[CUDA] Lost debug information when compiling CUDA code

Project Infrastructure LLVM Dev List Archives
1

Hi,

I needed to debug some CUDA code in my project; however, although I used -g when compiling the source code, no source-level information is available in cuda-gdb or cuda-memcheck.

Specifically, below is what I did:

  1. For a CUDA file a.cu, generate IR files: clang++ -g -emit-llvm --cuda-gpu-arch=sm_35 -c a.cu;
  2. Instrument the device code a-cuda-nvptx64-nvidia-cuda-sm_35.bc (generated in the previous step), inserting a call to a hook function before each device memory access. The hook function is defined in another file, b.cu. Let’s say we get a file named intrumented-a-device.bc after this step;
  3. Generate IR files for b.cu: clang++ -g -emit-llvm --cuda-gpu-arch=sm_35 -c b.cu;
  4. Link instrumented-a.device.bc with the device code generated for b.cu: llvm-link intrumented-a-device.bc b-cuda-nvptx64-nvidia-cuda-sm_35.bc -o ab-device.bc;
  5. Use llc, ptxas & fatbinary on ab-device.bc to get ab-device.ptx, ab-device.o & ab-device.fatbin;
  6. Call clang again the generate the host object file ab.o, with ab-device.o & ab-device.fatbin embedded;
  7. Link against libraries and get the final binary: a.out.

The binary a.out fails with an exception I when run it; but when I try to debug it with cuda-gdb or cuda-memcheck, no source information was available. Why?

Thanks!
Yuanfeng Peng

2

Is source level information present in generated IR files?

3

Yes, I’ve inspected all the intermediate IR files and source level information is present in them.

------------------ Original ------------------

4

Try dumping .out file using cuobjdump and see if it has .debug_info section.
If it does not have then PTX compilation skipped debug information generation.
If it does have then there is some other issue.

5

Hi,

I needed to debug some CUDA code in my project; however, although I used -g when compiling the source code, no source-level information is available in cuda-gdb or cuda-memcheck.

The NVPTX backend in LLVM does not support DWARF emission.

-Tom

6

Hi Tom,

Thanks for the response! I wonder whether there’s any other way that I can get the source level information when using cuda-gdb to debug a binary compiled by LLVM? Or is there any way to feed the IR files to NVCC and get a binary with the debug info I need?

Thanks!
Yuanfeng

------------------ Original ------------------

7

Hi Tom,

Thanks for the response! I wonder whether there's any other way that I can get the source level information when using cuda-gdb to debug a binary compiled by LLVM? Or is there any way to feed the IR files to NVCC and get a binary with the debug info I need?

I don't know, I think you will have to ask NVIDIA about this.

-Tom