Debugging a core dump

[sei-jgwohlbier]

Hi,
I have successfully run the simple_abs and mnist examples on a gpu, but my own model dumps core. Any hints on how to debug this?

Also, I'm not sure how to specify --input for my network as it has two input tensors. That might be my next issue, but I don't think the runtime is getting that far at the moment. My function signature is func.func @forward(%arg0: tensor<1x2xf32>, %arg1: tensor<1x1xf32>) -> tensor<1xf32>

iree-run-module \
--device=cuda \
--module=mymodel.vmfb \
--function=forward \
--input="1x2xf32,1x1xf32=0.5,0.4"
Aborted (core dumped)
make: *** [Makefile:37: run] Error 134

[ScottTodd]

Each input should be specified separately, like --input=1x2xf32 --input=1x1xf32=0.5,0.4. See https://iree.dev/developers/general/developer-overview/#iree-run-module for more examples

For core dumps, you'll want to build with debug symbols then consider running under a debugger. Either of those should get you useful error messages and stack traces.

[sei-jgwohlbier]

Thanks! That indeed fixed the core dump and the model appears to execute correctly!

I'm planning to use iree to generate ptx and link that ptx to another application that uses nvcc and clang to generate GPU code. If you have any thoughts about that I'd love to hear them. I'm still mulling over how I'll make the function call from the host code to get the proper signature for linking to the iree generated ptx.

Thanks!