[Mesa-dev] [PATCH 22/22] meta/blit: Use GL_EXT_shader_samples_identical in MSAA-SS resolve blit

[Neil Roberts]

I made a pathological test case (attached) which repeatedly renders into
an MSAA FBO and then blits it to the screen and measures the framerate.
It checks it with a range of different sample counts. The rendering is
done either by rendering two triangles to fill the framebuffer or by
calling glClear.

The percentage increase in framerate after applying the patch is like
this:

With triangles to fill buffer:
	
16      62,27%
8       48,59%
4       27,72%
2       5,34%
0       0,58%
        
With glClear:   
        
16      -5,20%
8       -7,08%
4       -2,45%
2       -20,76%
0       3,71%

It seems like a pretty convincing win for the triangle case but the
clear case makes it slightly worse. Presumably this is because we don't
do anything to detect the value stored in the MCS buffer when doing a
fast clear so the fast path isn't taken and the shader being more
complicated makes it slower.

Not sure if we want to try and do anything about that because
potentially the cleared pixels aren't very common in a framebuffer from
a real use case so it might not really matter.

Currently we don't use SIMD16 for 16x MSAA because we can't allocate the
registers well enough to make it worthwhile. This patch makes that
problem a bit more interesting because even if we end up spilling a lot
it might still be worth doing SIMD16 because the cases where the spilled
instructions are hit would be much less common.

- Neil

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: msaa-time.c
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160218/f5a1b452/attachment-0001.c>
-------------- next part --------------

Ian Romanick <idr at freedesktop.org> writes:

> From: Ian Romanick <ian.d.romanick at intel.com>
>
> Somewhat surprisingly, this didn't have any affect on performance in the
> benchmarks that Martin tried for me.
>
> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
> ---
>  src/mesa/drivers/common/meta_blit.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
>
> diff --git a/src/mesa/drivers/common/meta_blit.c b/src/mesa/drivers/common/meta_blit.c
> index 28aabd3..c0ec51f 100644
> --- a/src/mesa/drivers/common/meta_blit.c
> +++ b/src/mesa/drivers/common/meta_blit.c
> @@ -530,6 +530,7 @@ setup_glsl_msaa_blit_shader(struct gl_context *ctx,
>           fs_source = ralloc_asprintf(mem_ctx,
>                                       "#version 130\n"
>                                       "#extension GL_ARB_texture_multisample: require\n"
> +                                     "#extension GL_EXT_shader_samples_identical: enable\n"
>                                       "#define gvec4 %svec4\n"
>                                       "uniform %ssampler2DMS%s texSampler;\n"
>                                       "in %s texCoords;\n"
> @@ -569,7 +570,14 @@ setup_glsl_msaa_blit_shader(struct gl_context *ctx,
>                                       "   i%s tc = i%s(texCoords);\n"
>                                       "   int i;\n"
>                                       "\n"
> -                                     "   for (i = 0; i < SAMPLES; i++)\n"
> +                                     "   S[0] = texelFetch(texSampler, tc, 0);\n"
> +                                     "#if defined(GL_EXT_shader_samples_identical) && SAMPLES > 1\n"
> +                                     "   if (textureSamplesIdenticalEXT(texSampler, tc)) {\n"
> +                                     "      emit2(S[0]);\n"
> +                                     "      return;\n"
> +                                     "   }\n"
> +                                     "#endif\n"
> +                                     "   for (i = 1; i < SAMPLES; i++)\n"
>                                       "      S[i] = texelFetch(texSampler, tc, i);\n"
>                                       "\n"
>                                       "   REDUCE(s16, s32);\n"
> -- 
> 2.5.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev