A64FX vs Apple M1

a64fx

julia> versioninfo()
Julia Version 1.10.0-DEV.635
Commit 12d329b6e3d (2023-02-17 20:01 UTC)
Platform Info:
  OS: Linux (aarch64-linux-gnu)
  CPU: 50 × unknown
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, a64fx)
  Threads: 1 on 50 virtual cores

julia> using BenchmarkTools

julia> v = randn(Float64, 2 ^ 20);

julia> function sum_no_simd(v)
           sum = zero(eltype(v))
           for x in v
               sum += x
           end
           return sum
       end;

julia> function sum_simd(v)
           sum = zero(eltype(v))
           @simd for x in v
               sum += x
           end
           return sum
       end;

julia> run(`fcc -shared -fPIC -O3 -o libsum_no_simd.so sum.c`);

julia> run(`fcc -shared -fPIC -O3 -Keval,fp_contract -KSVE -o libsum_simd.so sum.c`);

julia> sum_no_simd_c(v::Vector{Float64}) = @ccall "./libsum_no_simd.so".sum(v::Ptr{Cdouble}, length(v)::Csize_t)::Cdouble;

julia> sum_simd_c(v::Vector{Float64}) = @ccall "./libsum_simd.so".sum(v::Ptr{Cdouble}, length(v)::Csize_t)::Cdouble;

julia> @btime sum_no_simd($v);
  4.738 ms (0 allocations: 0 bytes)

julia> @btime sum_no_simd_c($v);
  4.725 ms (0 allocations: 0 bytes)

julia> @btime sum_simd($v);
  180.370 μs (0 allocations: 0 bytes)

julia> @btime sum_simd_c($v);
  183.500 μs (0 allocations: 0 bytes)

m1

julia> versioninfo()
Julia Version 1.10.0-DEV.635
Commit 12d329b6e3 (2023-02-17 20:01 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin21.6.0)
  CPU: 8 × Apple M1
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, apple-m1)
  Threads: 1 on 4 virtual cores

julia> using BenchmarkTools

julia> v = randn(Float64, 2 ^ 20);

julia> function sum_no_simd(v)
           sum = zero(eltype(v))
           for x in v
               sum += x
           end
           return sum
       end;

julia> function sum_simd(v)
           sum = zero(eltype(v))
           @simd for x in v
               sum += x
           end
           return sum
       end;

julia> @btime sum_no_simd($v);
  981.750 μs (0 allocations: 0 bytes)

julia> @btime sum_simd($v);
  246.291 μs (0 allocations: 0 bytes)

sum.c

#include <stddef.h>
double sum(double *v, size_t n) {
    double sum = 0.0;
    for(size_t i = 0; i < n; i++) {
        sum += v[i];
    }
    return sum;
}