Adds arg for exec file dump dir, instruction counter script

[monorimet]

Adds the --exec_dump_dir flag to the gemmbench script and creates a script for generating a .csv with instruction counts (naive) from dumped .rocmasm files.

By naive, I mean it starts counting lines at "Kernarg preload header" and stops at "---" to count instructions.
@kuhar is the stopping criteria here too late? it includes what seems to be metadata, so maybe it should stop at "s_endpgm" or "func_end0"?

Perhaps some of the counter utilities should be further integrated into the benchmarking tooling.

The instruction counter generates a file like this (also attached):

Filename,Instruction Count
./dumps/gemm_4096_640_2560_bf16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1074
./dumps/gemm_4096_5120_640_f16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1074
./dumps/gemm_8192_640_2560_bf16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1074
./dumps/gemm_1024_1280_1280_bf16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1067
./dumps/gemm_2048_1280_1280_bf16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1067
./dumps/gemm_1024_10240_1280_bf16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1064
./dumps/gemm_128_1280_2048_bf16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1064
./dumps/gemm_64_640_2048_bf16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1064
./dumps/gemm_4096_4096_8192_f16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1034
./dumps/gemm_1024_1280_5120_bf16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,1003
./dumps/gemm_4096_4096_8192_bf16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,997
./dumps/gemm_64_1280_2048_bf16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,997
./dumps/gemm_2048_10240_1280_bf16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,997
./dumps/gemm_2048_1280_5120_bf16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,993
./dumps/gemm_8192_640_640_bf16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,993
./dumps/gemm_4096_640_640_bf16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,993
./dumps/gemm_8192_5120_640_bf16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,983
./dumps/gemm_4096_640_2560_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,931
./dumps/gemm_8192_640_2560_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,931
./dumps/gemm_4096_4096_8192_bf16_tA/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,929
./dumps/gemm_1024_1280_1280_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,914
./dumps/gemm_2048_1280_1280_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,914
./dumps/gemm_2048_2048_65536_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,909
./dumps/gemm_2048_8192_8192_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,909
./dumps/gemm_8192_8192_8192_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,909
./dumps/gemm_8192_2048_65536_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,909
./dumps/gemm_2_7168_8192_bf16_tB/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,908
./dumps/gemm_2048_2048_1024_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,901
./dumps/gemm_8192_2048_1024_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,901
./dumps/gemm_128_1280_2048_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,899
./dumps/gemm_1024_10240_1280_f16/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,899
./dumps/gemm_5120_8_3456_f16_tA/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,890
./dumps/gemm_3456_8_5120_f16_tA/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,890
./dumps/gemm_7680_8_5120_f16_tA/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,890
./dumps/gemm_8000_8_8192_f16_tA/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,890
./dumps/gemm_1920_8_5120_f16_tA/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,890
./dumps/gemm_5120_8_8192_f16_tA/module_main_dispatch_0_rocm_hsaco_fb.rocmasm,890
<truncated>

rocmasm_line_counts.csv

[kuhar]

Nice! We should be able to add more stats to this easily

[monorimet]

Updated results with the counter stopping fixed:
rocmasm_instr_counts.csv

Potential follow-ups:

• Cleanup dictionary entries i.e. gemm_<shape>_<dtype>_<variant> instead of relative path
• Detailed breakdown of instruction counts or other useful metrics

[kuhar]

LGTM % nit