopyGmToL1TestAtlasA2.RowMajorTozNTestOneRow_GMMPTD/0 [ SKIPPED ] CopyGmToL1/TileCopyGmToL1TestAtlasA2.RowMajorTozNTestOneRow_GMMPTD/2 ...` - Arch3510: `bash ... [ RUN ] Ascend950/TypedTileMmadA950Suite.DisableGemvMmad/14 [ OK ] Ascend950/TypedTileMmadA950Suite.DisableGemvMmad/14 (0 ms) [----------] 120 tests from Ascend950/TypedTileMmadA950Suite (2 ms total) [----------] Global test environment tear-down [==========] 383 tests from 20 test suites ran. (32 ms total) [ PASSED ] 383 tests.` ### 覆盖率 <span style="color:green">FULL</span> bash Overall coverage rate: lines......: 100.0% (2486 of 2486 lines) functions..: 99.3% (1480 of 1490 functions) Reading tracefile /catlass/build/coverage/coverage.info \|Lines \|Functions\|Branches Filename \|Rate Num\|Rate Num\|Rate Num ================================================================================ [/catlass/include/] catlass/coord.hpp \| 100% 17\| 100% 23\| - 0 catlass/detail/alignment.hpp \| 100% 10\| 100% 25\| - 0 catlass/epilogue/tile/copy_gm_to_ub.hpp \| 100% 68\| 100% 6\| - 0 catlass/epilogue/tile/copy_gm_to_ub_tla.hpp \| 100% 19\| 100% 2\| - 0 catlass/epilogue/tile/copy_ub_to_gm.hpp \| 100% 45\| 100% 5\| - 0 catlass/epilogue/tile/copy_ub_to_gm_tla.hpp \| 100% 10\| 100% 1\| - 0 catlass/gemm/tile/ascend950/copy_gm_to_l1.hpp \| 100% 417\| 100% 46\| - 0 catlass/gemm/tile/ascend950/copy_l0c_to_gm.hpp \| 100% 99\| 100% 10\| - 0 catlass/gemm/tile/ascend950/copy_l0c_to_ub.hpp \| 100% 44\| 100% 3\| - 0 catlass/gemm/tile/ascend950/copy_l1_to_bt.hpp \| 100% 12\| 100% 4\| - 0 catlass/gemm/tile/ascend950/copy_l1_to_l0a.hpp \| 100% 98\| 100% 16\| - 0 catlass/gemm/tile/ascend950/copy_l1_to_l0b.hpp \| 100% 162\| 100% 20\| - 0 catlass/gemm/tile/atlasa2/copy_gm_to_l1.hpp \| 100% 611\| 100% 68\| - 0 catlass/gemm/tile/atlasa2/copy_gm_to_ub.hpp \| 100% 12\| 100% 2\| - 0 catlass/gemm/tile/atlasa2/copy_l0c_to_gm.hpp \| 100% 69\| 100% 9\| - 0 catlass/gemm/tile/atlasa2/copy_l1_to_l0a.hpp \| 100% 171\| 100% 22\| - 0 catlass/gemm/tile/atlasa2/copy_l1_to_l0b.hpp \| 100% 219\| 100% 26\| - 0 catlass/gemm/tile/atlasa2/copy_ub_to_gm.hpp \| 100% 11\| 100% 2\| - 0 catlass/gemm/tile/tile_mmad.hpp \| 100% 91\| 100% 20\| - 0 catlass/layout/matrix.hpp \| 100% 139\| 100% 63\| - 0 catlass/layout/vector.hpp \| 100% 5\| 100% 3\| - 0 catlass/numeric_size.hpp \| 100% 2\| 100% 1\| - 0 tla/int_tuple.hpp \| 100% 16\| 100% 47\| - 0 tla/layout.hpp \| 100% 85\|99.6% 244\| - 0 tla/numeric/integral_constant.hpp \| 100% 4\| 100% 17\| - 0 tla/numeric/math.hpp \| 100% 4\| 100% 3\| - 0 tla/tensor.hpp \| 100% 24\|99.2% 241\| - 0 tla/tuple.hpp \| 100% 22\|98.8% 561\| - 0 ================================================================================ Total:\| 100% 2486\|99.3% 1k\| - 0 [100%] Built target coverage_collect ### 变异测试（部分） 1. src和dst typo错误混杂 Ref bugfix commit: 02d1078 Mutation: diff --- a/include/catlass/gemm/tile/ascend950/copy_l1_to_l0a.hpp +++ b/include/catlass/gemm/tile/ascend950/copy_l1_to_l0a.hpp @@ -626,7 +626,7 @@ struct CopyL1ToL0A<Arch::Ascend950, Gemm::GemmType<Element, layout::nZ, AscendC: const uint32_t L0M = layoutDst.shape(0) * layoutDst.shape(1); const uint32_t L0K = layoutDst.shape(2) * layoutDst.shape(3); const uint32_t srcOuterStrideRow = layoutSrc.stride(1); - const uint32_t dstOuterStrideCol = layoutDst.stride(3); + const uint32_t dstOuterStrideCol = layoutSrc.stride(3); AscendC::LoadData2DParamsV2 loadDataParams; loadDataParams.mStartPosition = 0; Result: <span style="color:green">KILLED</span> `bash [ FAILED ] 3 tests, listed below: [ FAILED ] CopyL1ToL0A/TileCopyL1ToL0ATestAscend950.nZTozNTestBasic/0, where GetParam() = 8-byte object <80-00 00-00 40-00 00-00> [ FAILED ] CopyL1ToL0A/TileCopyL1ToL0ATestAscend950.nZTozNTestBasic/2, where GetParam() = 8-byte object <40-00 00-00 2A-00 00-00> [ FAILED ] CopyL1ToL0A/TileCopyL1ToL0ATestAscend950.nZTozNTestBasic/3, where GetParam() = 8-byte object <7B-00 00-00 08-00 00-00> 3 FAILED TESTS` 2. 不对齐的情况 Ref bugfix commit: 0439b54e Mutation: diff --- a/include/catlass/gemm/tile/ascend950/copy_l0c_to_ub.hpp +++ b/include/catlass/gemm/tile/ascend950/copy_l0c_to_ub.hpp @@ -148,7 +148,7 @@ struct CopyL0CToUBTla< AscendC::FixpipeParamsC310<AscendC::CO2Layout::ROW_MAJOR> intriParams; // Fixpipe layout information - intriParams.nSize = RoundUp(tla::get<1>(dstTensor.originShape()), 32); // n must be multiple of 32 when split n + intriParams.nSize = tla::get<1>(dstTensor.originShape()); // n must be multiple of 32 when split n intriParams.mSize = tla::get<0>(dstTensor.originShape()); intriParams.srcStride = tla::get<1, 1>(srcTensor.stride()) / tla::get<0, 0>(srcTensor.stride()); intriParams.dstStride = tla::get<0>(dstTensor.stride()); Result: <span style="color:green">KILLED</span> `[ FAILED ] 1 test, listed below: [ FAILED ] CopyL0CToUbTlaAscend950/TileCopyL0CToUbTlaAscend950Test.L0CToRowMajorTestSplitN/1, where GetParam() = 8-byte object <41-00 00-00 7F-00 00-00> 1 FAILED TEST` 3. 常量错误 Ref bugfix commit: 81e1129e Mutation: diff --- a/include/catlass/gemm/tile/ascend950/copy_l1_to_bt.hpp +++ b/include/catlass/gemm/tile/ascend950/copy_l1_to_bt.hpp @@ -80,7 +80,7 @@ struct CopyL1ToBT<Arch::Ascend950, Catlass::Gemm::GemmType<ElementSrc, layout::V using LayoutDst = layout::VectorLayout; using LayoutSrc = layout::VectorLayout; - static constexpr uint32_t ELE_NUM_PER_C0 = BytesToBits(BYTE_PER_C0) / SizeOfBits<ElementSrc>::value; + static constexpr uint32_t ELE_NUM_PER_C2 = BytesToBits(BYTE_PER_C2) / SizeOfBits<ElementSrc>::value; CATLASS_DEVICE CopyL1ToBT(){} @@ -93,7 +93,7 @@ struct CopyL1ToBT<Arch::Ascend950, Catlass::Gemm::GemmType<ElementSrc, layout::V ){ AscendC::DataCopyParams intriParams; intriParams.blockCount = 1; - intriParams.blockLen = CeilDiv(layoutDst.shape(0), ELE_NUM_PER_C0); + intriParams.blockLen = CeilDiv(layoutDst.shape(0), ELE_NUM_PER_C2); if (sizeof(ElementSrc) == 4) { Result: <span style="color:yellow">SURVIVED</span> <span style="color:green">KILLED</span>(commit:b1d60262) bash [ FAILED ] 6 tests, listed below: [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestHalf/0, where GetParam() = 8-byte object <80-00 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestHalf/1, where GetParam() = 8-byte object <43-00 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestHalf/2, where GetParam() = 8-byte object <00-01 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestFloat/0, where GetParam() = 8-byte object <80-00 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestFloat/1, where GetParam() = 8-byte object <43-00 00-00 01-00 00-00> [ FAILED ] Ascend950/TypedL1ToBTA950Suite.VectorToVectorTestFloat/2, where GetParam() = 8-byte object <00-01 00-00 01-00 00-00> 6 FAILED TESTS ## 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 - [ ] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他，请描述： See merge request: cann/catlass!730	1 个月前