Force CPU, useful when OpenCL bugs are present and prevent running #113
Conversation
Removed macOS specific files and allowed for mulitple venvs with different endings (venv-3.x)
Some requirements did not work with python3.12, configobj was missing
Even when available, bugs in the OpenCL code can render the whole module useless if we cannot bypass OpenCL. The user can disable it: 1. with the environement variable PYTISSUE_FORCE_CPU = 1 2. with the function `forceCalculationOnCPU` It is not perfect, but at least it works.
|
There's already a |
…not the installed version.
The current bug on M2 with OpenCL is related to the smoothNormal kernel. I added a check for an environment variable to avoid smoothing, but I also had to modify intersection.c and now some other tests don't pass. The problem in intersection.c is related to passing __global pointers around but one function creates a __local INtersection, so passing the pointer makes no sense and the compiler complains. We need to find a way to either pass Intersection directly( which I tried to do) or always refer to the __global intersections array. I tried to make it work by passing Intersections around instead of passing the pointers, but not sure if this is functional (the tests don't pass but I dont have a proof they passed before).
Somewhat: it is a parameter to the method |
|
I have doubts about the |
1. CLProgram now checks immediately to see if program file exists 2. the fillRandomFloatBuffer kernel does not work as expected: same seeds give different resuts when using CLObject base classes, but fine when not using them. 3. New Buffer base class from which EmptyBuffer derives. 4. PLenty of OpenCL tests added (testGeneralOpenCL.py) to understand OpenCL
I figured out how to get the smoothNormals kernel to work: assing a local variable then assgning back to global array. The tests is fine now.
I fixed it, the kernel for smoothing is fine now. (See other branch) |
Test OpenCL code added to force_cpu branch
dccote
changed the title
| @@ -12,9 +13,16 @@ | |||
| else: | |||
| CONFIG = None | |||
|
|
|||
| def forceCalculationOnCPU(): | |||
| os.environ["PYTISSUE_FORCE_CPU"] = "1" | |||
| # warnings.warn("You can define PYTISSUE_FORCE_CPU=1 in your profile to avoid this call.") | |||
There was a problem hiding this comment.
This makes a total of 3 different endpoints to disable OpenCL, and I prefer this new apprach, so I think we should remove the parameter useHardwareAcceleration from Source.
I don't think "forceCPU" is the proper formulation, OpenCL is hardware-agnostic.
What about "PYTISSUE_DISABLE_OPENCL" and disableOpenCL() ?
Commented lines could be removed.
| def test_call_random_value_with_same_seeds_buffer_should_give_same_results(self): | ||
| """ | ||
| If we use the same seeds, then the two calculations should be the same. | ||
| This tests fails and should not, but the impact is minimal: the cude still runs fine |
| 'vtk>=9.2.2', | ||
| 'mayavi==4.8.1', | ||
| 'mayavi>=4.8.1', |
There was a problem hiding this comment.
We need to bump mayavi >= 4.8.2 for traitsui 8 compatibility.
https://docs.enthought.com/mayavi/mayavi/auto/changes.html
| for i, value in enumerate(valueBuffer): | ||
| self.assertEqual(value, i) | ||
|
|
||
| self.assertEqual(np.sum(local_sizes.hostBuffer), nWorkUnits) |
There was a problem hiding this comment.
I run into a few issues with these tests on windows and linux with OpenCL 3.0. Same results on an old python 3.8 env and fresh 3.10.
- Line 491:
AttributeError: 'Array' object has no attribute 'hostBuffer' - Line 425:
--> self.assertEqual(i%32, value)
AssertionError: 0 != cl.Array(32, dtype=uint32)
- And on windows only:
Line 427: test_trivial_kernel_local_id_set_workgroup_size :exit code 0xC0000094 (Integer division by zero)
General
The OpenCL code is fairly complex and needs better testing. This is a first pass at it, with a bug fix on the smoothNormal kernel
I started a new series of tests
testsCLProgram.pyandtestGeneralOpenCL.py.Three main bugs fixed:
propagate, but to be able to run the examples as they are, we need an option to disable OpenCL completely without modifying the code. In my case, the smoothingNormal kernel did not compile and everything failed (see below):forceCalculationOnCPU()testGeneralOpenCL.pywhen the basic PyOpenCL structures are used. There may be an error in CLObject.Smooth kernel bug details:
The kernel for
smoothNormalandsmoothNormalspasses__global Intersection*pointers around, but one function (_findClosestPolygonIntersection) actually does not generate an intersection in global space and on my machine, the OpenCL kernel does not compile. By assigning to a local variable and assigning back to the global array worked.