[Bug-Report|缺陷反馈]: ENABLE_CV_COMM_VIA_SSBUF在950上是不是默认打开？-asc-devkit-AtomGit

已关闭

[Bug-Report|缺陷反馈]: ENABLE_CV_COMM_VIA_SSBUF在950上是不是默认打开？ #940

zjun创建于 6月12日关闭于 11 天前

6月12日创建

Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.

Describe the current behavior / 问题描述 (Mandatory / 必填)

在体验cv融合算子
https://gitcode.com/cann/asc-devkit/blob/master/examples/01_simd_cpp_api/00_introduction/03_fusion_operation/matmul_leakyrelu_advanced_api/README.md
时，发现了一个问题，就是950上新增的L0C → FixPipe → UB的直接通路，在不添加编译选项ENABLE_CV_COMM_VIA_SSBUF的情况下，是默认不能使能该特性的。
前期问题具体见Issue：https://gitcode.com/cann/asc-devkit/issues/923

搜索ops-nn/ops-math/ops-transformer/ops-cv这4个算子仓库发现，均存在相关配置，举例ops-nn仓库，如：

$ grep 'ENABLE_CV_COMM_VIA_SSBUF'  *  -rn
matmul/fused_mat_mul/op_kernel/CMakeLists.txt:14:    OPTIONS "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true"
matmul/mat_mul_v3/op_kernel/CMakeLists.txt:32:    OPTIONS "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true"
matmul/quant_batch_matmul_v3/op_kernel/CMakeLists.txt:20:    OPTIONS "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true"
matmul/rotate_quant/op_kernel/CMakeLists.txt:32:    OPTIONS "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true"
matmul/transpose_quant_batch_mat_mul/op_kernel/CMakeLists.txt:19:    OPTIONS "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true"
matmul/weight_quant_batch_matmul_v2/op_kernel/CMakeLists.txt:19:    OPTIONS "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true"
scripts/kernel/binary_config/ascendc_config.json:79:    {"name":"DynamicRNN", "compute_units": ["ascend910b", "ascend910_93"], "auto_sync" : false, "compile_options": {"ascend950": ["-DENABLE_CV_COMM_VIA_SSBUF=true"]}},
scripts/kernel/binary_config/ascendc_config.json:82:    {"name":"DynamicRNNV2", "compute_units": ["ascend910b", "ascend910_93"], "auto_sync" : false, "compile_options": {"ascend950": ["-DENABLE_CV_COMM_VIA_SSBUF=true"]}},
scripts/kernel/binary_config/ascendc_config.json:411:    {"name":"Conv2DV2", "compute_units": ["ascend950", "mc62"], "auto_sync" : false, "impl_mode" : "", "compile_options": {"ascend950": ["-DENABLE_CV_COMM_VIA_SSBUF=true", "--cce-no-dcache-flush", "-DFORCE_TILING_CONST_PROPAGATION"], "mc62": ["-DFORCE_TILING_CONST_PROPAGATION", "--cce-no-dcache-flush"]}},
scripts/kernel/binary_config/ascendc_config.json:456:    {"name":"FlatQuant", "compute_units": ["ascend910b", "ascend910_93", "ascend950"], "auto_sync" : false, "compile_options": {"ascend950": ["-DENABLE_CV_COMM_VIA_SSBUF=true"]}},
scripts/kernel/binary_config/ascendc_config.json:500:    {"name":"QuantBatchMatmulV3", "compute_units": ["ascend910b", "ascend910_93", "ascend950", "kirinx90", "kirin9030"], "auto_sync": false, "compile_options": {"ascend950":["--cce-no-dcache-flush", "-DENABLE_CV_COMM_VIA_SSBUF=true"]}},
scripts/kernel/binary_config/ascendc_config.json:513:    {"name":"WeightQuantBatchMatmulV2", "compute_units": ["ascend910b", "ascend910_93", "ascend310p", "ascend950"], "auto_sync": false, "compile_options": {"ascend950":["--cce-no-dcache-flush", "-DENABLE_CV_COMM_VIA_SSBUF=true"]}},
scripts/kernel/binary_config/ascendc_config.json:524:    {"name":"Conv3DBackpropFilterV2", "compute_units": ["ascend910b", "ascend910_93", "ascend950"], "auto_sync" : false, "compile_options": {"ascend950": ["--cce-no-dcache-flush","-DENABLE_CV_COMM_VIA_SSBUF=true"]}},
scripts/kernel/binary_config/ascendc_config.json:600:    {"name":"TransposeQuantBatchMatMul", "compute_units": ["ascend950"], "auto_sync": false, "compile_options": {"ascend950":["--cce-no-dcache-flush", "-DENABLE_CV_COMM_VIA_SSBUF=true"]}},

算子中确实有针对相关算子添加编译选项的说明，问题是：
1、各个仓库，我发现添加DENABLE_CV_COMM_VIA_SSBUF写法不统一，有的是CMakeLists.txt，有的是json。就是算CMakeLists.txt，写法也不统一：
比如：ops-transformer/mc2/allto_all_matmul/op_host/CMakeLists.txt

ops-nn/matmul/rotate_quant/op_kernel/CMakeLists.txt

2、关键问题：对于一个刚入门的算子开发者，除了在编译选项中说明外，是不是提供更好的指导？否则都发现不了这个问题。

Environment / 环境信息 (Mandatory / 必填)

(base) root@50cfeb89dca6:.../03_matrix_vector_fusion/matmul_leakyrelu/build# npu-smi info
+-------------------------------------------------------------------------------------------------+
| npu-smi 25.7.rc1 Version: 25.7.rc1 |
+--------+------------------+---------------+-----------------------------------------------------+
| NPU ID | Name | Health | Power(W) Temp(C) Hugepages-Usage(page) |
| | | Bus-Id | NPU Util(%) Memory-Usage(MB) HBM-Usage(MB) |
+========+==================+===============+=====================================================+
| 0 | Ascend950PR | OK | 203.1 59 0 / 0 |
| | | 0000:81:00.0 | 0 0 / 0 5393 / 131072 |
+========+==================+===============+=====================================================+
+---------------------------+---------------+-----------------------------------------------------+
| NPU ID | Process id | Process name | Process memory(MB) |
+===========================+===============+=====================================================+
| No running processes found in NPU 0 |
+===========================+===============+=====================================================+

cann : v9.0.0