Theora On Nokia N900 with optimized ARM and DSP code

blog.mjg.im

35 pointsmbrubeck4 comments

Does anyone have a reference on the slow fragment shader they ran into? I've not done much with the OpenGL ES hardware out there, but even the lowest end desktop GPU can handle such things without blinking. I wonder if it's a driver issue more so than a hardware capability issue.

http://github.com/kinetiknz/plogg/blob/master/plogg.cpp#L495

The "obvious color conversion" shader is 18 cycles on SGX530, the "slightly optimized routine" shader is 20 cycles, and if you take the "slightly optimized" shader, remove the lowp and mediump qualifiers and stick a "precision highp float;" at the top of the shader, you get a 7 cycle shader, which by my calculations should be able to meet your fill rate goal.

Thanks for the suggestion. I made the suggested changes (check github for the latest) but it doesn't seem to be significantly faster than the other fragment shaders (within 0.1 FPS over a 150 second video). I'm at a bit of a loss as to how to investigate the performance/behaviour of the OpenGL pipeline on the N900.

How'd you come up with the cycle counts? I'm new to GLSL, so any documentation beyond the spec would be super useful to me.

Reply on news.ycombinator.com