Z-Buffer reads completely unoptimized

[aargueta]

Possible Optimizations

• [ ] Write/read in memory in 4x4 blocks instead of linearly
• [ ] Have buffer for these blocks
• [ ] Pre-calculate Z-buffer address to allow for more time to fetch from memory

[aargueta]

Not worth it at this point, to complex and it's unclear where the optimization needs to happen. Will revisit once Rasterizer Good milestone is reached.

[aargueta]

Intermediate idea, simpler cache that will still have some benefits at the cost of high latency spikes:

• Large cache
  • 4x36k bRAM = 96x96 z-buffer values
• Cache block replaced all at once
  • Avoids keeping track of many inflight requests to avoid duplicates
• Updates from pipeline written in cache and around cache
  • No need for eviction writes, memory always coherent
• Updates from memory put lock on cache, stalling pipeline
  • Avoids potentially nasty coherence issues
  • Avoids complex logic for arbitrating write ports
• Caching a block causes a lock to be placed on it
  • Avoids coherence issues
  • If fragments properly sorted from triangle pipeline, it shouldn't create stalls with independent fragment shaders waiting on each other to unlock