Optimizing the creation of many small structures during execution typically comes down to either cleverly eliminating the need to allocate them in the first place (via escape analysis, and the like) or making the first generation of the garbage collector wickedly fast. I understand both of these are being worked. On April 19, 2015 at 8:08:53 PM, Dahua Lin (linda...@gmail.com) wrote:
My benchmark shows that element indexing has been as fast as it can be for array views (or subarrays in Julia 0.4). Now the problem is actually the construction of views/subarrays. To optimize the overhead of this part, the compiler may need to introduce additional optimization. Dahua On Monday, April 20, 2015 at 6:39:35 AM UTC+8, Sebastian Good wrote: —track-allocation still requires guesswork, as optimizations can move the allocation to a different place than you would expect. On April 19, 2015 at 4:36:19 PM, Peter Brady (peter...@gmail.com) wrote: So I discovered the --track-allocation option and now I am really confused: Here's my session: $ julia --track-allocation=all _ _ _ _(_)_ | A fresh approach to technical computing (_) | (_) (_) | Documentation: http://docs.julialang.org _ _ _| |_ __ _ | Type "help()" for help. | | | | | | |/ _` | | | | |_| | | | (_| | | Version 0.3.8-pre+13 (2015-04-17 18:08 UTC) _/ |\__'_|_|_|\__'_| | Commit 0df962d* (2 days old release-0.3) |__/ | x86_64-redhat-linux julia> include("test.jl") test_all (generic function with 1 method) julia> test_unsafe(5) And here's the relevant part of the resulting test.jl.mem file. Note that I commented out some calls to 'size' and replaced with the appropriate hard-coded values but the resulting allocation is the same... Can anyone shed some light on this while I wait for 0.4 to compile? - function update(a::AbstractArray, idx, off) 8151120 for i=1:320 #size(a, idx) 0 a[i] = -10*off+i - end 0 a - end - - function setk_UnSafe{T}(a::Array{T,3}) 760 us = UnsafeSlice(a, 3) 0 for j=1:size(a,2),i=1:size(a,1) 8151120 us.start = (j-1)*320+i #size(a,1)+i - #off = sub2ind(size(a), i, j, 1) 0 update(us, 3, us.start) - end 0 a - end - function test_unsafe(n) 0 a = zeros(Int, (320, 320, 320)) - # warmup 0 setk_UnSafe(a); 0 clear_malloc_data() - #@time ( 0 for i=1:n; setk_UnSafe(a); end - end On Sunday, April 19, 2015 at 2:21:56 PM UTC-6, Peter Brady wrote: @Dahua, thanks for adding an unsafeview! I appreciate how quickly this community responds. I've added the following function to my test.jl script function setk_unsafeview{T}(a::Array{T,3}) for j=1:size(a,2),i=1:size(a,1) off = sub2ind(size(a), i, j, 1) update(unsafe_view(a, i, j, :), 3, off) end a end But I'm not seeing the large increase in performance I was expecting. My timings are now julia> test_all(5); test_stride elapsed time: 2.156173128 seconds (0 bytes allocated) test_view elapsed time: 9.30964534 seconds (94208000 bytes allocated, 0.47% gc time) test_unsafe elapsed time: 2.169307471 seconds (16303000 bytes allocated) test_unsafeview elapsed time: 8.955876793 seconds (90112000 bytes allocated, 0.41% gc time) To be fair, I am cheating a bit with my custom 'UnsafeSlice' since I make only one instance and simply update the offset on each iteration. If I make it immutable and create a new instance on every iteration (as I do for the view and unsafeview), things slow down a little and the allocation goes south: julia> test_all(5); test_stride elapsed time: 2.159909265 seconds (0 bytes allocated) test_view elapsed time: 9.029025282 seconds (94208000 bytes allocated, 0.43% gc time) test_unsafe elapsed time: 2.621667854 seconds (114606240 bytes allocated, 2.41% gc time) test_unsafeview elapsed time: 8.888434466 seconds (90112000 bytes allocated, 0.44% gc time) These are all with 0.3.8-pre. I'll try compiling master and see what happens. I'm still confused about why allocating a single type with a pointer, 2 ints and a tuple costs so much memory though. On Sunday, April 19, 2015 at 11:38:17 AM UTC-6, Tim Holy wrote: It's not just escape analysis, as this (new) issue demonstrates: https://github.com/JuliaLang/julia/issues/10899 --Tim On Sunday, April 19, 2015 12:33:51 PM Sebastian Good wrote: > Their size seems much decreased. I’d imagine to totally avoid allocation in > this benchmark requires an optimization that really has nothing to do with > subarrays per se. You’d have to do an escape analysis and see that Aj never > left sumcols. Not easy in practice, since it’s passed to slice and length, > and you’d have to make sure they didn’t squirrel it away or pass it on to > someone else. Then you could stack allocate it, or even destructure it into > a bunch of scalar mutations on the stack. After eliminating dead code, > you’d end up with a no-allocation loop much like you’d write by hand. This > sort of optimization seems to be quite tricky for compilers to pull off, > but it’s a common pattern in numerical code. > > In Julia is such cleverness left entirely to LLVM, or are there optimization > passes in Julia itself? On April 19, 2015 at 6:49:21 AM, Tim Holy > (tim....@gmail.com) wrote: > > Sorry to be slow to chime in here, but the tuple overhaul has landed and > they are still not zero-cost: > > function sumcols(A) > s = 0.0 > for j = 1:size(A,2) > Aj = slice(A, :, j) > for i = 1:length(Aj) > s += Aj[i] > end > end > s > end > > Even in the latest 0.4, this still allocates memory. On the other hand, > while SubArrays allocate nearly 2x more memory than ArrayViews, the speed > of the two (replacing `slice` with `view` above) is, for me, nearly > identical. > > --Tim > > On Friday, April 17, 2015 08:30:27 PM Sebastian Good wrote: > > This was discussed a few weeks ago > > > > https://groups.google.com/d/msg/julia-users/IxrvV8ABZoQ/uWZu5-IB3McJ > > > > I think the bottom line is that the current implementation *should* be > > 'zero-cost' once a set of planned improvements and optimizations take > > place. One of the key ones is a tuple overhaul. > > > > Fair to say it can never be 'zero' cost since there is different inherent > > overhead depending on the type of subarray, e.g. offset, slice, > > re-dimension, etc. however the implementation is quite clever about > > allowing specialization of those. > > > > In a common case (e.g. a constant offset or simple stride) my > > understanding > > is that the structure will be type-specialized and likely stack allocated > > in many cases, reducing to what you'd write by hand. At least this is what > > they're after. > > > > On Friday, April 17, 2015 at 4:24:14 PM UTC-4, Peter Brady wrote: > > > Thanks for the links. I'll check out ArrayViews as it looks like what I > > > was going to do manually without wrapping it in a type. > > > > > > By semi-dim agnostic I meant that the differencing algorithm itself only > > > cares about one dimension but that dimension is different for different > > > directions. Only a few toplevel routines actually need to know about the > > > dimensionality of the problem. > > > > > > On Friday, April 17, 2015 at 2:04:39 PM UTC-6, René Donner wrote: > > >> As far as I have measured it sub in 0.4 is still not cheap, as it > > >> provides the flexibility to deal with all kinds of strides and offsets, > > >> and > > >> the view object itself thus has a certain size. See > > >> https://github.com/rened/FunctionalData.jl#efficiency for a simple > > >> analysis, where the speed is mostly dominated by the speed of the > > >> "sub-view" mechanism. > > >> > > >> To get faster views which require strides etc you can try > > >> https://github.com/JuliaLang/ArrayViews.jl > > >> > > >> What do you mean by semi-dim agnostic? In case you only need indexing > > >> along the last dimension (like a[:,:,i] and a[:,:,:,i]) you can use > > >> > > >> https://github.com/rened/FunctionalData.jl#efficient-views-details > > >> > > >> which uses normal DenseArrays and simple pointer updates internally. It > > >> can also update a view in-place, by just incrementing the pointer. > > >> > > >> Am 17.04.2015 um 21:48 schrieb Peter Brady <peter...@gmail.com>: > > >> > Inorder to write some differencing algorithms in a semi-dimensional > > >> > > >> agnostic manner the code I've written makes heavy use of subarrays > > >> which > > >> turn out to be rather costly. I've noticed some posts on the cost of > > >> subarrays here and that things will be better in 0.4. Can someone > > >> comment > > >> on how much better? Would subarray (or anything like it) be on par with > > >> simply passing an offset and stride (constant) and computing the index > > >> myself? I'm currently using the 0.3 release branch.
