The Results Of Optimizing Radeon's VRAM Behavior

At the beginning of the year Lauri Kasanen had issued a call for help via Phoronix for collecting Radeon video memory statistics as he was looking to optimize Radeon's vRAM handling for his university thesis. The code and thesis are now available.

The current draft of Lauri's research on the Radeon vRAM behavior can be found via this PDF. Lauri's findings and work basically came down to no change as expected when there's plenty of video memory, under video RAM pressure he's scored a 1~2% frame-rate improvement and small to medium smoothness improvements, and there's some trade-offs of Urban Terror for instance showing 10% lower FPS while 30% better peaks.

For those wishing to test out the new Radeon code, Lauri's repositories for this work are this kernel repository and this Mesa repository. He said in an email this morning to me, "The code won't be changing beyond cleanups, there might be small edits to the thesis draft. As the main target was VRAM pressure, it will be pointless to test ioq3 games on 2 GB VRAM, for example - they will show no difference, as they fit completely into VRAM. You can use the radeon.vramlimit=256 kernel parameter to limit VRAM for testing different amounts. The kernel is fully backwards compatible with old mesa, so you should be able to compare just by changing mesa and the vram limit. I should note that there's a big ioq3 regression currently in mesa git[3], so if your comparison mesa is too far back, it could seem like it was caused by my work, when it's in reality in master too."