[Mesa-dev] [PATCH 2/9] i965: Fix textureGrad with cube samplers

[Roland Scheidegger]

Am 05.05.2015 um 00:16 schrieb Matt Turner:
> On Tue, Feb 24, 2015 at 10:39 AM, Roland Scheidegger <sroland at vmware.com> wrote:
>> As a side note (not really directly related to the patch) I think this
>> could benefit from some optimization, unless there's some passes
>> somewhere which can already do this.
>>
>> The non-scalar (non-cube) calculation does this:
>> lod_info.lod = log2(max(sqrt(dot1), sqrt(dot2)))
>> The second sqrt can be trivially eliminated:
>> = log2(sqrt(max(dot1, dot2)))
> 
> I don't think that's valid, since sqrt(x) is greater than x for x < 1.0.
That relationship is true, but not relevant, since
when (x >= y)
-> sqrt(x) >= sqrt(y) still holds true (for non-negative numbers anyway)

Roland



> 
>> And furthermore the remaining sqrt can be killed off too thanks to
>> n*log2(x) = log2(x^n)
>> = 0.5*log2(max(dot1, dot2)))
>>
>> I don't really know anything about the hw, but I would think this should
>> be faster not only on the 1 mathbox-per-gpu ancient igps ;-).
> 
> FWIW, I implemented that optimization in NIR based on this email in
> 
> commit 099c729b4cb6863e8ddbdb9afe7fd7bd53c11ee1
> Author: Matt Turner <mattst88 at gmail.com>
> Date:   Thu Mar 26 10:07:58 2015 -0700
> 
>     nir: Add identities for the log function.
> 
> and it didn't find any instances of log(sqrt(x)) in our shader-db. :(
> 
> It did find some rcp(log(x)) though!
>