mirror of
https://github.com/stevenrobertson/cuburn.git
synced 2025-02-05 11:40:04 -05:00
45 lines
1.7 KiB
Plaintext
45 lines
1.7 KiB
Plaintext
|
Status: currently broken (syntax errors, incomplete sections)
|
||
|
|
|
||
|
|
Current goals:
|
||
|
|
- Test DeviceStream, and get it working. Bugs are expected.
|
||
|
|
- Test allocator
|
||
|
|
- Test statement evaluator
|
||
|
|
- Test packing correctly
|
||
|
|
- Test that device instructions get injected correctly
|
||
|
|
- Test in working implementation
|
||
|
|
- Load a set of genomes and calculate a bare minimum `Feature` set (no xforms,
|
||
|
|
no filters, no oversample)
|
||
|
|
- Get frames loaded for rendering
|
||
|
|
- Get IterThread running in device kernel
|
||
|
|
- For now, implement as `PTXTest`
|
||
|
|
- For each frame, loop for FUSE times, then loop through expected number of
|
||
|
|
points for each CP. Keep a count of number of times looped, and number of
|
||
|
|
stores that would be done. Verify against expected counts.
|
||
|
|
|
||
|
|
|
||
|
|
Things to do (rather severely incomplete):
|
||
|
|
- LaunchContext thread distribution based on generated code register count and
|
||
|
|
shared memory size
|
||
|
|
- qlocal storage
|
||
|
|
- Performance implications of different state spaces
|
||
|
|
- Shared / cache projected usage and its effect on above
|
||
|
|
- Implement qlocal storage, and hide the complexity
|
||
|
|
- The `Feature` class
|
||
|
|
- Transform count and per-transform code layout
|
||
|
|
- Filter size, oversample, final buffer size
|
||
|
|
- Palette storage
|
||
|
|
- Performance implications of different state spaces
|
||
|
|
- Performance and quality of 2D texture interpolation
|
||
|
|
- Buffer allocation, clearing, reading from device
|
||
|
|
- Preview window
|
||
|
|
- When/how to sample?
|
||
|
|
- OpenGL interop worth it?
|
||
|
|
- Implement
|
||
|
|
- Implement xforms
|
||
|
|
- Shuffle
|
||
|
|
- State space implications, you know the drill
|
||
|
|
- Implement
|
||
|
|
- Test effects on quality by masking off writes on all but one lane and
|
||
|
|
boosting the sample density to compensate (muuuuuch later on)
|
||
|
|
|