[Mesa-dev] [PATCH 1/2] nir: add a compiler option for disabling float comparison simplifications

[Ian Romanick]

On 11/30/2018 01:29 PM, Jason Ekstrand wrote:
> On Fri, Nov 30, 2018 at 3:18 PM Ian Romanick <idr at freedesktop.org
> <mailto:idr at freedesktop.org>> wrote:
> 
>     On 11/29/2018 07:47 AM, Connor Abbott wrote:
>     > On Thu, Nov 29, 2018 at 4:22 PM Jason Ekstrand
>     <jason at jlekstrand.net <mailto:jason at jlekstrand.net>> wrote:
>     >>
>     >> Can you provide some context for this?  Those rules are already
>     flagged "inexact" (that's what the ~ means) so they won't apply to
>     anything that's "precise" or "invariant".
>     >
>     > I think the concern is that this isn't allowed in SPIR-V, even without
>     > exact or invariant. We even go out of our way to do the correct thing
>     > in the frontend by inserting an "&& a == a" or "|| a != a", but then
> 
>     If you're that paranoid about it, why not just mark the operations are
>     precise?  That's literally why it exists.
> 
>     > opt_algebraic removes it with another rule and then this rule can flip
>     > it from ordered to unordered. The spec says that operations don't have
>     > to produce NaN, but it doesn't say anything on comparisons other than
>     > the generic "everything must follow IEEE rules" and an entry in the
>     > table that says "produces correct results." Then again, I can't find
>     > anything in GLSL allowing these transforms either, so maybe we just
>     > need to get rid of them.
> 
>     What I hear you saying is, "The behavior isn't defined."  Unless you can
>     point to a CTS test or an application that has incorrect behavior, I'm
>     going to oppose removing this pretty strongly.  *Every* GLSL compiler
>     does this.
> 
> 
> The test case came from VKD3D which does D3D12 on Vulkan.  Someone
> (Samuel, maybe?) was going to ask around and see if we can figure out
> what D3D12's rules are.  It's possible that it requires IEEE or
> something close.  If that's the case, as I said to Samuel on IRC, we're
> probably looking at an extension.  I don't think we want a flag like
> this that's set per-API.

Why isn't it sufficient to mark the operation as precise?  Was that on
IRC, and I missed it?

It is also possible that we could improve the handling of 'if
(!condition)' in the backend to make these transformations less
important.  I have some patches for that somewhere.  They didn't really
help with these transformations in place, and I never measured the
result without these transformations.