[Mesa-dev] [PATCH 06/12] nir/opt_peephole_select: Don't try to remove flow control around indirect loads

[Eric Anholt]

Rob Clark <robdclark at gmail.com> writes:

> On Thu, Jun 28, 2018 at 12:46 AM, Ian Romanick <idr at freedesktop.org> wrote:
>> From: Ian Romanick <ian.d.romanick at intel.com>
>>
>> That flow control may be trying to avoid invalid loads.  On at least
>> some platforms, those loads can also be expensive.
>
> So for adreno, indirect uniform loads aren't really that expensive (it
> takes a few cycles to get the value from an alu instruction into the
> address register for indirect register addressing but that latency can
> usually be hidden).. and invalid access won't cause a fault or
> anything.  I think it just wraps around, so it will be some undefined
> value but won't cause a fault.
>
> This is completely different from UBO's were we could cause a fault,
> and loads are real memory access, so more expensive.
>
> So as long as this is just about uniforms (I think i965 calls them
> 'push constants'?), and not UBOs, then maybe we want
> nir_shader_compiler_options flag?

For VC4, I have to clamp the indirect load addresses anyway, so I don't
need this (though avoiding a load is probably worth it anyway!).  For
V3D I'm not clamping, so it would also want this I think.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180628/ad192231/attachment.sig>