[Mesa-users] Trouble compiling MesaGL : llvm intrinsics seem to be missing.

[Brian Paul]

On 11/04/2016 07:34 AM, Chuck Atkins wrote:
>     So you should actually do (without removing anything out of files like
>     I said to): --with-gallium-drivers=swrast,llvmpipe
>
>
> Still no.  llvmpipe isn't a driver option.
> --with-gallium-drivers=swrast specifies to use the gallium software
> rasterizer.  This will eable softpipe AND, if llvm is enabled (i.e.
> --enable-gallium-llvm) also llvmpipe.  If you want the fastest software
> render usable with osmesa then use swr.  OSMesa requires that at least
> swrast be enabled, hence enabling swrast and swr.  This will give you
> OSMesa with softpipe, llvmpipe, and swr, with llvmpipe being the default.
>
>
>     I actually don't know terribly much about openswr I just assumed it
>     would be optimized in someway for distributed systems that hinders non
>     distributed ones or designed with distributed as priority leaving out
>     minor optimizations or something.
>
>
> Nope.  It's a common misunderstanding though since large distributed
> systems without GPUs was the target as most other environments will have
> GPUs.  It's developed entirely though for a single node with no
> awareness to distributed systems.  Anything beyond that is the
> applications responsibilty.
>
>
>     And I am also not sure which version
>     of OpenGL it supports or the extensions.+ the fact that it is the
>     cause of a compilation error on the compilers system.
>
>
> They're vary close to feature parity.  From the current git / master
> (just past the 13.0 release):
>
> llvmpipe:
> OpenGL renderer string: Gallium 0.4 on llvmpipe (LLVM 3.8, 256 bits)
> OpenGL core profile version string: 3.3 (Core Profile) Mesa 13.1.0-devel
> (git-e3e5b1a)
> OpenGL version string: 3.0 Mesa 13.1.0-devel (git-e3e5b1a)
>
> swr:
> OpenGL renderer string: Gallium 0.4 on SWR
> OpenGL core profile version string: 3.3 (Core Profile) Mesa 13.1.0-devel
> (git-e3e5b1a)
> OpenGL version string: 3.0 Mesa 13.1.0-devel (git-e3e5b1a)
>
> As for extensions, llvmpipe -> swr:
> -GL_AMD_vertex_shader_viewport_index
> -GL_ARB_copy_image
> -GL_ARB_enhanced_layouts
> -GL_ARB_fragment_layer_viewport
> -GL_ARB_texture_buffer_range
> -GL_ARB_texture_cube_map_array
> -GL_ARB_texture_gather
> -GL_ARB_texture_view
> -GL_ARB_viewport_array
> +GL_ARB_pipeline_statistics_query
>
>
>     For mesa 12.0.1
>     swr is behind llvmpipe with extensions (at least from 4.0 core). So I
>     still say llvmpipe is probably preferable but perhaps swr does have
>     some performance advantage or is less buggy.
>
> It's come a long way since 12.0.1.  The current 13.0.0 release is a
> major step forward since the inital release.  While I'm not one of the
> swr developers we do make extensive use of it in production
> environments.  Regarding performance diffferences, you'll likely see an
> increase by a factor of 10x - 100x over llvmpipe, that's typically what
> we see in VTK and ParaView.

Are your workloads heavy on vertex transformation?  llvmpipe does vertex 
processing serially, while, I believe, swr has parallelism.  On 
workloads dominated by fragment processing, I'd expect swr and llvmpipe 
to fairly close.

Thanks for helping out on this thread, Chuck.  I've only had time to 
skim the thread.

-Brian