OpenACC supported version

[Imran]

Hi accull team,

I would like to ask you what version of openACC version is supported by accull?
are all the features of that version or supported?

I could not find it on the webiste or overview on the repository. May be its a good to mention
the supported/un-supported feature and the specification version.

Kind Regards,

[Jose Lucas Grillo Lorenzo]

Hi Imran,

Currently, accULL last revisions support the most relevant features of OpenACC 1.0 standard. The tables bellow show some known features that are and are not supported:

Kind Regards,

Table: Compliance with the OpenACC 1.0 standard (constructs)

Construct	Status	Description
kernels	Implemented	Kernels for and CUDA are generated for each loop inside the scope
loop	Implemented	Indicates a potential accelerator kernel. Some restrictions apply (e.g., no external definitions)
kernels loop	Implemented	A kernel will be extracted. Dependency analysis is used to check and allocate RO variables if possible.
parallel	Implemented	Code inside the parallel region will be translated in to a kernel.
parallel loop	Implemented	Loops inside the parallel region will be unrolled on to a kernel. (Support for large arrays)
update	Implemented	-
copy, copyin,copyout,	Implemented	Runtime dynamically handles memory transfers
pcopy, pcopyin,pcopyout ,	Implemented	Runtime dynamically handles memory transfers when required
async	Implemented	Working for OpenCL, implementation ongoing for CUDA
deviceptr	Not implemented	-
host	Implemented	-
name	Not in standard	Optional clause to name a particular acc region or loop and refer it from an external optimization file at compile time.
acc_malloc,free	Implemented	Return valid CUDA or pointers depending on the active platform. Runtime keep track of these user-requested memory, and can free it if user forgets.

Table: Compliance with the OpenACC 1.0 standard (clauses)

Clause	Status	Description
collapse	Implemented	Up to three nested loop can be used to generate a 3D kernel
gangs , workers	Implemented	Runtime sets appropriate values for these parameters internally
vector	Not implemented	Requires compiler driver support for vectorization
deviceptr	Not implemented	Requires compiler driver support
reduction	Partialy	Only available in CUDA backend



     J. Lucas Grillo

--
You received this message because you are subscribed to the Google Groups "accULL Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to accull+un...@googlegroups.com.
To post to this group, send email to acc...@googlegroups.com.
Visit this group at http://groups.google.com/group/accull.
For more options, visit https://groups.google.com/d/optout.

[imran]

Thanks Lucas. That is helpful.

So cache directive for placing data in shared memory is not yet supported.
Is it on your to-do list as i dont see it in this table at all?

Kind Regards,
imran