mmad算子尝试改为int8类型

一、问题现象（附报错日志上下文）：
将mmad示例（链接为operator/ascendc/0_introduction/20_mmad_kernellaunch/MmadBiasInvocation）中half类型改成int8_t类型后，计算结果出错。
我后面参见mmad的api文档尝试修改splitA、splitB这些里面的分型，因为我看到似乎A矩阵分型是16*32的，B矩阵的32*16，但是整来整去似乎还是不对，这个nd、nz格式的以及随路格式转换的整的很迷惑。所以如果要调整为int8类型的mmad，需要如何调整代码？
整个运行结果如下：
-- The C compiler identification is GNU 11.4.0
-- The CXX compiler identification is GNU 11.4.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Configuring done
-- Generating done
-- Build files have been written to: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build
[  2%] Creating directories for 'ascendc_kernels_npu_precompile'
[  4%] No download step for 'ascendc_kernels_npu_precompile'
[  6%] No update step for 'ascendc_kernels_npu_precompile'
[  9%] No patch step for 'ascendc_kernels_npu_precompile'
[ 11%] Performing configure step for 'ascendc_kernels_npu_precompile'
-- The C compiler identification is GNU 11.4.0
-- The CXX compiler identification is GNU 11.4.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Configuring done
-- Generating done
-- Build files have been written to: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_precompile-prefix/src/ascendc_kernels_npu_precompile-build
[ 13%] Performing build step for 'ascendc_kernels_npu_precompile'
[100%] Building CXX object CMakeFiles/precompile_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o
[100%] Built target precompile_obj
[100%] Built target check_src_template
[ 15%] No install step for 'ascendc_kernels_npu_precompile'
[ 18%] Completed 'ascendc_kernels_npu_precompile'
[ 18%] Built target ascendc_kernels_npu_precompile
[ 20%] Creating directories for 'ascendc_kernels_npu_preprocess'
[ 22%] No download step for 'ascendc_kernels_npu_preprocess'
[ 25%] No update step for 'ascendc_kernels_npu_preprocess'
[ 27%] No patch step for 'ascendc_kernels_npu_preprocess'
[ 29%] Performing configure step for 'ascendc_kernels_npu_preprocess'
-- The C compiler identification is GNU 11.4.0
-- The CXX compiler identification is GNU 11.4.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Configuring done
-- Generating done
-- Build files have been written to: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_preprocess-prefix/src/ascendc_kernels_npu_preprocess-build
[ 31%] Performing build step for 'ascendc_kernels_npu_preprocess'
[100%] Building CXX object CMakeFiles/aic_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o
[100%] Building CXX object CMakeFiles/preprocess_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o
[100%] Building CXX object CMakeFiles/aiv_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o
[100%] Built target preprocess_obj
[100%] Built target aic_obj
/usr/local/Ascend/ascend-toolkit/latest/tools/ccec_compiler/bin/ld.lld -m aicorelinux -Ttext=0 /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_preprocess-prefix/src/ascendc_kernels_npu_preprocess-build/CMakeFiles/aic_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o -o /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_preprocess-prefix/src/ascendc_kernels_npu_preprocess-build/CMakeFiles/aic_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o[100%] Built target merge_aic_obj_text
[100%] Built target aiv_obj
/usr/local/Ascend/ascend-toolkit/latest/tools/ccec_compiler/bin/ld.lld -m aicorelinux -Ttext=0 /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_preprocess-prefix/src/ascendc_kernels_npu_preprocess-build/CMakeFiles/aiv_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o -o /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_preprocess-prefix/src/ascendc_kernels_npu_preprocess-build/CMakeFiles/aiv_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o[100%] Built target merge_aiv_obj_text
[100%] Built target _host_cpp
[ 34%] No install step for 'ascendc_kernels_npu_preprocess'
[ 36%] Completed 'ascendc_kernels_npu_preprocess'
[ 36%] Built target ascendc_kernels_npu_preprocess
[ 43%] Creating directories for 'ascendc_kernels_npu_aic_device'
[ 43%] Creating directories for 'ascendc_kernels_npu_host'
[ 43%] Creating directories for 'ascendc_kernels_npu_aiv_device'
[ 50%] No download step for 'ascendc_kernels_npu_aic_device'
[ 50%] No download step for 'ascendc_kernels_npu_aiv_device'
[ 50%] No download step for 'ascendc_kernels_npu_host'
[ 56%] No update step for 'ascendc_kernels_npu_aic_device'
[ 56%] No update step for 'ascendc_kernels_npu_host'
[ 56%] No update step for 'ascendc_kernels_npu_aiv_device'
[ 63%] No patch step for 'ascendc_kernels_npu_host'
[ 63%] No patch step for 'ascendc_kernels_npu_aic_device'
[ 63%] No patch step for 'ascendc_kernels_npu_aiv_device'
[ 70%] Performing configure step for 'ascendc_kernels_npu_host'
[ 70%] Performing configure step for 'ascendc_kernels_npu_aic_device'
[ 70%] Performing configure step for 'ascendc_kernels_npu_aiv_device'
-- The C compiler identification is GNU 11.4.0
-- The C compiler identification is GNU 11.4.0
-- The C compiler identification is GNU 11.4.0
-- The CXX compiler identification is GNU 11.4.0
-- The CXX compiler identification is GNU 11.4.0
-- The CXX compiler identification is GNU 11.4.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compiler ABI info - done
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Detecting CXX compile features - done
-- Detecting CXX compiler ABI info - done
-- Configuring done
-- Configuring done
-- Generating done
-- Build files have been written to: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_aiv_device-prefix/src/ascendc_kernels_npu_aiv_device-build
-- Generating done
-- Build files have been written to: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_aic_device-prefix/src/ascendc_kernels_npu_aic_device-build
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Configuring done
-- Generating done
-- Build files have been written to: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_host-prefix/src/ascendc_kernels_npu_host-build
[ 75%] Performing build step for 'ascendc_kernels_npu_aic_device'
[ 75%] Performing build step for 'ascendc_kernels_npu_aiv_device'
[ 77%] Performing build step for 'ascendc_kernels_npu_host'
[ 79%] No install step for 'ascendc_kernels_npu_aiv_device'
[100%] Building CXX object CMakeFiles/device_aic_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/auto_gen/ascendc_kernels_npu/auto_gen_mmad_custom.cpp.o
[100%] Building CXX object CMakeFiles/host_bisheng_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o
[ 81%] Completed 'ascendc_kernels_npu_aiv_device'
[ 81%] Built target ascendc_kernels_npu_aiv_device
[100%] Built target host_bisheng_obj
[ 84%] Performing install step for 'ascendc_kernels_npu_host'
Consolidate compiler generated dependencies of target host_bisheng_obj
[100%] Building CXX object CMakeFiles/host_bisheng_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o
[100%] Built target host_bisheng_obj
Install the project...
-- Install configuration: "Debug"
-- Installing: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_host_dir/./objects-Debug/host_bisheng_obj/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o
[ 86%] Completed 'ascendc_kernels_npu_host'
[ 86%] Built target ascendc_kernels_npu_host
[100%] Built target device_aic_obj
/usr/local/Ascend/ascend-toolkit/latest/tools/ccec_compiler/bin/ld.lld  -m aicorelinux -r  -Ttext=0 /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_aic_device-prefix/src/ascendc_kernels_npu_aic_device-build/CMakeFiles/device_aic_obj.dir/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/auto_gen/ascendc_kernels_npu/auto_gen_mmad_custom.cpp.o -static -o /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_aic_device_dir/device_aic.o
[100%] Built target merge_aic_device_obj
[ 88%] No install step for 'ascendc_kernels_npu_aic_device'
[ 90%] Completed 'ascendc_kernels_npu_aic_device'
[ 90%] Built target ascendc_kernels_npu_aic_device
/usr/local/Ascend/ascend-toolkit/latest/tools/ccec_compiler/bin/ld.lld  -m aicorelinux   -Ttext=0 /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_aic_device_dir/device_aic.o -static -o /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_merge_obj_dir/device_aic.o
[ 90%] Built target ascendc_kernels_npu_merge_obj
[ 93%] Building CXX object CMakeFiles/ascendc_kernels_npu_host_stub_obj.dir/auto_gen/ascendc_kernels_npu/host_stub.cpp.o
[ 93%] Built target ascendc_kernels_npu_host_stub_obj
[ 95%] Linking CXX shared library lib/libascendc_kernels_npu.so
/usr/local/Ascend/ascend-toolkit/latest/bin/ascendc_pack_kernel /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/CMakeFiles/ascendc_kernels_npu_host_stub_obj.dir/auto_gen/ascendc_kernels_npu/host_stub.cpp.o /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_merge_obj_dir/device_aic.o 2 /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/CMakeFiles/ascendc_kernels_npu_host_stub_obj.dir/auto_gen/ascendc_kernels_npu/host_stub.cpp.o
recompile: /usr/bin/c++ -fPIC -g -Wl,-z,relro -Wl,-z,now -Wl,-z,noexecstack -shared -Wl,-soname,libascendc_kernels_npu.so -o lib/libascendc_kernels_npu.so CMakeFiles/ascendc_kernels_npu_host_stub_obj.dir/auto_gen/ascendc_kernels_npu/host_stub.cpp.o /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/build/ascendc_kernels_npu_host_dir/objects-Debug/host_bisheng_obj/root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/mmad_custom.cpp.o    -L/usr/local/Ascend/ascend-toolkit/latest/lib64  -L/usr/local/Ascend/ascend-toolkit/latest/tools/simulator/Ascend910B2/lib  /usr/local/Ascend/ascend-toolkit/latest/lib64/libascendc_runtime.a -lascend_dump -lc_sec
[ 95%] Built target ascendc_kernels_npu
[ 97%] Building CXX object CMakeFiles/ascendc_kernels_bbit.dir/main.cpp.o
[100%] Linking CXX executable ascendc_kernels_bbit
[100%] Built target ascendc_kernels_bbit
-- Install configuration: "Debug"
-- Installing: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/out/lib/libascendc_kernels_npu.so
-- Installing: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/out/include
-- Installing: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/out/include/ascendc_kernels_npu
-- Installing: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/out/include/ascendc_kernels_npu/aclrtlaunch_triple_chevrons_func.h
-- Installing: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/out/include/ascendc_kernels_npu/aclrtlaunch_mmad_custom.h
-- Installing: /root/wja/project/samples/operator/ascendc/0_introduction/20_mmad_kernellaunch_my/MmadBiasInvocation/out/bin/ascendc_kernels_bbit

计算完成，最终结果为 int32 类型。

Golden result (first 4x4 block):
[[751 693 644 657]
 [792 727 634 724]
 [747 693 706 642]
 [673 651 627 661]]
opType=mmad_custom, DumpHead: AIC-0, CoreType=AIC, block dim=1, total_block_num=1, block_remain_len=1041504, block_initial_space=1048576, rsv=0, magic=5aa5bccd
CANN Version: 8.2.RC1.alpha002, TimeStamp: 20250522125202340
DumpTensor: desc=1, addr=0, data_type=int8, position=L1
[[3,9,6,5,4,1,7,6,2,8,1,9,9,1,2,6,7,7,4,9,3,2,7,1,1,3,6,6,8,7,9,5],
[9,1,5,6,1,5,9,8,2,2,5,4,8,7,9,3,5,8,8,3,1,3,8,5,1,4,1,2,6,9,9,4],
[9,6,6,7,5,1,5,3,6,1,2,3,5,6,1,5,3,2,9,7,7,6,2,1,8,4,3,7,9,2,8,3],
[6,4,1,2,8,5,1,6,1,4,5,4,7,4,8,1,7,1,3,7,5,2,7,1,5,8,8,3,9,3,4,1],
[8,8,9,1,5,1,5,4,4,6,3,5,2,6,7,7,4,2,8,1,7,9,9,5,8,5,8,5,2,1,5,2],
[1,5,9,5,7,6,3,1,5,5,5,7,9,5,1,6,3,9,2,4,5,1,7,4,8,6,1,6,5,4,1,8],
[2,2,6,5,8,3,2,2,9,9,3,5,3,7,7,3,2,3,6,6,7,3,5,9,2,8,1,7,3,5,2,7],
[6,4,9,5,2,4,9,8,8,3,4,7,6,9,3,4,3,2,8,8,2,6,3,2,9,9,7,1,6,7,1,4],
[9,4,3,9,8,5,6,5,9,7,5,4,3,2,9,7,7,9,7,8,3,6,4,2,8,8,7,1,5,5,7,4],
[3,9,1,7,1,3,3,2,2,9,1,4,9,5,1,9,8,4,9,5,4,6,7,4,9,6,8,1,7,3,7,5],
[9,2,1,4,9,2,8,5,2,3,6,2,5,3,8,4,3,5,5,7,7,8,5,8,2,5,9,9,4,1,5,3],
[7,3,9,2,4,8,2,9,3,6,6,3,3,3,1,7,9,5,2,8,7,6,9,7,7,8,8,5,2,8,5,2],
[2,4,9,8,4,1,7,3,4,8,5,2,7,7,1,3,2,7,3,8,4,2,2,9,9,3,6,3,3,1,1,1],
[3,3,6,5,7,9,7,9,4,6,2,3,3,7,9,3,9,8,6,8,7,5,9,2,7,4,6,7,9,9,2,4],
[2,1,8,5,4,8,7,8,9,6,3,5,9,2,9,2,5,9,8,3,5,5,6,9,9,6,4,4,4,8,2,9],
[2,3,1,6,1,8,5,3,2,7,2,3,1,9,5,3,4,5,4,3,2,1,4,4,4,9,6,7,2,3,5,6],
[4,9,6,2,7,4,7,3,7,5,4,7,6,1,7,8,6,7,8,8,5,3,2,8,8,4,8,7,7,5,3,9],
[4,8,3,9,1,1,2,6,2,5,8,6,6,9,3,7,2,4,9,8,5,9,2,2,1,3,2,3,5,1,7,2],
[7,6,4,4,4,4,9,5,2,8,3,9,4,6,6,4,3,4,1,1,9,2,4,7,1,7,6,2,4,8,3,9],
[7,2,3,1,6,9,3,8,5,9,6,1,9,4,5,5,2,9,5,1,2,1,2,9,7,4,7,4,2,1,8,6],
[5,3,2,4,6,8,8,8,6,3,8,5,9,4,1,7,7,7,4,3,1,2,8,3,2,1,1,2,7,6,5,5],
[9,4,6,2,5,7,9,8,4,9,6,7,5,9,9,8,2,4,4,1,5,1,9,5,6,6,6,5,8,6,4,2],
[8,9,2,1,1,2,4,3,7,2,9,5,5,8,6,7,9,6,1,3,3,7,1,2,8,9,1,1,6,3,1,5],
[3,5,6,6,7,6,5,9,4,5,5,9,8,8,7,6,3,9,6,7,5,1,4,1,9,7,7,3,8,5,7,9],
[8,3,4,8,1,8,5,3,2,8,4,1,4,1,6,2,9,3,5,8,2,1,2,5,2,7,3,2,4,2,3,3],
[4,7,8,4,6,3,2,2,6,6,1,6,1,2,1,8,5,6,1,1,5,7,4,6,1,4,3,4,9,8,8,9],
[1,7,2,5,8,7,3,1,9,9,8,5,4,9,2,1,6,2,4,7,3,9,1,1,2,5,7,7,8,6,5,3],
[4,8,4,3,1,6,7,4,1,3,2,9,9,4,3,7,7,2,3,8,5,7,9,6,3,4,2,7,7,2,8,8],
[8,9,2,4,7,5,7,5,1,1,3,3,8,6,6,7,7,8,2,2,2,8,5,1,7,7,6,9,8,9,8,3],
[5,1,2,8,2,5,6,1,5,2,2,7,6,7,9,1,3,2,5,5,1,7,6,3,8,5,5,6,1,8,5,6],
[9,9,3,4,2,9,1,9,4,7,9,5,8,1,1,4,1,4,6,5,4,4,1,4,1,5,8,1,8,5,2,5],
[1,5,3,3,4,6,1,9,4,5,6,5,1,8,3,9,4,4,1,1,6,5,9,7,4,4,3,5,1,1,1,9]]
DumpTensor: desc=2, addr=400, data_type=int8, position=L1
[[6,3,7,5,7,6,1,2,4,6,5,3,1,3,6,5,2,6,6,8,7,3,3,6,3,2,6,9,9,6,3,1],
[4,1,3,1,4,6,9,8,5,5,6,5,5,1,9,6,5,5,5,1,8,1,6,3,9,7,4,3,5,3,2,7],
[9,5,4,2,9,6,8,4,5,2,7,3,1,8,3,3,5,1,2,4,2,3,7,6,3,1,1,3,4,5,1,3],
[3,6,5,9,5,3,9,5,2,5,4,2,2,5,5,4,2,3,6,1,5,4,5,2,2,3,3,1,7,9,1,2],
[4,1,7,6,5,9,8,6,1,9,7,9,5,8,4,3,9,2,2,7,7,9,4,6,7,5,7,2,8,6,4,8],
[9,7,1,6,5,8,8,5,6,4,8,9,6,5,1,9,5,1,2,2,7,3,7,4,3,8,6,2,9,1,9,7],
[6,2,9,2,8,7,5,2,8,9,6,9,8,4,5,9,9,8,2,6,3,1,3,9,3,3,8,7,2,5,7,5],
[9,3,6,9,4,1,4,4,2,8,5,1,1,9,7,4,5,1,1,2,7,3,7,8,9,4,9,5,3,8,4,6],
[4,4,7,3,4,6,6,4,7,2,1,1,9,5,6,5,4,6,3,8,2,1,6,5,7,9,6,5,2,3,3,3],
[1,3,2,1,5,1,7,8,5,6,3,5,7,6,5,4,2,8,7,1,9,1,9,5,4,7,1,7,3,5,9,2],
[7,6,9,6,7,9,7,2,2,1,4,3,4,9,3,4,7,3,6,5,6,2,5,9,8,1,2,5,8,7,9,6],
[3,9,4,1,3,5,1,2,5,8,2,8,3,1,1,8,1,2,8,6,5,1,8,5,7,4,1,2,1,8,1,3],
[1,4,1,5,6,9,9,3,2,1,1,6,1,2,7,5,5,4,6,5,7,9,1,6,9,1,3,5,6,2,1,2],
[7,8,5,2,3,7,9,9,7,7,8,1,2,8,8,8,1,1,2,2,2,1,4,7,9,3,1,7,8,9,6,2],
[1,4,2,7,2,9,9,2,3,2,7,2,6,6,9,7,7,8,5,5,9,7,6,1,3,5,3,9,8,9,6,4],
[4,6,1,8,3,2,8,9,3,3,1,2,8,5,9,3,5,3,7,1,2,2,8,8,2,9,5,2,4,8,6,5],
[4,7,6,7,1,5,9,3,5,7,6,9,8,2,4,1,2,1,1,2,5,8,4,1,6,1,9,9,9,1,6,2],
[1,2,3,2,3,1,4,9,5,3,8,2,4,2,1,5,3,4,4,6,6,5,3,5,3,9,8,3,9,5,6,3],
[5,8,3,7,2,1,7,5,2,3,7,2,5,3,8,8,6,2,9,7,7,5,4,7,2,6,5,4,1,7,3,7],
[4,5,6,2,5,5,3,2,9,3,2,2,9,8,6,9,1,3,3,4,2,2,6,4,1,5,9,4,3,8,1,1],
[7,7,8,1,5,2,6,2,5,9,8,8,7,9,9,8,4,4,6,7,9,1,5,7,7,1,3,9,8,1,5,1],
[3,2,1,8,3,8,1,5,4,1,9,5,1,5,5,8,7,9,1,4,7,9,8,6,4,1,7,4,4,3,2,4],
[6,5,1,4,3,1,2,6,3,5,3,9,1,1,4,2,8,9,5,4,8,1,8,1,1,8,3,8,9,3,9,5],
[7,8,1,1,8,7,7,4,1,3,4,7,1,5,3,1,6,9,4,5,3,8,2,3,6,2,7,8,5,2,9,2],
[5,7,7,4,6,2,3,1,1,7,5,9,3,2,9,4,4,4,2,3,1,4,1,6,8,9,9,2,7,4,8,4],
[1,3,6,5,6,4,6,4,4,1,4,7,8,8,4,1,6,4,2,3,8,3,5,6,6,9,4,5,2,8,4,7],
[8,7,4,9,5,4,3,5,1,8,2,5,6,9,3,3,6,4,8,9,7,7,3,6,9,8,1,2,4,2,1,2],
[6,1,4,1,5,7,1,1,3,3,2,1,3,5,9,8,6,8,5,6,5,4,1,2,2,3,6,3,8,3,2,9],
[5,3,5,3,9,5,1,2,7,5,2,2,8,4,9,6,2,5,4,3,7,4,9,6,2,2,5,9,2,1,1,5],
[7,1,1,3,3,7,2,2,9,5,3,8,2,5,5,7,4,5,8,7,4,3,9,1,4,7,7,9,8,6,1,6],
[3,5,1,1,3,4,4,5,7,4,5,3,1,8,9,4,5,9,3,4,2,6,3,9,6,4,4,8,7,5,2,3],
[7,2,5,9,1,2,7,8,6,2,3,9,3,3,8,6,6,1,3,5,8,5,8,7,2,4,6,6,1,9,2,7]]
DumpTensor: desc=3, addr=800, data_type=int32, position=L1
[5, 7, 3, 5, 6, 4, 2, 3, 1, 8, 8, 5, 6, 4, 9, 2, 1, 8, 5, 4, 7, 4, 1, 9, 9, 3, 8, 1, 6, 8, 6, 
7]
a2Local: 
a2[0]: 0, a2[1]: 0, a2[2]: 0, a2[3]: 0
b2Local: 
b2[0]: 0, b2[1]: 0, b2[2]: 0, b2[3]: 0
bias2Local: 
bias2[0]: 0, bias2[1]: 0, bias2[2]: 0, bias2[3]: 0
DumpTensor: desc=2, addr=0, data_type=int32, position=L0C
[[712,796,852,978,799,705,866,930,709,703,684,901,844,734,880,704,737,737,854,849,707,686,776,917,696,721,736,899,773,745,869,697],
[639,656,815,946,674,659,733,888,700,671,694,860,730,682,893,672,694,656,766,805,656,682,644,768,623,567,663,778,694,620,753,594],
[767,770,894,886,756,732,735,905,612,766,769,964,830,677,923,685,705,695,868,915,715,702,747,822,694,657,668,856,835,668,823,637],
[745,691,879,875,708,774,702,820,674,629,702,894,798,659,854,669,713,784,924,989,738,819,777,966,760,702,825,892,852,747,970,707],
[833,855,1003,1088,783,823,854,991,803,770,856,1017,972,823,1024,795,627,755,860,913,805,697,833,892,677,752,708,924,813,686,858,761],
[862,723,869,922,739,738,707,856,623,648,647,923,760,676,949,680,835,816,935,950,814,740,869,924,695,757,769,940,884,737,1022,693],
[604,555,766,863,717,716,605,790,641,613,593,730,756,588,794,588,847,858,1006,1091,823,794,910,1024,815,797,843,1041,993,809,1053,749],
[879,877,1018,1073,769,808,867,995,793,768,860,978,1007,773,1023,772,575,546,721,736,611,588,613,736,501,569,543,763,641,562,712,549],
[891,865,964,1132,850,826,881,1032,780,758,811,1028,994,815,1082,776,629,637,782,810,725,683,770,794,655,618,602,875,665,674,810,745],
[777,816,897,844,704,769,682,866,632,681,707,849,800,657,918,692,771,700,836,881,739,703,740,820,612,662,664,843,744,631,901,668],
[697,729,790,853,701,649,793,766,650,626,644,836,703,657,852,647,849,861,994,986,848,874,823,1009,718,804,890,1006,881,773,1053,766],
[670,725,810,761,631,660,698,787,682,622,681,815,703,757,841,613,861,874,1004,1151,850,865,896,1065,851,768,836,1040,1012,808,1034,830],
[570,566,687,742,610,599,620,691,526,537,604,706,639,581,698,554,685,725,801,842,648,645,765,845,629,663,641,863,751,641,887,645],
[723,672,923,947,738,805,835,858,732,640,675,926,716,741,916,731,759,773,837,908,774,683,863,904,654,699,637,961,778,726,887,690],
[816,820,985,943,765,686,851,1001,752,766,720,968,787,854,993,681,678,641,832,815,612,646,659,841,618,585,617,795,764,684,776,619],
[715,730,816,827,703,735,752,771,629,615,707,788,680,646,924,715,692,684,725,731,680,607,675,686,517,568,544,830,707,547,820,586],
[780,656,673,645,747,852,754,909,813,718,840,811,695,998,828,639,743,625,632,704,699,748,722,896,834,642,789,787,760,954,838,710],
[684,667,625,688,710,782,752,807,724,672,836,722,758,875,733,584,699,517,588,556,592,658,597,834,713,638,682,682,649,848,727,587],
[798,685,741,643,778,817,803,861,791,764,875,723,698,998,731,605,706,624,642,580,659,840,657,864,777,726,839,758,667,879,768,662],
[673,601,652,647,619,808,734,858,807,760,775,692,725,878,713,691,756,718,723,697,796,830,860,905,812,731,871,881,774,1009,861,672],
[929,771,763,816,847,933,906,1047,874,833,968,966,899,1085,881,739,815,666,688,628,779,779,741,898,805,733,857,821,702,952,729,611],
[749,608,670,646,715,770,764,873,790,776,786,752,778,922,771,660,853,678,735,623,762,829,866,986,856,814,872,859,810,1050,896,697],
[619,496,573,493,681,718,661,737,676,652,683,699,639,796,664,564,920,746,795,746,820,914,892,1081,902,885,867,881,958,1097,973,762],
[857,750,814,720,798,860,891,1089,888,872,949,950,850,1002,955,823,615,527,523,529,609,643,566,755,631,606,586,638,620,823,598,553],
[856,770,846,732,823,978,877,1015,888,846,986,910,821,1088,901,791,680,597,574,629,649,789,659,696,768,628,786,691,733,814,641,551],
[746,624,640,651,691,763,730,863,783,728,772,764,660,957,746,629,773,570,596,587,721,767,695,922,742,720,784,819,654,877,749,663],
[740,619,589,607,646,764,675,913,767,651,788,755,672,867,796,616,889,657,770,712,838,887,841,1039,929,813,884,853,758,1132,878,701],
[723,650,644,577,663,815,639,766,706,574,758,687,687,902,678,616,906,754,777,785,829,971,829,1062,942,847,983,988,856,1089,935,758],
[587,515,505,549,555,617,604,728,578,544,584,633,642,803,674,611,713,634,587,646,683,809,683,837,739,714,805,700,660,871,701,579],
[769,670,593,663,755,884,716,892,790,773,801,799,777,934,722,616,763,732,711,648,689,801,734,910,810,734,850,750,712,959,823,708],
[879,699,719,709,835,869,768,956,855,777,889,846,790,1081,819,688,667,634,630,658,635,673,688,814,736,679,758,761,693,854,723,647],
[722,583,557,597,691,766,684,813,706,659,799,800,674,851,785,590,683,575,605,519,566,732,614,799,705,687,702,601,633,801,666,548]]
9364fb1f2445cdc7125f9ea01e30a902  output/golden.bin
ed0beff0e69ef4adb9d84c299d3235ed  output/output.bin
Found mismatched elements:
data index: 000000, expected: 751, actual: 712, diff: -39
data index: 000001, expected: 693, actual: 796, diff: 103
data index: 000002, expected: 644, actual: 852, diff: 208
data index: 000003, expected: 657, actual: 978, diff: 321
data index: 000004, expected: 748, actual: 799, diff: 51
data index: 000005, expected: 760, actual: 705, diff: -55
data index: 000006, expected: 843, actual: 866, diff: 23
data index: 000007, expected: 731, actual: 930, diff: 199
data index: 000008, expected: 779, actual: 709, diff: -70
data index: 000009, expected: 795, actual: 703, diff: -92
data index: 000010, expected: 681, actual: 684, diff: 3
data index: 000011, expected: 806, actual: 901, diff: 95
data index: 000012, expected: 733, actual: 844, diff: 111
data index: 000013, expected: 783, actual: 734, diff: -49
data index: 000014, expected: 962, actual: 880, diff: -82
data index: 000015, expected: 866, actual: 704, diff: -162
data index: 000016, expected: 725, actual: 780, diff: 55
data index: 000017, expected: 752, actual: 656, diff: -96
data index: 000018, expected: 761, actual: 673, diff: -88
data index: 000019, expected: 714, actual: 645, diff: -69
data index: 000020, expected: 919, actual: 747, diff: -172
data index: 000021, expected: 629, actual: 852, diff: 223
data index: 000022, expected: 901, actual: 754, diff: -147
data index: 000023, expected: 846, actual: 909, diff: 63
data index: 000024, expected: 780, actual: 813, diff: 33
data index: 000025, expected: 778, actual: 718, diff: -60
data index: 000026, expected: 811, actual: 840, diff: 29
data index: 000027, expected: 857, actual: 811, diff: -46
data index: 000028, expected: 839, actual: 695, diff: -144
data index: 000029, expected: 826, actual: 998, diff: 172
data index: 000030, expected: 578, actual: 828, diff: 250
data index: 000031, expected: 679, actual: 639, diff: -40
data index: 000032, expected: 792, actual: 737, diff: -55
data index: 000033, expected: 727, actual: 737, diff: 10
data index: 000034, expected: 634, actual: 854, diff: 220
data index: 000035, expected: 724, actual: 849, diff: 125
data index: 000036, expected: 738, actual: 707, diff: -31
data index: 000037, expected: 817, actual: 686, diff: -131
data index: 000038, expected: 866, actual: 776, diff: -90
data index: 000039, expected: 683, actual: 917, diff: 234
data index: 000040, expected: 741, actual: 696, diff: -45
data index: 000042, expected: 783, actual: 736, diff: -47
data index: 000043, expected: 749, actual: 899, diff: 150
data index: 000044, expected: 612, actual: 773, diff: 161
data index: 000045, expected: 792, actual: 745, diff: -47
data index: 000046, expected: 927, actual: 869, diff: -58
data index: 000047, expected: 866, actual: 697, diff: -169
data index: 000048, expected: 759, actual: 743, diff: -16
data index: 000049, expected: 758, actual: 625, diff: -133
data index: 000050, expected: 708, actual: 632, diff: -76
data index: 000051, expected: 734, actual: 704, diff: -30
data index: 000052, expected: 894, actual: 699, diff: -195
data index: 000053, expected: 652, actual: 748, diff: 96
data index: 000054, expected: 839, actual: 722, diff: -117
data index: 000055, expected: 849, actual: 896, diff: 47
data index: 000056, expected: 736, actual: 834, diff: 98
data index: 000057, expected: 713, actual: 642, diff: -71
data index: 000058, expected: 821, actual: 789, diff: -32
data index: 000059, expected: 953, actual: 787, diff: -166
data index: 000060, expected: 928, actual: 760, diff: -168
data index: 000061, expected: 885, actual: 954, diff: 69
data index: 000062, expected: 688, actual: 838, diff: 150
data index: 000063, expected: 666, actual: 710, diff: 44
data index: 000064, expected: 747, actual: 639, diff: -108
data index: 000065, expected: 693, actual: 656, diff: -37
data index: 000066, expected: 706, actual: 815, diff: 109
data index: 000067, expected: 642, actual: 946, diff: 304
data index: 000068, expected: 746, actual: 674, diff: -72
data index: 000069, expected: 740, actual: 659, diff: -81
data index: 000070, expected: 777, actual: 733, diff: -44
data index: 000071, expected: 621, actual: 888, diff: 267
data index: 000072, expected: 679, actual: 700, diff: 21
data index: 000073, expected: 720, actual: 671, diff: -49
data index: 000074, expected: 715, actual: 694, diff: -21
data index: 000075, expected: 666, actual: 860, diff: 194
data index: 000076, expected: 666, actual: 730, diff: 64
data index: 000077, expected: 780, actual: 682, diff: -98
data index: 000078, expected: 1003, actual: 893, diff: -110
data index: 000079, expected: 847, actual: 672, diff: -175
data index: 000080, expected: 674, actual: 684, diff: 10
data index: 000081, expected: 683, actual: 667, diff: -16
data index: 000082, expected: 654, actual: 625, diff: -29
data index: 000083, expected: 698, actual: 688, diff: -10
data index: 000084, expected: 792, actual: 710, diff: -82
data index: 000085, expected: 589, actual: 782, diff: 193
data index: 000086, expected: 746, actual: 752, diff: 6
data index: 000087, expected: 862, actual: 807, diff: -55
data index: 000088, expected: 736, actual: 724, diff: -12
data index: 000089, expected: 668, actual: 672, diff: 4
data index: 000090, expected: 783, actual: 836, diff: 53
data index: 000091, expected: 778, actual: 722, diff: -56
data index: 000092, expected: 801, actual: 758, diff: -43
data index: 000093, expected: 771, actual: 875, diff: 104
data index: 000094, expected: 527, actual: 733, diff: 206
data index: 000095, expected: 628, actual: 584, diff: -44
data index: 000096, expected: 673, actual: 694, diff: 21
data index: 000097, expected: 651, actual: 656, diff: 5
data index: 000098, expected: 627, actual: 766, diff: 139
data index: 000099, expected: 661, actual: 805, diff: 144
data index: 000100, expected: 671, actual: 656, diff: -15
... (more errors exist but not shown)
----------------------------------------
Error ratio: 0.9980, Tolerance: 0.0000
[ERROR] Verification failed: Result mismatch.

mmad_custom_cube_only.h改为如下形式（可以主要参见注释和调整部分）：

```
/**
 * @file mmad_custom_cube_only.h
 *
 * Copyright (C) 2023-2024. Huawei Technologies Co., Ltd. All rights reserved.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 */
#ifndef MMAD_CUSTOM_CUBE_ONLY_H
#define MMAD_CUSTOM_CUBE_ONLY_H
#include "kernel_operator.h"

// int8_t type, cube block: [16, 16]
constexpr uint32_t CUBE_BLOCK = 16;
constexpr uint32_t CUBE_BLOCK_SIZE = 16 * 16;

class KernelMmad {
public:
    __aicore__ inline KernelMmad()
    {
        aSize = m * k;
        bSize = k * n;
        cSize = m * n;
    }
    __aicore__ inline void Init(GM_ADDR a, GM_ADDR b, GM_ADDR bias, GM_ADDR c)
    {
        // set cube only
        KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIC_ONLY);

aGM.SetGlobalBuffer((__gm__ int8_t *)a);
        bGM.SetGlobalBuffer((__gm__ int8_t *)b);
        cGM.SetGlobalBuffer((__gm__ int32_t *)c);
        biasGM.SetGlobalBuffer((__gm__ int32_t *)bias);
        pipe.InitBuffer(inQueueA1, 1, aSize * sizeof(int8_t));
        pipe.InitBuffer(inQueueA2, 1, aSize * sizeof(int8_t));
        pipe.InitBuffer(inQueueB1, 1, bSize * sizeof(int8_t));
        pipe.InitBuffer(inQueueB2, 1, bSize * sizeof(int8_t));
        pipe.InitBuffer(outQueueCO1, 1, cSize * sizeof(int32_t));
        pipe.InitBuffer(inQueueC1, 1, n * sizeof(int32_t));
        pipe.InitBuffer(outQueueC2, 1, n * sizeof(int32_t));
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        SplitA();
        SplitB();
        SplitBias();
        Compute();
        CopyOut();
    }

private:
    __aicore__ inline uint32_t CeilCubeBlock(uint32_t len) {
        return (len + CUBE_BLOCK - 1) / CUBE_BLOCK;
    }

// 对于K维度，每个分型大小应当为32
    __aicore__ inline uint32_t CeilCubeBlock_K(uint32_t len) {
        return (len + 32 - 1) / 32;
    }

__aicore__ inline void CopyIn()
    {
        AscendC::LocalTensor<int8_t> a1Local = inQueueA1.AllocTensor<int8_t>();
        AscendC::LocalTensor<int8_t> b1Local = inQueueB1.AllocTensor<int8_t>();
        AscendC::LocalTensor<int32_t> bias1Local = inQueueC1.AllocTensor<int32_t>();

AscendC::Nd2NzParams nd2nzA1Params;
        nd2nzA1Params.ndNum = 1;
        nd2nzA1Params.nValue = m;
        nd2nzA1Params.dValue = k;
        nd2nzA1Params.srcNdMatrixStride = 0;
        nd2nzA1Params.srcDValue = k;
        // nd2nzA1Params.dstNzC0Stride = CeilCubeBlock(m) * CUBE_BLOCK;
        nd2nzA1Params.dstNzC0Stride = CeilCubeBlock(m) * CUBE_BLOCK;   //32，为啥这个跟ceilcubeblock有关？？？
        nd2nzA1Params.dstNzNStride = 1;
        nd2nzA1Params.dstNzMatrixStride = 0;
        AscendC::DataCopy(a1Local, aGM, nd2nzA1Params);

AscendC::Nd2NzParams nd2nzB1Params;
        nd2nzB1Params.ndNum = 1;
        nd2nzB1Params.nValue = k;
        nd2nzB1Params.dValue = n;
        nd2nzB1Params.srcNdMatrixStride = 0;
        nd2nzB1Params.srcDValue = n;
        nd2nzB1Params.dstNzC0Stride = CeilCubeBlock(k) * CUBE_BLOCK;
        nd2nzB1Params.dstNzNStride = 1;
        nd2nzB1Params.dstNzMatrixStride = 0;
        AscendC::DataCopy(b1Local, bGM, nd2nzB1Params);

AscendC::DataCopy(bias1Local, biasGM, n);
        inQueueA1.EnQue(a1Local);
        inQueueB1.EnQue(b1Local);
        inQueueC1.EnQue(bias1Local);
    }

__aicore__ inline void SplitA()
    {
        AscendC::LocalTensor<int8_t> a1Local = inQueueA1.DeQue<int8_t>();
        AscendC::LocalTensor<int8_t> a2Local = inQueueA2.AllocTensor<int8_t>();

// uint32_t dstOffset = CeilCubeBlock(k) * CUBE_BLOCK_SIZE;
        // uint32_t srcOffset = CUBE_BLOCK_SIZE;
        uint32_t dstOffset = 32*16;
        uint32_t srcOffset = 16*32;
 
        AscendC::LoadData2DParams loadDataParams;
        // loadDataParams.repeatTimes = CeilCubeBlock(k);
        // loadDataParams.srcStride = CeilCubeBlock(m);
        loadDataParams.repeatTimes = 1;
        loadDataParams.srcStride = 2;
        loadDataParams.dstGap = 0;
        loadDataParams.ifTranspose = false;
        // for (int i = 0; i < CeilCubeBlock(m); ++i) {
            for (int i = 0; i < 2; ++i) {
            AscendC::LoadData(a2Local[i * dstOffset], a1Local[i * srcOffset], loadDataParams);
        }

inQueueA2.EnQue<int8_t>(a2Local);
        uint32_t array[] = {static_cast<uint32_t>(32),static_cast<uint32_t>(32)};
        AscendC::ShapeInfo shapeInfo(2, array); 
        AscendC::DumpTensor(a1Local, 1,32*32,shapeInfo);
        inQueueA1.FreeTensor(a1Local);
    }
    __aicore__ inline void SplitB()
    {
        AscendC::LocalTensor<int8_t> b1Local = inQueueB1.DeQue<int8_t>();
        AscendC::LocalTensor<int8_t> b2Local = inQueueB2.AllocTensor<int8_t>();

// uint32_t dstOffset = CeilCubeBlock(n) * CUBE_BLOCK_SIZE;
        // uint32_t srcOffset = CUBE_BLOCK_SIZE;
        uint32_t dstOffset = 32*16;
        uint32_t srcOffset = 16*32;
        // Nz -> Zn
        AscendC::LoadData2DParams loadDataParams;
        // loadDataParams.repeatTimes = CeilCubeBlock(n);
        // loadDataParams.srcStride = CeilCubeBlock(k);
        loadDataParams.repeatTimes = 2;
        loadDataParams.srcStride = 1;
        loadDataParams.dstGap = 0;
        loadDataParams.ifTranspose = true;
        // for (int i = 0; i < CeilCubeBlock(k); ++i) {
            for (int i = 0; i < 1; ++i) {
            AscendC::LoadData(b2Local[i * dstOffset], b1Local[i * srcOffset], loadDataParams);
        }

inQueueB1.FreeTensor(b1Local);
        uint32_t array[] = {static_cast<uint32_t>(32),static_cast<uint32_t>(32)};
        AscendC::ShapeInfo shapeInfo(2, array); 
        AscendC::DumpTensor(b1Local, 2,32*32,shapeInfo);
        inQueueB2.EnQue<int8_t>(b2Local);
    }
    __aicore__ inline void SplitBias()
    {
        AscendC::LocalTensor<int32_t> bias1Local = inQueueC1.DeQue<int32_t>();
        AscendC::LocalTensor<int32_t> bias2Local = outQueueC2.AllocTensor<int32_t>();
        AscendC::DataCopy(bias2Local, bias1Local, { 1, (uint16_t)(n * sizeof(int32_t) / 64), 0, 0 });
        outQueueC2.EnQue<int32_t>(bias2Local);
        AscendC::DumpTensor(bias1Local, 3,32);
        inQueueC1.FreeTensor(bias1Local);
    }
    __aicore__ inline void Compute()
    {
        AscendC::LocalTensor<int8_t> a2Local = inQueueA2.DeQue<int8_t>();
        AscendC::LocalTensor<int8_t> b2Local = inQueueB2.DeQue<int8_t>();
        AscendC::LocalTensor<int32_t> bias2Local = outQueueC2.DeQue<int32_t>();
        AscendC::LocalTensor<int32_t> c1Local = outQueueCO1.AllocTensor<int32_t>();
        AscendC::MmadParams mmadParams;
        mmadParams.m = m;
        mmadParams.n = n;
        mmadParams.k = k;
        mmadParams.cmatrixInitVal = false;
        //打印一下a2/b2的数值，没用，全是0，无法打印
        AscendC::printf("a2Local: \n");
        AscendC::printf("a2[0]: %d, a2[1]: %d, a2[2]: %d, a2[3]: %d\n", a2Local(0), a2Local(1), a2Local(2), a2Local(3));
        AscendC::printf("b2Local: \n");
        AscendC::printf("b2[0]: %d, b2[1]: %d, b2[2]: %d, b2[3]: %d\n", b2Local(0), b2Local(1), b2Local(2), b2Local(3));
        AscendC::printf("bias2Local: \n");
        AscendC::printf("bias2[0]: %d, bias2[1]: %d, bias2[2]: %d, bias2[3]: %d\n", bias2Local(0), bias2Local(1), bias2Local(2), bias2Local(3));

AscendC::Mmad(c1Local, a2Local, b2Local, bias2Local, mmadParams);
        outQueueCO1.EnQue<int32_t>(c1Local);
        uint32_t array[] = {static_cast<uint32_t>(32),static_cast<uint32_t>(32)};
        AscendC::ShapeInfo shapeInfo(2, array); 
        AscendC::DumpTensor(c1Local, 2,32*32,shapeInfo);
        inQueueA2.FreeTensor(a2Local);
        inQueueB2.FreeTensor(b2Local);
        outQueueC2.FreeTensor(bias2Local);
    }
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<int32_t> c1Local = outQueueCO1.DeQue<int32_t>();
        AscendC::FixpipeParamsV220 fixpipeParams;
        fixpipeParams.nSize = n;
        fixpipeParams.mSize = m;
        fixpipeParams.srcStride = m;
        fixpipeParams.dstStride = n;

fixpipeParams.ndNum = 1;
        fixpipeParams.srcNdStride = 0;
        fixpipeParams.dstNdStride = 0;
        AscendC::Fixpipe(cGM, c1Local, fixpipeParams);
        outQueueCO1.FreeTensor(c1Local);
    }

private:
    AscendC::TPipe pipe;
    AscendC::TQue<AscendC::TPosition::A1, 1> inQueueA1;
    AscendC::TQue<AscendC::TPosition::A2, 1> inQueueA2;
    AscendC::TQue<AscendC::TPosition::B1, 1> inQueueB1;
    AscendC::TQue<AscendC::TPosition::B2, 1> inQueueB2;
    AscendC::TQue<AscendC::TPosition::CO1, 1> outQueueCO1;
    AscendC::TQue<AscendC::TPosition::C1, 1> inQueueC1;
    AscendC::TQue<AscendC::TPosition::C2, 1> outQueueC2;

AscendC::GlobalTensor<int8_t> aGM;
    AscendC::GlobalTensor<int8_t> bGM;
    AscendC::GlobalTensor<int32_t> cGM;
    AscendC::GlobalTensor<int32_t> biasGM;
    uint16_t m = 32, k = 32, n = 32;
    uint16_t aSize, bSize, cSize;
};

#endif // MMAD_CUSTOM_CUBE_ONLY_H

```

main.cpp改动如下（仅改动数据类型）：

```
    size_t aFileSize = M * K * sizeof(int8_t); // uint16_t represent half
    size_t bFileSize = K * N * sizeof(int8_t); // uint16_t represent half
    size_t biasFileSize = N * sizeof(int32_t); // uint16_t represent half
    size_t cFileSize = M * N * sizeof(int32_t);
    uint32_t blockDim = 1;
```
gen_data.py改为如下（主要是改动数据类型）：
```
import numpy as np
import os

def gen_golden_data():
    M = 32
    N = 32
    K = 32

x1_gm = np.random.randint(1, 10, [M, K]).astype(np.int8)
    x2_gm = np.random.randint(1, 10, [K, N]).astype(np.int8)
    bias_gm = np.random.randint(1, 10, [N]).astype(np.int32)

golden = np.matmul(x1_gm.astype(np.int32), x2_gm.astype(np.int32)) + bias_gm
    
    os.system("mkdir -p input")
    os.system("mkdir -p output")
    x1_gm.tofile("./input/x1_gm.bin")
    x2_gm.tofile("./input/x2_gm.bin")
    bias_gm.tofile("./input/bias_gm.bin")
    golden.tofile("./output/golden.bin")

# 确认最终结果确实是 int32
    assert golden.dtype == np.int32
    print("\n计算完成，最终结果为 int32 类型。")

# 打印一小部分结果查看
    print("\nGolden result (first 4x4 block):")
    print(golden[:4, :4])

if __name__ == "__main__":
    gen_golden_data()
```

二、软件版本:
-- CANN 版本 :  8.2.RC1.alpha002
--Tensorflow/Pytorch/MindSpore 版本:无
--Python 版本 :无
-- MindStudio版本 :无
--操作系统版本 (e.g., Ubuntu 18.04):Ubuntu 22.04.5 LTS

三、测试步骤：
bash ./run.sh -r npu -v Ascend910B2

Ascend/samples

内容风险标识

评论 (1)

Ascend/samples .gitee-modal { width: 500px !important; }

内容风险标识