[Mesa-dev] [RFC] NIR serialization
Eero Tamminen
eero.t.tamminen at intel.com
Fri Sep 15 15:44:57 UTC 2017
Hi,
On 13.09.2017 19:08, Jason Ekstrand wrote:
> On Wed, Sep 13, 2017 at 8:42 AM, Nicolai Hähnle <nhaehnle at gmail.com
> <mailto:nhaehnle at gmail.com>> wrote:
> On 13.09.2017 11:54, Eero Tamminen wrote:
> On 12.09.2017 09:55, Jordan Justen wrote:
> On 2017-09-11 21:44:32, Timothy Arceri wrote:
> On 12/09/17 14:23, Ian Romanick wrote:
> On 09/08/2017 01:59 AM, Kenneth Graunke wrote:
> We shouldn't use SPIR-V for the shader cache.
>
> The compilation process for GLSL is: GLSL ->
> GLSL IR -> NIR -> i965 IRs.
> Storing the content at one of those points, and
> later loading it and
> resuming the normal compilation process from
> that point...that's totally
> reasonable.
>
> Having a fallback for "some things in the cache
> but not all the variants
> we needed" suddenly take a different compilation
> pipeline, i.e. SPIR-V
> -> NIR -> ... seems risky. It's a different
> compilation path that we
> don't normally use. And one you'd only hit in
> limited circumstances.
> There's a lot of potential for really obscure bugs.
>
>
> Since we're going to expose exactly that path for
> GL_ARB_spirv / OpenGL
> 4.6, we'd better make sure it works always. Right?
>
> One nice thing about SPIR-V is that all of the
> handling of uniform
> layouts, initial uniform values, attribute
> locations, etc. is already
> serialized. If I'm not mistaken, that was one of
> the big pain points
> for all of the existing on-disk storage methods.
> All of that has been
> sorted out for SPIR-V, and we have to make it work
> anyway.
>
>
> Correct these are the main issues for the fallback path,
> however this is
> only used by i965 (exactly because an intermediate cache
> is missing).
> Using SPIR-V as the intermediate cache means we still
> need to convert to
> NIR and run all the opts, so I don't really see the
> advantage of caching
> to SPIR-V over NIR.
>
>
> For shader cache, hopefully we'll normally have the final
> program in
> the cache, which means the 're-run opt passes' is probably
> not a big
> factor. But, it still seems a fair point.
>
> I think the biggest advantage of having either nir or spir-v
> would be
> not having to fallback to running the glsl compiler, right?
>
>
> Shader cache is performance optimization aimed at reducing
> compile times.
>
> When I earlier profiled it, 2/3 of the shader compilation work
> is done at linking stage.
Sorry, I should have mentioned that this is ancient data from before NIR
came into picture. Most of the time in linking was due to the i965
backend code, not intermediate representation manipulation.
> Caching NIR would completely eliminate the linking stage for shader
> variant compiles, since the NIR is already linked.
Quick check of latest Mesa code with Valgrind/Callgrind & perf and
SynMark v7 DrvShComp shader compilation test shows most of the time
(1/3) going to copy propagation in
./src/intel/compiler/brw_fs_copy_propagation.cpp.
Btw. I think one of the main things that shader compiler cache should
fix, is the huge amount of time spent with spilling shaders. Here's
some 1/2-1 year old profiling info about where they spend their time:
https://bugs.freedesktop.org/show_bug.cgi?id=93840#c33
https://bugs.freedesktop.org/show_bug.cgi?id=98455#c3
https://bugs.freedesktop.org/show_bug.cgi?id=97035#c3
- Eero
> Yes, so there goes 2/3 of our compile time.
>
> Cheers,
> Nicolai
>
>
> Caching done at higher level than the linked binary shaders, may
> be of questionable performance value, because the caching itself
> has also some cost (extra code, disk access cache lookups etc).
>
> Some profiling showing that doing caching at NIR/SPIRV level
> could measurable help performance (compared to caching
> overhead), would be appreciated.
>
> However, if the main purpose of caching NIR/SPIRV is something
> else than performance[1], then that's a different matter. It
> just should be clear what is the aim of this change.
>
>
> There are a couple of reasons:
>
> 1) Caching the linked NIR should git rid of 2/3 to 3/4 of your shader
> compile time. That's very significant. Sure, we can do better, but 70%
> is a big perf boost. :-) I highly doubt that nir deserialization will
> be a significant cost.
>
> 2) Provide a better fall-back when we don't have an Intel GEN binary.
> Falling back all the way to GLSL is very expensive and potentially
> unreliable due to the way OpenGL lets you change bindings post-linking
> and the changes don't apply until you re-link. Falling back to linked
> NIR should have similar reliability to the TGSI cache which seems to be
> working very well.
>
> --Jason
>
> Maybe we'll see which becomes available first? :)
>
>
>
> - Eero
>
> [1] Such as:
> * better code validation, or
> * ability to replace / tweak the cached files to manually test
> impact of compiler optimizations before implementing them
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> <mailto:mesa-dev at lists.freedesktop.org>
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> <https://lists.freedesktop.org/mailman/listinfo/mesa-dev>
More information about the mesa-dev
mailing list