Hi everybody, I have written a script that evaluates a non-linear function feta over a 2D domain. The function returns eta that depends on other 2D fields (Exx, Eyy, Exy). I try to compute the gradient of eta w.r.t Exx, Eyy, Exy et each cell of the domain using ForwardDiff (function feta_grad). Th…

I’ve not tried your latest examples, but I’ll consolidate functions to get ForwardDiff and AbstractSIMDs to play well together in SIMDDualNumbers.jl . It includes functions like abs, max, etc. This package should be registered on Wednesday (there is a three day waiting period). using SIMDDualNumbe…

Yes, I did run many times after compilation (what I meant by “from second execution onwards”). I am now trying to put the advices together to make a more suitable MWE (one main function, no global vars) to discuss performance. So far I’m stuck trying to add more inputs arguments to feta such that…

[image] Gravlax: Yes, I did run many times after compilation (what I meant by “from second execution onwards”). I just tried on a different computer without AVX512, and got basically the same pattern of results: julia> println("Base --> ", t1, " s") Base --> 0.09175798 s julia> printl…

[image] Gravlax: eta = 0.0*ones(Float64,nx,ny) Exx, Eyy, Exy = Ebg.*ones(Float64,nx,ny), -Ebg.*ones(Float64,nx,ny), 0.2*Ebg.*ones(Float64,nx,ny) Not related to your performance issues, but I should point this out: Instead of 0.0 * ones(Floa…

Thanks a lot to both of you. It now runs with parsing arguments using the closure suggested by Elrod and fill of arrays suggested by DNF . It also into a main function. The picture has now changed quite a bit, here for 3 consecutive evaluations: Base --> 0.008191534 s @turbo --> 0.003956482 s @t…

[image] Gravlax: i5-3230M Release date: Q1’13 a blast from the past! XD

[image] Gravlax: EDIT: with 4 threads :smiley: Currently, LV will only use up to as many threads as you have cores (so 2). Glad it’s helping performance now. [image] jling: Release date: Q1’13 a blast from the past! XD ivybridge does not have FMA (fused multiply add) or AVX2 (basic …

Really really cool!! I’ve made the update and ran on my desktop computer which could also be seen as a ‘blast from the past’ :smiley: julia> versioninfo() Julia Version 1.6.1 Commit 6aaedecc44 (2021-04-23 05:59 UTC) Platform Info: OS: macOS (x86_64-apple-darwin18.7.0) CPU: Intel(R) Core(TM) i5…

Just to illustrate the last point of my previous message. In other codes where I use LV, I usually ‘get around’ conditional statements by replacing if statements with some boolean operation. For example, if I would replace the if statement that defines the stopping condition within the feta function…

[image] Gravlax: Would anyone have an idea about how to get around this? From the stacktrace, you can see that the problem is in abs. ForwardDiff defines abs for dual numbers using branches. So we need to define one without it: @generated function Base.abs(x::ForwardDiff.Dual{T,S,N}) where …

[image] Elrod: From the stacktrace, you can see that the problem is in abs . ForwardDiff defines abs for dual numbers using branches. Thanks for pointing this out. Indeed I could check that replacing abs(r) by sqrt(r*r) actually works there. [image] Elrod: Also, I’d suggest using IfElse.…

Making a fast evaluation of a function and its gradient

General Usage Performance

Elrod October 11, 2021, 5:32pm 33

Note that LoopVectorization 0.12.83 and newer should automatically using SIMDDualNumbers for you if you using LoopVectorization, ForwardDiff.

Sparse matrix coefficient computation with ForwardDiff and LoopVectorization

Topic		Replies	Views
Vectorized gradient using ForwardDiff Specific Domains forwarddiff	5	478	July 17, 2022
Gradient evaluation with ForwardDiff and LoopVectorization General Usage forwarddiff , loopvectorization	2	944	October 24, 2021
Need help using ForwardDiff gradient with numerical loop General Usage question	11	269	November 7, 2024
High order approximation of a multi-country trade model Finance and Economics performance , differentiation , forwarddiff , finitediff	22	1215	July 6, 2022
Vectorization of "complex" loops New to Julia	15	537	September 27, 2025

Making a fast evaluation of a function and its gradient

Related topics