--dump-assembly and --dump-optimized don't do anything

[jeremyherbert]

Hi,

I'm running the latest numbapro (numba: 0.15.1, numbapro: 0.16.0) from anaconda with python3.4. If I use either the --dump-assembly or the --dump-optimized flags, I don't actually get any assembly output, and according to my quick strace it is not writing files anywhere. I'm running https://github.com/ContinuumIO/numbapro-examples/blob/master/cuda_memory/pinned.py with

numba --dump-optimized pinned.py

What am I doing wrong?

[sklam]

The flags were not picked up by the cuda backend in 0.15.1. It is partially fixed in master (dump-assembly doesn't work #889).

In addition, the dump will print to screen if it works.

[sklam]

Status on master:
dump-optimized works (but numba does not use optimize cuda backend code; it is inside libnvvm)
dump-assembly is printing the final llvm IR, not the PTX. I will be fixing this in a PR that I am working on.

[stuartarchibald]

As of 0.37, using the noted sample, --dump-optimized and --dump-assembly both work. Closing. If this is still a problem please reopen.

[AndiH]

Using 0.39.0, there's still some issues with dumping LLVM information. Not all environment variables seem to work. Namely NUMBA_DUMP_ANNOTATION, NUMBA_DUMP_ASSEMBLY, NUMBA_DUMP_LLVM, and NUMBA_DUMP_OPTIMIZED.

Using same pinned.py example file from above:

$ for e in BYTECODE CFG IR ANNOTATION ASSEMBLY LLVM FUNC_OPT OPTIMIZED; do echo $e; eval "NUMBA_DUMP_$e"=1 numba pinned.py | grep -v regular | grep -v pinned | wc -l; done
BYTECODE
24
CFG
22
IR
182
ANNOTATION
0
ASSEMBLY
0
LLVM
0
FUNC_OPT
376
OPTIMIZED
0

Browsing the source code, I see that these exact environment variables are used in numba_entry.py:L255-257 and set to their command-lines alternatives. To me, this sound like a bug. IMHO they should only be set to their command-line counterparts if the counterparts themselves are defined – and even then this behaviour should be noted somewhere.

[stuartarchibald]

@AndiH thanks. Please open a new ticket to track the specific DUMP_ options that are not working. I believe they are ok for CPU targets so it's probably something in the CUDA backend that either needs to raise as they aren't supported and be declared as such, or actually work.

[AndiH]

Done! See #3105!

[stuartarchibald]

Great, thanks @AndiH