I tried to to execute the segmentation sample with my own network on NUC, and it can successfully generated the output. However when I tried to execute the program on an Intel(R) VCA node, it prompted me with the error "memory allocation failed: exceeded global device memory". The graphics model of my NUC is Intel(R) Iris Pro Graphics 580 (Skylake GT4e) with 256 MB of memory size, and the graphics model of the VCA node is Intel(R) Iris Pro Graphics P580 with the same memory size.
I don't think the issue is caused by the size of my network model, as the program still works even after I changed the size of input image from 640*480 to 2048*1024. I also googled the error code but there doesn't seem to be similar issues. I have attached the model I used, custom kernel I implemented, as well as the main.cpp that I modified from the original file in the segmentation sample.
The intention for me to modify the main program of segmentation sample is that, as I can see the ArgMax operation is included in the sample and executed on CPU, I implemented the OpenCL kernel for ArgMax and moved the operation to be part of the network model, and such change achieved significant performance boost from 220 ms to 160 ms on my NUC.
Any help will be highly appreciated.
This is a bug and it is being investigated. I will update you with the outcome of when the fix can be rolled in the next release or a workaround if available.
I have the same issue. Running on VCA2, CentOS.
Please note that the OpenCL driver is getting no mem allocation..
I believe the issue is not with the sample (all of them fail) but the driver.
"Global memory size 0"
Dear rongrong, wang
This is a rather old thread. Can you tell me exactly what problem you are having ? It sounds like an OpenCL or graphics driver issue, which would be better served by posting to the following communities:
This forum is dedicated to OpenVino (Model Optimizer and Inference Engine).