I still don’t have great intuition for quickly spotting my bottlenecks due to unneeded allocations and the associated work around. Here is some code I’m trying to optimize right now: function logpost(w,data,μ,Σ) post = 0.0 # rightWrong = data[:,1] # total = data[:,2] # xvec = [1.0,0.0…

You’ll notice that I am repeatedly defining and drawing from statistical distributions and the rand function returns an array instead of a number. Can I make rand return a scalar instead of a vector? Would that help?

[image] lancejnelson: data[:,2] This copies the column of data, which will cause lots of unnecessary allocations. You want @view data[:, 2] most likely. [image] lancejnelson: w' * [1.0,d] This constructs a brand new Vector just to hold [1.0, d], then immediately reduces that down to a…

[image] lancejnelson: Can I make rand return a scalar instead of a vector? Would that help? Yes, and yes! rand() or rand(distribution) should return a scalar, and this is much cheaper than returning a 1-element vector.

Help me understand this. I know about SVectors, but in your modification, it seems that a new SVector will have to be created every time. How is that cheaper than a standard vector?

I see that you are peppering your code with @timeit already. I’d recommend using @btime from BenchmarkTools for sufficiently encapsulated methods instead (it will also show you allocations then). Did you ever consider using a profiler to get a high level picture of what your code is doing?

SVector usually gets its memory from the stack or even registers instead of the heap [image] StaticArrays and allocations One example of this is that it will often not store StaticArrays anywhere in memory, and just use a few registers instead.

Calls to this function account for nearly 80% of my allocations: function logPriorGaussian(w,μ,Σ)::Float64 a = MvNormal(μ,Σ) return log(pdf(a,w)) end The parameters to the MVN don’t change so I suppose I could define it once outside the main loop and pass it in instead of defining it every…

[image] lancejnelson: The parameters to the MVN don’t change so I suppose I could define it once outside the main loop and pass it in instead of defining it every time. If no statistician opposes (and stupid me doesn’t see a reason), go ahead. [image] lancejnelson: Other than that, any w…

Instead of @timeit, a very nice tool for granular inspection of timings and allocations is this package: Timer Outputs

Reducing allocations (again)

New to Julia

Jeff_Emanuel March 24, 2022, 4:04pm 8

SVector usually gets its memory from the stack or even registers instead of the heap

2 Likes

Topic		Replies	Views
Reducing the allocations of my function General Usage question , memory-allocation , staticarrays	10	986	October 21, 2022
Pre-allocating a structure to reduce memory? New to Julia preallocation	14	2797	August 16, 2022
Common allocation mistakes Performance memory-allocation	47	7713	August 21, 2023
Help to reduce number of allocations General Usage	23	5615	January 6, 2019
Reducing allocations to optimize (to match Java speed) Performance	22	1471	December 10, 2018

Reducing allocations (again)

Related topics