[VTA] GEMM cycle count

Hello,

I was wondering how one can determine the number of cycles a GEMM operation needs based on the configuration file and input dimensions.

Thanks

1 Like

Did you mean the number of cycles for a single GEMM instruction? AKAIK, it depends on the implementation, as it for Chisel-based implementation, a single GEMM instruction takes 4 cycles to complete, since there are stages in the design that prepare the data stream for such execution.

Thank you! Just for me to clarify. If I look at the output of tutorial " Deploy Pretrained Vision Model from MxNet on VTA" it shows: Execution statistics: inp_load_nbytes : 5549568 wgt_load_nbytes : 12763136 acc_load_nbytes : 30720 uop_load_nbytes : 22832 out_store_nbytes: 1680896 gemm_counter : 6623232 alu_counter : 572320

in order to get the cycle count for the gemm, I multiply gemm_counter with 4. correct? Thanks!

Hi @liangfu

Based on your previous answer, HLS design have a different numbers of cycle for GEMM? If so, do you know how different it is in comparison to Chisel-based implementation?

Thank you,

ISS