Cuda array interface #43

SimonRit · 2024-12-19T17:15:08Z

Adding export code to #42

CMakeLists.txt

SimonRit · 2024-12-20T19:02:31Z

This is a detail but fc4cfa8 and 792d00e could be squashed into a single commit. This PR will be merged for CudaCommon v2, after a first release of CudaCommon v1.2, based on ITK v5.4.1 and/or v5.4.2.

LucasGandel · 2025-04-25T12:24:21Z

@SimonRit please review. The code has been updated to consume cuda_array_interface. Below is a quick code snippet to test:

import itk
import cupy as cp
x_gpu = cp.array([[0, 1, 2],[3, 4, 5]], dtype=cp.float32, order='C')

cuda_img = itk.cuda_image_from_cuda_array(x_gpu)
pixelIndex = itk.Index[2]()
pixelIndex[0] = 2
pixelIndex[1] = 1
print(f"5 = {cuda_img.GetPixel(pixelIndex)}")

SimonRit · 2025-09-04T08:51:37Z

TODO:

Doc page describing how to use cuda_image_from_cuda_array and the cuda array interface
Tests
Examples

SimonRit

Looks good in general. I think we need to be more general than CuPy and Torch as it can be used with many packages. We should add a few hyperlinks.

SimonRit · 2025-09-10T09:07:20Z

cuda_array_interface.md

+
+## Overview
+
+ITK's `CudaImage` implements the `__cuda_array_interface__` protocol, allowing it to be used directly as a CuPy array or PyTorch tensor without copying data to/from the GPU.


Maybe make __cuda_array_interface__ an hyperlink to https://numba.pydata.org/numba-doc/0.43.0/cuda/cuda_array_interface.html

SimonRit · 2025-09-10T09:10:39Z

cuda_array_interface.md

@@ -0,0 +1,26 @@
+# CUDA Array Interface in ITK
+
+The CUDA Array Interface allows seamless interoperability between ITK's `CudaImage` and GPU arrays from libraries like CuPy and PyTorch. This enables zero-copy data sharing, efficient GPU operations, and integration with deep learning pipelines.


I would indicate in parentheses the structure names in the libraries, i.e., CuPy (cupy.ndarray) and PyTorch (torch.Tensor).

SimonRit · 2025-09-10T09:12:22Z

cuda_array_interface.md

+
+ITK's `CudaImage` implements the `__cuda_array_interface__` protocol, allowing it to be used directly as a CuPy array or PyTorch tensor without copying data to/from the GPU.
+
+- **Zero-copy access**: Convert `CudaImage` to CuPy/Torch views instantly.


I would use cupy.ndarray / torch.Tensor.

SimonRit · 2025-09-23T14:51:57Z

@LucasGandel I changed your wrapping/itkCudaImageFromCudaArray.py code in the last forced push because I don't think it makes sense to set buffered region without setting the other regions.

@axel-grc Your examples and test, presumably based on this, have also been modified. Note that the same code with an itk.Image would crash with a seg fault.

Prior to this commit CudaDataManager::GetGPUBufferPointer() returned a pointer to the GPU buffer pointer. Its return type should have been void**. Every usage of this method across RTK and CudaCommon dereference the returned pointer pointer, so just return the buffer pointer directly as it required for future commits adding support for __cuda_array_interface__

Adapt ITK's PyBuffer approach to inject __cuda_array_interface__ in wrapped code. Closes RTKConsortium#10

This allows modules like RTK to easily reuse files from the CudaCommon sources, e.g. CudaImage.i.init and CudaImage.i.in

Add function to create a CudaImage from an object providing __cuda_array_interface__

Allows the CPU buffer to be allocated automatically when only the GPU pointer is set on the image.

Suppresses the message: ``` ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. ```

SimonRit · 2025-10-02T08:11:25Z

Thanks again @LucasGandel, time to merge!

SimonRit force-pushed the cuda-array-interface branch from 9e356cf to e176586 Compare December 19, 2024 17:47

SimonRit commented Dec 20, 2024

View reviewed changes

CMakeLists.txt Outdated Show resolved Hide resolved

LucasGandel mentioned this pull request Dec 20, 2024

Cuda array interface #42

Closed

6 tasks

SimonRit added this to the CudaCommon 2.0 milestone Mar 24, 2025

SimonRit force-pushed the cuda-array-interface branch from fc4cfa8 to a6729f8 Compare March 25, 2025 14:40

SimonRit force-pushed the cuda-array-interface branch 5 times, most recently from f7a5762 to 214cfbe Compare April 5, 2025 11:27

LucasGandel force-pushed the cuda-array-interface branch from 214cfbe to dd33430 Compare April 25, 2025 12:21

axel-grc force-pushed the cuda-array-interface branch from 7e8572b to 94dced1 Compare September 9, 2025 11:23

SimonRit commented Sep 10, 2025

View reviewed changes

axel-grc force-pushed the cuda-array-interface branch 2 times, most recently from ee768ae to 22bdfbb Compare September 12, 2025 14:52

SimonRit force-pushed the cuda-array-interface branch 3 times, most recently from 5b1a577 to e19ca53 Compare September 14, 2025 20:50

axel-grc force-pushed the cuda-array-interface branch 4 times, most recently from bb453c4 to 726ff1b Compare September 15, 2025 09:32

SimonRit force-pushed the cuda-array-interface branch 2 times, most recently from ba4828d to e19ca53 Compare September 17, 2025 06:12

SimonRit mentioned this pull request Sep 17, 2025

ENH: Upgrade CI CUDA Python packages to CudaCommon v2 RTKConsortium/RTK#813

Merged

axel-grc force-pushed the cuda-array-interface branch from e19ca53 to 726ff1b Compare September 18, 2025 12:16

SimonRit force-pushed the cuda-array-interface branch from 726ff1b to cb1e549 Compare September 23, 2025 11:00

SimonRit force-pushed the cuda-array-interface branch from cb1e549 to 611b407 Compare September 23, 2025 14:47

Simon Rit and others added 20 commits September 24, 2025 21:53

COMP: GHA: Use ITK v5.4 release to generate Python packages

05437b8

ENH: Add support for __cuda_array_interface__

0caa810

Adapt ITK's PyBuffer approach to inject __cuda_array_interface__ in wrapped code. Closes RTKConsortium#10

ENH: Add support for other dimensions in cuda array interface

4d88517

ENH: Add support for vector types in cuda array interface

57d2396

ENH: Add CudaDataManager::GetGPUBufferPointerPtr

89d9daa

ENH: Export CudaCommon_SOURCE_DIR in CudaCommon targets file

be21f9a

This allows modules like RTK to easily reuse files from the CudaCommon sources, e.g. CudaImage.i.init and CudaImage.i.in

ENH: Expose flag to release GPU memory when dirty

252d9e8

ENH: Define and export version in configuration file

3242053

ENH: Set CudaCommon version in the exported CMake code

a0fd540

STYLE: Use lowercase CudaCommon_VERSION variables

8ff52ee

ENH: Enable manual setting of the GPU buffer pointer

fbb2f8c

ENH: Add python function cuda_image_from_cuda_array

b2f2ae3

Add function to create a CudaImage from an object providing __cuda_array_interface__

ENH: Automatically allocate CPU buffer on update

b9c3162

Allows the CPU buffer to be allocated automatically when only the GPU pointer is set on the image.

COMP: Fix ambiguous call to ImageType::Superclass

06fd79b

ENH: Add examples for CudaArrayInterface

fa2a7d6

ENH: Add tests for CudaArrayInterface

fc2654e

DOC: Add documentation for CudaArrayInterface

8199f44

ENH: Install cupy and torch in the CI

66b9f89

BUG: Uninstall existing ITK packages in GHA env before testing

2bef773

Suppresses the message: ``` ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. ```

SimonRit force-pushed the cuda-array-interface branch from 611b407 to 2bef773 Compare September 24, 2025 20:54

SimonRit mentioned this pull request Sep 25, 2025

ENH: Copy .in and .init files of modules for use in other modules InsightSoftwareConsortium/ITKPythonPackage#296

Merged

SimonRit merged commit d8dfe12 into RTKConsortium:main Oct 2, 2025
12 checks passed

SimonRit deleted the cuda-array-interface branch October 14, 2025 09:53

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Cuda array interface #43

Cuda array interface #43

Uh oh!

SimonRit commented Dec 19, 2024

Uh oh!

Uh oh!

SimonRit commented Dec 20, 2024

Uh oh!

LucasGandel commented Apr 25, 2025

Uh oh!

SimonRit commented Sep 4, 2025

Uh oh!

SimonRit left a comment

Uh oh!

SimonRit Sep 10, 2025

Uh oh!

SimonRit Sep 10, 2025

Uh oh!

SimonRit Sep 10, 2025

Uh oh!

SimonRit commented Sep 23, 2025

Uh oh!

SimonRit commented Oct 2, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants


		## Overview

		ITK's `CudaImage` implements the `__cuda_array_interface__` protocol, allowing it to be used directly as a CuPy array or PyTorch tensor without copying data to/from the GPU.

		@@ -0,0 +1,26 @@
		# CUDA Array Interface in ITK

		The CUDA Array Interface allows seamless interoperability between ITK's `CudaImage` and GPU arrays from libraries like CuPy and PyTorch. This enables zero-copy data sharing, efficient GPU operations, and integration with deep learning pipelines.


		ITK's `CudaImage` implements the `__cuda_array_interface__` protocol, allowing it to be used directly as a CuPy array or PyTorch tensor without copying data to/from the GPU.

		- Zero-copy access: Convert `CudaImage` to CuPy/Torch views instantly.

Cuda array interface #43

Cuda array interface #43

Uh oh!

Conversation

SimonRit commented Dec 19, 2024

Uh oh!

Uh oh!

SimonRit commented Dec 20, 2024

Uh oh!

LucasGandel commented Apr 25, 2025

Uh oh!

SimonRit commented Sep 4, 2025

Uh oh!

SimonRit left a comment

Choose a reason for hiding this comment

Uh oh!

SimonRit Sep 10, 2025

Choose a reason for hiding this comment

Uh oh!

SimonRit Sep 10, 2025

Choose a reason for hiding this comment

Uh oh!

SimonRit Sep 10, 2025

Choose a reason for hiding this comment

Uh oh!

SimonRit commented Sep 23, 2025

Uh oh!

SimonRit commented Oct 2, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants