[Piglit] [PATCH v3 2/3] arb_shader_precision: add framework for calculating tolerances for complex functions

[Micah Fedke]

I dunno, we might be alone on this gang plank, matey ;)

I can take a swing at your interval arithmetic approach and see if it's 
workable - I've got a few cycles and I'd like to see this through.


-mf

On 03/02/2015 04:29 PM, Ilia Mirkin wrote:
> On Mon, Mar 2, 2015 at 5:14 PM, Micah Fedke <micah.fedke at collabora.co.uk> wrote:
>> In my approach, the error discovered by comparing the ~infinite precision
>> result to the finite precision result *is* the allowable error range.  It's
>> like saying "This function diverged from the true result by amount x when
>> simulated at finite precision on the CPU, so we need to give it x amount of
>> leeway when it is run on the GPU."  The finite precision result should be a
>> complete representation of how error truly propagated through the
>> intermediate stages of the equation, for the given inputs.
>
> As seen by the CPU. TBH, I have no clue what the precision guarantees
> are for the various functions are on CPUs, but it seems unlikely to be
> identical to the ones in ARB_shader_precision. I sort of assume that
> all the simpler ops are supposed to be within 1 ULP on the CPU, but...
> who knows. (x87 had a few colorful opcodes which I'm sure introduce
> tons of error... some taking over 1K cycles on the original 8087,
> which wasn't exactly a speed demon to begin with.) It seems like
> you're computing error due to fp32 representation limitations rather
> than due to calculation imprecision.
>
> I understand that it's annoying to have spent a bunch of time on
> something that seems so trivial, but having failing tests due to the
> test being wrong is _quite_ annoying (and I say this having debugged a
> few such tests). I may give the approach I suggested a shot shortly,
> at least for a few ops, to see what the code looks like. Note that I
> have about 30 side-projects, and this would join that list... so...
> don't want to set the wrong expectations :)
>
> FWIW I never actually ran the results of your tests on my nvc0,
> perhaps I should do that. If it passes, then maybe what you have in
> practice is good enough, despite not being what the spec wants. Would
> love to hear opinions from others as well.
>
>    -ilia
>

-- 

Micah Fedke
Collabora Ltd.
+44 1223 362967
https://www.collabora.com/
https://twitter.com/collaboraltd