# L-17 MCS 572 Fri 4 Oct 2024 : mmmulmeta.jl

# The CUDA version of the example is copied from the 
# Julia for High-Performance Scientific Computing web site.
# https://enccs.github.io/Julia-for-HPC/GPU
# and adjusted to use on M1 MacBook Air:
# (1) using Metal instead of using CUDA
# (2) work with Float32 instead of Float64
# (3) use MtlArray instead of CuArray

using Metal
using BenchmarkTools

dim = 2^10
A_h = rand(Float32, dim, dim);
A_d = MtlArray(A_h);

@btime $A_h * $A_h;
@btime $A_d * $A_d;