boyuanfeng / apnn-tc Goto Github PK

Hello, thanks for your wonderful work and available code! However, during read this paper, I have the following questions.

In the paper, the authors claim that "the int1 Tensor Core compute primitive can only generate 32 outputs". It seems that 32 bit-width output is relatively large for Int1 matric multiply-accumulate (MMA). So I want to check whether we can control the bit-width of outputs for Int1 MMA.

First, I try to access the white paper in reference [32] https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/nvidia-amperearchitecture-whitepaper.pdf but I got an error with "Even AI can't find this page!".

Then, I access https://www.nvidia.com/content/PDF/[nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.pdf](https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.pdf) to search relative information in this white paper. However, I still can not find the introduction about the setting of Tensor Core outputs' bitwidth.

So, my question is:
(1) Where can I find the description that can support or provide evidence for "the int1 Tensor Core compute primitive can only generate 32 outputs" in white papers or other documents?
(2) Do I have any way to control the outputs' bit-width of the int1 Tensor Core compute?

Looking forward to your reply.
Best wish to you！

Hi! I am learning your code, and firstly, I download the whole zip file, secondly, I put cutlass package into cutlass directory here. Then I cd to cutlass_kernel, I enter "make all". And it works! It shows something like this:
(base) C:\trash_can\APNN-TC-main\cutlass_kernel>make all nvcc -I../cutlass/include -I../cutlass/tools/util/include -I../cutlass/examples/common -std=c++11 -O3 -w -arch=sm_86 bench_gemm.cu -o bench_gemm.bin nvcc warning : The -std=c++11 flag is not supported with the configured host compiler. Flag will be ignored. bench_gemm.cu 正在创建库 bench_gemm.lib 和对象 bench_gemm.exp
And... I get a bench_gemm.bin. I am not sure...how to run this bin file? Previously I meet a.exe and a.out for win and linux as nvcc's return file. But never meet bin. Also find nothing on google...

Thank you!!!!