Valentin Churavy comments

Results 1413 comments of


                                            Valentin Churavy

feat: VJP utility based on `autodiff_thunk`

Only when looking at the examples I feel like we should fuse `Active, Seed(dresult),` maybe `Seed{Active}(dresult)` and then we could add `Seedx(::Float64) = Seed{Active}(x)`

feat: VJP utility based on `autodiff_thunk`

Sorry was on vacation. > automatic activity detection for the seed? Yes I think so. At least it is consistent with current activity deduction for return, but allows specification of...

feat: VJP utility based on `autodiff_thunk`

Ok last thing would be to bump the version in EnzymeCore since we are exporting new types.

uniform memory and `@synchronize` inside an if statement

> My if-statement is for filtering out invalid indices, Please take a look at https://juliagpu.github.io/KernelAbstractions.jl/stable/#0.9.34 if you are manually filtering out indices `unsafe_indicies=true` is likely what you want, but you...

Computing third order derivatives

This is likely the same as https://github.com/EnzymeAD/Enzyme.jl/issues/1173, the goal is that https://github.com/JuliaGPU/GPUCompiler.jl/pull/599 will fix this.

Computing third order derivatives

Finally fixed by https://github.com/EnzymeAD/Enzyme.jl/pull/2161

KernelAbstractions version of GPUArrays

Congratulations @leios !

Introduce `AsyncNumber` to lazily copy numeric `mapreduce` results to the host

> So it does look like these gaps are GC pauses. Can you set `--gcthreads` to be 4 or so?

Atomics: configurable scope (for multi-device unified memory)

Yeah https://github.com/JuliaGPU/CUDA.jl/pull/1790 and https://github.com/JuliaGPU/CUDA.jl/pull/1644 used inline assembly for the scoped operations.

use ScopedValues for Tracy Context

We could support 1.10 with ScopedValues.jl, but that lacks the optimizations mentioned in #35