[Mesa-dev] [PATCH] glsl: Optimize 1/exp(x) into exp(-x).

[Ian Romanick]

On 02/10/2015 01:11 PM, Jason Ekstrand wrote:
> Looks fine to me but why are we still adding algebraic optimizations to
> GLSL-IR?

Some cases, though this is not one, can slightly reduce memory usage.
GLSL IR will likely live on for some time as the pre-linked IR, so
that's a benefit we shouldn't discount.

While we are close to enabling NIR on the i965 scalar backend, that
leaves a lot of cases (including GS on all platforms, and VS on pre-GEN8
platforms) still needing love... not to mention non-Intel drivers.

> Reviewed-by: Jason Ekstrand <jason.ekstrand at intel.com
> <mailto:jason.ekstrand at intel.com>>

Also

Reviewed-by: Ian Romanick <ian.d.romanick at intel.com>

> On Feb 10, 2015 12:36 PM, "Matt Turner" <mattst88 at gmail.com
> <mailto:mattst88 at gmail.com>> wrote:
> 
>     Lots of shaders divide by exp2(...) which we turn into a multiplication
>     by the reciprocal. We can avoid the reciprocal by simply negating exp2's
>     argument.
> 
>     total instructions in shared programs: 5947154 -> 5946695 (-0.01%)
>     instructions in affected programs:     118661 -> 118202 (-0.39%)
>     helped:                                380
>     ---
>      src/glsl/opt_algebraic.cpp | 6 ++++++
>      1 file changed, 6 insertions(+)
> 
>     diff --git a/src/glsl/opt_algebraic.cpp b/src/glsl/opt_algebraic.cpp
>     index c6f4a9c..616ed37 100644
>     --- a/src/glsl/opt_algebraic.cpp
>     +++ b/src/glsl/opt_algebraic.cpp
>     @@ -747,6 +747,12 @@
>     ir_algebraic_visitor::handle_expression(ir_expression *ir)
>            if (op_expr[0] && op_expr[0]->operation == ir_unop_rcp)
>              return op_expr[0]->operands[0];
> 
>     +      if (op_expr[0] && (op_expr[0]->operation == ir_unop_exp2 ||
>     +                         op_expr[0]->operation == ir_unop_exp)) {
>     +         return new(mem_ctx) ir_expression(op_expr[0]->operation,
>     ir->type,
>     +                                         
>      neg(op_expr[0]->operands[0]));
>     +      }
>     +
>            /* While ir_to_mesa.cpp will lower sqrt(x) to rcp(rsq(x)), it
>     does so at
>             * its IR level, so we can always apply this transformation.
>             */
>     --
>     2.0.5
> 
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>