Chris Elrod

That's great, so we could replace one piece at a time.

> GraphBLAS is a sparse linear algebra library where the arithmetic semiring (+, *) is replaced with an arbitrary one like (max, +) for instance. Have you seen https://github.com/TensorBFS/TropicalGEMM.jl ?...

> No I had not, that is quite interesting. I would, however, like to take the alternate route of generalizing operations, rather than eltypes. It is much more flexible imho,...

With: https://github.com/JuliaSIMD/CPUSummary.jl/commit/d93cf1c1765c37c9fbe809b68a3e5f10fb6bb458 It should support dual sockets. However, Octavian also does not support more than 64 threads: https://github.com/JuliaLinearAlgebra/Octavian.jl/blob/ccd903373524827e92ddb4f68967529e4761626b/src/matmul.jl#L378 This will have to be replaced with the more usual `PolyesterWeave.request_threads`, which...

Yeah, from the [model zoo](https://github.com/FluxML/model-zoo/blob/d074b649f6b8a98f0c15afd866368d33d3fd0145/vision/conv_mnist/conv_mnist.jl#L18-L34): ```julia # LeNet5 "constructor". # The model can be adapted to any image size # and any number of output classes. function LeNet5(; imgsize=(28,28,1), nclasses=10)...

Yes, in the example if `f === tanh` then `g(x) = x * muladd(-x, x, 1)`. Although of course `f` and `g` would be arbitrary (so long as they're defined...

Yes and yes. The goal would be to just launch threads once per pass (i.e., once for forward, and once for back). This would reduce threading overhead, and also let...

I'm also working on multithreading in PaddedMatrices, so I would be inclined to add that and Tullio.jl as well. The machines I have access to, in order of speed, are:...

So all I have without AVX512 is the Haswell laptop. I was experimenting with threading in PaddedMatrices. Unfortunately, the threading overhead -- at least of my implementation -- is extremely...

I have code [here](https://github.com/chriselrod/PaddedMatrices.jl/blob/master/benchmark/blasbench.jl) in PaddedMatrices. It's descended from something @MasonProtter wrote, the `(StructVector ∘ map)(sizes) do sz` at least is from his original script.