GPU Memory Leak

[weichengkai]

Hi, After running the following code,GPU memory leak. silx version=0.13.2

import numpy
import silx.opencl.sift as sift

def siftTest(image):
    gpuSiftPlan = sift.SiftPlan(template=image, devicetype="GPU")
    gpuSiftPlan.keypoints(image)

image = numpy.zeros((100, 100, 3), dtype=numpy.uint8)
for i in range(500000):
    siftTest(image)
    print(i)

[pierrepaleo]

Hi @weichengkai

The purpose of a SiftPlan is to be created once for a given image shape, and then to be used multiple times over different images. It is similar to the "FFTW plan" concept. Of course creating it each time will consume a lot of memory.

So instead, I would rather use:

import numpy
import silx.opencl.sift as sift

image = numpy.zeros((100, 100, 3), dtype=numpy.uint8)
gpuSiftPlan = sift.SiftPlan(template=image, devicetype="GPU")
for i in range(500000):
    gpuSiftPlan.keypoints(image)
    print(i)

[weichengkai]

image = numpy.zeros((100, 100, 3), dtype=numpy.uint8) just for testing Actually, Images come from local files and come in different image shape. Function siftTest get a image and return image's keypoint.Like following code

def siftTest(image):
    gpuSiftPlan = sift.SiftPlan(template=image, devicetype="GPU")
    kps = gpuSiftPlan.keypoints(image)
    return kps

When I called the function siftTest many times, I found that GPU memory could not be released. What can I do to free memory before I return kps?

[kif]

If your image size differs, you will have to destroy the plan each time ...

[weichengkai]

The plan instantiated within function siftTest，after the function runs, the plan is automatically destroyed.
The problem now is that GPU memory is not released after the plan is destroyed.

[kif]

pyopencl keeps the context in caches but it should recycle them. We could destroy the CommandQueue but it is not enough.