[Mesa-dev] [PATCH 2/3] glsl: Optimize clamp(x, 0.0, b), where b < 1.0 as saturate(min(x, b))

[Matt Turner]

On Tue, Jun 24, 2014 at 6:27 AM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> On Tue, Jun 24, 2014 at 1:32 AM, Abdiel Janulgue
> <abdiel.janulgue at linux.intel.com> wrote:
>> Signed-off-by: Abdiel Janulgue <abdiel.janulgue at linux.intel.com>
>> ---
>>  src/glsl/opt_algebraic.cpp | 4 ++++
>>  1 file changed, 4 insertions(+)
>>
>> diff --git a/src/glsl/opt_algebraic.cpp b/src/glsl/opt_algebraic.cpp
>> index 8d7609d..0d694b6 100644
>> --- a/src/glsl/opt_algebraic.cpp
>> +++ b/src/glsl/opt_algebraic.cpp
>> @@ -589,6 +589,10 @@ ir_algebraic_visitor::handle_expression(ir_expression *ir)
>>              /* Found a min (max(x, 0), 1.0) */
>>              if (outer_const->is_one() && inner_val_a->is_zero())
>>                 return saturate(inner_val_b);
>> +
>> +            /* Found a min (max(x, 0.0) b), where b < 1.0 */
>> +            if ((outer_const->get_float_component(0) < 1.0f) && inner_val_b->is_zero())
>> +               return saturate(expr(ir_binop_min, inner_val_a, outer_const));
>
> Not sure about other GPU's, but on reasonably modern nvidia gpu's, in
> addition to being able to just saturate a single var, some other
> instructions take a "saturate" modifier (like fp add/multiply). But
> not min/max. If you were to write this as
>
> min(saturate(a), b)
>
> then it would be easier to propagate the saturate as a modifier to
> whatever expression produces 'a' without having a clever optimization
> that moves it through min/max.

Interesting. I guess it's (basically) all the same to us. It's pretty
rare that a i965 instruction can't do saturate.