Do our own fp16 conversion in ResampleEffect.
authorSteinar H. Gunderson <sgunderson@bigfoot.com>
Sun, 9 Mar 2014 17:22:18 +0000 (18:22 +0100)
committerSteinar H. Gunderson <sgunderson@bigfoot.com>
Sun, 9 Mar 2014 17:25:37 +0000 (18:25 +0100)
commitf62661c1bef3acac7b98631970700dfa5ac94768
treea446d1a22bbc9307293714b49120f617b73872e4
parentc2371eea5dadf97add7c1a71b8ea660973c9c6de
Do our own fp16 conversion in ResampleEffect.

This not only fixes issues with poor downconversion on ATI, but also
allows us to normalize while being aware of fp16 roundoff issues.
Seems to about cut the error in half in the HeavyResampleGetsSumRight
test, which as far as I can see would take us up to 10-bit accuracy.
resample_effect.cpp
resample_effect_test.cpp