New f1600x4 #14
New f1600x4 #14
Conversation
|
This also means it no longer works on Windows. |
shake256-avx2/f1600x4.s
Outdated
| #endif /* !__APPLE */ | ||
| movq $6, %rax | ||
| 17: # loop.begin: | ||
| vmovdqa 0(%rdi), %ymm8 |
There was a problem hiding this comment.
This doesn't compile with MSVC
There was a problem hiding this comment.
I don't have a windows machine available to me. What is the error?
There was a problem hiding this comment.
MSVC has completely different syntax; there is not a chance to write assembly code that works for both Windows and gnu-style assemblers without going to a portable assembler like NASM.
There was a problem hiding this comment.
peachpy (which I used to generate the assembler, see f1600x4.py) does support nasm as output.
There was a problem hiding this comment.
Could you not generate microsoft assembly as .asm files? Then those could be used on Windows (with a nmake Makefile (see PQClean) or you could consider contributing an appropriate CMakeList.txt).
There was a problem hiding this comment.
Unfortunately peachpy doesn't support microsoft asm. It can directly generate the windows object file. It can also generate nasm. What would we prefer? @joostrijneveld
Instead of using intrinsics and full unrolling, this uses a
four-round unrolled version adapted from the one I wrote for
Cloudflare's CIRCL library:
github.com/cloudflare/circl/simd/keccakf1600
| keccak4x/KeccakP-1600-times4-SIMD256.o | ||
| *.s |
There was a problem hiding this comment.
with the file having .S (capital letter) now this probably also needs to be updated?
There was a problem hiding this comment.
We wouldn't want to ignore the file itself, right? The .s is generated form the .S when compiling.
There was a problem hiding this comment.
ah you're not running peachpy in Make of course.
There was a problem hiding this comment.
Indeed. Try to keep dependencies to compile to a minimum.
|
@thomwiggers Any suggestions how to advance this? (Given you are facing the same troubles with Kyber & Dilithium.) |
|
The Kyber/Dilithium AVX2 implementations are simply marked as not supported on Windows. It would still be a shame to drop Windows support from this otherwise mostly (bar VLAs) portable plain-C implementation. |
Instead of using intrinsics and full unrolling, this uses a
four-round unrolled version adapted from the one I wrote for
Cloudflare's CIRCL library:
This is about 10-20% faster on my notebook. We'd better check whether that's also true for other systems as well before merging.