Magnitudes for openCL

[Carolina710]

Hello, do you have any idea when the magnitude calculation for the openCL version will be available?

I'm using the y_error and the ns values that are returned from bfastfinaldetailed, but it takes some time to do the calculation on the CPU, as well as a large allocation of GPU memory necessary to perform the bfastfinaldetailed. I have already looked at the bfastfinal and bfastfinaldetailed files a lot, trying to understand which kernels do what, but they are very unreadable.
I saw that you made commits to the opencl-backend-debug branch with work for magnitude calculation for the openCL, but I was unable to use the updates.

Thank you!

[mortvest]

Hi Carolina,

There has been a lot of changes to the OpenCL version, and we are working on tuning its performance and it would take a while.
In a meantime you can use opencl-backend-debug branch. It supports the magnitudes for both the OpenCL and Python backends. What problems are you experiencing with the update?

[Carolina710]

Thank you for the rapid response.

I just used your peru_small_opencl example. The error occurs in the openCL version:

I wonder if it is not related to the _get_futhark_params function, that is returning null?

Thank you! :)

[mortvest]

That looks like an internal OpenCL problem. _get_futhark_params() should not be a problem, since returns an empty dictionary here, hence no tuning is applied. I don't get this error when running the peru_small_opencl.py and it runs fine for me.
What GPU are you using?

[Carolina710]

Hello,

My GPU is quit of lower performance:

After your response, I take some time to run the code on a remote PC that I have access and all works fine. 👍

Thank you!

[mortvest]

No problem. I have also experienced similar issues with my old laptop after we have changed the indexing to 64-bit integers.

[mortvest]

OpenCL magnitude support has been added to develop