GPU support image format but it return `0` or `CL_IMAGE_FORMAT_NOT_SUPPORTED`

[StRigaud]

in OpenCL, the cl_image is defined by a cl_image_format composed of cl_channel_order and cl_channel_type.

• cl_channel_type: correspond to the data type float, int8 etc,
• cl_channel_order: correspond to the composition of the image CL_RGB for example.

Clesperanto only manage single channel image, which point to the following channel_order: CL_R, CL_A, CL_INTENSITY, or CL_LUMINANCE.

Eventhough they all represente a single channel image, their behaviour differe based on opencl version and hardware. Quick tests shows the following.

On my local NVIDIA card:

• CL_A return 0 with CLIj framework
• CL_R work with all and does not replicate the values
• CL_INTENSITY work with all, but the value is replicated on the with channel RGB
• CL_LUMINANCE work with all but the value is replicated on the with channel RGB

• On my local M2 Apple:

• CL_A return 0 with CLIj framework
• CL_R work with all and does not replicate the values
• CL_INTENSITY work only with CL_FLOAT, and the value is replicated on the with channel RGB
• CL_LUMINANCE is not supported

This point toward the usage of the CL_R. I will set a PR to ensure maximum compatibility.

Classic OpenCL error that could be related to this issue is CL_IMAGE_FORMAT_NOT_SUPPORTED

[StRigaud]

@haesleinhuepf

With #318, cl_image is fully compatible on my M2, this include the interpolation. This will be available on the python side with the next release (soon).

[haesleinhuepf]

That's super cool! And it might be an advantage over clij2 and pyclesperanto_prototype

[StRigaud]

Its possible that it is the same issue for the prototype and clij2 somewhere in the code?

Rigth now I am trying to make the CI skip tests over images because it does not support them. Once I manage, I will merge and make a release to have both chan-vese and affine transform operational

[StRigaud]

Test involving images are always skip. To enable them for local test, one must comment GTEST_SKIP() in the test code.

Its not great but will do for now