On Thu, Aug 25, 2016 at 5:19 PM, Chris Rackauckas <[email protected]>
wrote:

> Oh it's that same issue? Thanks for pointing it out. That means every one
> of my performance problems are due to the same issue... that's both a good
> thing and a bad thing I guess. So would your suggestion just be to move
> forward doing things like this, knowing that this issue will get fixed? If
> it will be possible to have good performance with closures on a type like
> that, then I'll be moving forward with a larger design change
> <https://github.com/ChrisRackauckas/DifferentialEquations.jl/issues/41#issuecomment-242310914>
>  in
> hopes that it will improve. But right now I'm in limbo because I don't know
> what to do about this!
>

That's why I also linked the workaround.


>
> On Thursday, August 25, 2016 at 1:43:56 AM UTC-7, Yichao Yu wrote:
>
>>
>>
>> On Thu, Aug 25, 2016 at 4:12 PM, Chris Rackauckas <[email protected]>
>> wrote:
>>
>>> It seems like a closure on a non bitstype still has issues, or is there
>>> something wrong with doing this?
>>>
>>> immutable Params
>>>   α::Float64
>>> end
>>>
>>> function test()
>>>   h = (t,u,du,p) -> @inbounds begin
>>>     du[1] = 1.01u[1]
>>>     du[2] = p.α*u[2]
>>>   end
>>>   h2 = (t,u,du) -> @inbounds begin
>>>     du[1] = 1.01u[1]
>>>     du[2] = 1.0*u[2]
>>>   end
>>>   const u = [0.1,0.2]
>>>   const du = zeros(2)
>>>   const p = Params(1.0)
>>>   l = (t,u,du) -> h(t,u,du,p)
>>>   println("Pass In Parameters")
>>>   @code_native h(1.0,u,du,p)
>>>   println("No Parameters")
>>>   @code_native h2(1.0,u,du)
>>>   println("Closures")
>>>   @code_native l(1.0,u,du)
>>> end
>>> test()
>>>
>>> Pass In Parameters
>>> .text
>>> Filename: replacement_macros_tests.jl
>>> pushq %rbp
>>> movq %rsp, %rbp
>>> Source line: 29
>>> movq (%rdi), %rax
>>> movsd (%rax), %xmm0           # xmm0 = mem[0],zero
>>> movabsq $140006850762016, %rcx  # imm = 0x7F55E29B1520
>>> mulsd (%rcx), %xmm0
>>> movq (%rsi), %rcx
>>> movsd %xmm0, (%rcx)
>>> Source line: 30
>>> movsd (%rdx), %xmm0           # xmm0 = mem[0],zero
>>> mulsd 8(%rax), %xmm0
>>> movsd %xmm0, 8(%rcx)
>>> popq %rbp
>>> retq
>>>
>>> No Parameters
>>> .text
>>> Filename: replacement_macros_tests.jl
>>> pushq %rbp
>>> movq %rsp, %rbp
>>> Source line: 33
>>> movq (%rdi), %rax
>>> movsd (%rax), %xmm0           # xmm0 = mem[0],zero
>>> movabsq $140006850762144, %rcx  # imm = 0x7F55E29B15A0
>>> mulsd (%rcx), %xmm0
>>> movq (%rsi), %rcx
>>> movsd %xmm0, (%rcx)
>>> Source line: 34
>>> movq 8(%rax), %rax
>>> movq %rax, 8(%rcx)
>>> popq %rbp
>>> retq
>>> nopw (%rax,%rax)
>>>
>>>
>>> Closures
>>> .text
>>> Filename: replacement_macros_tests.jl
>>> pushq %rbp
>>> movq %rsp, %rbp
>>> pushq %r15
>>> pushq %r14
>>> pushq %r13
>>> pushq %r12
>>> pushq %rbx
>>> subq $72, %rsp
>>> movq %rdx, %r14
>>> movq %rsi, %r15
>>> movsd %xmm0, -112(%rbp)
>>> movabsq $140007647692112, %rbx  # imm = 0x7F56121B4550
>>> movq %fs:0, %r12
>>> addq $-2672, %r12            # imm = 0xFFFFFFFFFFFFF590
>>> xorpd %xmm0, %xmm0
>>> movupd %xmm0, -64(%rbp)
>>> movupd %xmm0, -80(%rbp)
>>> movq $0, -48(%rbp)
>>> movq $12, -104(%rbp)
>>> movq (%r12), %rax
>>> movq %rax, -96(%rbp)
>>> leaq -104(%rbp), %rax
>>> movq %rax, (%r12)
>>> movq $0, -88(%rbp)
>>> Source line: 39
>>> movq (%rdi), %rax
>>> movq (%rax), %r13
>>> testq %r13, %r13
>>> je L232
>>> movq %r13, -88(%rbp)
>>> movq %rbx, -80(%rbp)
>>> movabsq $jl_gc_pool_alloc, %rax
>>> movl $1432, %esi             # imm = 0x598
>>> movl $16, %edx
>>> movq %r12, %rdi
>>> callq *%rax
>>> addq $1072064, %rbx          # imm = 0x105BC0
>>> movq %rbx, -8(%rax)
>>> movsd -112(%rbp), %xmm0       # xmm0 = mem[0],zero
>>> movsd %xmm0, (%rax)
>>> movq %rax, -72(%rbp)
>>> movq %r15, -64(%rbp)
>>> movq %r14, -56(%rbp)
>>> movq %r13, -48(%rbp)
>>> movabsq $jl_apply_generic, %rax
>>> movl $5, %esi
>>> leaq -80(%rbp), %rdi
>>> callq *%rax
>>> movq -96(%rbp), %rax
>>> movq %rax, (%r12)
>>> addq $72, %rsp
>>> popq %rbx
>>> popq %r12
>>> popq %r13
>>> popq %r14
>>> popq %r15
>>> popq %rbp
>>> retq
>>> L232:
>>> addq $1448, %rbx             # imm = 0x5A8
>>> movabsq $jl_throw, %rax
>>> movq %rbx, %rdi
>>> callq *%rax
>>> nop
>>>
>>>
>>> Clearly the 3rd method (using a closure on a parameter type) has a lot
>>> more steps. My untrained eye says that means it's bad... is that simple
>>> analysis wrong?
>>>
>>>
>> The issue is https://github.com/JuliaLang/julia/issues/15276 See
>> https://github.com/JuliaLang/julia/issues/17395#issuecomment-241911387
>> for a workaround that works for me most of the time.
>>
>>

Reply via email to