[Mesa-dev] [PATCH] util: import sha1 implementation from OpenBSD

[Vladislav Egorov]

16.01.2017 16:13, Emil Velikov пишет:
> Hi Vladislav,
>
> On 14 January 2017 at 01:50, Vladislav Egorov <vegorov180 at gmail.com> wrote:
>> 14.01.2017 01:45, Timothy Arceri пишет:
>>> I'm asking for a chance to test before we jump in, its probably not a
>>> big deal and I may even still be able to reduce my use of hashing but
>>> it would be nice to be given a few days to test and even explore
>>> alternatives before jumping on this implementation.
>> A very quick and very dirty simple benchmark. I took shader-cache from
>> github, branch shader-cache39. Then I've applied my preprocessor patch on
>> top (because shader-cache still uses preprocessor even if the shader is
>> cached and it was painful to see preprocessor taking more than half of the
>> whole time). Then I've compiled it with openssl and with the Emil's patch.
>> Full run on shader-db (300Mb+ of shaders) with shader-cache warmed up. It
>> takes 78s, spends in libcrypto 0.27%. With OpenBSD SHA1 it runs
>> approximately the same time, spends 0.53% in SHA1Transform() and other SHA1*
>> functions. Subtest - 46Mb of shaders from Total War: Attila - 3.10s (for
>> some reason, the cache works much faster on smaller subsets than on full
>> shader-db). 1.08% were spent in libcrypto, 1.04% in
>> sha1_block_data_order_avx2(). With OpenBSD 3.07s - 2.27% in SHA1Transform()
>> and other SHA1* functions.
>>
> Did you mean "shader-db" with the "shader-cache" references above ? I
> cannot find any projects with the latter name on github.

I meant Timothy's github: https://github.com/tarceri/Mesa

>> Overall not that significant in context of shader-cache, but as expected, on
>> Haswell it's twice slower than OpenSSL's AVX2 implementation.
> If I understood you correctly you're saying that despite this
> implementation being slower the actual total runtime remains
> unchanged.
> In the Attila case we even get a minimal improvement, although I'm
> inclined to discard that as within the error margin.
>
> That sounds interesting, but in all honesty I don't believe it's worth
> worrying unless we have usecase(s) where the runtime difference is
> noticeable.
>
> Thanks
> Emil

Sure, it's just noise, it should be slower by 1%. When using hot shader 
cache it would be at most 1% difference, and at most ~2% difference on 
CPUs with hardware SHA like Goldmont. I am not sure about Vulkan. But 
yeah, it doesn't seem like there is something to worry about so far.