[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: fsm-lite -- should binaries down from sse4.2 to baseline be compiled on amd64?

Hi Nilesh,

Nilesh Patra, on 2021-03-10 17:19:40 +0530:
> fsm-lite seems to pass in -msse4.2 for amd64. From the conversation I
> had in debian-cross@l.d.o this is probably a baseline violation --
> please check this mail[1]

If there are SSE4.2 instructions, then the program will crash
with Illegal Instruction error on early versions of the AMD
Athlon 64 CPU for instance.  So yes, that would be a baseline

> Probably the best way forward is to build sse4.2 down to no vectorized
> operations, like it is done in simde-based builds?
> I'm not sure about this, please let me know.

I think the easier first step would be to just drop the build
flag.  Since upstream added -msse4.2 to build flags, it is quite
possible it makes a differences in terms of performances.  In
which case, it may be worth using a selector program checking
CPU flags, like is done with SIMDe, and launch the appropriate
binary depending on the support, in the present case SSE4.2 or
generic x86_64.

Maybe other SIMD flags would make further difference in terms of
performances, but it would be interesting to have a proper
benchmark to measure this.  I am not too sure however that
adding a selector on CPU flags would apply as a "minimal
change".  With the hard freeze beginning tomorrow, the package
would probably need substantial autopkgtest in addition to the
wrapper, to make it's way in the 20 days time frame.

> In any case this is likely a case for RC bug -- thought?

I'm afraid it is an RC bug, yes.

> [1]: https://lists.debian.org/debian-cross/2021/03/msg00007.html

Thanks for catching this,
Étienne Mollier <etienne.mollier@mailoo.org>
Fingerprint:  8f91 b227 c7d6 f2b1 948c  8236 793c f67e 8f0d 11da
Sent from /dev/pts/1, please excuse my verbosity.

Attachment: signature.asc
Description: PGP signature

Reply to: