[Mesa-dev] [PATCH 1/2] nir: optimize min(min(a, b), c) to min3(a, b, c) under an option

[Timothy Arceri]

On 31/05/18 19:37, Bas Nieuwenhuizen wrote:
> On Thu, May 31, 2018 at 11:35 AM, Timothy Arceri <tarceri at itsqueeze.com> wrote:
>>
>>
>> On 31/05/18 19:26, Samuel Pitoiset wrote:
>>>
>>>
>>>
>>> On 05/31/2018 11:25 AM, Bas Nieuwenhuizen wrote:
>>>>
>>>> On Thu, May 31, 2018 at 11:13 AM, Samuel Pitoiset
>>>> <samuel.pitoiset at gmail.com> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 05/31/2018 01:55 AM, Timothy Arceri wrote:
>>>>>>
>>>>>>
>>>>>> On 31/05/18 03:06, Samuel Pitoiset wrote:
>>>>>>>
>>>>>>>
>>>>>>> Similar for max().
>>>>>>>
>>>>>>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>>>>>>> ---
>>>>>>>     src/compiler/nir/nir.h                | 3 +++
>>>>>>>     src/compiler/nir/nir_opt_algebraic.py | 8 ++++++++
>>>>>>>     2 files changed, 11 insertions(+)
>>>>>>>
>>>>>>> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
>>>>>>> index f6086bd6c0..04991b7d04 100644
>>>>>>> --- a/src/compiler/nir/nir.h
>>>>>>> +++ b/src/compiler/nir/nir.h
>>>>>>> @@ -1897,6 +1897,9 @@ typedef struct nir_shader_compiler_options {
>>>>>>>        /* lower b2f to iand */
>>>>>>>        bool lower_b2f;
>>>>>>> +   /* lower min(min(a, b), c) to min3(a, b, c) (same for max()). */
>>>>>>> +   bool lower_minmax3;
>>>>>>> +
>>>>>>>        /* Does the native fdot instruction replicate its result for
>>>>>>> four
>>>>>>>         * components?  If so, then opt_algebraic_late will turn all
>>>>>>> fdotN
>>>>>>>         * instructions into fdot_replicatedN instructions.
>>>>>>> diff --git a/src/compiler/nir/nir_opt_algebraic.py
>>>>>>> b/src/compiler/nir/nir_opt_algebraic.py
>>>>>>> index 909ea3daf4..1d67e2d88c 100644
>>>>>>> --- a/src/compiler/nir/nir_opt_algebraic.py
>>>>>>> +++ b/src/compiler/nir/nir_opt_algebraic.py
>>>>>>> @@ -224,6 +224,14 @@ optimizations = [
>>>>>>>        (('imax', a, a), a),
>>>>>>>        (('umin', a, a), a),
>>>>>>>        (('umax', a, a), a),
>>>>>>> +
>>>>>>> +   (('fmin', ('fmin', a, b), c), ('fmin3', a, b, c),
>>>>>>> 'options->lower_minmax3'),
>>>>>>> +   (('imin', ('imin', a, b), c), ('imin3', a, b, c),
>>>>>>> 'options->lower_minmax3'),
>>>>>>> +   (('umin', ('umin', a, b), c), ('umin3', a, b, c),
>>>>>>> 'options->lower_minmax3'),
>>>>>>> +   (('fmax', ('fmax', a, b), c), ('fmax3', a, b, c),
>>>>>>> 'options->lower_minmax3'),
>>>>>>> +   (('imax', ('imax', a, b), c), ('imax3', a, b, c),
>>>>>>> 'options->lower_minmax3'),
>>>>>>> +   (('umax', ('umax', a, b), c), ('umax3', a, b, c),
>>>>>>> 'options->lower_minmax3'),
>>>>>>
>>>>>>
>>>>>>
>>>>>> These look like they would be better suited in late_optimizations
>>>>>> rather
>>>>>> than optimizations?
>>>>>
>>>>>
>>>>>
>>>>> Totals from affected shaders:
>>>>> SGPRS: 104 -> 104 (0.00 %)
>>>>> VGPRS: 64 -> 64 (0.00 %)
>>>>> Spilled SGPRs: 0 -> 0 (0.00 %)
>>>>> Spilled VGPRs: 0 -> 0 (0.00 %)
>>>>> Private memory VGPRs: 0 -> 0 (0.00 %)
>>>>> Scratch size: 0 -> 0 (0.00 %) dwords per thread
>>>>> Code Size: 4652 -> 4644 (-0.17 %) bytes
>>>>> LDS: 4 -> 4 (0.00 %) blocks
>>>>> Max Waves: 24 -> 24 (0.00 %)
>>>>> Wait states: 0 -> 0 (0.00 %)
>>>>>
>>>>> When moved to late_optimizations.
>>>>
>>>>
>>>> Is that compared to no min3/max3 lowering at all or compared to doing
>>>> those non-late?
>>>
>>>
>>> Compared to "default" for both.
>>
>>
>> It might be worth looking at the difference in the shaders. It would be
>> interesting to understand why we need to do this early and if it means there
>> are other opportunities for lowering these that this pattern doesn't find.
> 
> The reason why late does not make any real diff is because only Intel
> calls the before_ffma/late variants of this optimization. I.e. if you
> move it to late the optimization does not happen.
> 
> Not sure we want to add late though, given
> 
> # Most of these optimizations aren't quite safe when you get infinity or
> # Nan involved but the first one should be fine.
> 
> comments in the late opts?

Patch sent to fix the comment:

https://patchwork.freedesktop.org/patch/226215/

>>
>>
>>>
>>>>>
>>>>>
>>>>>>
>>>>>>
>>>>>>> +
>>>>>>>        (('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))),
>>>>>>>        (('imin', a, ('ineg', a)), ('ineg', ('iabs', a))),
>>>>>>>        (('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))),
>>>>>>>
>>>>> _______________________________________________
>>>>> mesa-dev mailing list
>>>>> mesa-dev at lists.freedesktop.org
>>>>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev