Re: [PATCH 6/9] [nvptx] Force vl32 if calling vector-partitionable routines -- test-cases

[Tom de Vries <tdevries@suse.de>]

On 10/30/20 5:16 PM, Thomas Schwinge wrote:
> Hi Tom!
> 
> While working on something completely different, I had to dig deeper, and
> noticed a thing there, and deeper, and notice another thing, and deeper,
> and noticed this other thing here...  (So, business as usual...)  ;-)
> 
> On 2019-01-12T23:21:28+0100, Tom de Vries <tdevries@suse.de> wrote:
>> --- /dev/null
>> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/pr85486.c
> 
>> +#pragma acc routine vector
>> +void __attribute__((noinline, noclone))
>> +Vector (int *ptr, int n, const int inc)
>> +{
> 
>> +#pragma acc parallel copy (ary) vector_length (128) /* { dg-warning "using vector_length \\(32\\) due to call to vector-partitionable routine, ignoring 128" } */
>> +  {
>> +    Vector (&ary[0][0], m * n, (1 << 24) - (1 << 16));
> 
> This works as diagnosed/expected.
> 
> On 2019-01-12T23:21:31+0100, Tom de Vries <tdevries@suse.de> wrote:
>> --- /dev/null
>> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/pr85486-2.c
>> @@ -0,0 +1,52 @@
>> +/* { dg-do run { target openacc_nvidia_accel_selected } } */
>> +/* { dg-additional-options "-fopenacc-dim=::128" } */
> 
> Via '-fopenacc-dim', we here request a default 'vector_length(128)'.
> 
>> +#pragma acc parallel copy (ary)
>> +  {
>> +    Vector (&ary[0][0], m * n, (1 << 24) - (1 << 16));
> 
> As above, 'vector_length(128)' must be demoted to 'vector_length(32)'
> (and in fact, it is) -- but we're not getting a diagnostic for that.  Is
> this expected?
> 

I think it would be good to have.  I don't know whether it's implemented.

> On 2019-01-12T23:21:28+0100, Tom de Vries <tdevries@suse.de> wrote:
>> --- /dev/null
>> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/pr85486-3.c
>> @@ -0,0 +1,54 @@
>> +/* { dg-do run { target openacc_nvidia_accel_selected } } */
>> +/* { dg-set-target-env-var "GOMP_OPENACC_DIM" "::128" } */
> 
> This testcase needs 'dg-additional-options "-fopenacc-dim=::-"' (or
> similar), but support for that is still missing in master branch (I'm
> working on porting over the corresponding patch), so this currently
> defaults to 'vector_length(32)', and...
> 
>> +#pragma acc parallel copy (ary)
>> +  {
>> +    Vector (&ary[0][0], m * n, (1 << 24) - (1 << 16));
> 
> ... thus no diagnostic here, and...
> 
>> +/* { dg-prune-output "using vector_length \\(32\\), ignoring runtime setting" } */
> 
> ... we're in fact not seeing this diagnostic.
> 
> 
> In addition to the (presumedly unexpected) missing diagnostic for
> '-fopenacc-dim=::128' mentioned above -- OK to simplify and enhance the
> testcases as attached, "Simplify and enhance
> 'libgomp.oacc-c-c++-common/pr85486*.c' [PR85486]"?
> 

Yep, looks good.

Thanks,
- Tom