Support for cupyx.scipy.linalg by quasiben · Pull Request #7563 · dask/dask

quasiben · 2021-04-15T22:30:59Z

Closes #Linalg lstsq fails on cupy backed arrays #7537
Tests added / passed
Passes black dask / flake8 dask / isort dask

dask/array/linalg.py

dask/array/utils.py

dask/array/linalg.py

dask/array/utils.py

jrbourbeau

Just wanted to check in here. Is this something we'd like to get into the release tomorrow, or are we okay waiting until the one after that?

quasiben · 2021-04-22T20:07:50Z

No, this can wait. But thanks for checking in

dask/array/core.py

jakirkham · 2021-04-26T20:57:10Z

dask/array/utils.py

+    if is_cupy_type(a):
+        import cupyx.scipy.linalg
+
+        func = getattr(cupyx.scipy.linalg, func_name)
+    else:
+        import scipy.linalg
+
+        func = getattr(scipy.linalg, func_name)


I wonder if we could use Dispatch for this kind of thing

dask/dask/sizeof.py

Lines 17 to 28 in c5633c2

sizeof = Dispatch(name="sizeof")

@sizeof.register(object)

def sizeof_default(o):

return getsizeof(o)

@sizeof.register(bytes)

@sizeof.register(bytearray)

def sizeof_bytes(o):

return len(o)

We could also then do lazy registration with this

dask/dask/sizeof.py

Lines 97 to 106 in c5633c2

@sizeof.register_lazy("numpy")

def register_numpy():

import numpy as np

@sizeof.register(np.ndarray)

def sizeof_numpy_ndarray(x):

if 0 in x.strides:

xs = x[tuple(slice(None) if s != 0 else slice(1) for s in x.strides)]

return xs.nbytes

return int(x.nbytes)

I think that's a good idea if this grows more. As it stands now, this PR only adds solve_triangular. I tried adding in lu but ran into a number of issues to due to lack of int64 support in the cupy implementation

FWIW I had some code here that might be relevant ( jakirkham@81adbc8 ). Though feel free to skip it if you would rather

Also thinking about this in the context of dispatch refactoring work going on around Arrays and DataFrames ( #7505 ) ( #7503 )

I tried adding in lu but ran into a number of issues to due to lack of int64 support in the cupy implementation

Just to clarify is that suppose to be float64? While we may be able to provide int64 arrays, expect these would likely need to be casted to float64 and then the result would also be in float64. Though maybe I'm misunderstanding something here

No, I think we need int64 support for lu functionality. Or maybe a conversion is what scipy lu does and I'm unaware of it

pentschev

Overall LGTM, thanks @quasiben .

jsignell · 2021-05-03T14:18:41Z

Seems like this is good to merge right @quasiben?

quasiben · 2021-05-03T14:31:32Z

Yeah, this can be merged now

quasiben added 2 commits April 15, 2021 15:28

add safe function for cupyx/scipy linalg solve_triangular

0f9452b

lint

381ff29

github-actions bot added the array label Apr 15, 2021

quasiben changed the title ~~add safe function for cupyx/scipy linalg solve_triangular~~ CuPy support for solve_triangular Apr 15, 2021

quasiben requested a review from pentschev April 15, 2021 22:31

jrbourbeau reviewed Apr 15, 2021

View reviewed changes

dask/array/linalg.py Outdated Show resolved Hide resolved

quasiben added 2 commits April 15, 2021 17:13

use array_safe instead of np.array

3261173

lint

dfe857a

pentschev reviewed Apr 16, 2021

View reviewed changes

dask/array/utils.py Outdated Show resolved Hide resolved

dask/array/utils.py Outdated Show resolved Hide resolved

add support for cupy cholesky and matmul

23cb130

pentschev suggested changes Apr 19, 2021

View reviewed changes

remove zeros_safe and cleanup

8679072

jrbourbeau mentioned this pull request Apr 20, 2021

2021.04.1 release dask/community#150

Closed

3 tasks

jrbourbeau reviewed Apr 22, 2021

View reviewed changes

quasiben added 2 commits April 26, 2021 09:20

add lu and generalize cupyx linalg scipy funcs

dc35ded

remove lu and _zeros_safe

a0d65ce

jrbourbeau reviewed Apr 26, 2021

View reviewed changes

dask/array/core.py Outdated Show resolved Hide resolved

remove breakpoint

357cfcb

jakirkham requested review from jrbourbeau and pentschev April 26, 2021 20:48

jakirkham reviewed Apr 26, 2021

View reviewed changes

quasiben added 3 commits April 26, 2021 14:04

fix all test in test_cupy

85791ea

move matmul back to routines

534d1a5

Merge remote-tracking branch 'upstream/main' into fix/cupy-lstsq

b13833c

pentschev approved these changes Apr 27, 2021

View reviewed changes

quasiben changed the title ~~CuPy support for solve_triangular~~ Support for cupyx.scipy.linalg Apr 29, 2021

jsignell merged commit a81fed4 into dask:main May 3, 2021

quasiben deleted the fix/cupy-lstsq branch May 3, 2021 19:09

pentschev mentioned this pull request May 6, 2021

NEP 35: Finalize like= argument behaviour before 1.21 release numpy/numpy#17075

Closed

quasiben mentioned this pull request May 7, 2021

Linalg lstsq fails on cupy backed arrays #7537

Closed

jsignell mentioned this pull request Jul 20, 2021

Linalg solve_triangular fails on cupy backed arrays #7536

Closed

ravwojdyla mentioned this pull request Dec 2, 2021

improved the efficiency of matmul() by completely removing concatenation #8423

Merged

3 tasks

	sizeof = Dispatch(name="sizeof")


	@sizeof.register(object)
	def sizeof_default(o):
	return getsizeof(o)


	@sizeof.register(bytes)
	@sizeof.register(bytearray)
	def sizeof_bytes(o):
	return len(o)

	@sizeof.register_lazy("numpy")
	def register_numpy():
	import numpy as np

	@sizeof.register(np.ndarray)
	def sizeof_numpy_ndarray(x):
	if 0 in x.strides:
	xs = x[tuple(slice(None) if s != 0 else slice(1) for s in x.strides)]
	return xs.nbytes
	return int(x.nbytes)

Uh oh!

Conversation

quasiben commented Apr 15, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jrbourbeau left a comment

Choose a reason for hiding this comment

Uh oh!

quasiben commented Apr 22, 2021

Uh oh!

Uh oh!

jakirkham Apr 26, 2021

Choose a reason for hiding this comment

Uh oh!

quasiben Apr 26, 2021

Choose a reason for hiding this comment

Uh oh!

jakirkham Apr 26, 2021

Choose a reason for hiding this comment

Uh oh!

jakirkham Apr 27, 2021

Choose a reason for hiding this comment

Uh oh!

jakirkham Apr 27, 2021

Choose a reason for hiding this comment

Uh oh!

quasiben May 3, 2021

Choose a reason for hiding this comment

Uh oh!

pentschev left a comment

Choose a reason for hiding this comment

Uh oh!

jsignell commented May 3, 2021

Uh oh!

quasiben commented May 3, 2021

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

quasiben commented Apr 15, 2021 •

edited

Loading