simd.jl

len = 10000
function nosimd1(x,y,z,w)
  @inbounds for i in eachindex(x)
    x[i] += z[i]
    y[i] += w[i]
  end
end
function nosimd2(x,y,z,w)
  @inbounds for i in eachindex(x)
    x[i] += z[i]
  end
  @inbounds for i in eachindex(x)
    y[i] += w[i]
  end
end

function simd1(x,y,z,w)
  @inbounds @simd for i in eachindex(x)
    x[i] += z[i]
    y[i] += w[i]
  end
end
function simd2(x,y,z,w)
  @inbounds @simd for i in eachindex(x)
    x[i] += z[i]
  end
  @inbounds @simd for i in eachindex(x)
    y[i] += w[i]
  end
end

using BenchmarkTools
srand(100)
x = rand(len)
y = rand(len)
z = rand(len)
w = rand(len)
@benchmark nosimd1(x,y,z,w)

BenchmarkTools.Trial:
  memory estimate:  0.00 bytes
  allocs estimate:  0
  --------------
  minimum time:     7.909 μs (0.00% GC)
  median time:      7.997 μs (0.00% GC)
  mean time:        8.096 μs (0.00% GC)
  maximum time:     20.351 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     4
  time tolerance:   5.00%
  memory tolerance: 1.00%

srand(100)
x = rand(len)
y = rand(len)
z = rand(len)
w = rand(len)
@benchmark nosimd2(x,y,z,w)

BenchmarkTools.Trial:
  memory estimate:  0.00 bytes
  allocs estimate:  0
  --------------
  minimum time:     8.122 μs (0.00% GC)
  median time:      8.215 μs (0.00% GC)
  mean time:        8.353 μs (0.00% GC)
  maximum time:     17.972 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     3
  time tolerance:   5.00%
  memory tolerance: 1.00%

srand(100)
x = rand(len)
y = rand(len)
z = rand(len)
w = rand(len)
@benchmark simd1(x,y,z,w)

BenchmarkTools.Trial:
  memory estimate:  0.00 bytes
  allocs estimate:  0
  --------------
  minimum time:     8.838 μs (0.00% GC)
  median time:      8.993 μs (0.00% GC)
  mean time:        9.260 μs (0.00% GC)
  maximum time:     45.303 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     3
  time tolerance:   5.00%
  memory tolerance: 1.00%

srand(100)
x = rand(len)
y = rand(len)
z = rand(len)
w = rand(len)
@benchmark simd2(x,y,z,w)

BenchmarkTools.Trial:
  memory estimate:  0.00 bytes
  allocs estimate:  0
  --------------
  minimum time:     8.379 μs (0.00% GC)
  median time:      8.461 μs (0.00% GC)
  mean time:        8.585 μs (0.00% GC)
  maximum time:     18.038 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     3
  time tolerance:   5.00%
  memory tolerance: 1.00%

@inline function test_inbounds(x,y,z,w,i)
  x[i] += z[i]
  y[i] += w[i]
end

function nosimd3(x,y,z,w)
  @inbounds for i in eachindex(x)
    test_inbounds(x,y,z,w,i)
  end
end

srand(100)
x = rand(len)
y = rand(len)
z = rand(len)
w = rand(len)
@benchmark nosimd3(x,y,z,w)

BenchmarkTools.Trial:
  memory estimate:  0.00 bytes
  allocs estimate:  0
  --------------
  minimum time:     8.260 μs (0.00% GC)
  median time:      8.324 μs (0.00% GC)
  mean time:        8.446 μs (0.00% GC)
  maximum time:     19.281 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     3
  time tolerance:   5.00%
  memory tolerance: 1.00%


function nosimd4(x,y,z,w)
  for i in eachindex(x)
    test_inbounds(x,y,z,w,i)
  end
end

srand(100)
x = rand(len)
y = rand(len)
z = rand(len)
w = rand(len)
@benchmark nosimd4(x,y,z,w)

BenchmarkTools.Trial:
  memory estimate:  0.00 bytes
  allocs estimate:  0
  --------------
  minimum time:     18.800 μs (0.00% GC)
  median time:      18.992 μs (0.00% GC)
  mean time:        19.449 μs (0.00% GC)
  maximum time:     49.651 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%

function nosimd5(x,y,z,w)
  @inbounds @fastmath for i in eachindex(x)
    test_inbounds(x,y,z,w,i)
  end
end

srand(100)
x = rand(len)
y = rand(len)
z = rand(len)
w = rand(len)
@benchmark nosimd5(x,y,z,w)

BenchmarkTools.Trial:
  memory estimate:  0.00 bytes
  allocs estimate:  0
  --------------
  minimum time:     7.833 μs (0.00% GC)
  median time:      7.919 μs (0.00% GC)
  mean time:        7.997 μs (0.00% GC)
  maximum time:     15.480 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     4
  time tolerance:   5.00%
  memory tolerance: 1.00%


function simd3(x,y,z,w)
  @inbounds @fastmath for i in eachindex(x)
    x[i] += z[i]
  end
  @inbounds @fastmath for i in eachindex(x)
    y[i] += w[i]
  end
end

srand(100)
x = rand(len)
y = rand(len)
z = rand(len)
w = rand(len)
@benchmark simd3(x,y,z,w)

BenchmarkTools.Trial:
  memory estimate:  0.00 bytes
  allocs estimate:  0
  --------------
  minimum time:     9.557 μs (0.00% GC)
  median time:      10.067 μs (0.00% GC)
  mean time:        11.115 μs (0.00% GC)
  maximum time:     23.846 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     3
  time tolerance:   5.00%
  memory tolerance: 1.00%