Oscar Smith
Oscar Smith
no reason not to.
Putting the code in a function means that you get type stability dropping the time down to 145 ms on my computer.
it would be good to add some cpu and gpu benchmarks for a simple broadcast and a matmul/convolution compared to alternatives.
both are around 100x slower than float64. `exp` in particular should be able to be done relatively quickly since `exp(hi+lo)=exp(hi)exp(lo)`, and since `hi` can easily be reduced to [1,2), `exp(lo)`...
Part 1 of 2 solution to https://github.com/JuliaMath/DoubleFloats.jl/issues/135. This still needs some testing, but I think it is pretty good as is. I benchmark it to be 10x faster and more...
https://github.com/JuliaMath/Bessels.jl has pure Julia bessel functions for real arguments and orders that are frequently significantly faster than the ones currently used (provided by AMOS). The package was just registered, but...
New paper that allegedly gives global newton convergence.
based on discussion https://github.com/JuliaDiff/Diffractor.jl/pull/54 and in slack. This is an initial implementation of forward and reverse chunked mode AD.
This is a general issue, but for a specific incarnation, https://github.com/JuliaDiff/ChainRules.jl/blob/8073c7c4638bdd46f4e822d2ab72423c051c5e4b/src/rulesets/Base/array.jl#L40 ``` function rrule(::typeof(Base.vect), X::Vararg{T, N}) where {T, N} vect_pullback(ȳ) = (NoTangent(), NTuple{N}(ȳ)...) return Base.vect(X...), vect_pullback end ``` This rule...
useful for tests. I'm not sure whether this is a good idea, but it seems reasonable enough. @Keno