Filtered backproject on AMD GPU

[kif]

Well, I am not sure we do support this platform, but it is good to know.

• Intel/CPU is OK,
• Apple/gpu is OK
• PoCL/CPU (1.5-6, amd64) is OK
• AMD/GPU (Vega56) Fails with:

  
  FAIL: test_fbp (silx.opencl.test.test_backprojection.TestFBP)
  tests FBP
  ----------------------------------------------------------------------
  Traceback (most recent call last):
  File "/home/jerome/workspace/silx/build/lib.linux-x86_64-3.9/silx/opencl/test/test_backprojection.py", line 155, in test_fbp
  self.assertTrue(err < self.tol, "Max error is too high")
  AssertionError: False is not true : Max error is too high

Stderr: WARNING:py.warnings:/usr/lib/python3/dist-packages/pytools/py_codegen.py:26: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses import imp

====================================================================== FAIL: test_fbp_filters (silx.opencl.test.test_backprojection.TestFBP) Test the different available filters of silx FBP.

Traceback (most recent call last): File "/home/jerome/workspace/silx/build/lib.linux-x86_64-3.9/silx/opencl/test/test_backprojection.py", line 198, in test_fbp_filters self.assertTrue( AssertionError: False is not true : Something wrong with FBP(filter=ramlak)

====================================================================== FAIL: test_fbp_oddsize (silx.opencl.test.test_backprojection.TestFBP)

Traceback (most recent call last): File "/home/jerome/workspace/silx/build/lib.linux-x86_64-3.9/silx/opencl/test/test_backprojection.py", line 214, in test_fbp_oddsize self.assertLess( AssertionError: 258.20255 not less than 0.1 : Something wrong with FBP on odd-sized sinogram

[jcesardasilva]

Hi everyone, If possible, please keep supporting the openCL packages in SILX or, if not in your plans, move all, or at least, the Backprojection and Projection to Nabu (or any other place to be maintained), please. Nabu only offers the former, but not the latter.

In any case, there is a feature in those packages, which only work (at least for me) with pyopencl<=2020.2. If I upgrade pyopencl to the newest version, the packages get broken.

Could anyone help me with this, please?

Thank you very much in advance.