Operation not permitted with HDDL plugin

[gdumont82]

Hi,

I am trying to convert YOLOv5 models to ONNX and then to OpenVINO to use them with the HDDL plugin.

I noticed that if my images are larger than 352 x 352 I get errors like:

[ion_ioctl][82]ioctl c0484900 failed with code -1: Operation not permitted, fd(17)
ion alloc failed size=4915200
[ion_ioctl][82]ioctl c0484900 failed with code -1: Operation not permitted, fd(17)
ion alloc failed size=13107200

The output seems right but I wonder if I should be worried about these error messsages?

I made a small repro example where I create a conv net with a single layer and vary the input size from 320 to 640. All networks have the error above except the one with a 320 x 320 input.

Running test_conv_320.xml
Running test_conv_480.xml
[ion_ioctl][82]ioctl c0484900 failed with code -1: Operation not permitted, fd(17)
ion alloc failed size=7372800
Running test_conv_512.xml
[ion_ioctl][82]ioctl c0484900 failed with code -1: Operation not permitted, fd(17)
ion alloc failed size=8388608
Running test_conv_640.xml
[ion_ioctl][82]ioctl c0484900 failed with code -1: Operation not permitted, fd(17)
ion alloc failed size=4915200
[ion_ioctl][82]ioctl c0484900 failed with code -1: Operation not permitted, fd(17)
ion alloc failed size=13107200

Is this normal?

You can find the code of the example attached. 

Thanks

Guillaume

[gdumont82]

Ok some more investigation leads me to think that the problem is related to the size of the output. If we change the number of filters in the convolution layer from 32 to 3 then only the model at 640 x 640 generates the error.

[gdumont82]

I was able to reproduce this error with the SSD512 model by removing everything except for the first few layers up to "conv1_1/WithoutBiases". 

[Zulkifli_Intel]

Hello Guillaume,

Greetings to you.

The error message indicates that the ION buffer is not large enough to handle large tensors, so it will use shared memory which causes such problems.

Here are some suggestions:

-            Release the shared memory by referring to this documentation.

-            Use the lower size of the model when the HDDL device is selected.

Hope this information helps.

Sincerely,

Zulkifli 

[Zulkifli_Intel]

Hello Guillaume,

This thread will no longer be monitored since we have provided a solution. If you need any additional information from Intel, please submit a new question.

Sincerely,

Zulkifli