CUDA: Add overloads generated by specialization to the current dispatcher. #9106
Conversation
Dispatcher specialization takes actual arguments, not types. So although these tests did technically test the specialization mechanism, the way in which they were doing it was extremely weird because a kernel would never actually be launched with a type as an argument.
Adding overloads to the current dispatcher when specializing saves recreating the overload for subsequent calls to the unspecialized dispatcher that would have used the overload. This also has the side effects of making overloads available after a call to `ForAll()` on a dispatcher.
|
gpuci run tests |
|
gpuci run tests |
|
The PR still works with my cuDF example for |
| f_f32a_f32a = f.specialize(f_arg, f_arg) | ||
| self.assertEqual(len(f.overloads), 1) | ||
| self.assertIs(f_f32a_f32a.overloads[f_arg_ty, f_arg_ty], |
There was a problem hiding this comment.
| f_f32a_f32a = f.specialize(f_arg, f_arg) | |
| self.assertEqual(len(f.overloads), 1) | |
| self.assertIs(f_f32a_f32a.overloads[f_arg_ty, f_arg_ty], | |
| f_f32f_f32f = f.specialize(f_arg, f_arg) | |
| self.assertEqual(len(f.overloads), 1) | |
| self.assertIs(f_f32f_f32f.overloads[f_arg_ty, f_arg_ty], |
suggest naming scheme consistency with F-ordered input.
| f_f32a_f32a = f.specialize(float32[:], float32[:]) | ||
| # 'F' order specialization | ||
| f_arg = np.zeros((2, 2), order='F') | ||
| f_f32a_f32a = f.specialize(f_arg, f_arg) |
There was a problem hiding this comment.
| f_f32a_f32a = f.specialize(f_arg, f_arg) | |
| f_f32f_f32f = f.specialize(f_arg, f_arg) |
suggestion to match ordering F.
stuartarchibald
added
4 - Waiting on author
|
This pull request is marked as stale as it has had no activity in the past 3 months. Please respond to this comment if you're still interested in working on this. Many thanks! |
|
Still planning to finish this when I get a moment. |
|
This pull request is marked as stale as it has had no activity in the past 3 months. Please respond to this comment if you're still interested in working on this. Many thanks! |
github-actions
bot
added
the
abandoned - stale
|
Missed this due to being on PTO, going to reopen it to keep it relevant (to get round to one day). |
gmarkall
removed
the
abandoned - stale
Adding overloads to the current dispatcher when specializing saves recreating the overload for subsequent calls to the unspecialized dispatcher that would have used the overload.
This also has the side effects of making overloads available after a call to
ForAll()on a dispatcher.This provides an alternative to the implementation in #9057, which also aimed to make overloads available after a
ForAll()call, but did it by changing the wayForAll()worked and potentially increasing latency for calls on specialized dispatchers (of which there are some use cases in cuDF).Whilst looking into this I also noticed that all the CUDA dispatcher specialize tests worked in an "unexpected" way and didn't provide a good pattern to follow for this test added in this PR, so they are all fixed up in this PR too.