[Bug]: Behaviour of here.gpus[x] when x is out of bounds

[Guillaume-Helbecque]

Descritpion

While playing with the built-in array here.gpus, I discovered that the behaviour of here.gpus[x] when x is out of bounds is somehow not clearly defined. For example, I executed this program on a system with 8 GPUs, but only one enabled (CHPL_RT_NUM_GPUS_PER_LOCALE=1):

config const gpuID = 0;

proc main() {
  writeln(here.gpus);
  writeln(here.gpus.domain);
  writeln(here.gpus[gpuID], "\n");

  var A: [1..10] int;
  on here.gpus[gpuID] {
    var B: [1..10] int;
    @assertOnGpu
    foreach i in B.domain {
      B[i] = i;
    }

    A = B;
  }

  writeln(A);
}

By default (--gpuID 0), this program returns as expected:

LOCALE0-GPU0
{0..0}
LOCALE0-GPU0

1 2 3 4 5 6 7 8 9 10

But, for any --gpuID x values greater than 1, we get:

LOCALE0-GPU0
{0..0}
nil

1 2 3 4 5 6 7 8 9 10

While nil is probably expected because only one GPU is enabled, it seems that assertOnGPU is not triggered. Is on nil possible?
I also extended the experiment to negative numbers, and the results seem unpredictable as I encountered at least four different outputs:

• For -1, here.gpus[-1] returns here.gpus[0] and assertOnGPU is not triggered:

LOCALE0-GPU0
{0..0}
LOCALE0-GPU0

1 2 3 4 5 6 7 8 9 10

• For -2 and -3, here.gpus[x] returns here.id and assertOnGPU is triggered:

LOCALE0-GPU0
{0..0}
LOCALE0

sandbox.chpl:12: error: assertOnGpu() failed

• For -4, here.gpus[-4] returns segfault:

LOCALE0-GPU0
{0..0}
Segmentation fault

• For -5, here.gpus[-5] returns nil and assertOnGPU is triggered:

LOCALE0-GPU0
{0..0}
nil

sandbox.chpl:12: error: assertOnGpu() failed

etc.

Of course here.gpus is not expected to be used that way, and these experiments are a bit sadistic, but first I'd like to report this just in case this is not a known behaviour, and then I wonder if there is any interesting explanation behind that.

[bradcray]

@Guillaume-Helbecque : I don't think this behavior is intentional and strongly suspect that it's one of the impacts of the following warning printed when doing GPU compilations:

warning: The prototype GPU support implies --no-checks. This may impact debuggability. To suppress this warning, compile with --no-checks explicitly

Specifically, the gpus array is a normal array in Chapel, and accesses to it would normally be bounds-checked; but since GPU compilations use --no-checks, that bounds-checking is disabled. If I compile similar programs for the flat (non-GPU) locale model, I get out-of-bounds errors as expected.

In saying this, I'm only providing a likely explanation, not saying that this is as we'd like things to be. Coming up with a way to enable checks for GPU compilations is definitely something that we'd consider to be important to Chapel's long-term productivity for GPU programming. Sorry for any hassle in the meantime.