tuner: Fix missing config.h include

[rajachan]

Without this, we were incorrectly including neuron headers instead of CUDA ones when HAVE_CUDA was not defined. Also fail compilation if neither definition was found rather than having a fall-through case.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[aws-nslick]

Note that #351 has the same fix and integrates with clang-tidy to prevent this in perpetuity.

[aws-nslick]

Why did this pass distcheck given the unguarded usage of ncclInvalidUsage here?

[rauteric]

Why did this pass distcheck given the unguarded usage of ncclInvalidUsage here?

I think it's because nccl_ofi_tuner.c includes nccl-headers/nvidia/tuner.h (which includes nccl-headers/net.h which includes nccl-headers/neuron/net.h which includes nccl-headers/neuron/error.h) before including nccl_ofi_tuner.h (which includes config.h which defines HAVE_CUDA) 🙃🙃🙃

tl;dr nccl_ofi_tuner.c should include nccl_ofi_tuner.h first.

[bwbarrett]

tl;dr nccl_ofi_tuner.c should include nccl_ofi_tuner.h first.

Incorrect. The first include in any .c file should be #include "config.h". I missed this in the PR review, but that needs to be fixed.

[bwbarrett]

The failure looks like a real issue:

In file included from ../include/nccl-headers/nvidia/tuner.h:12,
                 from tuner/nccl_ofi_model.c:3:
../include/nccl-headers/error.h:13:2: error: #error "Neither CUDA nor Neuron support is available"
   13 | #error "Neither CUDA nor Neuron support is available"
      |  ^~~~~
make[2]: *** [Makefile:559: tuner/nccl_ofi_model.lo] Error 1
make[1]: *** [Makefile:439: all-recursive] Error 1
make: *** [Makefile:371: all] Error 2