merayen
commented
4 years ago - [ ] Output C-code (?)
- C-code?
- Probably easy to write
- Probably well optimized
- LLVM IR?
- I know nothing about it
- Too much work to write nodes due to being low-level?
- May compile faster than C due to less overhead
- We get to learn about LLVM, deeply!
- Can we still run on GPUs
- [ ] Have LLVM JIT create binary
- How is it to compile single group-nodes, not needing to JIT everything again upon node changes?
- [ ] Have Java run the binary
- [ ] Re-implement the nodes in this backend
- Fixed buffer widths? Loops could have fixed iteration count, easily SIMD-unrolling too, if that is needed
- How do we solve "tight" feedbacks?
- Automatic detection on how big each group-node buffer should be?
- How should we work with delay compensation? Like if any node has an offset at e.g +100 samples, meaning it requests 100 samples in the future
- Implement offset on each node, telling how many samples it needs in the future vs its outputs
- Need to calculate all these upon compilation, I guess
- Output 1 mega method for each group node?
- [ ] Test threading?
- Device drivers, should they be built by nodes, only?
- Reinventing programming language? Maybe not... Hmm