Function passed as an argument of a function is slow to broadcast()

[wsshin]

For example, execution of apply_f(f, A) = broadcast(f, A) is much slower than direct direct execution of broadcast(f, A):

julia> apply_f(f, A) = broadcast(f, A);

julia> f(x) = 2x;

julia> A = rand(10,10);

julia> using BenchmarkTools

julia> @benchmark apply_f($f, $A)
BenchmarkTools.Trial:
  memory estimate:  1.97 kb
  allocs estimate:  25
  median time:      18.385 μs (0.00% GC)

julia> @benchmark broadcast($f, $A)
BenchmarkTools.Trial:
  memory estimate:  896.00 bytes
  allocs estimate:  1
  median time:      118.314 ns (0.00% GC)

Here is the version info:

julia> versioninfo()
Julia Version 0.6.0-dev.2338
Commit 6f9fe0bd1 (2017-01-25 01:55 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin16.4.0)
  CPU: Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.9.1 (ORCJIT, skylake)

[KristofferC]

Sometimes you need to force specialization to take place:

julia> apply_f2{F}(f::F, A) = broadcast(f, A);

julia> @benchmark apply_f2($f, $A)
BenchmarkTools.Trial: 
  memory estimate:  896.00 bytes
  allocs estimate:  1
  --------------
  minimum time:     95.070 ns (0.00% GC)
  median time:      112.688 ns (0.00% GC)
  mean time:        132.656 ns (12.22% GC)
  maximum time:     845.149 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     957
  time tolerance:   5.00%
  memory tolerance: 1.00%

[Sacha0]

Ah, good catch @KristofferC! This is #19137. Best!

[wsshin]

Thanks! The explanation in #19137 is clear.

However, there is still a problem. When the dot syntax is used instead of broadcast(), execution remains slow even if specialization is forced:

julia> apply_fdot(f, A) = f.(A)
apply_fdot (generic function with 1 method)

julia> apply_fdot2{F}(f::F, A) = f.(A)
apply_fdot2 (generic function with 1 method)

julia> @benchmark apply_fdot(f, A)
BenchmarkTools.Trial:
  memory estimate:  1.19 kb
  allocs estimate:  25
  median time:      10.514 μs (0.00% GC)

julia> @benchmark apply_fdot2(f, A)
BenchmarkTools.Trial:
  memory estimate:  1.19 kb
  allocs estimate:  25
  median time:      11.395 μs (0.00% GC)

Maybe this has been discussed already as well?

[Sacha0]

Interpolating f and A into the benchmarks eliminates the discrepancy:

julia> using BenchmarkTools

julia> f(x) = 2x;

julia> A = rand(10,10);

julia> apply_fdot(f, A) = f.(A)
apply_fdot (generic function with 1 method)

julia> apply_fdot2{F}(f::F, A) = f.(A)
apply_fdot2 (generic function with 1 method)

julia> apply_fdot3{F}(f::F, A) = broadcast(f, A)
apply_fdot3 (generic function with 1 method)

julia> @benchmark apply_fdot($f, $A)
BenchmarkTools.Trial:
  samples:          10000
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  1.97 kb
  allocs estimate:  25
  minimum time:     21.51 μs (0.00% GC)
  median time:      22.23 μs (0.00% GC)
  mean time:        23.46 μs (0.00% GC)
  maximum time:     129.07 μs (0.00% GC)

julia> @benchmark apply_fdot2(f, A)
BenchmarkTools.Trial:
  samples:          10000
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  1.97 kb
  allocs estimate:  25
  minimum time:     22.18 μs (0.00% GC)
  median time:      23.02 μs (0.00% GC)
  mean time:        24.76 μs (0.00% GC)
  maximum time:     118.79 μs (0.00% GC)

julia> @benchmark apply_fdot2($f, $A)
BenchmarkTools.Trial:
  samples:          10000
  evals/sample:     886
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  896.00 bytes
  allocs estimate:  1
  minimum time:     130.00 ns (0.00% GC)
  median time:      135.00 ns (0.00% GC)
  mean time:        154.72 ns (8.82% GC)
  maximum time:     1.43 μs (84.16% GC)

julia> @benchmark apply_fdot3($f, $A)
BenchmarkTools.Trial:
  samples:          10000
  evals/sample:     888
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  896.00 bytes
  allocs estimate:  1
  minimum time:     128.00 ns (0.00% GC)
  median time:      133.00 ns (0.00% GC)
  mean time:        153.16 ns (8.80% GC)
  maximum time:     1.40 μs (85.15% GC)

julia> versioninfo()
Julia Version 0.6.0-dev.2374
Commit 876549f (2017-01-25 22:41 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin15.6.0)
  CPU: Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Sandybridge)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.9.1 (ORCJIT, ivybridge)

Best!

[wsshin]

Of course. I hate myself for making this same mistake again!