How to compile OpenCV with CUDA support?

[EmilPi]

When trying to use mxnet.gluon.data.vision.transforms I get error
MXNetError: Operator _cvimresize is not implemented for GPU..

As a workaround, I am sending data to CPU, augmenting it and then sending back to GPU. This is horribly slow.
Do I need to compile OpenCV separately or should I modify config.cmake, or both? Is there any guide on this? What I found is very old.

[szha]

@EmilPi it would require integrating the opencv GPU support in mxnet image API to use it natively with GPU arrays. For now you can consider using the OpenCV python API with GPU support for data augmentation, and then use from_dlpack and to_dlpack_for_write for zero-copy loading and sharing of ndarray.

[EmilPi]

Ok, I will try. 谢谢！

[EmilPi]

Unluckily, I could not use OpenCV for that, but that means, I have to use other augmentation library.

[EmilPi]

Is there any example of sharing mxnet <-> OpenCV GpuMat data? What am I doing wrong?

I am trying to

1. get data
2. pass it to opencv GpuMat and augment
3. get MXNet GPU array from opencv using mxnet.ndarray.from_dlpack(gpu_mat)

But this code

import cv2
import mxnet as mx
import mxnet.ndarray as nd

ctx = mx.gpu()

w, h, d = 8, 8, 3
angle, scale = 45, .9
M = cv2.getRotationMatrix2D((w / 2, h / 2), angle, scale)
image = nd.random.uniform(shape=(w, h, d)).as_in_context(ctx)

image_ = image.asnumpy()  # Let's say this is loaded from a disk
cv_gpumat = cv2.cuda_GpuMat(image_)  # CPU -> GPU
cv_gpumat_augmented = cv2.cuda.warpAffine(cv_gpumat, M, (w, h))  # GPU -> GPU
# next line fails :(
nd_arr = nd.from_dlpack(cv_gpumat_augmented)  # GPU -> GPU

Fails with AssertionError: Invalid DLPack Tensor. DLTensor capsules can be consumed only once. error.

I also tried

image_cv = nd.to_dlpack_for_write(image)
cv_gpumat = cv2.cuda_GpuMat()
cv_gpumat.upload(image_cv)
image_cv_augmented = cv2.cuda.warpAffine(cv_gpumat, M, (w, h))
image_augmented = nd.from_dlpack(image_cv_augmented)

but this fails with

cv2.error: OpenCV(4.5.3-dev) :-1: error: (-5:Bad argument) in function 'upload'
> Overload resolution failed:
>  - arr is not a numpy array, neither a scalar
>  - Expected Ptr<cv::cuda::GpuMat> for argument 'arr'
>  - Expected Ptr<cv::UMat> for argument 'arr'