Iterate directly over uniforms when applying them

Iterate directly over uniforms when applying them

Directly iterate over uniforms, not over uniform locations, so we don't waste time looking at locations in the middle of arrays. Gets about 1 fps on a NaCl benchmark.


BUG=
TEST=webgl conformance tests

Committed: http://code.google.com/p/angleproject/source/detail?r=792

[nicolas, 2011-10-06 13:28:14]

So you're getting measurably better performance from applying the uniforms in a
different order? Are you certain the exact same set of uniforms get applied (no
less and no fewer)? It's been a while since I looked at uniforms...

FYI, we have a patch (currently in review) which eliminates many getDevice()
calls, which affects the performance of applying uniforms as well.

[John Bauman, 2011-10-06 13:30:36]

On 2011/10/06 13:28:14, nicolas wrote:
> So you're getting measurably better performance from applying the uniforms in
a
> different order? Are you certain the exact same set of uniforms get applied
(no
> less and no fewer)? It's been a while since I looked at uniforms...

It's the loop itself that's expensive. With big arrays of matrices or whatever,
we loop through elements 1-n, but don't actually do anything with them. It
surprising how much CPU that takes.

> 
> FYI, we have a patch (currently in review) which eliminates many getDevice()
> calls, which affects the performance of applying uniforms as well.
Oh, good to hear.

[nicolas, 2011-10-06 15:40:59]

Ok, I've verified that it's functionally equivalent. The patch makes things more
elegant and faster at the same time so looks great to me!

[dgkoch, 2011-10-07 04:27:59]

One minor issue then LGTM.

http://codereview.appspot.com/5210042/diff/1/src/libGLESv2/Program.cpp
File src/libGLESv2/Program.cpp (right):

http://codereview.appspot.com/5210042/diff/1/src/libGLESv2/Program.cpp#newcod...
src/libGLESv2/Program.cpp:923: unsigned int numUniforms = mUniformIndex.size();
I believe numUniforms is unused now?

[John Bauman, 2011-10-07 15:43:34]

On 2011/10/07 04:27:59, dgkoch wrote:
> One minor issue then LGTM.
> 
> http://codereview.appspot.com/5210042/diff/1/src/libGLESv2/Program.cpp
> File src/libGLESv2/Program.cpp (right):
> 
>
http://codereview.appspot.com/5210042/diff/1/src/libGLESv2/Program.cpp#newcod...
> src/libGLESv2/Program.cpp:923: unsigned int numUniforms =
mUniformIndex.size();
> I believe numUniforms is unused now?
Oops, thought I already removed that. Fixed.