[Mesa-dev] Backends and support for pow-instructions

[Thomas Helland]

Hi all,

I've been looking at some shaders from Portal recently.
A lot of them seem to hand-roll a pow-function with
log2, multiply, and exp2, like this:

r3.x = log2( r3.x );
r3.y = log2( r3.y );
r3.z = log2( r3.z );
r3.xyz = r3.xyz * vd2.zzz;
r3.x = exp2( r3.x );
r3.y = exp2( r3.y );
r3.z = exp2( r3.z );

The corresponding output from the shader-db run shows:

   (assign  (xyz) (var_ref r3)  (expression vec3 log2 (swiz xyz (var_ref
r3) )) )
0x000003c8: math log(8)     g27<1>.xyzF     g27<4,4,1>.xyzzF null
{ align16 WE_normal 1Q compacted };
   (assign  (xz) (var_ref r6)  (swiz yy (array_ref (var_ref vc) (constant
int (0)) ) ))
0x000003d0: mov(8)          g25<1>.xzF      g1<0,4,1>.yF
 { align16 WE_normal 1Q };
   (assign  (w) (var_ref r0)  (expression float + (expression float dot
(swiz xyz (array_ref (var_ref vc) (constant int (28)) ) )(expression vec3
neg (swiz xyz (var_ref r5) )) ) (expression float neg (swiz z (array_ref
(var_ref vc) (constant int (30)) ) )) ) )
0x000003e0: dp3(8)          g105<1>.xF      g5.4<0,4,1>.xyzzF
-g26<4,4,1>.xyzzF { align16 WE_normal 1Q };
   (assign  (xyz) (var_ref r0)  (expression vec3 + (expression vec3 * (swiz
xyz (var_ref r0) )(expression vec3 neg (swiz xxx (var_ref
inversesqrt_retval) )) ) (expression vec3 neg (swiz xyz (array_ref (var_ref
vc) (constant int (28)) ) )) ) )
0x000003f0: mad(8)          g30<1>.xyzF     g108<4,4,1>F.xyzz
g30<4,4,1>F.xyzz -g97<4,4,1>F.x { align16 WE_normal 1Q };
   (assign  (xyz) (var_ref r6)  (expression vec3 * (swiz yxx (var_ref r7)
)(swiz xyz (var_ref r6) )) )
0x00000400: mul(8)          g25<1>.xyzF     g24<4,4,1>.yxxxF
g25<4,4,1>.xyzzF { align16 WE_normal 1Q };
   (assign  (xyz) (var_ref r3)  (expression vec3 * (swiz xyz (var_ref r3)
)(constant vec3 (2.200000 2.200000 2.200000)) ) )
0x00000410: mul(8)          g27<1>.xyzF     g27<4,4,1>.xyzzF 2.2F
{ align16 WE_normal 1Q };
   (assign  (w) (var_ref r0)  (expression float + (expression float dot
(swiz xyz (array_ref (var_ref vc) (constant int (28)) ) )(expression vec3
neg (swiz xyz (var_ref r5) )) ) (expression float neg (swiz z (array_ref
(var_ref vc) (constant int (30)) ) )) ) )
0x00000420: add(8)          g107<1>.xF      g105<4,4,1>.xF  -g6.4<0,4,1>.zF
{ align16 WE_normal 1Q };
   (assign  (xyz) (var_ref r0)  (expression vec3 + (expression vec3 * (swiz
www (array_ref (var_ref vc) (constant int (27)) ) )(swiz xyz (var_ref r0)
)) (swiz xyz (var_ref r5) )) )
0x00000430: mad(8)          g30<1>.xyzF     g26<4,4,1>F.xyzz g5.3<0,1,0>F.w
g30<4,4,1>F.xyzz { align16 WE_normal 1Q };
   (assign  (x) (var_ref dot_retval)  (expression float dot (swiz xyz
(array_ref (var_ref vc) (constant int (31)) ) )(swiz xyz (var_ref r6) )) )
0x00000440: dp3(8)          g18<1>.xF       g7<0,4,1>.xyzzF
g25<4,4,1>.xyzzF { align16 WE_normal 1Q compacted };
   (assign  (xyz) (var_ref r3)  (expression vec3 exp2 (swiz xyz (var_ref
r3) )) )
0x00000448: math exp(8)     g27<1>.xyzF     g27<4,4,1>.xyzzF null
{ align16 WE_normal 1Q compacted };


As far as I've been able to understand, the i965 hardware
has a pow() instruction, how is this with other hardware?

I've considered writing an algebraic optimization to convert
this into an ir_binop_pow. If my understanding is correct the backend
will then implement this in a similar fashion as above if it does not
have a native pow() instruction.

If, on the other hand, we have a pow() instruction, my guess is
we'd see reduced instruction-counts.

Is my understanding correct? Is this something that's worth doing?

regards,
Thomas Helland
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20140713/d5357409/attachment.html>