mlochbaum / rhsort Goto Github PK

I was checking ./runwolfbench 512 0 0 1 and it crashed on rhsort. Offending code:

      for (; i<n-4; i+=4) { x[i+4] += x[i+3] += x[i+2] += x[i+1] += x[i]; }
      for (; i<n-1; i++) { x[i+1] += x[i]; }

Seemingly working fix:

      for (; i+4<n; i+=4) { x[i+4] += x[i+3] += x[i+2] += x[i+1] += x[i]; }
      for (; i+1<n; i++) { x[i+1] += x[i]; }

Btw, have you had any luck improving performance on the 1-31 range?

@jariji shared the existence of rhsort with me here LilithHafner/InterLanguageSortingComparisons#10. I ran the benchmarks in gcc bench.c -o bench && ./bench and what I think are equivalent benchmarks in Julia. I found that Julia's default sorting is 2x faster than rhsort in that particular case. If you're willing I'd appreciate your confirmation to refutation of those results. I benchmarked Julia's sorting with this code snippet that can be pasted into the Julia REPL:

using BenchmarkTools, Random, Printf
for n in 10 .^ (3:6)
    v = rand(Int32, n)
    result = @benchmark sort($v) setup=(rand!($v)) evals=1 gctrial=false samples=3e7/n
    @printf "Testing size %8i: best: %6.3f avg: %6.3f ns/v\n" n minimum(result.times)/n mean(result.times)/n
end