Allow separate queue for fft and ifft #17

isuruf · 2022-06-09T19:02:51Z

No description provided.

inducer · 2022-06-09T22:54:38Z

pyvkfft/opencl.py

@@ -155,20 +155,24 @@ def _make_config(self):
                                         int(self.use_lut), int(self.keepShaderCode),
                                         n_batch, skipx, skipy, skipz)

-    def fft(self, src: cla.Array, dest: cla.Array = None):
+    def fft(self, src: cla.Array, dest: cla.Array = None, queue: cl.CommandQueue = None):


Note that the array also carries a queue.

Which can be None right? Any suggestions on what to do?

Yes, it can be None, but I think it's reasonable to expect an array passed in here to have a valid queue.

If you were designing this API from scratch, I think the following fallback order would make sense:

queue (if passed)

src.queue and dest.queue (if given, must be the same)

I would not consider self.queue.

For backward compatibility, I would warn if self.queue disagrees with queue resulting from the above and say that self.queue is being used for now, but that that behavior is deprecated, with a certain fuse.

Note that the 'simple' fft API (pyvkfft.fft.fftn, pyvkfft.fft.ifftn...) already allows to transform just by passing an array - its queue will be used as default, and a different one can be supplied. This already allows to use different queues for different transforms.

If you were designing this API from scratch, I think the following fallback order would make sense:

queue (if passed)

src.queue and dest.queue (if given, must be the same)

I would not consider self.queue.

For backward compatibility, I would warn if self.queue disagrees with queue resulting from the above and say that self.queue is being used for now, but that that behavior is deprecated, with a certain fuse.

There can be some strange results when different queues are used, e.g. when using a different queue from the array's, the get() method will not wait for the fft to be finished.

I would still keep VkFFTApp's self.queue as the default (assuming it's a deliberate choice from the developer), but a warning about when different queues are used will be useful.

The queue choice is indeed debatable.. Preferring the array queues could also make sense.. It really depends on the use case.

The main use case is probably not a single array which is transformed with different queues at different times, but rather different arrays which are transformed by the same VkFFTApp. In which case using the array queues would indeed be easier...

vincefn · 2022-06-11T14:26:41Z

pyvkfft/test/test_fft.py

+                vd = []
+                vapp = []
+
+                with cl.CommandQueue(ctx) as cq:


Creating the queue using with may be dangerous here. This means the queue should not be used anymore outside of the withblock, but a reference to that queue will still persist in the VkFFTApp object. It is not used later in the code so it is fine here, but I think it's giving a dangerous example.

Yes, I agree that it's dangerous to keep persisting the queue in VkFFTApp. Latest pyopencl will give a warning if the queue is used outside of the with block though. inducer/pyopencl#561

vincefn · 2022-06-11T14:29:18Z

pyvkfft/test/test_fft.py

+        Test multiple FFT with queues different from the queue used
+        in creating the VkFFTApp
+        """
+        for dtype in (np.complex64, np.complex128):


You need to check for complex128/fp64 support

I don't understand. I'm already checking for complex128.

pyvkfft/opencl.py

vincefn · 2022-10-24T08:52:37Z

Hi @isuruf , @inducer - sorry for the long delay, as you can see I'm still undecided regarding which queue policy to follow, but it would be good to finalise this soon.

Using the queue supplied as a parameter to fft or ifft as the first choice is very clear.

However when no queue is passed to fft/ifft, I don't like the idea of having another two (or three) extra choices (the app, source or destiantion queues) - that will be confusing.

So in the end the two possibilities:

use the FFT app queue and ignore the array queues. This has the advantage of using the same approach as the cuda API (the arrays don't have a stream attribute AFAIK), and avoids issues with mismatched source and destination arrays queue, and has no backward-compatibility issues
use the source and destination arrays (hopefully matching) and deprecate the app's queue (as in the current PR). This allows to use the embedded queues and rely in with_queue but I don't use this personally

So I prefer 1) which seems clearer but if there is a widespread reliance on the embedded arrays queues then maybe 2) would be better. Let me know you thoughts.

inducer · 2022-12-28T06:03:43Z

I think the key point is that a single app object could validly be used, concurrently, by work going on in multiple queues, and in that case, the app's queue is an unhappy default. The queue that comes with the array is much more likely to be right.

As for input and output having separate queues, pyopencl only ever considers queues on inputs, at least for now. inducer/pyopencl#668. But if the information is used, I think it's wise to check that things match, at least in __debug__ mode.

vincefn · 2023-01-05T13:12:18Z

I think the key point is that a single app object could validly be used, concurrently, by work going on in multiple queues, and in that case, the app's queue is an unhappy default. The queue that comes with the array is much more likely to be right.

Ok, you probably have a broader view of usage of pyopencl queues, so let's do it your way (#17 (comment)). I'll try to add warning if queues don't match the app's - with a way to disable it.

I need to finish handling strides first...

* strides: Update notebooks Add support for F-ordered array in the simple fft interface (C2C and R2C). Fox use of opencl_platform in tests Fix tests using F-strided arrays. Allow to select the backend also for the basic test, and add the option to select the opencl platform. Check also for OSError in fftp.py. Should fix vincefn#21 Also display the platforma and device name along vkfft error check Non-C-contiguous arrays: add systematic test, fix DCT check, upodate some doc and outputs Add support for non C-ordered arrays (F-order or different types of strides). Supports both C2C and R2C (as long as the fast axis is transformed). Adresses vincefn#20 Update changelog More elegant re-ordering array during accuracy test, when necessary During accuracy tests, make sure the array after an R2C transform still has the fastest axis along the last dimension (X) - since this is not the case after a 2D numpy rfftn transform. Fixes vincefn#19

Fix backend check in pyvkfft-test

vincefn · 2023-01-06T08:58:54Z

pyvkfft/opencl.py

+                queue = src.queue
+
+        if queue != self.queue:
+            self.queue.finish()


I don't think it's a good idea to include a finish() here , as it would break asynchronous execution.
It may create issues but anyone juggling multiple queues should manage synchronisation explicitly.

vincefn · 2023-01-06T09:00:03Z

pyvkfft/opencl.py

+                elif src.queue is None:
+                    queue = dest.queue
+                else:
+                    warnings.warn("queue is not given and the source/dest arrays are not equal. "


I will switch to a RuntimeError instead, no queue supplied and different src and destination queues is a recipe for disasters

vincefn · 2023-01-06T09:01:57Z

Note that I am now editing this in https://github.com/vincefn/pyvkfft/tree/queue

…nd destination arrays differ and no queue is supplied to fft() or ifft() ; use array queues also for the simple fft interface ; expand tests. Fix use of config variables.

isuruf added 2 commits June 9, 2022 14:02

support separate queue for fft and ifft

7836edc

Add a test to check that different queues work

140545f

isuruf mentioned this pull request Jun 9, 2022

FFT using pyvkfft and use loopy callables inducer/sumpy#114

Merged

2 tasks

inducer reviewed Jun 9, 2022

View reviewed changes

vincefn reviewed Jun 11, 2022

View reviewed changes

Use array queues as a fallback first

767d1db

isuruf commented Jun 11, 2022

View reviewed changes

pyvkfft/opencl.py Outdated Show resolved Hide resolved

isuruf added 2 commits June 11, 2022 21:37

fix warnings

93169fa

Fix tests

1ade3d2

isuruf force-pushed the queue branch from 312751f to 1ade3d2 Compare June 12, 2022 02:44

finish queue

b81f9bc

vincefn added 2 commits January 6, 2023 08:57

Update tests for multiple opencl queues.

6af6505

Fix backend check in pyvkfft-test

vincefn reviewed Jan 6, 2023

View reviewed changes

OpenCL queues: update warnings ; switch to a RuntimeError if source a…

ca04b50

…nd destination arrays differ and no queue is supplied to fft() or ifft() ; use array queues also for the simple fft interface ; expand tests. Fix use of config variables.

vincefn merged commit 52a45f7 into vincefn:master Jan 6, 2023

This pull request was closed.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Allow separate queue for fft and ifft #17

Allow separate queue for fft and ifft #17

isuruf commented Jun 9, 2022

inducer Jun 9, 2022

isuruf Jun 9, 2022

inducer Jun 9, 2022

vincefn Jun 11, 2022

vincefn Jun 11, 2022

vincefn Jun 11, 2022

isuruf Jun 11, 2022

vincefn Jun 11, 2022

isuruf Jun 22, 2022

vincefn commented Oct 24, 2022

inducer commented Dec 28, 2022

vincefn commented Jan 5, 2023

vincefn Jan 6, 2023

vincefn Jan 6, 2023

vincefn commented Jan 6, 2023

Allow separate queue for fft and ifft #17

Allow separate queue for fft and ifft #17

Conversation

isuruf commented Jun 9, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vincefn commented Oct 24, 2022

inducer commented Dec 28, 2022

vincefn commented Jan 5, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vincefn commented Jan 6, 2023