Broadcast the propagation_expr for vector mode AD
#93
Conversation
YingboMa
force-pushed
the
myb/scalar_rule
branch
from
fe8d1b7
to
65b2ce2
Compare
| # avoiding the extra `+` operation, it is potentially | ||
| # expensive for vector mode AD | ||
| sumed_∂_mul_Δs = if n∂s > 1 | ||
| :(@. +($(∂_mul_Δs...))) |
There was a problem hiding this comment.
Why use the macro?
.+($(...))
should be identical
There was a problem hiding this comment.
There was a problem hiding this comment.
And it is more readable.
There was a problem hiding this comment.
I find it less readable.
And this is kind of why -- it's not as obvious what is being broadcast.
E.g. This case where we are also broadcasting the stuff from the other line.
I'ld just make both explicit broadcasts.
--
I know the DiffEq code base uses this a ton.
I have never used it except during the transition from implict broadcast back in 0.5ish
There was a problem hiding this comment.
| :(@. +($(∂_mul_Δs...))) | |
| :(.+($(∂_mul_Δs...))) |
There was a problem hiding this comment.
Oh, now I remember. Using explicit broadcast broke tests on older versions of Julia.
There was a problem hiding this comment.
It looks right to me when I do macroexpand but indeed it doesn't work in 1.0 for me locally either.
There was a problem hiding this comment.
| :(@. +($(∂_mul_Δs...))) | |
| # we use `@.` to broadcast the . and + | |
| # Note: we don't want to do broadcasting to occur if only 1 multiply (no +), | |
| # because some arrays overload multiply with scalar, and those that don't fall back to broadcasting anyway. | |
| # Note also: due to bugs in Julia 1.0 can't use `.+` or `.*` inside expression literals anyway | |
| :(@. +($(∂_mul_Δs...))) |
This reverts commit b7d1da2.
…ChainRulesCore (#93) * =Remove uses of add and mul from tests because they have been removed from ChainRulesCore * remove ref to Cassette. Slightly improve talk of ChainRulesCore * Update getting_started.md * Update Project.toml
I also added
@muladdto optimizepropagation_expr. It rewritesa*b + c*dtomuladd(a, b, c*d).