Illegal instruction when unsing debugger runtime

Questions
1

I tried to use the debugger to see the bottleneck of the module. I followed the instructions in tutorial.

set the USE_GRAPH_RUNTIME_DEBUG flag to ON make In frontend script file instead of from tvm.contrib import graph_runtime import the debug_runtime from tvm.contrib.debugger import debug_runtime as graph_runtime

But I encountered an error

Illegal instruction (core dumped)

I use the example code in the tutorials. I tried https://docs.tvm.ai/_downloads/83dedc6352b4016772e17480ef01345d/deploy_model_on_rasp.py and https://docs.tvm.ai/_downloads/f83f0c3da8a2ab10657c61e034b7218d/from_pytorch.py, and my own code. All have the same error as above. But all of them work if from tvm.contrib import graph_runtime rather than debugger runtime.

I have no idea how to solve it. Thanks

2

This maybe a problem with dependent shared libraries. Try the following. In your ~/.gdbinit add the line:

handle SIGILL nostop noprint

3

Thanks for your help. But it still raise the same error after I undating gdbinit file.

I think it may be something wrong when make and build tvm. Even if I use tvm.contrib import graph_runtime, it still raise the error.

Then I set set(USE_GRAPH_RUNTIME_DEBUG OFF) back, and cmake, then make again. But I still encountered illegal instruction error.

4

Try with a fresh build in debug mode, i.e.

cd build
cmake -DCMAKE_BUILD_TYPE=Debug ..
make

Then in GDB, if you encounter the illegal exception run again after explicitly disabling the sigill signal.

$ gdb <however you call gdb>
... exception
gdb> handle SIGILL nostop noprint
gdb> run
5

If you’re doing this on a raspberry pi over RPC make sure that USE_GRAPH_RUNTIME_DEBUG is on on both your host, and on the RPC server (e.g. rpi device).