4 ms·

Hey cool, someone pointed me here -- surprised to find my toy side-project on HN! I'm not actively hacking on autopiper at the moment (busy with other things) b

by cfallin 11y ago

Hey cool, someone pointed me here -- surprised to find my toy side-project on HN! I'm not actively hacking on autopiper at the moment (busy with other things) but I still believe the semantics/ideas are sound and I'd love to find more time for it someday. In its current state it's more of a proof-of-concept. I think it might do better as a DSL that lowers to the pipeline representation directly (embedded in, say, Haskell or Scala), rather than inferring intent from a CFG -- benefit of hindsight...

FullyFunctional 11y ago
This is very cool. I have been/am toying with very same ideas, but using a Lava variant, all in the context of Reduceron. I'm also planning to redo my RISC-V implementation as a working example. Silvia Muller in "Computer Architecture: Complexity and Correctness" describes work on auto-pipelining. Are you basing your work on this or something else?
- cfallin 11y ago
  Interesting! I saw Lava briefly and understood it as basically a DSL to describe logic; additional abstractions to build pipelines on top of that would be very cool. Looking forward to seeing what comes out of your work! And yup, there's been a lot of past work on auto-pipelining, but my understanding of what's been done before is that it handles splitting logic into stages, but doesn't necessarily encapsulate more complex patterns -- e.g. no one has built a tool that understands stall logic or bypass networks or clears/restarts in a pipeline and can synthesize those patterns flexibly. So you could see autopiper's main contribution as the "lowering" of a conventional SSA-based IR into pipelined logic, along with the primitives necessary to describe the actual semantic intent behind design patterns (ask/provide for bypass, killyounger and CFG backedges for pipeline restarts), and the insight that macros can use those primitives to build things like speculation. If you know of any other past work I'd be really interested to see it!

by cfallin 11y ago

FullyFunctional 11y ago

This is very cool. I have been/am toying with very same ideas, but using a Lava variant, all in the context of Reduceron. I'm also planning to redo my RISC-V implementation as a working example. Silvia Muller in "Computer Architecture: Complexity and Correctness" describes work on auto-pipelining. Are you basing your work on this or something else?

cfallin 11y ago
Interesting! I saw Lava briefly and understood it as basically a DSL to describe logic; additional abstractions to build pipelines on top of that would be very cool. Looking forward to seeing what comes out of your work! And yup, there's been a lot of past work on auto-pipelining, but my understanding of what's been done before is that it handles splitting logic into stages, but doesn't necessarily encapsulate more complex patterns -- e.g. no one has built a tool that understands stall logic or bypass networks or clears/restarts in a pipeline and can synthesize those patterns flexibly. So you could see autopiper's main contribution as the "lowering" of a conventional SSA-based IR into pipelined logic, along with the primitives necessary to describe the actual semantic intent behind design patterns (ask/provide for bypass, killyounger and CFG backedges for pipeline restarts), and the insight that macros can use those primitives to build things like speculation. If you know of any other past work I'd be really interested to see it!