libera/#clasp - IRC Chatlog
Search
13:59:39
drmeister
That presentation has taught me more about automatic differentiation than any others I've seen.
14:00:38
drmeister
I want to incorporate this into Cando. I want to generate native code on the fly that generates first and second partial derivatives of energy functions with vector inputs.
14:01:05
drmeister
I'd like that code to use vector instructions as much as possible and be optimized for whatever processor we find ourselves on.
14:32:02
bike
going through llvm with enzyme would probably be our best bet for processor specific optimizations
14:33:43
drmeister
Our energy functions have constant parameters and variables that we want to calculate partial derivatives on.
14:37:37
bike
we want to differentiate with respect to constants? i think enzyme want everything it differentiates with respect to to be a parameter
14:38:35
bike
if it's a problem we could just rewrite the code to make constants into parameters probably
15:14:10
drmeister
We don't want to differentiate with respect to constants. I was worried that enzyme might differentiate against every argument - but I don't think it's a worry.
15:52:24
drmeister
Can we have enzyme create LLVM-IR occasionally for new math functions and ship that IR with clasp.
15:54:29
drmeister
That tutorial I posted above describes how they calculate the derivative of a multidimensional normalization function - that's similar to what I need.
15:56:09
bike
i guess it depends on how critical the derivatives are to cando. are they like, optional? right now they don't seem very optional