[Mesa-dev] [PATCH v2 6/7] nir: Teach nir_opt_algebraic about adding and subtracting the same thing

Matt Turner mattst88 at gmail.com
Wed Dec 16 17:38:34 PST 2015


On Wed, Dec 16, 2015 at 3:47 PM, Kristian Høgsberg <krh at bitplanet.net> wrote:
> From: Kristian Høgsberg Kristensen <krh at bitplanet.net>
>
> This optimizes a + b - b to just a. Modest shader-db results (BDW):
>
>   total instructions in shared programs: 7842452 -> 7841862 (-0.01%)
>   instructions in affected programs:     61938 -> 61348 (-0.95%)
>   total loops in shared programs:        2131 -> 2131 (0.00%)
>   helped:                                263
>   HURT:                                  0
>   GAINED:                                0
>   LOST:                                  0
>
> but the optimization turns
>
>   gl_VertexID - gl_BaseVertexARB
>
> into just a reference to SYSTEM_VALUE_VERTEX_ID_ZERO_BASE, which the
> i965 hardware supports natively. That means we can avoid using the
> internal vertex buffer for gl_BaseVertexARB in this case.
>
> Reviewed-by: Eduardo Lima Mitev <elima at igalia.com>
> Reviewed-by: Ian Romanick <ian.d.romanick at intel.com>
> ---
>  src/glsl/nir/nir_opt_algebraic.py | 4 ++++
>  1 file changed, 4 insertions(+)
>
> diff --git a/src/glsl/nir/nir_opt_algebraic.py b/src/glsl/nir/nir_opt_algebraic.py
> index cb715c0..1fdad3d 100644
> --- a/src/glsl/nir/nir_opt_algebraic.py
> +++ b/src/glsl/nir/nir_opt_algebraic.py
> @@ -62,6 +62,10 @@ optimizations = [
>     (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
>     (('fadd', ('fneg', a), a), 0.0),
>     (('iadd', ('ineg', a), a), 0),
> +   (('iadd', ('ineg', a), ('iadd', a, b)), b),
> +   (('iadd', a, ('iadd', ('ineg', a), b)), b),
> +   (('fadd', ('fneg', a), ('fadd', a, b)), b),
> +   (('fadd', a, ('fadd', ('fneg', a), b)), b),

a + b - b isn't equal to a for floating point numbers in general.
There are other technically unsafe algebraic optimizations we do today
that we'll probably have to predicate on some flag.

I think it's fine to add them at this point.

Reviewed-by: Matt Turner <mattst88 at gmail.com>


More information about the mesa-dev mailing list