[Mesa-dev] Optimizations in mesa shader compiler

[Ian Romanick]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Segovia, Benjamin wrote:
> Actually, even the computation of rad_angle is ignored at compile time.
> 
> I was also thinking about more optimizations in the final GPU code i.e. contuining what has been started with the _mesa_optimize function. rather working directly on the ir tree. Not sure of what is the best strategy. I actually output random shaders from warsow, openarena and nexuiz. For both GPU code and native Intel Gen code, there are clear room for improvements. I am not sure if there are easy to do (compared to pure scalar code, I imagine the main problem is to handle the swizzling thing).
> 
> So, even with glsl2, optimizing the final GPU code optimization pass would be still usefull (except if glsl2 properly optimizes everything before)

There will absolutely still be room for chip-specific peephole
optimizations.  That the way most C compilers work, after all.  There
are multiple layers of optimization from language and target generic all
the way through very, very target specific.  At this point we're focused
on the target-independent parts of the compiler.  We're still generating
Mesa's low-level assembly IR as our output.  We plan to fix that in
later work.

> ________________________________________
> From: Brian Paul [brianp at vmware.com]
> Sent: Wednesday, July 14, 2010 7:54 AM
> To: Segovia, Benjamin
> Cc: mesa-dev at lists.freedesktop.org
> Subject: Re: [Mesa-dev] Optimizations in mesa shader compiler
> 
> On 07/13/2010 05:46 PM, Segovia, Benjamin wrote:
>> To be more precise,
>> - The shader compiler present in src/mesa/slang is used.
>>
>> - The glsl shader (randomly taken on the net) is:
>> void main(){
>>    const float PI = 3.14159265358979323846264;
>>    const float angle = 45.0;
>>    const float rad_angle = angle*PI/180.0;
>>    vec4 a = gl_Vertex;
>>    vec4 b = a;
>>
>>    b.x = a.x*cos(rad_angle) - a.y*sin(rad_angle);
>>    b.y = a.y*cos(rad_angle) + a.x*sin(rad_angle);
>>    gl_Position = gl_ModelViewProjectionMatrix*b;
>> }

Ken's recent work in the glsl2 branch causes all of the sin/cos
computations to be replaced with constants.  The generated IR below is
from passing the --dump-lir option to src/glsl/glsl_compiler.  All of
the built-in functions and variables also appear in the output (not
included below), but that should be fixed either this week or next week.

During generation of the Mesa assembly IR, I'm pretty sure the
(a*b)+(c*d) sequences get turned into a MUL followed by a MAD.  Eric was
doing some work in that area recently.

(function main
(signature void
  (parameters
  )
((declare () vec4 b at 0x8a97910)

(assign (constant bool (1)) (var_ref b at 0x8a97910)  (var_ref
gl_Vertex at 0x88816f0) )

(declare () float assignment_tmp at 0x8a981b8)

(assign (constant bool (1)) (var_ref assignment_tmp at 0x8a981b8)
(expression float - (expression float * (swiz x (var_ref
gl_Vertex at 0x88816f0) )(constant float (0.707107)) ) (expression float *
(swiz y (var_ref gl_Vertex at 0x88816f0) )(constant float (0.707107)) ) ) )

(assign (constant bool (1)) (swiz x (var_ref b at 0x8a97910) ) (var_ref
assignment_tmp at 0x8a981b8) )

(declare () float assignment_tmp at 0x8a988f0)

(assign (constant bool (1)) (var_ref assignment_tmp at 0x8a988f0)
(expression float + (expression float * (swiz y (var_ref
gl_Vertex at 0x88816f0) )(constant float (0.707107)) ) (expression float *
(swiz x (var_ref gl_Vertex at 0x88816f0) )(constant float (0.707107)) ) ) )

(assign (constant bool (1)) (swiz y (var_ref b at 0x8a97910) ) (var_ref
assignment_tmp at 0x8a988f0) )
(declare () vec4 assignment_tmp at 0x8a98c38)

(assign (constant bool (1)) (var_ref assignment_tmp at 0x8a98c38)
(expression vec4 * (var_ref gl_ModelViewProjectionMatrix at 0x8882a00)
(var_ref b at 0x8a97910) ) )

(assign (constant bool (1)) (var_ref gl_Position at 0x887e4e0)  (var_ref
assignment_tmp at 0x8a98c38) )
))

)

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAkw/XJUACgkQX1gOwKyEAw8spgCeKfrjl6NSAEfVnpbi9j8b2L11
arsAnRHCoDH05BmDj2pmWvA/X2k29vZz
=K7Uc
-----END PGP SIGNATURE-----