[Mesa-dev] [PATCH 2/2] glsl: Only call do_common_optimization once

[Timothy Arceri]

On Sun, 2016-11-13 at 02:29 +0100, Thomas Helland wrote:
> It turns out that only 8% of shaders benefit from more than one run
> of the do_common_optimization function, now that it has the loop.
> Doing only one run makes no difference to instruction count with NIR.
> 
> 8324 shaders were run

This is a small amount of shaders I take it this is just the public
shader-db?

While the series is interesting I think there are more robust ways to
deal with the issue of slow compile times. Landing the shader-cache
(once I rebase it on the recent changes I've made) is obviously one but
we are also getting very close to being able to use the NIR
optimisation passes from the GLSL IR linker (at least for things that
have a NIR backend). Once we do that we can work towards only calling
the GLSL IR passes from the compiler. 

Since I landed changes to switch i965 to using nir_gather_shader_info()
this week the only big blocker for using NIR opts in the linker is a
varying packing pass for NIR, once we have that is should be fairly
easy to wire things up like so in brw_create_nir() for i965:

   nir = brw_preprocess_nir(brw->screen->compiler, nir);
 
   //TODO: remove from GLSL IR the varyings/uniform that the nir opt
   // passes removed. This is safe because at this point all we want
   // from the GLSL IR is to assign locations and do validation on
   // varyings and uniforms we don't care if we make the IR invalid.

   //TODO: call GLSL assign varying/uniform location/validation passes

   //TODO: now remove from nir any more varyings that GLSL IR removed  
   // (this is safe because these will just be varyings that are only 
   //  used in one of the stages) and add locations that were assigned
   // in GLSL to NIR.

   //TODO: free GLSL IR

   //TODO: add a NIR varying packing pass here

We also need theses patches I'm still waiting on review for [1][2].

Anyway the point is IMO your time would probably be better spent
working towards improving deficiencies in the NIR opts so we can switch
to only using them faster as opposed to spending to much time improving
the GLSL IR opts.


[1] https://patchwork.freedesktop.org/patch/112044/
[2] https://patchwork.freedesktop.org/patch/112045/


> 8128 benefit from 1 calls to do_common_optimization
>  679 benefit from 2 calls to do_common_optimization
>   69 benefit from 3 calls to do_common_optimization
> ---
>  src/compiler/glsl/glsl_parser_extras.cpp | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/src/compiler/glsl/glsl_parser_extras.cpp
> b/src/compiler/glsl/glsl_parser_extras.cpp
> index a5a0837..a1a1f90 100644
> --- a/src/compiler/glsl/glsl_parser_extras.cpp
> +++ b/src/compiler/glsl/glsl_parser_extras.cpp
> @@ -1944,9 +1944,8 @@ _mesa_glsl_compile_shader(struct gl_context
> *ctx, struct gl_shader *shader,
>        /* Do some optimization at compile time to reduce shader IR
> size
>         * and reduce later work if the same shader is linked multiple
> times
>         */
> -      while (do_common_optimization(shader->ir, false, false,
> options,
> -                                    ctx->Const.NativeIntegers))
> -         ;
> +      do_common_optimization(shader->ir, false, false, options,
> +                                    ctx->Const.NativeIntegers);
>  
>        validate_ir_tree(shader->ir);
>