Michael Abbott

I have never used pyJulia, but did briefly manage to get this error directly from the repl: ``` julia> TensorCast.transmute(rand(2,2), Val((2, 1))) ERROR: MethodError: no method matching _trex(::Symbol, ::Type{Matrix{Float64}}, ::Tuple{Int64,...

Depends on the expression, for both. Any `transmute(M, ::Val)` call is generated, I think, and `@pretty` will show you what the macro calls.

A better test, timing both package loading and time to run the macro. On master, Julia 1.5, best of a few: ``` EscBook:TensorCast me$ time julia -e '@time (using TensorCast;...

I think this is similar to #25, which is https://github.com/JuliaGPU/CUDA.jl/issues/228. When the broadcast has one naked CuArray, and others reshaped / permuted, then it hits CUDA's broadcasting. But when all...

On the CPU, `orient(...)` usually reshapes, but sometimes makes a copy, since things like `ReshapedArray(Transpose(...))` are very slow. This was disabled on the GPU in #10, and one possibility would...

I haven't tried it out, it means diving deep into the belly of this package's logic, as it still needs to call `reshape` in other contexts. And TransmuteDims needs a...

Thanks for testing. I think these might need `CUDA.@sync` in there, and a few more `$` signs? But if these are accurate, then I'm a little surprised that "trick" costs...

Oh that's embarrassing, my example of `@matmul` seems to hit a bug. And ought not to work anyway, as it intends to only support single `*` calls, no batch indices,...

It might be worth opening an issue with OMEinsum.jl, I see Leo was working to update CUDA things recently. I'm out of ideas for now, it just seems a bit...

All the examples from the first message run without scalar indexing now, TensorCast v0.4.5, which uses TransmuteDims, #31. The expansions from [here](https://github.com/mcabbott/TensorCast.jl/issues/28#issuecomment-707961529) have become: ```julia julia> @pretty @reduce C[a,b,d] :=...