Good question. I’m no developer unfortunately.
I’m running dt on a AMD system with rusicl. i have the feeling that the new version is even a tiny bit slower than the previous one. Maybe a driver problem. I don’t know. I will wait untill there is a build in the repos of debian unstable and will try it then again. If it’S not better I will try to do some research.
I installed Ansel and for me is slower than darktable.
Not sure what is exactly measured here, UI responsiveness when chaning parameters in modules, export processing, moving aroubd the canvas … Many points could contribute
- When processing a pipe dt always runs through some “preliminary state” taking care of the rois. I think ansel prepares that once and keeps that “plan”. Dt also misses a strict order of pipe runs, it can happen, that the full pipe requires data from preview pipe …
- In dt we have different rois for the demosaicer instead of processing full data until the first demosaicer part and starting the pipe from cached-date and just scale for roi. (This is important for moving canvas around/zooming in/out)I once did a similar PR … that takes a bit longer for the first pixelpipe run but after that is clearly faster.
- The OpenCL caching and tiling for ansel got a major overhaul that can lead to benefits.
- @Popanz on master or 5.6 (there was some performance penalty now fixed on master)?
- EDIT: also minimimum cpu architecture has been bumped to V3 - some distributions do that too for performance. There is also - iirc - a report in phoronix measuring dt with perf gain …
- …
Nvidia.
This. Exporrting and moving around is no problem.
I run 5.6 not master. But will give master a try. Thanx for the hint.
I think I have to correct my statement regarding speed. I totally forgot some standard settings on my side. Demosaic sharpening for example. So in the end I did the same as the TO and compared apples with pears.
If I switch some things off, which are not available in Ansel, dt wins ground. I still have the feeling, that Ansel is a touch faster, but that’s just a feeling I can’t really prove and in the end it doesn’t really matter. I’m simply in love with dt. ![]()
Ansel’s UI is definitely faster. A particular metric which you can find in his (very rude) performance blog posts is when switching from lighttable to darktable, this is pretty slow in darktable, at least one second, independent of hardware, and nearly instant in Ansel. I am not saying Ansel is better, just that in this particular example, and some other UI bits, a lot of optimization has been done to ease out a few delays which are pretty uncomfortable in dt. Obviously the pipeline is likely faster in darktable after all the optimizations.
From his post, although it is not clear if he is comparing with the latest version or 4.0 which Ansel is compatible with.
The pixel pipeline has been completely overhauled: modules run about 1.8× faster on CPU and 1.6× faster on GPU than darktable, and — more importantly — the pipeline cache was rewritten so that disabling a module, using a color picker or re-exporting an already-exported image does not recompute the whole thing. That makes Ansel 4 to 100× faster than darktable to refresh the image. Memory management was rewritten too: Ansel knows exactly how much RAM it uses and stays within the budget you set, instead of darktable’s hand-wavy “performance profiles” that have no predictability.
The whole lighttable, thumbnail and mipmap-cache stack was rewritten. Concretely: starting the app and drawing a full collection is roughly 3.5× faster, switching between lighttable and darkroom is near-instant, and scrolling a large collection no longer stutters. The right sidebar was removed (its content moved to the global menu and the left panel), culling became the Restrict to selection filter, the filtering toolbar was rebuilt as plain inclusive toggle buttons, the zoomable lighttable came back (Columns + Zoom with barycentric auto-pan), and selection became explicit and What You See Is What You Get — no more metadata loss from a dangling mouse cursor.
Interesting to see the speed difference between Ansel and Darktable. In images with Darktable’s AgX enabled, rendering a full quality image is very slow in 5.7.0 (I did not experience that at all in the 5.5.x nightlies) compared to the 5.6.0 release. Maybe it is just my laptop system because that does not hav a proper GPU.
The Ansel rendering feels faster (although that is subjective) on the laptop.
i will try the latest nightly on my older desktop with nvidia dinosaur GPU (1060 ti)
Interesting. The 5.7.0 is for me as quick as it should. There is some known issue in 5.6.0. See here: Speed of darktable 5.6
- There will definitely be more and important improvements for 5.8, some related to “compute less”, some for improved OpenCL tiling, some just for UI responsiveness. A fair amount of that is already discussed intensively.
- I have read about the mentioned “ansel being xyz times faster” comments. Some parts of that just make me shudder or make me grinning wide depending on pre-reading mood. For me it’s like my new car can accelerate to 70mph is xxx seconds. I just give a very big shit about such claims (besides never driving cars - for my whole life been bike only)
- A general remark about benchmarking - what exactly and how exactly are you measuring … you want me to create a benchmark showing dt being 10times faster for certain situations? No problem at all.
Both programs have developed in very different directions, speaking for dt i think we are in a very good path also regarding performance.
I absolutely second this. If 5.6.0 wouldn’t have been VERY noticeable slower than dt 5.4.1. I wouldn’t have mentioned anything in the comparison to Ansel.
As written already, your fix in 5.7. solved my issues. And I don’t care if Ansel or whatever Editor is a bit faster or slower. A Ferrari is faster than an Unimog, but it is of no doubt, that you can do much more things with an Unimog. And on an acre it is even much faster. ![]()
About benchmarking: I don’t benchmark at all. What I don’t really can see and feel isn’t important for me. And there has to be a big difference before I even start thinking about this difference.
I think anybody here is without any doubt appreciating your work and that of the other devs! ![]()
By the way who has created this thread title? It wasn’t me, even though it seems like, because I’m the first in this thread. But this thread was generated by splitting it. My question was: Why is 5.6.0 noticable slower than dt nightly on my side. The question is answered.
I think the nightly is quite fast and wonder who has asked the question of the thread title.