New Issue: Support '-g' with GPU support

e-kayrakli1 · May 9, 2022, 3:00pm

19774, "e-kayrakli", "Support '-g' with GPU support", "2022-05-09T14:58:44Z"

Support `-g` with GPU support

opened 02:58PM - 09 May 22 UTC

area: Compiler area: GPU Support

**This is moved from https://github.com/Cray/chapel-private/issues/2718.** T…hese are all my notes trying to get debugging support right: --- Currently we can't compile with `-g` because LLVM generates DWARF lines that are not compatible with `ptxas` until version 7.5. The incompatible lines cause syntax errors while running `ptxas`. As a long shot, I tried to change the dwarf language we use from C99 to C89 to see if it helps, but it doesn't. See "PTX ISA Notes" [in section 11.5.2 of PTX ISA reference](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#debugging-directives-section), we need both `label1-label2` syntax and negative numbers to be able to compile with `-g`. --- #### Notes about tools: - `cuda-memcheck` is deprecated and `compute-sanitizer` is the replacement for it. - We need `--report-api-errors no` to suppress warnings from failed API calls. There are some that are supposed to fail and we handle their errors in the runtime. - `cuda-gdb` can use the `memcheck` tool with `set cuda memcheck on`. You need to also do `set cuda api_failures hide` for the same reason as above. - We can improve these and turn them into an individual note once we have a better story for debugging. --- Hmm looks like PTX supports DWARF2: https://docs.nvidia.com/cuda/ptx-writers-guide-to-interoperability/index.html#generation-of-debug and https://lists.llvm.org/pipermail/llvm-dev/2017-November/118859.html The default DWARF for LLVM is 4. I wonder if we can drop that. --- I think the statement that the debug info are standard DWARF2 is a bit misleading. Because going back to the link in the OP, it clearly says that negative numbers are supported in PTX 7.5. And my initial attempt to fall back to C89 was right, because apparently C89 uses DWARF2. [Here is the DWARF version](https://github.com/chapel-lang/chapel/blob/main/third-party/llvm/llvm-src/include/llvm/BinaryFormat/Dwarf.def#L675). But it doesn't help.

This is moved from https://github.com/Cray/chapel-private/issues/2718.

These are all my notes trying to get debugging support right:

Topic		Replies	Views	Activity
Chapel on an open-science Nvidia GPU + ARM system Users	6	324	September 20, 2021
Cannot make gpu-enabled chapel Users	17	40	September 10, 2024

New Issue: Support '-g' with GPU support

Related Topics