Reduce the minimum scratch size to 16, don't look for next pow2 larger sizes.

Reduce the minimum scratch size to 16, don't look for next pow2 larger sizes.
Committed: https://code.google.com/p/skia/source/detail?r=8007

[bsalomon, 2013-03-06 16:09:34]

Guillem Gangolells in IMG DevTech showed that smaller scratch sizes give
better/same performance on a wide variety of mobile GPUs. I don't think there
was a strong rationale for the current code. I'll watch the perf bots to see if
there are any unexpected regressions.

[robertphillips, 2013-03-06 16:17:37]

lgtm - it might be useful to paste the performance numbers here.

[bsalomon, 2013-03-06 16:35:51]

Committed patchset #1 manually as r8007 (presubmit successful).

[bsalomon, 2013-03-06 16:49:25]

On 2013/03/06 16:17:37, robertphillips wrote:
> lgtm - it might be useful to paste the performance numbers here.

Here is the part of the report for two canvas2d tests on Android:

Change 1 - Lowering the scratchpad floor down to 16:
From MIN_SIZE = 256 to MIN_SIZE = 16
Change 2 - Preventing textures to double the size while looking for a cached
scratch:
Comment the lines: //desc.fWidth *= 2; and //desc.fHeight *= 2;
Some numbers from two benchmarks (Cvperf and Asteroids) after the changes:

Galaxy Nexus (PowerVR SGX540):
ContentShell_vanilla.apk:
Cvperf: ~0.348
Asteroids bench: ~163

ContentShell_16x16.apk (Change 1):
Cvperf: ~0.382
Asteroids bench: ~228

ContentShell_16x16_and_no_doubling.apk (Change 1 + 2):
Cvperf: ~0.425
Asteroids bench: ~292
Nexus 4 (Ardeno 320):
ContentShell_vanilla.apk:
Cvperf: ~1.06
Asteroids bench: ~420

ContentShell_16x16_and_no_doubling.apk (Change 1 + 2):
Cvperf: ~1.18
Asteroids bench: ~512