v0.3.0 To-do List

[xrq-phys]

• Waiting for more Arm64 works in the upstream (by myself, though).
• Introduce more tests (sizes, transpose, etc.)
• Unsure Option to not override the LinearAlgebra module.
• Mixed precision

[xrq-phys]

Just noticed that since 0.1.0, gemmt is introduced while upstream has ≥0.8.0 as requirement.
Hence I should tag the current HEAD as 0.2.0 and make this roadmap 0.3.0.

[xrq-phys]

And another thing:

• Overriding gemm! alone seems forwarding generic strides but not mixed precision. I'd be good to also override somehow mul!{A::Array{TA},B::Array{TB},C::Array{TC}}{TA≠TB≠TC} for mixed precision support.

cf. log:

# Generic striding is good:
julia> @benchmark view(A, 1:2:1000, :)*B setup=(A=rand(Float64, 1000, 1000); B=rand(Float64, 1000, 1000))
BenchmarkTools.Trial: 22 samples with 1 evaluation.
 Range (min … max):  228.803 ms … 233.134 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     229.311 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   229.954 ms ±   1.277 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

julia> using BLIS; @benchmark view(A, 1:2:1000, :)*B setup=(A=rand(Float64, 1000, 1000); B=rand(Float64, 1000, 1000))
BenchmarkTools.Trial: 213 samples with 1 evaluation.
 Range (min … max):  20.919 ms …  21.676 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     21.161 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   21.137 ms ± 109.760 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

# Mixed precision not reaching BLIS directly:
julia> A=rand(Float64, 1000, 1000); B=rand(Float32, 1000, 1000);
julia> using Cthulhu; @descend A*B
...
   %7  = invoke copyto_axcheck!(::Matrix{Float64},::Matrix{Float32})::Matrix{Float64}
   %8  = invoke gemm!(::Char,::Char,::Float64,::Matrix{Float64},::Matrix{Float64},::Float64,::Matrix{Float64})::Matrix{Float64}
 • ↩