blas tflops support

[shuailong616]

PR Category

Benchmark

Type of Change

New Feature

Description

The total flops is calculated by judging the shape information by the op name, the time is calculated by triton dobench, and then the TFLOPS is calculated, and finally the TFLOPS is printed out by modifying the attri_util.py

op mm shape(m,n)(n,p) total_flops= mxnx2k

op addmm shape(m,n)(n,p) total_flops= mxpx(2n+1)

op bmm shape(b,n,m), (b,m,p) total_flops= bxnxpx2m

op mv shape(n,m)(m,) total_flops = n*2m

Issue

275

Performance

[shuailong616]

code format

[shuailong616]

pelase tell me some details, I don't konw why model-test and op-unit-test no run