Very slow convolution

[BrendanC]

Hi,

I've noticed since the latest v7 that convolution takes ages. I mean, just previewing a postage-stamp size of image, it's taking 4 minutes to process, same with the full screen. My machine isn't high-end but it's good for everything else I've thrown at it (Lenovo T440P, 8GB RAM, 2.5GHz i5 CPU). I'm sure convolution used to be faster.

Any ideas? Or, it has occurred to me, that perhaps the process finishes before the circling logo stops circling?

Thanks
Brendan

[admin]

Hi Brendan,

That sounds strange...
Which version are you using? The GPU version or the non-GPU version?
If using the GPU version, what does it say in the About dialog with regards to what device it is using?

1.7's deconvolution is indeed more taxing than 1.6's, as Tracking Propagation is now always set to "During Regularization" (it is no longer an option as it was in 1.6 - it always on). 1.7 also defaults to a Moffat primary PSF, which is again more taxing than the default Gaussian PSF in 1.6.

Finally, 1.7 has the number of iterations increased to 11 instead of 6, which will take longer.

Running 1.7's decon (with all the aforementioned changes) on a non-GPU accelerated dual core laptop processor will be asking a lot from such an underpowered processor. That said, 4 minutes for a tiny preview definitely sounds excessive, even for that CPU!

[Burly]

Same here for deconvoloution , and auto -stars do in the mask ages well seems longer than before maybe I just got more stars in image .

[admin]

Same here for deconvoloution , and auto -stars do in the mask ages well seems longer than before maybe I just got more stars in image .

Same question for you then
Which version are you using? The GPU version or the non-GPU version?
If using the GPU version, what does it say in the About dialog with regards to what device it is using?

Star Mask generation now uses Tracking data to better separate stars from noise (but you can override this by selecting a different source).

[BrendanC]

Hi Ivo,

I'm running StarTools 1.7.417 alpha, using 4 logical CPU cores.

I did wonder about the number of iterations but that doesn't make a huge difference.

Which of the decon algorithms should be fastest? If you could let me know, then I can take a look and play around with iterations etc to see what difference it makes.

I actually don't mind the delay, completely accept that my laptop could do with more power. I was just wondering whether this is expected behaviour and what if anything I could do about it. Given that I don't really want to replace the CPU in the laptop, how much difference would, say, more RAM or installing an SSD make, do you think?

Thanks
Brendan

[Burly]

Same here for deconvoloution , and auto -stars do in the mask ages well seems longer than before maybe I just got more stars in image .

Same question for you then
Which version are you using? The GPU version or the non-GPU version?
If using the GPU version, what does it say in the About dialog with regards to what device it is using?

Star Mask generation now uses Tracking data to better separate stars from noise (but you can override this by selecting a different source).

Non gpu ivo 64 bit version 1.7.419 alpha using 4 logical cores

i5 processor 4gb ram , cheers for heads up on star mask.

[admin]

Can you two try the GPU version of 1.7.422? It should use the iGPU on your laptops and should yield some speed improvements.

[BrendanC]

Much better - down to about 1.5 minutes now.

Great to hear. The improved processing speeds using the GPU are why I upped the quality and (more) brute force number crunching in 1.7. There is so much more you can do with more processing power! Expect this to get cut down drastically with even a modest (by 2020 standards) new machine with dedicated graphics solution.

the delay is really just for the initial process, if I then change any parameters it's just a few seconds (unless I go crazy with the number of iterations).

At first launch, the de-ringing information is being generated, which takes a while (but should not be too long).
After that it is pure decon being performed. Once you compare ST's performance here with other software, you hopefully will no longer think the wait is that long (and that is not even taking into account the sophistication of the regularization that goes up and down the processing chain to squash artifact development!)

And it never ceases to amaze me how you pack something as powerful as StarTools into a 6.5MB download...

Programming close to the metal helps keep things optimized, clean and dependency-less (ST does not rely on any external libraries). The worst thing that can happen to a project is to get bogged down in hard to fix/trace bugs or problems that occur in some component you do not have control over! Being master of my own destiny and fundamentally understand each and every aspect, lets me research, iterate and release much faster.