search

clEsperanto / pyclesperanto_prototype

GPU-accelerated bio-image analysis focusing on 3D+t microscopy image data
http://clesperanto.net
BSD 3-Clause "New" or "Revised" License
201 stars 44 forks source link

clImages versus clBuffers #29

Open haesleinhuepf opened 4 years ago

haesleinhuepf commented 4 years ago

So far we are only working with [OCLArrays]() and I'm not sure what it internally is: clBuffers or clImages.

It might be important to be able to choose. For interpolation-related filters (scaling, affine transform, image warping ...) and filters accessing local pixel neighborhood, we should be able to choose. In clij, an additional buffer-to-image-copy pays off because there is apparently hardware-optimization for accessing neigboring pixels and interpolation.

I could imagine that pyopencl has some internal magic for converting buffers to images transparent for the API user.

tlambert03 commented 4 years ago

cl.array.Arrays use cl.Buffers. I have some logic in napari-ndtiffs that provides a numpy-to-image function and does some nice device inspection to maximize chances of compatibility depending on what the device supports (using cl.get_supported_image_formats). That was after some trial and error getting the code to pass tests on CI. I can port some of that code here if you want to have a cl.Image object to start working with.

tlambert03 commented 4 years ago

actually that whole file has a nicely working affine transform... and tries to match the scipy.ndimage.interpolation API as closely as possible (drop in replacement). I know you have those kernels elsewhere in clij, but feel free to look through it for ideas.

haesleinhuepf commented 4 years ago

That was after some trial and error getting the code to pass tests on CI.

Awesome! I struggled with the same (actually reviewer #2 struggled with it more than myself). :-) I've implemented a pretty fail-safe alternative as it tests on the GPU and disables image support in case it crashes.

I can port some of that code here if you want to have a cl.Image object to start working with.

If you find time for that, that would be awesome. I would say your solution looks more sophisticated than mine.

actually that whole file has a nicely working affine transform...

Hehe, that looks pretty much like we both were inspired by gputools. I've got a buffer variant using nearest-neighbor and a image counter part with interpolation.

It's so good that we work together on this. Obviously, we spend time on doing the same things in Java and Python =-)