[Mesa-dev] [PATCH v3 000/104] nir: Move to using instructions for derefs

[Jason Ekstrand]

On Mon, Apr 9, 2018 at 4:58 PM, Caio Marcelo de Oliveira Filho <
caio.oliveira at intel.com> wrote:

> Hi,
>
> Given the fixes you already made based on my comments. Patches 1-20,
> 22-27, 29-43, and 61 (multiview!) are
>
> Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira at intel.com>
>
> Patches 46-47 and 49 seem to be valid regardless the rest of the code,
> so I'd consider getting them in independently. They are also R-b'ed.
>

Thanks!


> I've skipped 21 and 28 because I wanted to give a deeper look at the
> originals.
>
> From the perspective of someone that is living with deref_vars for
> just a short time, I like the idea of removing one special
> construction (derefs) and rely on instructions instead.
>
> Which made me wonder: was there a special factor that led NIR to start
> with the "old-school derefs" in the first place? Other day Curro asked
> about one of the "selling points" of NIR being it did not have all
> those nodes representing dereferences. I digged up an old comment to
> what I think he was referring to
>
>     https://lists.freedesktop.org/archives/mesa-dev/2014-
> February/053477.html
>
>     - All the ir_dereference chains blow up the memory usage, and the
>     constant pointer chasing in the recursive algorithms needed to handle
>     them is not just cache-unfriendly but "cache-mean."
>
> How does deref_instructions avoid being "cache-mean" as their
> "predecessors"? Was the blow up more a result of how the instructions
> were structured than the fact it had those dereferences nodes?
>

The blow up was mostly due to the fact that GLSL IR uses dereferences for
*everything*.  NIR, in contrast, uses SSA defs for 95% of all temporary
values so there simply aren't as many deref chains in play.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180409/7035827d/attachment.html>