Brian Chen (Jun 24 2022 at 23:21): Brian Chen (Jun 24 2022 at 23:21):Should I refrain from calling it in a hot loop?
Brian Chen (Jun 24 2022 at 23:23):Benchmarking suggests the overhead is non-negligible, which makes sense:
julia> function f1(x)
total = 0
for i in 1:x
total += i
end
return total
end
f1 (generic function with 1 method)
julia> @code_native f1(1000)
.text
; ┌ @ REPL[37]:3 within `f1`
; │┌ @ range.jl:5 within `Colon`
; ││┌ @ range.jl:354 within `UnitRange`
; │││┌ @ range.jl:359 within `unitrange_last`
testq %rdi, %rdi
; │└└└
jle L43
; │ @ REPL[37] within `f1`
movq %rdi, %rax
sarq $63, %rax
andnq %rdi, %rax, %rax
; │ @ REPL[37]:4 within `f1`
leaq -1(%rax), %rdx
leaq -2(%rax), %rcx
mulxq %rcx, %rcx, %rdx
shldq $63, %rcx, %rdx
leaq (%rdx,%rax,2), %rax
decq %rax
; │ @ REPL[37]:6 within `f1`
retq
; │ @ REPL[37] within `f1`
L43:
xorl %eax, %eax
; │ @ REPL[37]:6 within `f1`
retq
nop
; └
julia> function f2(x)
total = 0
for i in 1:x
@gensym a b c d e f g
total += i
end
return total
end
f2 (generic function with 1 method)
julia> @code_native f2(1000)
.text
; ┌ @ REPL[40]:3 within `f2`
; │┌ @ range.jl:5 within `Colon`
; ││┌ @ range.jl:354 within `UnitRange`
; │││┌ @ REPL[40]:1 within `unitrange_last`
pushq %rbp
pushq %r15
pushq %r14
pushq %r13
pushq %r12
pushq %rbx
subq $40, %rsp
; │││└
; │││┌ @ range.jl:359 within `unitrange_last`
testq %rdi, %rdi
; │└└└
jle L253
; │ @ REPL[40] within `f2`
movq %rdi, %rax
sarq $63, %rax
andnq %rdi, %rax, %rax
movabsq $139821397918949, %r12 # imm = 0x7F2AB4C144E5
; │ @ REPL[40]:5 within `f2`
negq %rax
movq %rax, 32(%rsp)
movl $1, %r13d
xorl %ebx, %ebx
leaq -178798797(%r12), %rax
movq %rax, 24(%rsp)
leaq -178798637(%r12), %rax
movq %rax, 16(%rsp)
leaq -178798477(%r12), %rax
movq %rax, 8(%rsp)
leaq -178798317(%r12), %rax
movq %rax, (%rsp)
leaq -178798157(%r12), %rbp
leaq -178797997(%r12), %r14
leaq -178797837(%r12), %r15
nopl (%rax,%rax)
; │ @ REPL[40]:4 within `f2`
; │┌ @ expr.jl:12 within `gensym`
L144:
movl $1, %esi
movq 24(%rsp), %rdi
callq *%r12
movl $1, %esi
movq 16(%rsp), %rdi
callq *%r12
movl $1, %esi
movq 8(%rsp), %rdi
callq *%r12
movl $1, %esi
movq (%rsp), %rdi
callq *%r12
movl $1, %esi
movq %rbp, %rdi
callq *%r12
movl $1, %esi
movq %r14, %rdi
callq *%r12
movl $1, %esi
movq %r15, %rdi
callq *%r12
; │└
; │ @ REPL[40]:5 within `f2`
; │┌ @ int.jl:87 within `+`
addq %r13, %rbx
movq 32(%rsp), %rax
; │└
; │┌ @ range.jl:837 within `iterate`
; ││┌ @ promotion.jl:468 within `==`
addq %r13, %rax
incq %rax
; ││└
incq %r13
; ││┌ @ promotion.jl:468 within `==`
cmpq $1, %rax
; │└└
jne L144
jmp L255
; │ @ REPL[40] within `f2`
L253:
xorl %ebx, %ebx
; │ @ REPL[40]:7 within `f2`
L255:
movq %rbx, %rax
addq $40, %rsp
popq %rbx
popq %r12
popq %r13
popq %r14
popq %r15
popq %rbp
retq
nopw %cs:(%rax,%rax)
; └
julia> @benchmark f1(1000)
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 1.152 ns … 69.760 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 1.160 ns ┊ GC (median): 0.00%
Time (mean ± σ): 1.171 ns ± 0.712 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
█
▂▁▁▆▁▁▁▄▁▁▁▄▁▁▁▃▁▁▁▂▁▁▁▂▁▁▁▃▁▁█▁▁▁▅▁▁▁▂▁▁▁▄▁▁▁█▁▁▁▆▁▁▁▇▁▁▂ ▂
1.15 ns Histogram: frequency by time 1.17 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
julia> @benchmark f2(1000)
BenchmarkTools.Trial: 641 samples with 1 evaluation.
Range (min … max): 7.365 ms … 10.314 ms ┊ GC (min … max): 0.00% … 0.00%
Time (median): 7.726 ms ┊ GC (median): 0.00%
Time (mean ± σ): 7.795 ms ± 346.021 μs ┊ GC (mean ± σ): 0.00% ± 0.00%
▁ ▁▂█▆▆▆▆▁▃▁▃▂▂
▅▆█▆▅█▅█████████████▆▅▃▃▃▃▂▃▁▃▃▃▃▁▄▄▃▃▃▃▃▃▂▁▁▂▁▂▂▁▁▁▂▂▁▁▂▁▃ ▄
7.36 ms Histogram: frequency by time 9.05 ms <
Memory estimate: 0 bytes, allocs estimate: 0.
You answered the question by benchmarking both ways and noting the relative slowdown exceeds 5000x.
You will be happier reorganizing the code, lift the gensym out of the loop -- pregen batches of these symbols.
Michael Fiano (Jun 24 2022 at 23:51):gensym comes from Common Lisp where it is only supposed to be used at macro-expansion time. string-interning symbols as Julia calls it, is not cheap.
Brian Chen (Jun 24 2022 at 23:51):I was trying to get a better sense of the absolute unit overhead. Some more local benchmarking seemed to suggest 1-2us, but that's subject to noise and possibly other factors
Michael Fiano (Jun 24 2022 at 23:56):If it's anything like CL, there is more than string-interning involved too, since it has to keep a generation counter to ensure uniqueness.
Brian Chen (Jun 24 2022 at 23:56):Yup, this one specifically: https://github.com/JuliaLang/julia/blob/87ded5a9aa502cfc4e03cbf230cb9bba86c85cc1/src/symbol.c#L126
Brian Chen (Jun 24 2022 at 23:57):I think I can fulfill my use case using a custom macro (basically @static_gensym x, y, ...), but will have to try. Edit: it worked!
Notification Bot (Jun 25 2022 at 00:18):Brian Chen has marked this topic as resolved.
Sukera (Jun 25 2022 at 04:48):oh boy, did not know that @gensym does a ccall
Sukera (Jun 25 2022 at 04:49):I would expect this particular benchmark to be fast, since the gensymmed variables are never used, but the ccall may have sideeffects the compiler can't see, so it can't be eliminated
Brian Chen (Jun 25 2022 at 05:35):Isn't side effect inference for ccall a nightly only feature?
Sukera (Jun 25 2022 at 05:37):I don't think there's any sideeffect inference on ccalls
Sukera (Jun 25 2022 at 05:37):are you talking about the @assume_effects macro?
Sukera (Jun 25 2022 at 05:38):I was more thinking about a being a dead variable, since it's never used
Mosè Giordano (Jun 25 2022 at 08:32):I'm not sure the benchmarks were realistic, I think f1 was constant propagated, no? Summing 1000 elements took 1 ns
Sukera (Jun 25 2022 at 08:37):oh yeah, 100% - it got folded to eulers gauss formula, there isn't even a jump there anymore
Sukera (Jun 25 2022 at 08:37):to me at least the @gensym should not have made a difference for that case
Mosè Giordano (Jun 25 2022 at 08:42):did you mean to say Gauss or I'm missing a new formula? :sweat_smile:
Sukera (Jun 25 2022 at 08:43):ah, yes of course
Sukera (Jun 25 2022 at 08:44):always hard to remember which of the gazillion famous formulas is from which of the two
Brian Chen (Jun 25 2022 at 13:45):@Mosè Giordano the goal was just to have the function body not optimized away entirely. I wanted to use donotdelete but didn't want to switch versions :P
Brian Chen (Jun 25 2022 at 13:46):Ended up just testing a single iteration and explicit calls because those are kept around still
Mosè Giordano (Jun 25 2022 at 14:52):Brian Chen said:
Mosè Giordano the goal was just to have the function body not optimized away entirely. I wanted to use donotdelete but didn't want to switch versions :P
on nightly I get
julia> @benchmark f1(1000)
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 2.460 ns … 20.945 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 2.472 ns ┊ GC (median): 0.00%
Time (mean ± σ): 2.511 ns ± 0.566 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
█▄ ▃ ▃ ▁
██▃▃▁█▁▃▁▁█▇▄▁▁▁▃▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▇ █
2.46 ns Histogram: log(frequency) by time 3.26 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
which would also suggests it isn't doing much in practice. these damned compilers eliminating everything...
chriselrod (Jun 25 2022 at 20:35):Mosè Giordano said:
I'm not sure the benchmarks were realistic, I think f1 was constant propagated, no? Summing 1000 elements took 1 ns
LLVM will turn that sum into binomial(x+1,2).
Last updated: Dec 28 2024 at 04:38 UTC