[Mesa-dev] [PATCH 02/16] radeonsi: Initial geometry shader support

[Marek Olšák]

On Sat, Jan 25, 2014 at 2:21 AM, Marek Olšák <maraeo at gmail.com> wrote:
> On Fri, Jan 24, 2014 at 10:19 PM, Marek Olšák <maraeo at gmail.com> wrote:
>> 1) Do we really have to have so much code inside #if...#endif? Can we
>> always compile everything and just not set the pipe_context functions
>> if LLVM < 3.5 and the same for get_(shader_)param? You can even do it
>> without the preprocessor, e.g. if (HAVE_LLVM >= 0x0305) ...
>>
>> 2) Sooner or later we'll have GL_ARB_shader_storage_buffer_object,
>> which adds shader read-write resources. Maybe it would be better to
>> put the ring buffers in a new array of descriptors and leave the
>> constant buffers with the READ-only usage. The new array of
>> descriptors could later be used for OpenGL read-write resources in
>> addition to being used for ring buffers.
>
> Actually, my original plan was to put stream-out buffers after OpenGL
> shader-storage buffers. We'll probably want to use the same array for
> OpenGL shader atomic counters as well.
>
> 3) In si_init_gs_rings:
> - could you please use readable decimal numbers for specifying the
> sizes? Like 1024 * 1024 * ...
> - could you please add an assertion to make sure the function is not
> called twice to prevent potential memory leaks
> - isn't 64 MB too many for a ring buffer?
>
> 4) If the GS is disabled, GS_OUT_PRIM_TYPE should be compatible with
> the draw_info primitive for transform feedback to work properly. It
> was a requirement on R600-Evergreen. Not sure if it applies to SI too.
>
> Please bear with me, I still haven't fully reviewed the whole file
> si_shader.c, but it's too late to do it today.

FYI, I'm done reviewing this.

Marek