After converting my custom ConvNet to .dlc format, I ran inference on it using "snpe-net-run" via adb on my SD820 device. Everything works when running on the CPU, I verified that the outputs are as expected. But when I used the "--use_gpu" flag (while keeping everything else exactly the same) I got the following error:
error_code=811; error_message=GPU tensor dimensions are invalid. Host tensor dimensions do not match.; error_component=GPU Runtime; line_no=463; thread_id=-220814028
I did not get this error when I ran inference on the Inception_v3 model from the tutorial.
Hi. I'm not sure why it would work on the CPU and not on the GPU with everythig else the same, but that error message is usually a result of passing an input tensor to the execute() function that is different than what the model expects (although it can occur in other places as well). Here is one thing you can try. Call the (SNPE::)getInputDimensions() function before calling execute and see if it returns the same dimensions that you are setting in your input tensor. If not (and the getInputDimensions() results are what you expect), set the dimensions of your input tensor to these values and then try calling execute().
Let me know how that goes. Thanks.
Hi bspinar,
It sounds like your recommendation is for me add a line in a C++ program. I'm not at that stage yet: I'm just running everything via the binaries and using adb to communicate with the device.
I'm guessing there's no way to troubleshoot without messing with the C++ API?
Sorry for the slow response. Unfortunately for that kind error, it's difficult to debug without doind some of the things I suggested.
Hi, I got the same problem when I trying to execute my network on the GPU of a snapdragon 835 chip. Here is the stack trace:
is anyone know the SNPE error message table?
where can i get the table, Developer need debug activity/fragment and SNPE Runtime error,
but i can't find any information about the error code and how to fix.
for example:
error_code=910; error_message=undefined error. failed to execute network
and there are 835 cpu phones but can't run snpe in gpu/dsp mode.
Hi, I also have the samp issue with Snapdragon 820 by running snpe-net-run command with --use-gpu options. In case of CPU or DSP, it works fine. Reproducible version of SNPE: 1.19.2, 1.18.0
Error message: error_code=811; error_message=GPU tensor dimensions are invalid.; error_component=GPU Runtime; line_no=94; thread_id=548252692480
Does SNPE team work on this issue?