hipudding commented 5 months ago

[x] 解决当前运行时的错误。
[x] 解决推理精度问题[LLAMA系FP16精度问题已解决]。
[x] 使用内存池管理临时使用的dev或者host内存（buddy system）。还需要考虑及时释放已经使用完的内存（stream 同步时释放， dst重用时释放，需要优化成更优雅的方式），避免OOM。目前支持单卡，如果内存存在问题，考虑是否先同步运行。
[ ] 自定义算子，使用模板减少重复代码。
[x] 接口中的alloc padding，tensor alloc等函数，重新审视下是否符合NPU的使用场景。
[ ] 组合算子替换成融合算子，提高效率，提需求，包括要支持bs的mm算子。
[x] 对所有的代码完善文档，所有的函数完善注释。
[x] ollama适配，0代码0配置使用npu后端的ollama服务。
[x] 考虑q5_x量化的可行性。
[x] 模型运行过程中的精度对比，建议保留对比工具函数。
[ ] 添加tensor自定义内存布局的测试用例。
[x] cont，dup，cpy算子统一。
[x] ascendc_rope_init_cache函数使用多核。
[x] kernel函数中param拷贝耗时长，需优化(加载模型时统一拷贝需考虑ne,nb在transpose/permute过程中改变的问题)
[ ] 清理代码中的TODO。
[x] rope_init_cache函数param拷贝input_ne踩内存问题定位，后续使用tiling进行参数传递
[x] llama_kv_cache_update_internal()中进行k-shfit时存在问题，模型输出错误
[ ] 算子异步发射
[x] 文档
[x] 多卡切图不正确，执行会卡住

hipudding commented 5 months ago

llama.cpp 算子性能对比

[toc]

环境信息

推理后端	内存/显存
Intel(R) Xeon(R) Gold 6348 CPU @ 2.60GHz 8核 16线程	64G
Ascend	64G

构建命令

mkdir build
cd build
cmake .. -DLLAMA_CANN=ON
cmake --build . --config Release

算子性能对比

GGML_OP_ADD 😄

Backend 1/1 (CPU)
  Backend name: CPU
  ADD(type=f32,ne=[1,1,8,1],nr=[1,1,1,1]):                       8192 runs -     1.62 us/run -        0 kB/run -    0.06 GB/s
  ADD(type=f32,ne=[1,1,1,1],nr=[32,1,1,1]):                      8192 runs -     1.60 us/run -        0 kB/run -    0.22 GB/s
  ADD(type=f32,ne=[1,1,320,320],nr=[1,1,1,1]):                   6991 runs -   449.66 us/run -     1200 kB/run -    2.55 GB/s
  ADD(type=f32,ne=[16,10,1,1],nr=[1,1,1,1]):                     8192 runs -     1.73 us/run -        1 kB/run -    1.04 GB/s
  ADD(type=f32,ne=[16,10,10,1],nr=[1,1,1,1]):                    8192 runs -     2.08 us/run -       18 kB/run -    8.58 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,1,1,1]):                   8192 runs -     6.03 us/run -      187 kB/run -   29.67 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[2,1,1,1]):                   8192 runs -     6.99 us/run -      375 kB/run -   51.16 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,2,1,1]):                   8192 runs -    10.77 us/run -      375 kB/run -   33.19 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,1,2,1]):                   8192 runs -    10.64 us/run -      375 kB/run -   33.63 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,1,1,2]):                   8192 runs -    13.36 us/run -      375 kB/run -   26.77 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,1,2,2]):                   8192 runs -    19.62 us/run -      750 kB/run -   36.46 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,2,2,2]):                   5593 runs -    37.53 us/run -     1500 kB/run -   38.12 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[2,2,2,2]):                   2797 runs -    44.91 us/run -     3000 kB/run -   63.70 GB/s
  ADD(type=f32,ne=[1280,1,1,1],nr=[1,1,1,1]):                    8192 runs -     1.58 us/run -       15 kB/run -    9.04 GB/s
  ADD(type=f32,ne=[1280,1,1,1],nr=[1,16,16,1]):                  2185 runs -    22.45 us/run -     3840 kB/run -  163.10 GB/s
  ADD(type=f32,ne=[1280,16,16,1],nr=[1,1,1,1]):                  2185 runs -    15.38 us/run -     3840 kB/run -  238.15 GB/s
  ADD(type=f32,ne=[1280,1,1,1],nr=[1,256,1,1]):                  2185 runs -    19.36 us/run -     3840 kB/run -  189.19 GB/s
  ADD(type=f32,ne=[1,1,1280,1],nr=[16,16,1,1]):                  2185 runs -   279.71 us/run -     3840 kB/run -   13.09 GB/s
  ADD(type=f32,ne=[16,16,1280,1],nr=[1,1,1,1]):                  2185 runs -    95.61 us/run -     3840 kB/run -   38.30 GB/s
  ADD(type=f32,ne=[1,1,1920,1],nr=[16,16,1,1]):                  1457 runs -   418.49 us/run -     5760 kB/run -   13.13 GB/s
  ADD(type=f32,ne=[1,1,2560,1],nr=[16,16,1,1]):                  1093 runs -   572.52 us/run -     7680 kB/run -   12.79 GB/s
  ADD(type=f32,ne=[1,1,1280,1],nr=[32,32,1,1]):                   547 runs -   918.11 us/run -    15360 kB/run -   15.95 GB/s
  ADD(type=f32,ne=[1,1,1920,1],nr=[32,32,1,1]):                   365 runs -  1404.75 us/run -    23040 kB/run -   15.64 GB/s
  ADD(type=f32,ne=[1,1,640,1],nr=[32,32,1,1]):                   1093 runs -   459.20 us/run -     7680 kB/run -   15.95 GB/s
  ADD(type=f32,ne=[5120,1,1,1],nr=[1,256,1,1]):                   547 runs -   116.24 us/run -    15360 kB/run -  126.02 GB/s
  ADD(type=f32,ne=[640,1,1,1],nr=[1,1,1,1]):                     8192 runs -     1.55 us/run -        7 kB/run -    4.61 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  ADD(type=f32,ne=[1,1,8,1],nr=[1,1,1,1]):                       8192 runs -     7.77 us/run -        0 kB/run -    0.01 GB/s
  ADD(type=f32,ne=[1,1,1,1],nr=[32,1,1,1]):                      8192 runs -     8.75 us/run -        0 kB/run -    0.04 GB/s
  ADD(type=f32,ne=[1,1,320,320],nr=[1,1,1,1]):                   8192 runs -     8.14 us/run -     1200 kB/run -  140.57 GB/s
  ADD(type=f32,ne=[16,10,1,1],nr=[1,1,1,1]):                     8192 runs -     8.16 us/run -        1 kB/run -    0.22 GB/s
  ADD(type=f32,ne=[16,10,10,1],nr=[1,1,1,1]):                    8192 runs -     8.59 us/run -       18 kB/run -    2.08 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,1,1,1]):                   8192 runs -     8.13 us/run -      187 kB/run -   21.98 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[2,1,1,1]):                   8192 runs -     8.15 us/run -      375 kB/run -   43.87 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,2,1,1]):                   8192 runs -     8.92 us/run -      375 kB/run -   40.08 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,1,2,1]):                   8192 runs -     8.05 us/run -      375 kB/run -   44.45 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,1,1,2]):                   8192 runs -     8.03 us/run -      375 kB/run -   44.51 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,1,2,2]):                   8192 runs -     8.08 us/run -      750 kB/run -   88.49 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[1,2,2,2]):                   8192 runs -     8.54 us/run -     1500 kB/run -  167.48 GB/s
  ADD(type=f32,ne=[16,10,10,10],nr=[2,2,2,2]):                   8192 runs -     9.44 us/run -     3000 kB/run -  303.13 GB/s
  ADD(type=f32,ne=[1280,1,1,1],nr=[1,1,1,1]):                    8192 runs -     7.93 us/run -       15 kB/run -    1.80 GB/s
  ADD(type=f32,ne=[1280,1,1,1],nr=[1,16,16,1]):                  8192 runs -     7.90 us/run -     3840 kB/run -  463.72 GB/s
  ADD(type=f32,ne=[1280,16,16,1],nr=[1,1,1,1]):                  8192 runs -     7.91 us/run -     3840 kB/run -  463.18 GB/s
  ADD(type=f32,ne=[1280,1,1,1],nr=[1,256,1,1]):                  8192 runs -     8.49 us/run -     3840 kB/run -  431.20 GB/s
  ADD(type=f32,ne=[1,1,1280,1],nr=[16,16,1,1]):                  8192 runs -     7.85 us/run -     3840 kB/run -  466.67 GB/s
  ADD(type=f32,ne=[16,16,1280,1],nr=[1,1,1,1]):                  8192 runs -     7.90 us/run -     3840 kB/run -  463.79 GB/s
  ADD(type=f32,ne=[1,1,1920,1],nr=[16,16,1,1]):                  5826 runs -     7.89 us/run -     5760 kB/run -  696.16 GB/s
  ADD(type=f32,ne=[1,1,2560,1],nr=[16,16,1,1]):                  4370 runs -     7.88 us/run -     7680 kB/run -  929.89 GB/s
  ADD(type=f32,ne=[1,1,1280,1],nr=[32,32,1,1]):                  2185 runs -     8.32 us/run -    15360 kB/run - 1759.78 GB/s
  ADD(type=f32,ne=[1,1,1920,1],nr=[32,32,1,1]):                  1457 runs -     9.81 us/run -    23040 kB/run - 2240.63 GB/s
  ADD(type=f32,ne=[1,1,640,1],nr=[32,32,1,1]):                   4370 runs -     7.85 us/run -     7680 kB/run -  932.55 GB/s
  ADD(type=f32,ne=[5120,1,1,1],nr=[1,256,1,1]):                  2185 runs -     9.26 us/run -    15360 kB/run - 1581.05 GB/s
  ADD(type=f32,ne=[640,1,1,1],nr=[1,1,1,1]):                     8192 runs -     7.88 us/run -        7 kB/run -    0.91 GB/s
  Backend CANN0: OK

GGML_OP_ACC 😄

Backend 1/1 (CPU)
  Backend name: CPU
  ACC(type=f32,ne_a=[1024,577,1,1],ne_b=[1024,576,1,1]):      1213 runs -   244.46 us/run -     6920 kB/run -   27.00 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  ACC(type=f32,ne_a=[1024,577,1,1],ne_b=[1024,576,1,1]):      4849 runs -    12.37 us/run -     6920 kB/run -  533.53 GB/s
  Backend CANN0: OK

GGML_OP_MUL 😄

Backend 1/1 (CPU)
  Backend name: CPU
  MUL(type=f32,ne=[1,1,8,1],nr=[1,1,1,1]):                       8192 runs -     2.26 us/run -        0 kB/run -    0.04 GB/s
  MUL(type=f32,ne=[1,1,1,1],nr=[32,1,1,1]):                      8192 runs -     1.83 us/run -        0 kB/run -    0.20 GB/s
  MUL(type=f32,ne=[1,1,320,320],nr=[1,1,1,1]):                   6991 runs -   517.90 us/run -     1200 kB/run -    2.21 GB/s
  MUL(type=f32,ne=[16,10,1,1],nr=[1,1,1,1]):                     8192 runs -     1.98 us/run -        1 kB/run -    0.90 GB/s
  MUL(type=f32,ne=[16,10,10,1],nr=[1,1,1,1]):                    8192 runs -     2.28 us/run -       18 kB/run -    7.84 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,1,1,1]):                   8192 runs -    18.86 us/run -      187 kB/run -    9.48 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[2,1,1,1]):                   8192 runs -     8.04 us/run -      375 kB/run -   44.49 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,2,1,1]):                   8192 runs -    12.15 us/run -      375 kB/run -   29.44 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,1,2,1]):                   8192 runs -    14.02 us/run -      375 kB/run -   25.51 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,1,1,2]):                   8192 runs -    12.11 us/run -      375 kB/run -   29.54 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,1,2,2]):                   8192 runs -    25.10 us/run -      750 kB/run -   28.50 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,2,2,2]):                   5593 runs -    98.92 us/run -     1500 kB/run -   14.46 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[2,2,2,2]):                   2797 runs -    76.57 us/run -     3000 kB/run -   37.37 GB/s
  MUL(type=f32,ne=[1280,1,1,1],nr=[1,1,1,1]):                    8192 runs -     1.79 us/run -       15 kB/run -    7.98 GB/s
  MUL(type=f32,ne=[1280,1,1,1],nr=[1,16,16,1]):                  2185 runs -    16.44 us/run -     3840 kB/run -  222.70 GB/s
  MUL(type=f32,ne=[1280,16,16,1],nr=[1,1,1,1]):                  2185 runs -    37.19 us/run -     3840 kB/run -   98.48 GB/s
  MUL(type=f32,ne=[1280,1,1,1],nr=[1,256,1,1]):                  2185 runs -    25.18 us/run -     3840 kB/run -  145.43 GB/s
  MUL(type=f32,ne=[1,1,1280,1],nr=[16,16,1,1]):                  2185 runs -   346.94 us/run -     3840 kB/run -   10.56 GB/s
  MUL(type=f32,ne=[16,16,1280,1],nr=[1,1,1,1]):                  2185 runs -   141.48 us/run -     3840 kB/run -   25.88 GB/s
  MUL(type=f32,ne=[1,1,1920,1],nr=[16,16,1,1]):                  1457 runs -   547.23 us/run -     5760 kB/run -   10.04 GB/s
  MUL(type=f32,ne=[1,1,2560,1],nr=[16,16,1,1]):                  1093 runs -   678.00 us/run -     7680 kB/run -   10.80 GB/s
  MUL(type=f32,ne=[1,1,1280,1],nr=[32,32,1,1]):                   547 runs -  1182.22 us/run -    15360 kB/run -   12.39 GB/s
  MUL(type=f32,ne=[1,1,1920,1],nr=[32,32,1,1]):                   365 runs -  1821.39 us/run -    23040 kB/run -   12.06 GB/s
  MUL(type=f32,ne=[1,1,640,1],nr=[32,32,1,1]):                   1093 runs -   604.43 us/run -     7680 kB/run -   12.12 GB/s
  MUL(type=f32,ne=[5120,1,1,1],nr=[1,256,1,1]):                   547 runs -    96.47 us/run -    15360 kB/run -  151.84 GB/s
  MUL(type=f32,ne=[640,1,1,1],nr=[1,1,1,1]):                     8192 runs -     1.87 us/run -        7 kB/run -    3.83 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  MUL(type=f32,ne=[1,1,8,1],nr=[1,1,1,1]):                       8192 runs -     8.21 us/run -        0 kB/run -    0.01 GB/s
  MUL(type=f32,ne=[1,1,1,1],nr=[32,1,1,1]):                      8192 runs -     7.80 us/run -        0 kB/run -    0.05 GB/s
  MUL(type=f32,ne=[1,1,320,320],nr=[1,1,1,1]):                   8192 runs -     7.79 us/run -     1200 kB/run -  146.95 GB/s
  MUL(type=f32,ne=[16,10,1,1],nr=[1,1,1,1]):                     8192 runs -     7.98 us/run -        1 kB/run -    0.22 GB/s
  MUL(type=f32,ne=[16,10,10,1],nr=[1,1,1,1]):                    8192 runs -     8.41 us/run -       18 kB/run -    2.13 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,1,1,1]):                   8192 runs -     7.73 us/run -      187 kB/run -   23.14 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[2,1,1,1]):                   8192 runs -     7.81 us/run -      375 kB/run -   45.77 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,2,1,1]):                   8192 runs -     7.82 us/run -      375 kB/run -   45.73 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,1,2,1]):                   8192 runs -     7.81 us/run -      375 kB/run -   45.81 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,1,1,2]):                   8192 runs -     7.84 us/run -      375 kB/run -   45.63 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,1,2,2]):                   8192 runs -     7.88 us/run -      750 kB/run -   90.79 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[1,2,2,2]):                   8192 runs -     8.53 us/run -     1500 kB/run -  167.70 GB/s
  MUL(type=f32,ne=[16,10,10,10],nr=[2,2,2,2]):                   8192 runs -     9.47 us/run -     3000 kB/run -  302.22 GB/s
  MUL(type=f32,ne=[1280,1,1,1],nr=[1,1,1,1]):                    8192 runs -     7.76 us/run -       15 kB/run -    1.84 GB/s
  MUL(type=f32,ne=[1280,1,1,1],nr=[1,16,16,1]):                  8192 runs -     7.77 us/run -     3840 kB/run -  471.41 GB/s
  MUL(type=f32,ne=[1280,16,16,1],nr=[1,1,1,1]):                  8192 runs -     7.75 us/run -     3840 kB/run -  472.52 GB/s
  MUL(type=f32,ne=[1280,1,1,1],nr=[1,256,1,1]):                  8192 runs -     7.78 us/run -     3840 kB/run -  470.90 GB/s
  MUL(type=f32,ne=[1,1,1280,1],nr=[16,16,1,1]):                  8192 runs -     7.70 us/run -     3840 kB/run -  475.52 GB/s
  MUL(type=f32,ne=[16,16,1280,1],nr=[1,1,1,1]):                  8192 runs -     7.75 us/run -     3840 kB/run -  472.57 GB/s
  MUL(type=f32,ne=[1,1,1920,1],nr=[16,16,1,1]):                  5826 runs -     7.73 us/run -     5760 kB/run -  710.99 GB/s
  MUL(type=f32,ne=[1,1,2560,1],nr=[16,16,1,1]):                  4370 runs -     7.76 us/run -     7680 kB/run -  943.46 GB/s
  MUL(type=f32,ne=[1,1,1280,1],nr=[32,32,1,1]):                  2185 runs -     8.27 us/run -    15360 kB/run - 1771.86 GB/s
  MUL(type=f32,ne=[1,1,1920,1],nr=[32,32,1,1]):                  1457 runs -     9.79 us/run -    23040 kB/run - 2244.40 GB/s
  MUL(type=f32,ne=[1,1,640,1],nr=[32,32,1,1]):                   4370 runs -     7.67 us/run -     7680 kB/run -  954.52 GB/s
  MUL(type=f32,ne=[5120,1,1,1],nr=[1,256,1,1]):                  2185 runs -     9.29 us/run -    15360 kB/run - 1577.62 GB/s
  MUL(type=f32,ne=[640,1,1,1],nr=[1,1,1,1]):                     8192 runs -     7.79 us/run -        7 kB/run -    0.92 GB/s
  Backend CANN0: OK

GGML_OP_DIV 😄

Backend 1/1 (CPU)
  Backend name: CPU
  DIV(type=f32,ne=[1,1,8,1],nr=[1,1,1,1]):                       8192 runs -     1.58 us/run -        0 kB/run -    0.06 GB/s
  DIV(type=f32,ne=[1,1,1,1],nr=[32,1,1,1]):                      8192 runs -     1.56 us/run -        0 kB/run -    0.23 GB/s
  DIV(type=f32,ne=[1,1,320,320],nr=[1,1,1,1]):                   6991 runs -   524.49 us/run -     1200 kB/run -    2.18 GB/s
  DIV(type=f32,ne=[16,10,1,1],nr=[1,1,1,1]):                     8192 runs -     1.95 us/run -        1 kB/run -    0.92 GB/s
  DIV(type=f32,ne=[16,10,10,1],nr=[1,1,1,1]):                    8192 runs -     1.99 us/run -       18 kB/run -    8.97 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,1,1,1]):                   8192 runs -    15.88 us/run -      187 kB/run -   11.26 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[2,1,1,1]):                   8192 runs -     8.08 us/run -      375 kB/run -   44.26 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,2,1,1]):                   8192 runs -    12.05 us/run -      375 kB/run -   29.68 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,1,2,1]):                   8192 runs -    11.83 us/run -      375 kB/run -   30.23 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,1,1,2]):                   8192 runs -    11.96 us/run -      375 kB/run -   29.91 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,1,2,2]):                   8192 runs -    23.32 us/run -      750 kB/run -   30.67 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,2,2,2]):                   5593 runs -   110.49 us/run -     1500 kB/run -   12.95 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[2,2,2,2]):                   2797 runs -    74.99 us/run -     3000 kB/run -   38.15 GB/s
  DIV(type=f32,ne=[1280,1,1,1],nr=[1,1,1,1]):                    8192 runs -     1.58 us/run -       15 kB/run -    9.05 GB/s
  DIV(type=f32,ne=[1280,1,1,1],nr=[1,16,16,1]):                  2185 runs -    17.70 us/run -     3840 kB/run -  206.89 GB/s
  DIV(type=f32,ne=[1280,16,16,1],nr=[1,1,1,1]):                  2185 runs -    22.97 us/run -     3840 kB/run -  159.46 GB/s
  DIV(type=f32,ne=[1280,1,1,1],nr=[1,256,1,1]):                  2185 runs -    17.21 us/run -     3840 kB/run -  212.78 GB/s
  DIV(type=f32,ne=[1,1,1280,1],nr=[16,16,1,1]):                  2185 runs -   358.86 us/run -     3840 kB/run -   10.20 GB/s
  DIV(type=f32,ne=[16,16,1280,1],nr=[1,1,1,1]):                  2185 runs -   140.07 us/run -     3840 kB/run -   26.14 GB/s
  DIV(type=f32,ne=[1,1,1920,1],nr=[16,16,1,1]):                  1457 runs -   556.51 us/run -     5760 kB/run -    9.87 GB/s
  DIV(type=f32,ne=[1,1,2560,1],nr=[16,16,1,1]):                  1093 runs -   769.92 us/run -     7680 kB/run -    9.51 GB/s
  DIV(type=f32,ne=[1,1,1280,1],nr=[32,32,1,1]):                   547 runs -  1277.11 us/run -    15360 kB/run -   11.47 GB/s
  DIV(type=f32,ne=[1,1,1920,1],nr=[32,32,1,1]):                   365 runs -  1849.03 us/run -    23040 kB/run -   11.88 GB/s
  DIV(type=f32,ne=[1,1,640,1],nr=[32,32,1,1]):                   1093 runs -   599.96 us/run -     7680 kB/run -   12.21 GB/s
  DIV(type=f32,ne=[5120,1,1,1],nr=[1,256,1,1]):                   547 runs -    95.95 us/run -    15360 kB/run -  152.67 GB/s
  DIV(type=f32,ne=[640,1,1,1],nr=[1,1,1,1]):                     8192 runs -     1.50 us/run -        7 kB/run -    4.77 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  DIV(type=f32,ne=[1,1,8,1],nr=[1,1,1,1]):                       8192 runs -     8.19 us/run -        0 kB/run -    0.01 GB/s
  DIV(type=f32,ne=[1,1,1,1],nr=[32,1,1,1]):                      8192 runs -     7.77 us/run -        0 kB/run -    0.05 GB/s
  DIV(type=f32,ne=[1,1,320,320],nr=[1,1,1,1]):                   8192 runs -     7.83 us/run -     1200 kB/run -  146.23 GB/s
  DIV(type=f32,ne=[16,10,1,1],nr=[1,1,1,1]):                     8192 runs -     7.70 us/run -        1 kB/run -    0.23 GB/s
  DIV(type=f32,ne=[16,10,10,1],nr=[1,1,1,1]):                    8192 runs -     8.56 us/run -       18 kB/run -    2.09 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,1,1,1]):                   8192 runs -     7.75 us/run -      187 kB/run -   23.06 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[2,1,1,1]):                   8192 runs -     7.74 us/run -      375 kB/run -   46.19 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,2,1,1]):                   8192 runs -     7.76 us/run -      375 kB/run -   46.08 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,1,2,1]):                   8192 runs -     7.71 us/run -      375 kB/run -   46.41 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,1,1,2]):                   8192 runs -     7.70 us/run -      375 kB/run -   46.44 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,1,2,2]):                   8192 runs -     7.89 us/run -      750 kB/run -   90.65 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[1,2,2,2]):                   8192 runs -     8.12 us/run -     1500 kB/run -  176.08 GB/s
  DIV(type=f32,ne=[16,10,10,10],nr=[2,2,2,2]):                   8192 runs -     9.09 us/run -     3000 kB/run -  314.91 GB/s
  DIV(type=f32,ne=[1280,1,1,1],nr=[1,1,1,1]):                    8192 runs -     7.76 us/run -       15 kB/run -    1.84 GB/s
  DIV(type=f32,ne=[1280,1,1,1],nr=[1,16,16,1]):                  8192 runs -     7.81 us/run -     3840 kB/run -  468.76 GB/s
  DIV(type=f32,ne=[1280,16,16,1],nr=[1,1,1,1]):                  8192 runs -     7.77 us/run -     3840 kB/run -  471.38 GB/s
  DIV(type=f32,ne=[1280,1,1,1],nr=[1,256,1,1]):                  8192 runs -     7.65 us/run -     3840 kB/run -  478.88 GB/s
  DIV(type=f32,ne=[1,1,1280,1],nr=[16,16,1,1]):                  8192 runs -     7.64 us/run -     3840 kB/run -  479.47 GB/s
  DIV(type=f32,ne=[16,16,1280,1],nr=[1,1,1,1]):                  8192 runs -     7.79 us/run -     3840 kB/run -  470.21 GB/s
  DIV(type=f32,ne=[1,1,1920,1],nr=[16,16,1,1]):                  5826 runs -     7.70 us/run -     5760 kB/run -  713.00 GB/s
  DIV(type=f32,ne=[1,1,2560,1],nr=[16,16,1,1]):                  4370 runs -     7.84 us/run -     7680 kB/run -  934.53 GB/s
  DIV(type=f32,ne=[1,1,1280,1],nr=[32,32,1,1]):                  2185 runs -     8.70 us/run -    15360 kB/run - 1683.77 GB/s
  DIV(type=f32,ne=[1,1,1920,1],nr=[32,32,1,1]):                  1457 runs -    10.50 us/run -    23040 kB/run - 2092.84 GB/s
  DIV(type=f32,ne=[1,1,640,1],nr=[32,32,1,1]):                   4370 runs -     7.77 us/run -     7680 kB/run -  943.01 GB/s
  DIV(type=f32,ne=[5120,1,1,1],nr=[1,256,1,1]):                  2185 runs -     9.16 us/run -    15360 kB/run - 1598.50 GB/s
  DIV(type=f32,ne=[640,1,1,1],nr=[1,1,1,1]):                     8192 runs -     7.82 us/run -        7 kB/run -    0.91 GB/s
  Backend CANN0: OK

GGML_UNARY_OP_GELU 😄

Backend 1/1 (CPU)
  Backend name: CPU
  GELU(type=f32,ne=[128,10,10,10]):                              8192 runs -   166.72 us/run -     1000 kB/run -    5.72 GB/s
  GELU(type=f32,ne=[7,13,19,23]):                                8192 runs -    48.31 us/run -      310 kB/run -    6.13 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  GELU(type=f32,ne=[128,10,10,10]):                              8192 runs -    11.09 us/run -     1000 kB/run -   86.01 GB/s
  GELU(type=f32,ne=[7,13,19,23]):                                8192 runs -    11.08 us/run -      310 kB/run -   26.74 GB/s
  Backend CANN0: OK

GGML_UNARY_OP_SILU 😄

Backend 1/1 (CPU)
  Backend name: CPU
  SILU(type=f32,ne=[128,10,10,10]):                              8192 runs -    23.57 us/run -     1000 kB/run -   40.46 GB/s
  SILU(type=f32,ne=[7,13,19,23]):                                8192 runs -    13.51 us/run -      310 kB/run -   21.93 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  SILU(type=f32,ne=[128,10,10,10]):                              8192 runs -    11.09 us/run -     1000 kB/run -   85.98 GB/s
  SILU(type=f32,ne=[7,13,19,23]):                                8192 runs -    11.08 us/run -      310 kB/run -   26.73 GB/s
  Backend CANN0: OK

GGML_UNARY_OP_GELU_QUICK 😄

实际上调用的是GELU算子，未实GELU_QUICK。

Backend 1/1 (CPU)
  Backend name: CPU
  GELU_QUICK(type=f32,ne=[128,10,10,10]):                        8192 runs -    31.43 us/run -     1000 kB/run -   30.35 GB/s
  GELU_QUICK(type=f32,ne=[7,13,19,23]):                          8192 runs -    13.40 us/run -      310 kB/run -   22.10 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  GELU_QUICK(type=f32,ne=[128,10,10,10]):                        8192 runs -    11.26 us/run -     1000 kB/run -   84.71 GB/s
  GELU_QUICK(type=f32,ne=[7,13,19,23]):                          8192 runs -    11.07 us/run -      310 kB/run -   26.77 GB/s
  Backend CANN0: OK

GGML_UNARY_OP_TANH 😄

Backend 1/1 (CPU)
  Backend name: CPU
  TANH(type=f32,ne=[128,10,10,10]):                              8192 runs -  1118.90 us/run -     1000 kB/run -    0.85 GB/s
  TANH(type=f32,ne=[7,13,19,23]):                                8192 runs -   391.71 us/run -      310 kB/run -    0.76 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  TANH(type=f32,ne=[128,10,10,10]):                              8192 runs -     7.10 us/run -     1000 kB/run -  134.32 GB/s
  TANH(type=f32,ne=[7,13,19,23]):                                8192 runs -     7.16 us/run -      310 kB/run -   41.39 GB/s
  Backend CANN0: OK

GGML_UNARY_OP_RELU 😄

Backend 1/1 (CPU)
  Backend name: CPU
  RELU(type=f32,ne=[128,10,10,10]):                              8192 runs -    14.06 us/run -     1000 kB/run -   67.82 GB/s
  RELU(type=f32,ne=[7,13,19,23]):                                8192 runs -   185.79 us/run -      310 kB/run -    1.59 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  RELU(type=f32,ne=[128,10,10,10]):                              8192 runs -     7.01 us/run -     1000 kB/run -  136.13 GB/s
  RELU(type=f32,ne=[7,13,19,23]):                                8192 runs -     7.57 us/run -      310 kB/run -   39.15 GB/s
  Backend CANN0: OK

GGML_UNARY_OP_HARDSIGMOID 😄

Backend 1/1 (CPU)
  Backend name: CPU
  HARDSIGMOID(type=f32,ne=[128,10,10,10]):                       8192 runs -   461.39 us/run -     1000 kB/run -    2.07 GB/s
  HARDSIGMOID(type=f32,ne=[7,13,19,23]):                         8192 runs -   145.02 us/run -      310 kB/run -    2.04 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  HARDSIGMOID(type=f32,ne=[128,10,10,10]):                       8192 runs -     7.44 us/run -     1000 kB/run -  128.11 GB/s
  HARDSIGMOID(type=f32,ne=[7,13,19,23]):                         8192 runs -     7.38 us/run -      310 kB/run -   40.16 GB/s
  Backend CANN0: OK

GGML_UNARY_OP_HARDSWISH 😄

Backend 1/1 (CPU)
  Backend name: CPU
  HARDSWISH(type=f32,ne=[128,10,10,10]):                         8192 runs -   463.49 us/run -     1000 kB/run -    2.06 GB/s
  HARDSWISH(type=f32,ne=[7,13,19,23]):                           8192 runs -   146.70 us/run -      310 kB/run -    2.02 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  HARDSWISH(type=f32,ne=[128,10,10,10]):                         8192 runs -     7.09 us/run -     1000 kB/run -  134.48 GB/s
  HARDSWISH(type=f32,ne=[7,13,19,23]):                           8192 runs -     7.12 us/run -      310 kB/run -   41.63 GB/s
  Backend CANN0: OK

GGML_OP_NORM 😄

Backend 1/1 (CPU)
  Backend name: CPU
  NORM(type=f32,ne=[64,10,10,10],eps=0.000001):                  8192 runs -    53.92 us/run -      500 kB/run -    8.84 GB/s
  NORM(type=f32,ne=[64,10,10,10],eps=0.000010):                  8192 runs -    53.88 us/run -      500 kB/run -    8.85 GB/s
  NORM(type=f32,ne=[64,10,10,10],eps=0.001000):                  8192 runs -    53.85 us/run -      500 kB/run -    8.86 GB/s
  NORM(type=f32,ne=[64,10,10,10],eps=0.100000):                  8192 runs -    54.25 us/run -      500 kB/run -    8.79 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  NORM(type=f32,ne=[64,10,10,10],eps=0.000001):                  8192 runs -    20.17 us/run -      500 kB/run -   23.64 GB/s
  NORM(type=f32,ne=[64,10,10,10],eps=0.000010):                  8192 runs -    19.72 us/run -      500 kB/run -   24.17 GB/s
  NORM(type=f32,ne=[64,10,10,10],eps=0.001000):                  8192 runs -    19.74 us/run -      500 kB/run -   24.16 GB/s
  NORM(type=f32,ne=[64,10,10,10],eps=0.100000):                  8192 runs -    19.68 us/run -      500 kB/run -   24.23 GB/s
  Backend CANN0: OK

GGML_OP_GROUP_NORM 😄

Backend 1/1 (CPU)
  Backend name: CPU
  GROUP_NORM(type=f32,ne=[64,64,320,1],num_groups=32):            820 runs -   621.33 us/run -    10240 kB/run -   15.72 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  GROUP_NORM(type=f32,ne=[64,64,320,1],num_groups=32):           3277 runs -    24.46 us/run -    10240 kB/run -  399.22 GB/s
  Backend CANN0: OK

GGML_OP_UPSCALE ❓ （算子执行出错）

Backend 1/1 (CPU)
  Backend name: CPU
  UPSCALE(type=f32,ne=[512,512,3,1],scale_factor=2):              547 runs -  3120.34 us/run -    15360 kB/run -    4.69 GB/s
  Backend CPU: OK

perf test 在NPU上执行异常
The error from device(chipId:0, dieId:0), serial number is 472, there is an aivec error exception, core id is 20, error code = 0x800000, dump info: pc start: 0x1240c005e5d8, current: 0x1240c005f43c, vec error info: 0xea1159162f, mte error info: 0x7d060000f8, ifu error info: 0x2d93518c01f80, ccu error info: 0x8ce05b7e7b4f1703, cube error info: 0, biu error info: 0, aic error mask: 0x6500020bd000288, para base: 0x1241000ed000.[FUNC:ProcessStarsCoreErrorInfo][FILE:device_error_proc.cc][LINE:1169]

GGML_OP_PAD 😄

Backend 1/1 (CPU)
  Backend name: CPU
  PAD(type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1):              4089 runs -   101.08 us/run -     2052 kB/run -   19.36 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  PAD(type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1):              8192 runs -    23.06 us/run -     2052 kB/run -   84.86 GB/s
  Backend CANN0: OK

GGML_OP_LEAKY_RELU 😄

Backend 1/1 (CPU)
  Backend name: CPU
  LEAKY_RELU(type=f32,ne_a=[10,10,10,10],negative_slope=0.100000):  8192 runs -     8.35 us/run -       78 kB/run -    8.92 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  LEAKY_RELU(type=f32,ne_a=[10,10,10,10],negative_slope=0.100000):  8192 runs -     7.60 us/run -       78 kB/run -    9.80 GB/s
  Backend CANN0: OK

GGML_OP_SCALE 😄

Backend 1/1 (CPU)
  Backend name: CPU
  SCALE(type=f32,ne=[10,10,10,10],scale=2.000000):               8192 runs -     9.93 us/run -       78 kB/run -    7.50 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  SCALE(type=f32,ne=[10,10,10,10],scale=2.000000):               8192 runs -     7.32 us/run -       78 kB/run -   10.18 GB/s
  Backend CANN0: OK

GGML_OP_CLAMP 😄

Backend 1/1 (CPU)
  Backend name: CPU
  CLAMP(type=f32,ne=[10,10,10,10],min=-0.500000,max=0.500000):     8192 runs -    10.59 us/run -       78 kB/run -    7.03 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  CLAMP(type=f32,ne=[10,10,10,10],min=-0.500000,max=0.500000):     8192 runs -     7.38 us/run -       78 kB/run -   10.09 GB/s
  Backend CANN0: OK

GGML_OP_ARGSORT 😄

Backend 1/1 (CPU)
  Backend name: CPU
  ARGSORT(type=f32,ne=[8,1,1,1],order=0):                        8192 runs -     1.73 us/run -        0 kB/run -    0.03 GB/s
  ARGSORT(type=f32,ne=[16,10,10,10],order=0):                    8192 runs -   330.08 us/run -      125 kB/run -    0.36 GB/s
  ARGSORT(type=f32,ne=[60,10,10,10],order=0):                    8192 runs -  1717.51 us/run -      468 kB/run -    0.26 GB/s
  ARGSORT(type=f32,ne=[8,1,1,1],order=1):                        8192 runs -     1.21 us/run -        0 kB/run -    0.05 GB/s
  ARGSORT(type=f32,ne=[16,10,10,10],order=1):                    8192 runs -   334.36 us/run -      125 kB/run -    0.36 GB/s
  ARGSORT(type=f32,ne=[60,10,10,10],order=1):                    8192 runs -  1671.17 us/run -      468 kB/run -    0.27 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  ARGSORT(type=f32,ne=[8,1,1,1],order=0):                        8192 runs -    31.20 us/run -        0 kB/run -    0.00 GB/s
  ARGSORT(type=f32,ne=[16,10,10,10],order=0):                    8192 runs -    37.65 us/run -      125 kB/run -    3.17 GB/s
  ARGSORT(type=f32,ne=[60,10,10,10],order=0):                    8192 runs -    32.66 us/run -      468 kB/run -   13.69 GB/s
  ARGSORT(type=f32,ne=[8,1,1,1],order=1):                        8192 runs -    30.82 us/run -        0 kB/run -    0.00 GB/s
  ARGSORT(type=f32,ne=[16,10,10,10],order=1):                    8192 runs -    37.14 us/run -      125 kB/run -    3.21 GB/s
  ARGSORT(type=f32,ne=[60,10,10,10],order=1):                    8192 runs -    33.49 us/run -      468 kB/run -   13.35 GB/s
  Backend CANN0: OK

GGML_OP_CONCAT 🌚

执行慢的原因：算子本身执行慢。

Backend 1/1 (CPU)
  Backend name: CPU
  CONCAT(type=f32,ne=[10,10,10,10],b_ne2=10):                    8192 runs -     4.32 us/run -      156 kB/run -   34.48 GB/s
  CONCAT(type=i32,ne=[10,10,10,10],b_ne2=10):                    8192 runs -     4.39 us/run -      156 kB/run -   33.91 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  CONCAT(type=f32,ne=[10,10,10,10],b_ne2=10):                    8192 runs -     8.53 us/run -      156 kB/run -   17.47 GB/s
  CONCAT(type=i32,ne=[10,10,10,10],b_ne2=10):                    8192 runs -     8.22 us/run -      156 kB/run -   18.12 GB/s
  Backend CANN0: OK

GGML_OP_REPEAT 🌚

执行慢的原因：算子本身执行慢。

Backend 1/1 (CPU)
  Backend name: CPU
  REPEAT(type=f32,ne=[10,10,10,10],nr=[1,1,1,1]):                8192 runs -     5.38 us/run -       78 kB/run -   13.85 GB/s
  REPEAT(type=f32,ne=[10,10,10,10],nr=[2,1,1,1]):                8192 runs -     8.08 us/run -      156 kB/run -   18.45 GB/s
  REPEAT(type=f32,ne=[10,10,10,10],nr=[1,2,1,1]):                8192 runs -    10.64 us/run -      156 kB/run -   14.00 GB/s
  REPEAT(type=f32,ne=[10,10,10,10],nr=[1,1,2,1]):                8192 runs -    10.85 us/run -      156 kB/run -   13.73 GB/s
  REPEAT(type=f32,ne=[10,10,10,10],nr=[1,1,1,2]):                8192 runs -    10.83 us/run -      156 kB/run -   13.76 GB/s
  REPEAT(type=i32,ne=[10,10,10,10],nr=[2,1,1,1]):                8192 runs -     8.12 us/run -      156 kB/run -   18.35 GB/s
  REPEAT(type=i16,ne=[10,10,10,10],nr=[1,1,1,2]):                8192 runs -    21.40 us/run -       78 kB/run -    3.48 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  REPEAT(type=f32,ne=[10,10,10,10],nr=[1,1,1,1]):                8192 runs -    62.14 us/run -       78 kB/run -    1.20 GB/s
  REPEAT(type=f32,ne=[10,10,10,10],nr=[2,1,1,1]):                8192 runs -   102.18 us/run -      156 kB/run -    1.46 GB/s
  REPEAT(type=f32,ne=[10,10,10,10],nr=[1,2,1,1]):                8192 runs -    98.47 us/run -      156 kB/run -    1.51 GB/s
  REPEAT(type=f32,ne=[10,10,10,10],nr=[1,1,2,1]):                8192 runs -   101.23 us/run -      156 kB/run -    1.47 GB/s
  REPEAT(type=f32,ne=[10,10,10,10],nr=[1,1,1,2]):                8192 runs -   101.48 us/run -      156 kB/run -    1.47 GB/s
  REPEAT(type=i32,ne=[10,10,10,10],nr=[2,1,1,1]):                8192 runs -    94.11 us/run -      156 kB/run -    1.58 GB/s
  REPEAT(type=i16,ne=[10,10,10,10],nr=[1,1,1,2]):                8192 runs -    86.27 us/run -       78 kB/run -    0.86 GB/s
  Backend CANN0: OK

GGML_OP_ARANGE 🌚

执行慢的原因：算子本身执行慢。

Backend 1/1 (CPU)
  Backend name: CPU
  ARANGE(type=f32,start=0.000000,stop=10.000000,step=1.000000):  8192 runs -     1.65 us/run -        0 kB/run -    0.02 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  ARANGE(type=f32,start=0.000000,stop=10.000000,step=1.000000):  8192 runs -     7.05 us/run -        0 kB/run -    0.01 GB/s
  Backend CANN0: OK

GGML_OP_TIMESTEP_EMBEDDING 🌚 （需实现融合算子）

执行慢的原因：调用了多个算子组合，需要融合算子

Backend 1/1 (CPU)
  Backend name: CPU
  TIMESTEP_EMBEDDING(type=f32,ne_a=[2,1,1,1],dim=320,max_period=10000):    8192 runs -     9.87 us/run -        2 kB/run -    0.24 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  TIMESTEP_EMBEDDING(type=f32,ne_a=[2,1,1,1],dim=320,max_period=10000):    8192 runs -    81.09 us/run -        2 kB/run -    0.03 GB/s
  Backend CANN0: OK

GGML_OP_RMS_NORM 🌚（需修改算子接口）

执行慢的原因：gamma是必传参数，需要构造全1矩阵，rstd输出是必传参数，需要申请设备上内存。

Backend 1/1 (CPU)
  Backend name: CPU
  RMS_NORM(type=f32,ne=[64,10,10,10],eps=0.000001):              8192 runs -    31.75 us/run -      500 kB/run -   15.02 GB/s
  RMS_NORM(type=f32,ne=[64,10,10,10],eps=0.000010):              8192 runs -    31.93 us/run -      500 kB/run -   14.93 GB/s
  RMS_NORM(type=f32,ne=[64,10,10,10],eps=0.001000):              8192 runs -    32.23 us/run -      500 kB/run -   14.80 GB/s
  RMS_NORM(type=f32,ne=[64,10,10,10],eps=0.100000):              8192 runs -    31.99 us/run -      500 kB/run -   14.90 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  RMS_NORM(type=f32,ne=[64,10,10,10],eps=0.000001):              8192 runs -    78.59 us/run -      500 kB/run -    6.07 GB/s
  RMS_NORM(type=f32,ne=[64,10,10,10],eps=0.000010):              8192 runs -    76.18 us/run -      500 kB/run -    6.26 GB/s
  RMS_NORM(type=f32,ne=[64,10,10,10],eps=0.001000):              8192 runs -    76.55 us/run -      500 kB/run -    6.23 GB/s
  RMS_NORM(type=f32,ne=[64,10,10,10],eps=0.100000):              8192 runs -    75.31 us/run -      500 kB/run -    6.33 GB/s
  Backend CANN0: OK

GGML_OP_SQR 🌚

执行慢的原因：算子本身执行慢。

Backend 1/1 (CPU)
  Backend name: CPU
  SQR(type=f32,ne=[10,10,10,10]):                                8192 runs -     2.94 us/run -       78 kB/run -   25.37 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  SQR(type=f32,ne=[10,10,10,10]):                                8192 runs -     7.91 us/run -      117 kB/run -   14.13 GB/s
  Backend CANN0: OK

GGML_OP_CONT 🌚

执行慢的原因：算子本身执行慢。

Backend 1/1 (CPU)
  Backend name: CPU
  CONT(type=f32,ne=[10,10,10,1]):                                8191 runs -     2.44 us/run -        7 kB/run -    3.05 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  CONT(type=f32,ne=[10,10,10,1]):                                8191 runs -    12.55 us/run -        7 kB/run -    0.59 GB/s
  Backend CANN0: OK

GGML_OP_DIAG_MASK_INF 🌚 （需实现融合算子）

执行慢的原因：调用了多个算子组合，需要融合算子

Backend 1/1 (CPU)
  Backend name: CPU
  DIAG_MASK_INF(type=f32,ne=[10,10,1,1],n_past=5):               8192 runs -     1.59 us/run -        0 kB/run -    0.47 GB/s
  DIAG_MASK_INF(type=f32,ne=[10,10,10,1],n_past=5):              8192 runs -     2.07 us/run -        7 kB/run -    3.60 GB/s
  DIAG_MASK_INF(type=f32,ne=[10,10,10,10],n_past=5):             8192 runs -     5.20 us/run -       78 kB/run -   14.32 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  DIAG_MASK_INF(type=f32,ne=[10,10,1,1],n_past=5):               8192 runs -    59.52 us/run -        0 kB/run -    0.01 GB/s
  DIAG_MASK_INF(type=f32,ne=[10,10,10,1],n_past=5):              8192 runs -    61.12 us/run -        7 kB/run -    0.12 GB/s
  DIAG_MASK_INF(type=f32,ne=[10,10,10,10],n_past=5):             8192 runs -    91.92 us/run -       78 kB/run -    0.81 GB/s
  Backend CANN0: OK

GGML_OP_IM2COL 🌚 （需实现融合算子）

执行慢的原因：调用了多个算子组合，需要融合算子

Backend 1/1 (CPU)
  Backend name: CPU
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1):            8192 runs -     5.10 us/run -       11 kB/run -    2.20 GB/s
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1):            8192 runs -     4.54 us/run -        6 kB/run -    1.36 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1):            8192 runs -    23.89 us/run -       11 kB/run -    0.47 GB/s
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1):            8192 runs -    35.48 us/run -        6 kB/run -    0.17 GB/s
  Backend CANN0: OK

GGML_OP_POOL_2D 🌚 （需实现融合算子）

执行慢的原因：max pool 调用了多个算子组合，需要融合算子，算子本身也慢

Backend 1/1 (CPU)                                                                                                              Backend name: CPU
  POOL_2D(pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=0):              8192 runs -     1.12 us/run -        2 kB/run -    1.99 GB/s
  POOL_2D(pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=1):              8192 runs -     1.31 us/run -        2 kB/run -    1.88 GB/s
  POOL_2D(pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=1,p1=0):              8192 runs -     1.30 us/run -        2 kB/run -    1.88 GB/s
 POOL_2D(pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=0):              8192 runs -     1.33 us/run -        2 kB/run -    1.69 GB/s
  POOL_2D(pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=1):              8192 runs -     1.49 us/run -        2 kB/run -    1.65 GB/s
  POOL_2D(pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=1,p1=0):              8192 runs -     1.49 us/run -        2 kB/run -    1.65 GB/s
... ...

Backend 2/2 (CANN0)                                                                                                            Backend name: CANN0
  POOL_2D(pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=0):              8192 runs -    77.98 us/run -        2 kB/run -    0.03 GB/s
  POOL_2D(pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=1):              8192 runs -    76.85 us/run -        2 kB/run -    0.03 GB/s
  POOL_2D(pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=1,p1=0):              8192 runs -    78.23 us/run -        2 kB/run -    0.03 GB/s
  POOL_2D(pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=0):              8192 runs -    37.53 us/run -        2 kB/run -    0.06 GB/s
  POOL_2D(pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=1):              8192 runs -    41.14 us/run -        2 kB/run -    0.06 GB/s
  POOL_2D(pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=1,p1=0):              8192 runs -    43.39 us/run -        2 kB/run -    0.06 GB/s
... ...

GGML_OP_SUM_ROWS 🌚

执行慢的原因：算子本身执行慢。

Backend 1/1 (CPU)
  Backend name: CPU
  SUM_ROWS(type=f32,ne=[10,10,10,10]):                           8192 runs -     4.85 us/run -       42 kB/run -    8.44 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  SUM_ROWS(type=f32,ne=[10,10,10,10]):                           8192 runs -    11.48 us/run -       42 kB/run -    3.57 GB/s
  Backend CANN0: OK

GGML_OP_DUP 🌚

执行慢的原因：算子本身执行慢。

Backend 1/1 (CPU)
  Backend name: CPU
  DUP(type=f32,ne=[10,10,10,1]):                                 8192 runs -     1.71 us/run -        7 kB/run -    4.35 GB/s
  DUP(type=f16,ne=[10,10,10,1]):                                 8192 runs -     1.63 us/run -        3 kB/run -    2.29 GB/s
  DUP(type=i32,ne=[10,10,10,1]):                                 8192 runs -     1.67 us/run -        7 kB/run -    4.46 GB/s
  DUP(type=i16,ne=[10,10,10,1]):                                 8192 runs -     1.63 us/run -        3 kB/run -    2.29 GB/s
  DUP(type=i16,ne=[10,8,3,1],permute=[0,2,1,3]):                 8191 runs -     1.89 us/run -        0 kB/run -    0.47 GB/s
  DUP(type=i16,ne=[10,8,3,1],permute=[1,2,0,3]):                 8191 runs -     2.09 us/run -        0 kB/run -    0.43 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  DUP(type=f32,ne=[10,10,10,1]):                                 8192 runs -    10.04 us/run -        7 kB/run -    0.74 GB/s
  DUP(type=f16,ne=[10,10,10,1]):                                 8192 runs -    10.02 us/run -        3 kB/run -    0.37 GB/s
  DUP(type=i32,ne=[10,10,10,1]):                                 8192 runs -    10.07 us/run -        7 kB/run -    0.74 GB/s
  DUP(type=i16,ne=[10,10,10,1]):                                 8192 runs -    10.04 us/run -        3 kB/run -    0.37 GB/s
  DUP(type=i16,ne=[10,8,3,1],permute=[0,2,1,3]):                 8191 runs -    26.69 us/run -        0 kB/run -    0.03 GB/s
  DUP(type=i16,ne=[10,8,3,1],permute=[1,2,0,3]):                 8191 runs -    26.70 us/run -        0 kB/run -    0.03 GB/s
  Backend CANN0: OK

GGML_OP_SOFTMAX （需实现融合算子）

GGML_OP_ROPE （需实现融合算子）

GGML_OP_ALIBI 🌚 （需实现融合算子）

执行慢的原因：max pool 调用了多个算子组合，需要融合算子

Backend 1/1 (CPU)
  Backend name: CPU
  ALIBI(type=f32,ne_a=[16,2,10,1],n_past=0,n_head=10,bias_max=-0.500000):  8192 runs -  0.38 us/run -   2 kB/run -    6.29 GB/s
  ALIBI(type=f32,ne_a=[16,2,32,1],n_past=0,n_head=32,bias_max=-0.500000):  8192 runs -  1.08 us/run -   8 kB/run -    7.08 GB/s
  ALIBI(type=f32,ne_a=[128,4,10,1],n_past=0,n_head=10,bias_max=-0.500000): 8192 runs -  4.29 us/run -   40 kB/run -   8.90 GB/s
  ALIBI(type=f32,ne_a=[128,4,32,1],n_past=0,n_head=32,bias_max=-0.500000): 8192 runs -  13.77 us/run -  128 kB/run -  8.87 GB/s
  ALIBI(type=f32,ne_a=[16,2,10,1],n_past=0,n_head=10,bias_max=0.500000):   8192 runs -  0.38 us/run -   2 kB/run -    6.30 GB/s
  ALIBI(type=f32,ne_a=[16,2,32,1],n_past=0,n_head=32,bias_max=0.500000):   8192 runs -  1.07 us/run -   8 kB/run -    7.10 GB/s
  ALIBI(type=f32,ne_a=[128,4,10,1],n_past=0,n_head=10,bias_max=0.500000):  8192 runs -  4.29 us/run -   40 kB/run -   8.88 GB/s
  ALIBI(type=f32,ne_a=[128,4,32,1],n_past=0,n_head=32,bias_max=0.500000):  8192 runs -  13.73 us/run -  128 kB/run -  8.89 GB/s
  Backend CPU: OK

Backend 2/2 (CANN0)
  Backend name: CANN0
  ALIBI(type=f32,ne_a=[16,2,10,1],n_past=0,n_head=10,bias_max=-0.500000):  8192 runs -  94.75 us/run -   2 kB/run -    0.03 GB/s
  ALIBI(type=f32,ne_a=[16,2,32,1],n_past=0,n_head=32,bias_max=-0.500000):  8192 runs -  78.96 us/run -   8 kB/run -    0.10 GB/s
  ALIBI(type=f32,ne_a=[128,4,10,1],n_past=0,n_head=10,bias_max=-0.500000): 8192 runs -  99.79 us/run -   40 kB/run -   0.38 GB/s
  ALIBI(type=f32,ne_a=[128,4,32,1],n_past=0,n_head=32,bias_max=-0.500000): 8192 runs -  81.76 us/run -   128 kB/run -  1.49 GB/s
  ALIBI(type=f32,ne_a=[16,2,10,1],n_past=0,n_head=10,bias_max=0.500000):   8192 runs -  98.16 us/run -   2 kB/run -    0.02 GB/s
  ALIBI(type=f32,ne_a=[16,2,32,1],n_past=0,n_head=32,bias_max=0.500000):   8192 runs -  80.33 us/run -   8 kB/run -    0.09 GB/s
  ALIBI(type=f32,ne_a=[128,4,10,1],n_past=0,n_head=10,bias_max=0.500000):  8192 runs -  100.40 us/run -  40 kB/run -   0.38 GB/s
  ALIBI(type=f32,ne_a=[128,4,32,1],n_past=0,n_head=32,bias_max=0.500000):  8192 runs -  81.56 us/run -   128 kB/run -  1.50 GB/s
  Backend CANN0: OK

hipudding commented 5 months ago

量化算法需求

[toc]

计划支持常用量化算法

GGML_TYPE_Q4_0（最常用，ollama latest默认类型）

量化分组格式

#define QK4_0 32 // 每组32个f32数据
typedef struct {
    ggml_half d; // 公共系数
    uint8_t qs[QK4_0 / 2]; // 4bit存储的数据
} block_q4_0;

量化算法描述

void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int64_t k) {
    static const int qk = QK4_0;

    assert(k % qk == 0);

    const int nb = k / qk;

    // 1. 找到绝对值最大的数的值
    for (int i = 0; i < nb; i++) {
        float amax = 0.0f; // absolute max
        float max  = 0.0f;

        for (int j = 0; j < qk; j++) {
            const float v = x[i*qk + j];
            if (amax < fabsf(v)) {
                amax = fabsf(v);
                max  = v;
            }
        }

        // 2. 公共系数是第一步的值除以 -8
        const float d  = max / -8;
        const float id = d ? 1.0f/d : 0.0f;

        y[i].d = GGML_FP32_TO_FP16(d);

        // 3. 对组内的所有数据，除以公共系数，然后按以下顺序存储
        // 量化前： 1,2,3,4,5,6,7,8 ...... 30,31
        // 量化后： 1,16,2,17,3,18 ...... 15,31
        // 也就是数据按顺序先填充量化后组的低4位，然后再填充高4位。
        for (int j = 0; j < qk/2; ++j) {
            const float x0 = x[i*qk + 0    + j]*id;
            const float x1 = x[i*qk + qk/2 + j]*id;
                        // 量化的值+8.5，溢出后取最大值。
            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));

            y[i].qs[j]  = xi0;
            y[i].qs[j] |= xi1 << 4;
        }
    }
}

反量化算法描述

void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int64_t k) {
    static const int qk = QK4_0;

    assert(k % qk == 0);

    const int nb = k / qk;

    for (int i = 0; i < nb; i++) {
        const float d = GGML_FP16_TO_FP32(x[i].d);

                // 1. 量化后的4bit数据-8
        for (int j = 0; j < qk/2; ++j) {
            const int x0 = (x[i].qs[j] & 0x0F) - 8;
            const int x1 = (x[i].qs[j] >>   4) - 8;
                        // 2. 数据乘以公共系数后按以下顺序存储：
            // 反量化前： 1,16,2,17,3,18 ...... 15,31
            // 反量化后： 1,2,3,4,5,6,7,8 ...... 30,31
            y[i*qk + j + 0   ] = x0*d;
            y[i*qk + j + qk/2] = x1*d;
        }
    }
}

GGML_TYPE_Q4_1（实现方式相似，可一同实现，下同）

量化分组格式

#define QK4_1 32
typedef struct {
    union {
        struct {
            ggml_half d; // 公共系数
            ggml_half m; // 公共偏移
        } GGML_COMMON_AGGR;
        ggml_half2 dm;
    };
    uint8_t qs[QK4_1 / 2];

量化算法和反量化算法与Q4_0基本一致，额外多了公共偏移。

GGML_TYPE_Q5_0

量化分组格式

#define QK5_0 32
typedef struct {
    ggml_half d;           // 公共系数
    uint8_t qh[4];         // 5-th bit of quants
    uint8_t qs[QK5_0 / 2];
} block_q5_0;

量化算法和反量化算法与Q4_0基本一致，额外记录了每个数据的第五个bit位的值，在反量化的时候恢复。

GGML_TYPE_Q5_1

量化分组格式

#define QK5_1 32
typedef struct {
    union {
        struct {
            ggml_half d; // delta
            ggml_half m; // min
        } GGML_COMMON_AGGR;
        ggml_half2 dm;
    };
    uint8_t qh[4];         // 5-th bit of quants
    uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;

量化算法和反量化算法与Q5_0基本一致，额外多了公共偏移。

GGML_TYPE_Q8_0

#define QK8_0 32
typedef struct {
    ggml_half d;       // delta
    int8_t  qs[QK8_0];
} block_q8_0;

量化算法和反量化算法与Q4_0基本一致，区别是q8_0使用8位记录一个数据

Ascend实现难点

位操作和数据搬运无法使用向量计算。
一个分组是32个f32数据，需要按组处理，无法充分利用向量计算能力。

带量化的融合算子

GGML_OP_MUL_MAT (核心算子，性能对整个推理影响非常大)

输入2个矩阵（src0：量化后的tensor， src1：f32或f16），量化后的数据直接相乘，结果是float类型。

GGML_OP_MUL_MAT_ID

支持多个Moe专家一起计算，计算方法与GGML_OP_MUL_MAT类似。

GGML_OP_GET_ROWS

该算子实际上做了Embedding操作，但是允许输入的tensor是量化的tensor。输出F32的tensor。

GGML_OP_CPY

Tensor拷贝，允许从F16或者F32的tensor拷贝到一个量化的tensor中。

hipudding commented 4 months ago

aclnn只能适配部分场景的算子： UPSCALE：需要支持非整数倍scale CONCAT：需要支持维度不同的tensor做concat DUP：需要支持量化，非连续，不同shape，不同数据类型之间做拷贝 ROPE：不支持n_dims < ne0的场景，不支持fs,ef等参数为非默认值场景，且算子复杂，应考虑ascendc方式实现 MUL_MAT：量化相乘不支持batch

算法有差异的算子： ARGSORT：排序使用了不稳定算法（千问在用）

不支持的算子 flash_attention：能提速 mul_mat_id：MOE需要使用 sigmoid: 新增算子，实现简单 conv_transpose_1d: 新增算子

TOP：

MUl_MAT ROPE SOFTMAX(alibi)

airMeng commented 4 months ago

excellent work! hope to get access and try it myself soon!

hipudding / llama.cpp

TODO List #2

llama.cpp 算子性能对比

环境信息

构建命令

算子性能对比

GGML_OP_ADD 😄

GGML_OP_ACC 😄

GGML_OP_MUL 😄

GGML_OP_DIV 😄

GGML_UNARY_OP_GELU 😄

GGML_UNARY_OP_SILU 😄

GGML_UNARY_OP_GELU_QUICK 😄

GGML_UNARY_OP_TANH 😄

GGML_UNARY_OP_RELU 😄

GGML_UNARY_OP_HARDSIGMOID 😄

GGML_UNARY_OP_HARDSWISH 😄

GGML_OP_NORM 😄

GGML_OP_GROUP_NORM 😄

GGML_OP_UPSCALE ❓ （算子执行出错）

GGML_OP_PAD 😄

GGML_OP_LEAKY_RELU 😄

GGML_OP_SCALE 😄

GGML_OP_CLAMP 😄

GGML_OP_ARGSORT 😄

GGML_OP_CONCAT 🌚

GGML_OP_REPEAT 🌚

GGML_OP_ARANGE 🌚

GGML_OP_TIMESTEP_EMBEDDING 🌚 （需实现融合算子）

GGML_OP_RMS_NORM 🌚（需修改算子接口）

GGML_OP_SQR 🌚

GGML_OP_CONT 🌚

GGML_OP_DIAG_MASK_INF 🌚 （需实现融合算子）

GGML_OP_IM2COL 🌚 （需实现融合算子）

GGML_OP_POOL_2D 🌚 （需实现融合算子）

GGML_OP_SUM_ROWS 🌚

GGML_OP_DUP 🌚

GGML_OP_SOFTMAX （需实现融合算子）

GGML_OP_ROPE （需实现融合算子）

GGML_OP_ALIBI 🌚 （需实现融合算子）

量化算法需求

计划支持常用量化算法

GGML_TYPE_Q4_0（最常用，ollama latest默认类型）

GGML_TYPE_Q4_1（实现方式相似，可一同实现，下同）

GGML_TYPE_Q5_0

GGML_TYPE_Q5_1

GGML_TYPE_Q8_0

Ascend实现难点

带量化的融合算子

GGML_OP_MUL_MAT (核心算子，性能对整个推理影响非常大)

GGML_OP_MUL_MAT_ID

GGML_OP_GET_ROWS

GGML_OP_CPY