Re: [PATCH] Add OpenACC acc_get_property support for AMD GCN

[Thomas Schwinge <thomas@codesourcery.com>]

[-- Attachment #1: Type: text/plain, Size: 3178 bytes --]

Hi Frederik!

On 2020-01-31T13:17:52+0100, "Harwath, Frederik" <frederik@codesourcery.com> wrote:
> On 30.01.20 17:08, Thomas Schwinge wrote:
>
>> I understand correctly that the only reason for:
>> 
>> On 2020-01-29T10:52:57+0100, "Harwath, Frederik" <frederik@codesourcery.com> wrote:
>>> 	* testsuite/libgomp.oacc-c-c++-common/acc_get_property-aux.c
>>> 	(expect_device_properties): Split function into ...
>>> 	(expect_device_string_properties): ... this new function ...
>>> 	(expect_device_memory): ... and this new function.
>> 
>> ... this split is that we can't test 'expect_device_memory' here:
>
>> [...]
>
>> ..., because that one doesn't (re-)implement the 'acc_property_memory'
>> interface?
>
> Correct.

OK, thanks for confirming.

> But why "re-"? It has not been implemented before.

Well, yes, it has already been implemented: in
'libgomp/plugin/plugin-gcn.c' ;-) -- that's why I meant it's
re-implemented in the test case(s): in a similar yet slightly different
way compared to the respective plugin(s).  And for avoidance of doubt: I
agree that's a good approach to verify that we're getting some consistent
and meaningful results.)

>>> --- a/libgomp/plugin/plugin-gcn.c
>>> +++ b/libgomp/plugin/plugin-gcn.c

>>> +    case GOACC_PROPERTY_MEMORY:
>>> +      {
>>> +	size_t size;
>>> +	hsa_region_t region = agent->data_region;
>>> +	hsa_status_t status =
>>> +	  hsa_fns.hsa_region_get_info_fn (region, HSA_REGION_INFO_SIZE, &size);
>>> +	if (status == HSA_STATUS_SUCCESS)
>>> +	  propval.val = size;
>>> +	break;
>>> +      }
>>> [...]
>>>  }
>> 
>> Here we got 'acc_property_memory' implemented, but not here:
>> 
>>> --- /dev/null
>>> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/acc_get_property-gcn.c
>
> Yes, there seems to be no straightforward way to determine the expected value through
> the runtime API. We might of course try to replicate the logic that is
> used in plugin-gcn.c.

No need to do that now; I was just curious whether that's the reason.


>>> --- a/libgomp/plugin/plugin-gcn.c
>>> +++ b/libgomp/plugin/plugin-gcn.c
>> 
>>> @@ -4115,12 +4141,37 @@ GOMP_OFFLOAD_openacc_async_dev2host (int device, void *dst, const void *src,
>>>  union goacc_property_value
>>>  GOMP_OFFLOAD_openacc_get_property (int device, enum goacc_property prop)
>>>  {
>>> [...]
>>> +  switch (prop)
>>> +    {
>>> +    case GOACC_PROPERTY_FREE_MEMORY:
>>> +      /* Not supported. */
>>> +      break;
>> 
>> (OK, can be added later when somebody feels like doing that.)
>
> Well, "not supported" means that there seems to be no (reasonable) way to obtain
> the necessary information from the runtime - in contrast to the nvptx plugin
> where it can be obtained easily through the CUDA API.

OK, I see, and again that's fine.  (The API explicitly permits such
"zero" returns.)

And, I don't know what a user is actually supposed to tell from
'acc_property_free_memory' given that this can change at any point in
time, arbitrarily, not under control of the executing processe, given
that multiple processes may be using the GPU concurrently.


Grüße
 Thomas

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 658 bytes --]