cult icon indicating copy to clipboard operation
cult copied to clipboard
verified publisher icon asmjit

[Feature Request] Extension to the instruction flag

Open billyauhk opened this issue 4 years ago • 4 comments

Is it possible to have an extension to the --instruction flag which also allow us to specify the instruction signature? Say I want to test --instruction="vsomevector zmm,zmm" directly without going over the xmm and ymm variant.

billyauhk avatar Mar 23 '21 06:03 billyauhk

I'm wondering, what is a use-case for this?

kobalicek avatar Mar 23 '21 22:03 kobalicek

I am trying to exclude that the long latency of the zmm variant is related to the the preceding instructions of the ymm variant...

"instructions":[ {"inst":"vfmadd231ps xmm, xmm, xmm" ,"lat": 2.96,"rcp": 0.43}, {"inst":"vfmadd231ps ymm, ymm, ymm" ,"lat": 2.96,"rcp": 0.44}, {"inst":"vfmadd231ps zmm, zmm, zmm" ,"lat": 55.17,"rcp": 0.45} ]

billyauhk avatar Mar 30 '21 09:03 billyauhk

I think this should be considered a bug - maybe we should run some code before testing anything AVX-512 to heat it up?

kobalicek avatar Mar 30 '21 09:03 kobalicek

It is probably not an AVX-512 issue as such situation seems only limited to AVX-512 FMA instructions: vpmaddubsw xmm, xmm, xmm : Lat: 3.70 Rcp: 0.53 vpmaddubsw ymm, ymm, ymm : Lat: 3.70 Rcp: 0.54 vpmaddubsw zmm, zmm, zmm : Lat: 4.17 Rcp: 0.62

billyauhk avatar Apr 02 '21 10:04 billyauhk