Tilemap optimisations #626
Conversation
Doesn't have a noticable effect, but is less instructions. (Maybe ~0.1ms saved)
Convert to using fixed-point and skip/combine pixels where possible Saves ~6ms in blit kart
Saves ~2ms in blit kart
32blit/graphics/tilemap.cpp
Outdated
| wc += dwc; | ||
| doff++; | ||
| c--; | ||
| } while(c && (wc.x >> fix_shift + 3) == wcx >> 3 && (wc.y >> fix_shift + 3) == wcy >> 3); |
There was a problem hiding this comment.
Whoops, warning on every compiler...
|
Hmm, very slightly worse for drawing a non-scaled fullscreen hires tilemap (< a millisecond). Though that's already a suboptimal case... (looking up the tile for every pixel when it isn't going to change for 8 most of the time) More numbers:
|
|
Thank you! I think the benefits to mode7 far outweigh any minor drawbacks, though I notice this still has optimised cases for fixed-size spritesheets and it would be interesting to see what happens if arbitrary limits like those are lifted. I think some chatter on Discord suggested this change gives enough headroom for that not to be a problem? I've a strong preference for simplicity over optimisation but if we're not hitting draw times that get us comfortably within 50fps.... well it doesn't matter how easy to use something is if it tanks your framerate. The lack of FPS counter on Tilemap Test/Demo is conspicuous when testing this change 😆 but I can't see any perceptible drawbacks in our - granted, very basic - examples. Nice work! |
|
At least for the mode7 case, removing the "power of two bounds" and "16 column spritesheet" restrictions cost ~.2ms each. I think it had a much larger impact on the non-scaled case though. (Since it's then affecting every pixel). (I kind of like the 16 col restriction though as it makes the max-size spritesheet square) |
Reduces "Super Blit Kart"
rendertimes by 8ms and makes the track rendering > 2.5x faster. First commit also means you can put something in the tile index you have defined as empty for other uses.(I also have some patches for sprite stretch blits, but I think those need more work/testing)