Geforce GT 730 Win10 CL_INVALID_COMMAND_QUEUE and CL_OUT_OF_RESOURCES errors about clij2

[haesleinhuepf]

Hey @chalkie666 ,

thanks for reporting this issue. I think a reviewer had a similar one...

Maybe its just me, maybe a bad card.... its been around a while....

Yes, I've had similar reports with the same card. It is 7 years old and has a bandwith lower than current integrated GPUs.
https://www.techpowerup.com/gpu-specs/geforce-gt-730.c1988

Furthermore, your card has 2 GB of memory limiting the image size in type float (32 bit) < 500 MB. How large is your image data?

Could you try running your code on another GPU? If the issue persists, I'll take a closer look :-)

One last thing.... is the Convolve 2 images together gadget working in real spac3 (slow) or via FFT (fast)? Just out of interest.... I will want to run stuff using the CL-FFT stuff eventually.

CLIJ2_convolve runs in real space. CLIJx_convolveFFT runs in FFT space :-)
https://github.com/clij/clij2-fft/blob/master/src/main/java/net/haesleinhuepf/clijx/plugins/ConvolveFFT.java

Cheers,
Robert

[chalkie666]

@haesleinhuepf OK, maybe i need to replace that old gpu.... it was cheap second hand.

i tested on a tiny old 2 in 1. levono miix, 4 GB RAM, with intel atom x5 and intel HD graphics, so totally weak and ancient.
Good news, the two clij2 functions i used, 3d gauss blur and convolve both worked, albeit with small kernels.
i guess this is a weaker gpu than the geforce gt 730... but maybe not....

i see a silent fail with no error message when using a large convolution kernel on this intel hd graphics machine.
sometimes i got an out of resources error, as i expected, but not always. just so you know.

in the same repos as my script are 2 psf images, a large and smaller one for testing.

thanks for the tips so far.
if its not worth trouble shooting the old geforce gt 730 card i will get my hands on something newer.... i though i might be able to use it for small tests and algorithm dev .... maybe not.

i dont see the CLIJx_convolveFFT in my menus.... only the clfft RL.
but i do see simpleITK things including Wiener Deconv.... this might be exactly what i need for inverse filtering residuals in the algorithm... is this real space or FFT? i see other deconv stuff in simpleITK too... i will have a play with it
cheers!

[chalkie666]

@haesleinhuepf Yikes! The great GPU shortage has really made even cheaper newer GPUs super expensive...
no problem, i will fall back to CPU (i have 12 threads at 2.8 GHx) and just use the small test data while i get things working in the code for the algorithm im working on.

I just tried
Ext.CLIJx_convolveFFT(gaussGuessGPU, psfGPU, convGuessGPU)
and it seems to work (on my CPU openCL), and is about as fast as 3D Gaussian blur with 1,1,1
so that's awesome!
I was confused that i couldn't see it in the menus, but its sure there and apparently working!
Next i will make use of whatever openCL image Math and intensity normalisation you have implemented,
to make my 2 flavours of deconv iterations work...
and figure out the regularized Wiener filter from simpleITK... i need that as the magic ingredient for this algorithm, in order for it to converge faster than RL.

thanks R!