This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH RFC v2] Add support for sparc cryptographic hash opcodes.
On Thu, Nov 15, 2012 at 2:06 AM, David Miller <davem@davemloft.net> wrote:
> From: "Carlos O'Donell" <carlos@systemhalted.org>
> Date: Wed, 14 Nov 2012 23:13:30 -0500
>
>> Nothing bad to report, just curious what kind of testing you did
>> and what kind of performance numbers you got?
>
> I made sure that all the crypt tests for md5/sha256/sha512 pass
> and that localedef generates correct md5 signatures.
>
> Performance wise those assembler routines have been in the Linux
> kernel crypto layer and the openssl CVS tree for a few months
> now and give performance on the order of:
>
> md5: 735MB/sec
> sha256: 700MB/sec
> sha512: 1.2GB/sec
>
> md5test-giant runs in about 1min16sec, and without these changes it
> takes 3min38sec.
>
> The biggest gain is probably sha512 on 32-bit, because gcc creates
> quite a mess with the C code we use. This is ironic because sha512
> has the potential the be the fastest, as shown above, due to it having
> the largest block size of the 3 hashes.
>
> It's partly the sparc backend's fault, because we don't take do
> 64-bit operations on 64-bit chips when in 32-bit as well as we could.
That's a great performance gain. Thanks for the datum.
Cheers,
Carlos.