[Mesa-dev] [PATCH 1/2] nir: add nir_opt_move_load_ubo() optimization pass

[Marek Olšák]

On Mon, Mar 12, 2018 at 6:18 AM, Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
>
>
> On 03/11/2018 04:41 PM, Marek Olšák wrote:
>>
>> On Thu, Mar 8, 2018 at 5:48 PM, Ian Romanick <idr at freedesktop.org> wrote:
>>>
>>> On 03/08/2018 06:50 AM, Samuel Pitoiset wrote:
>>>>
>>>> This pass moves load UBO operations just before their first use,
>>>> loosely based on nir_opt_move_comparisons.
>>>
>>>
>>> If I'm reading this correctly, it moves UBO loads closer to the first
>>> use in the same block.  My assumption is the benefit in the next patch
>>> occurs because live ranges are smaller.  It seems like this could also
>>> hurt performance since it may be harder for the schedule to hide the
>>> latency of the load when register pressure is not an issue.  Have you
>>> measured performance of running apps to see if this is an issue?
>>>
>>> I'm mostly asking because Jason had a series for global code motion that
>>> does, in some cases, the opposite of this patch by moving UBO loads up
>>> to earlier blocks.
>>
>>
>> The pass is OK for LLVM, because LLVM does CSE across basic blocks,
>> and it also does instruction scheduling within a block.
>>
>> radeonsi/tgsi does the same thing: it load uniforms from memory at
>> every use. It sounds inefficient, but we found out that it's the best
>> thing to do with LLVM. LLVM can move loads away from the use, but it
>> doesn't move loads close to the use.
>
>
> Exactly, RadeonSI does something similar. Though the shader-db result posted
> by Timothy doesn't look very good for NIR, what do you think?
>

The results are OK.

Marek