Unrolled loop results in Improbable @benchmark timing

General Usage
, ,
2

The last timing is probably not accurate. The compiler is really good at propagating things lately, which make proper benchmarks incredibly hard. Your case is solvable by letting the @benchmark macro initialise the input locally. This avoids the lookup time for global variables and doesn’t get constant-propped away (at least not on my machine):

julia> @benchmark co(A, B, UInt64(1), UInt32(0), UInt32(0)) setup = begin A = SVector{16, UInt32}([rand(UInt32) for x in 1:16]...); B = SVector{16, UInt32}([rand(UInt32) for x in 1:16]...) end
BenchmarkTools.Trial:
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     54.813 ns (0.00% GC)
  median time:      54.914 ns (0.00% GC)
  mean time:        56.826 ns (0.00% GC)
  maximum time:     165.957 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     987

julia> @benchmark co2(A, B, UInt64(1), UInt32(0), UInt32(0)) setup = begin A = SVector{16, UInt32}([rand(UInt32) for x in 1:16]...); B = SVector{16, UInt32}([rand(UInt32) for x in 1:16]...) end
BenchmarkTools.Trial:
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     51.064 ns (0.00% GC)
  median time:      51.165 ns (0.00% GC)
  mean time:        52.313 ns (0.00% GC)
  maximum time:     140.932 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     987