RFC: clacc: translating OpenACC to OpenMP in clang

[jdenny-ornl]

Hi,

We are working on a new project, clacc, that extends clang with OpenACC support. Clacc’s approach is to translate OpenACC (a descriptive language) to OpenMP (a prescriptive language) and thus to build on clang’s existing OpenMP support. While we plan to develop clacc to support our own research, an important goal is to contribute clacc as a production-quality component of upstream clang.

We have begun implementing an early prototype of clacc. Before we get too far into the implementation, we would like to get feedback from the LLVM community to help ensure our design would ultimately be acceptable for contribution. For that purpose, below is an analysis of several high-level design alternatives we have considered and their various features. We welcome any feedback.

Thanks.

Joel E. Denny
Future Technologies Group
Oak Ridge National Laboratory

Design Alternatives

[jeffhammond]

All of the usage of OpenACC outside of benchmarks/research that I know about is done in Fortran. Can you provide a list of C/C++ applications using OpenACC today and estimate the number of users that will benefit from this feature?

Thanks,

Jeff

[Finkel_Hal_J]

Such lists exist, although I don’t know what can be shared (and Oak Ridge likely has better lists in this regard than I do). I can tell you, from my own experience, that we’re seeing an increase in development using OpenACC, in both C/C++ and Fortran, over the last couple of years (essentially because the compiler technology has improved to the point where that is now a potentially-productive choice). Also, we have a strong desire to enable tooling over code bases using OpenACC. Among many other things, at some point we’ll likely want the option to automatically migrate much of this code to using OpenMP. Having an OpenACC-enabled Clang, with an implementation that maps to OpenMP, is an important step in that process. Â -Hal

[Finkel_Hal_J]

Great. I don’t think that we want this option because, if nothing else, it will preclude builting source-level tooling for OpenACC. My recommendation: We should think about the very best way we could refactor the code to implement (2), and if that is too ugly (or otherwise significantly degrades maintainability of the OpenMP code), then we should choose (3). We’ll end up doing this, but most of this falls within the scope of the “parallel IR” designs that many of us are working on. Doing this kind of analysis in the frontend is hard (because it essentially requires it to do inlining, simplification, and analysis akin to what the optimizer itself does). I’m fairly certain that not everything maps exactly. They’ll be some things we need to deal with explicitly in CodeGen. +1 This makes sense to me, and I think is most likely to leave the CodeGen code easiest to maintain (and has good separation of concerns). Nevertheless, I think we should go through the mental refactoring exercise for (2) to decide on the value of (3). Thanks again, Hal

[jdenny-ornl]

Hi Jeff, Hal,

Thanks for your feedback. My comments are inline below.

[jdenny-ornl]

Hi Hal,

Thanks for your feedback. It sounds like we’re basically in agreement, but I’ve added a few thoughts inline below.

[jeffhammond]

Hi Jeff, Hal,

Thanks for your feedback. My comments are inline below.

All of the usage of OpenACC outside of benchmarks/research that I know about is done in Fortran.

I agree that it’s easier to find real apps that use OpenACC in Fortran than those that use OpenACC in C/C++. However, the latter certainly exist. For example:

Two of the three examples you cite are primarily Fortran and using OpenACC exclusively in Fortran subroutines.

http://mrfil.github.io/PowerGrid/

/tmp/PowerGrid$ git grep -il “pragma acc”
PowerGrid/Gfft.hpp
PowerGrid/Gnufft.hpp
PowerGrid/ftCpu.hpp
PowerGrid/gridding.hpp
PowerGrid/griddingSupport.hpp

From http://mrfil.github.io/PowerGrid/docs/Installation:

We have experience with PGC++ 15.7 from NVIDIA/The Portland Group as the version we have used most extensively. There is a free license available as part of the OpenACC Toolkit for academic users.

GCC 6.1 has OpenACC support but has not yet been tested by the developers, we welcome reports of anyone trying to compile with it. We hope to support it alongside PGI compilers in the near future.

For those lucky enough to have access to Cray supercomputers, the Cray compiler does support OpenACC, but we have not tried to build with it. Because the Cray compilers are not available on desktops, workstations, or non-Cray branded clusters, we have not dedicated resources to testing PowerGrid on it.

So these folks support OpenACC, but haven’t bothered to try the GCC implementation in the 1+ year that it’s been available. How likely are they to use Clang’s?

https://nek5000.mcs.anl.gov/ (look at the openacc branch in github)

(on the openacc branch)

/tmp/Nek5000$ git grep -il "$acc "
core/acc.f
core/comm_mpi.f
core/gmres.f
core/hmholtz.f
core/hsmg.f
core/math.f
core/navier1.f
core/navier4.f
core/plan4.f
core/prepost.f
core/subs2.f

https://nekcem.mcs.anl.gov/

(on master)
/tmp/svn$ git grep -il “$acc”
branches/maxwell-experimental/src/cem_dg.F
branches/maxwell-experimental/src/dssum2.F
branches/maxwell-experimental/src/io.F
branches/maxwell-experimental/src/mat1.F
branches/maxwell-experimental/src/maxwell.F
branches/maxwell-experimental/src/maxwell_acc.F
branches/maxwell-experimental/src/mxm_acc.F
branches/trunkQu/src/quantum_csr.F
branches/trunkQu/src/quantum_setup.f
branches/trunkQu/src/quantum_time.F
trunk/examples/openacc_gpu=1/box.usr
trunk/examples/openacc_gpu=8/box.usr
trunk/src/acoustic.F
trunk/src/cem_dg2.F
trunk/src/complex.F
trunk/src/drift1.F
trunk/src/drift1_maud.F
trunk/src/drive.F
trunk/src/drive_maud.F
trunk/src/dssum2.F
trunk/src/hmholtz.F
trunk/src/io.F
trunk/src/mat1.F
trunk/src/maxwell.F
trunk/src/maxwell_acc.F
trunk/src/mg_r2204.F
trunk/src/mxm_acc.F
trunk/src/poisson.F
trunk/src/quantum2.F
www/examples/libs/phpThumb/phpthumb.functions.php
www/examples/phpthumb.functions.php

Can you provide a list of C/C++ applications using OpenACC today and estimate the number of users that will benefit from this feature?

Such lists exist, although I don’t know what can be shared (and Oak Ridge likely has better lists in this regard than I do).

I’ll look for a better list that I can share.

That would be helpful.

Best,

Jeff

[jdenny-ornl]

>
> Hi Jeff, Hal,
>
> Thanks for your feedback. My comments are inline below.
>
>>
>>
>> All of the usage of OpenACC outside of benchmarks/research that I know
about is done in Fortran.
>
> I agree that it's easier to find real apps that use OpenACC in Fortran
than those that use OpenACC in C/C++. However, the latter certainly
exist. For example:

Two of the three examples you cite are primarily Fortran and using OpenACC
exclusively in Fortran subroutines.

Are you saying that the occurrences of "pragma acc" in Nek5000 and NekCEM
are unused?

> PowerGrid - Develop Iterative MR Image Reconstructions and Run Anywhere - Home

/tmp/PowerGrid$ git grep -il "pragma acc"
PowerGrid/Gfft.hpp
PowerGrid/Gnufft.hpp
PowerGrid/ftCpu.hpp
PowerGrid/gridding.hpp
PowerGrid/griddingSupport.hpp

From http://mrfil.github.io/PowerGrid/docs/Installation:

*We have experience with PGC++ 15.7 from NVIDIA/The Portland Group as the
version we have used most extensively. There is a free license available as
part of the OpenACC Toolkit for academic users.GCC 6.1 has OpenACC support
but has not yet been tested by the developers, we welcome reports of anyone
trying to compile with it. We hope to support it alongside PGI compilers in
the near future.For those lucky enough to have access to Cray
supercomputers, the Cray compiler does support OpenACC, but we have not
tried to build with it. Because the Cray compilers are not available on
desktops, workstations, or non-Cray branded clusters, we have not dedicated
resources to testing PowerGrid on it.*

So these folks support OpenACC, but haven't bothered to try the GCC
implementation in the 1+ year that it's been available. How likely are
they to use Clang's?

I cannot answer that. Perhaps they were waiting for GCC support to mature?

Thanks.

Joel

[jeffhammond]

Hi Jeff, Hal,

Thanks for your feedback. My comments are inline below.

All of the usage of OpenACC outside of benchmarks/research that I know about is done in Fortran.

I agree that it’s easier to find real apps that use OpenACC in Fortran than those that use OpenACC in C/C++. However, the latter certainly exist. For example:

Two of the three examples you cite are primarily Fortran and using OpenACC exclusively in Fortran subroutines.

Are you saying that the occurrences of “pragma acc” in Nek5000 and NekCEM are unused?

The instances of “pragma acc” in those - it’s the same code in both projects - are either (1) only causing host-device data synchronization or (2) commented-out.

It’s unclear to me what actually happens in the code as currently written. The OpenACC C/C++ code does not more than copy data to/from the device. I didn’t trace the entire code execution but I can’t tell if any code touches the device data that OpenACC is updating. If it is updated, it is updated by Fortran OpenACC code somewhere else in the source tree.

What does the OpenACC standard say about interoperability of compilers+runtimes, as would be required if one used Clang OpenACC for C/C++ and Fortran OpenACC implemented by PGI, Cray, or GCC. OpenMP definitely does not support this, even if a subset of usage may work when one uses the same runtime library with different compilers.

/tmp/Nek5000$ git grep “pragma acc”
jl/gs.c:#pragma acc update host(sendbuf[0:unit_sizebufSize/2]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:unit_sizebufSize/2]) if(acc)
jl/gs.c:#pragma acc update host(sendbuf[0:unit_sizebufSize/2]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:unit_sizebufSize/2]) if(acc)
jl/gs.c:#pragma acc exit data delete(map0,map1)
jl/gs.c:#pragma acc update host(buf[0:unit_sizebufSize]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:unit_sizebufSize]) if(acc)
jl/gs.c:#pragma acc update host(buf[0:unit_sizebufSize]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:unit_sizebufSize]) if(acc)
jl/gs.c://#pragma acc enter data copyin(stage[0].scatter_map[0:stage[0].s_size],stage[0].scatter_mapf[0:stage[0].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage[i].scatter_map[i:stage[i].s_size],stage[i].scatter_mapf[i:stage[i].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage[i].gather_map[i:stage[i].g_size],stage[i].gather_mapf[i:stage[i].g_nt])
jl/gs.c://#pragma acc enter data copyin(stage[i].scatter_map[i:stage[i].s_size],stage[i].scatter_mapf[i:stage[i].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage[i].gather_map[i:stage[i].g_size],stage[i].gather_mapf[i:stage[i].g_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[0].scatter_map[0:stage2[0].s_size],stage2[0].scatter_mapf[0:stage2[0].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[i].scatter_map[i:stage2[i].s_size],stage2[i].scatter_mapf[i:stage2[i].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[i].gather_map[i:stage2[i].g_size],stage2[i].gather_mapf[i:stage2[i].g_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[i].scatter_map[i:stage2[i].s_size],stage2[i].scatter_mapf[i:stage2[i].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[i].gather_map[i:stage2[i].g_size],stage2[i].gather_mapf[i:stage2[i].g_nt])
jl/gs.c:#pragma acc exit data delete(map,mapf)
jl/gs.c:#pragma acc exit data delete(map,mapf)
jl/gs.c:#pragma acc exit data delete(map,mapf)
jl/gs.c:#pragma acc exit data delete(map,mapf)
jl/gs.c:#pragma acc update host(buf[0:vnunit_sizebufSize]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:vnunit_sizebufSize]) if(acc)
jl/gs.c: //#pragma acc exit data delete(ard->map_to_buf[0],ard->map_to_buf[1],ard->map_from_buf[0],ard->map_from_buf[1])
jl/gs.c: //#pragma acc enter data copyin(ard->map_to_buf[0][0:ard->mt_size[0]],ard->map_from_buf[0][0:ard->mf_size[0]],ard->map_to_buf_f[0][0:ard->mt_nt[0]],ard->map_from_buf_f[0][0:ard->mf_nt[0]],ard->map_to_buf[1][0:ard->mt_size[1]],ard->map_from_buf[1][0:ard->mf_size[1]],ard->map_to_buf_f[1][0:ard->mt_nt[1]],ard->map_from_buf_f[1][0:ard->mf_nt[1]])
jl/gs.c:#pragma acc update host(a[0:n])
jl/gs.c:#pragma acc update host(a[0:n])
jl/gs.c:#pragma acc exit data delete(bufPtr)
jl/gs.c:#pragma acc enter data create(bufPtr[0:vngs_dom_size[dom]gsh->r.buffer_size])
jl/gs.c:#pragma acc exit data delete(bufPtr)
jl/gs.c:#pragma acc enter data create(bufPtr[0:vngs_dom_size[dom]gsh->r.buffer_size])
jl/gs.c:#pragma acc exit data delete(map_local0,map_local1,flagged_primaries)
jl/gs.c:#pragma acc enter data pcopyin(map[0:m_size],mapf2[0:2mf_temp])
jl/gs_acc.c://#pragma acc data present(buf[0:l])
jl/gs_acc.c://#pragma acc host_data use_device(buf)
jl/gs_acc.c://#pragma acc data present(buf[0:l])
jl/gs_acc.c://#pragma acc host_data use_device(buf)
jl/gs_acc.c: //#pragma acc enter data copyin(t_mapf[0:t_m_nt2],mapf[0:m_nt2],snd_mapf[0:snd_m_nt2],rcv_mapf[0:rcv_m_nt2],fp_mapf[0:fp_m_nt2], t_map[0:t_m_size],map[0:m_size],fp_map[0:fp_m_size],snd_map[0:snd_m_size],rcv_map[0:rcv_m_size])
jl/gs_acc.c: //#pragma acc enter data copyin(t_mapf[0:t_m_nt2],mapf[0:m_nt2],snd_mapf[0:snd_m_nt2],rcv_mapf[0:rcv_m_nt2],fp_mapf[0:fp_m_nt2], t_map[0:t_m_size],map[0:m_size],fp_map[0:fp_m_size],snd_map[0:snd_m_size],rcv_map[0:rcv_m_size])
jl/gs_acc.c://#pragma acc enter data pcopyin(t_mapf[0:t_m_nt2],mapf[0:m_nt2],snd_mapf[0:snd_m_nt2],rcv_mapf[0:rcv_m_nt2],fp_mapf[0:fp_m_nt2], t_map[0:t_m_size],map[0:m_size],fp_map[0:fp_m_size],snd_map[0:snd_m_size],rcv_map[0:rcv_m_size])
jl/gs_acc.c://#pragma acc data present(u[0:uds],mapf[0:m_nt2],snd_mapf[0:snd_m_nt2],rcv_mapf[0:rcv_m_nt2],fp_mapf[0:fp_m_nt2],t_map[0:t_m_size],map[0:m_size],fp_map[0:fp_m_size],snd_map[0:snd_m_size],rcv_map[0:rcv_m_size])
jl/gs_acc.c://#pragma acc data create(sbuf[0:bl],rbuf[0:bl]) if(bl!=0)
jl/gs_acc.c://#pragma acc parallel loop gang vector present(u[0:uds],map[0:m_size],mapf[0:m_nt2]) private(i,j,t) async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc parallel loop gang vector present(u[0:uds],fp_map[0:fp_m_size],fp_mapf[0:fp_m_nt2]) private(i,j) async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc parallel loop gang vector present(u[0:uds],fp_map[0:fp_m_size]) private(i,k)
jl/gs_acc.c://#pragma acc parallel loop gang vector present(u[0:uds],snd_map[0:snd_m_size],snd_mapf[0:snd_m_nt2],sbuf[0:bl]) private(i,j,t) async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc parallel loop gang vector present(u[0:uds],snd_map[0:snd_m_size],sbuf[0:bl]) private(i,j,k)
jl/gs_acc.c://#pragma acc update host(sbuf[0:bl]) async(vn+2)
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc update device(rbuf[0:bl]) async(vn+2)
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc parallel loop gang vector present(u[0:uds],rcv_map[0:rcv_m_size],rcv_mapf[0:rcv_m_nt2],rbuf[0:bl]) private(i,j,t) async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c: //#pragma acc parallel loop gang vector present(u[0:uds],rcv_map[0:rcv_m_size],rbuf[0:bl]) private(i,j,k)
jl/gs_acc.c://#pragma acc parallel loop gang vector present(u[0:uds],t_map[0:t_m_size],t_mapf[0:t_m_nt2]) private(i,j,t) async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait

http://mrfil.github.io/PowerGrid/

/tmp/PowerGrid$ git grep -il “pragma acc”
PowerGrid/Gfft.hpp
PowerGrid/Gnufft.hpp
PowerGrid/ftCpu.hpp
PowerGrid/gridding.hpp
PowerGrid/griddingSupport.hpp

From http://mrfil.github.io/PowerGrid/docs/Installation:

We have experience with PGC++ 15.7 from NVIDIA/The Portland Group as the version we have used most extensively. There is a free license available as part of the OpenACC Toolkit for academic users.

GCC 6.1 has OpenACC support but has not yet been tested by the developers, we welcome reports of anyone trying to compile with it. We hope to support it alongside PGI compilers in the near future.

For those lucky enough to have access to Cray supercomputers, the Cray compiler does support OpenACC, but we have not tried to build with it. Because the Cray compilers are not available on desktops, workstations, or non-Cray branded clusters, we have not dedicated resources to testing PowerGrid on it.

So these folks support OpenACC, but haven’t bothered to try the GCC implementation in the 1+ year that it’s been available. How likely are they to use Clang’s?

I cannot answer that. Perhaps they were waiting for GCC support to mature?

Or maybe they aren’t interested using in OpenACC compiler support outside of PGI.

What I’m really getting at here is who is going to use OpenACC support in Clang, particularly if there is no compatible Fortran OpenACC compiler? In addition to justifying the code maintenance effort, users who are not developers are essential for implementation hardening.

Best,

Jeff

[jdenny-ornl]

>
>>
>>
>>
>> >
>> > Hi Jeff, Hal,
>> >
>> > Thanks for your feedback. My comments are inline below.
>> >
>> >>
>> >>
>> >> All of the usage of OpenACC outside of benchmarks/research that I
know about is done in Fortran.
>> >
>> > I agree that it's easier to find real apps that use OpenACC in
Fortran than those that use OpenACC in C/C++. However, the latter
certainly exist. For example:
>>
>> Two of the three examples you cite are primarily Fortran and using
OpenACC exclusively in Fortran subroutines.
>
>
> Are you saying that the occurrences of "pragma acc" in Nek5000 and
NekCEM are unused?
>

The instances of "pragma acc" in those - it's the same code in both
projects - are either (1) only causing host-device data synchronization or
(2) commented-out.

It's unclear to me what actually happens in the code as currently
written. The OpenACC C/C++ code does not more than copy data to/from the
device. I didn't trace the entire code execution but I can't tell if any
code touches the device data that OpenACC is updating. If it is updated,
it is updated by Fortran OpenACC code somewhere else in the source tree.

The point is that here is some evidence that compiler support for OpenACC
in C/C++ is useful.

What does the OpenACC standard say about interoperability of
compilers+runtimes, as would be required if one used Clang OpenACC for
C/C++ and Fortran OpenACC implemented by PGI, Cray, or GCC. OpenMP
definitely does not support this, even if a subset of usage may work when
one uses the same runtime library with different compilers.

Flang is under development. I see no reason to believe it cannot grow
OpenACC support eventually as well.

/tmp/Nek5000$ git grep "pragma acc"
jl/gs.c:#pragma acc update host(sendbuf[0:unit_size*bufSize/2]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:unit_size*bufSize/2]) if(acc)
jl/gs.c:#pragma acc update host(sendbuf[0:unit_size*bufSize/2]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:unit_size*bufSize/2]) if(acc)
jl/gs.c:#pragma acc exit data delete(map0,map1)
jl/gs.c:#pragma acc update host(buf[0:unit_size*bufSize]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:unit_size*bufSize]) if(acc)
jl/gs.c:#pragma acc update host(buf[0:unit_size*bufSize]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:unit_size*bufSize]) if(acc)
jl/gs.c://#pragma acc enter data copyin(stage[0].scatter_map[0:
stage[0].s_size],stage[0].scatter_mapf[0:stage[0].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage[i].scatter_map[i:
stage[i].s_size],stage[i].scatter_mapf[i:stage[i].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage[i].gather_map[i:
stage[i].g_size],stage[i].gather_mapf[i:stage[i].g_nt])
jl/gs.c://#pragma acc enter data copyin(stage[i].scatter_map[i:
stage[i].s_size],stage[i].scatter_mapf[i:stage[i].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage[i].gather_map[i:
stage[i].g_size],stage[i].gather_mapf[i:stage[i].g_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[0].scatter_map[
0:stage2[0].s_size],stage2[0].scatter_mapf[0:stage2[0].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[i].scatter_map[
i:stage2[i].s_size],stage2[i].scatter_mapf[i:stage2[i].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[i].gather_map[i:
stage2[i].g_size],stage2[i].gather_mapf[i:stage2[i].g_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[i].scatter_map[
i:stage2[i].s_size],stage2[i].scatter_mapf[i:stage2[i].s_nt])
jl/gs.c://#pragma acc enter data copyin(stage2[i].gather_map[i:
stage2[i].g_size],stage2[i].gather_mapf[i:stage2[i].g_nt])
jl/gs.c:#pragma acc exit data delete(map,mapf)
jl/gs.c:#pragma acc exit data delete(map,mapf)
jl/gs.c:#pragma acc exit data delete(map,mapf)
jl/gs.c:#pragma acc exit data delete(map,mapf)
jl/gs.c:#pragma acc update host(buf[0:vn*unit_size*bufSize]) if(acc)
jl/gs.c:#pragma acc update device(buf[0:vn*unit_size*bufSize]) if(acc)
jl/gs.c: //#pragma acc exit data delete(ard->map_to_buf[0],ard-
>map_to_buf[1],ard->map_from_buf[0],ard->map_from_buf[1])
jl/gs.c: //#pragma acc enter data copyin(ard->map_to_buf[0][0:
ard->mt_size[0]],ard->map_from_buf[0][0:ard->mf_size[0]]
,ard->map_to_buf_f[0][0:ard->mt_nt[0]],ard->map_from_buf_f[
0][0:ard->mf_nt[0]],ard->map_to_buf[1][0:ard->mt_size[1]],
ard->map_from_buf[1][0:ard->mf_size[1]],ard->map_to_buf_f[
1][0:ard->mt_nt[1]],ard->map_from_buf_f[1][0:ard->mf_nt[1]])
jl/gs.c:#pragma acc update host(a[0:n])
jl/gs.c:#pragma acc update host(a[0:n])
jl/gs.c:#pragma acc exit data delete(bufPtr)
jl/gs.c:#pragma acc enter data create(bufPtr[0:vn*gs_dom_
size[dom]*gsh->r.buffer_size])
jl/gs.c:#pragma acc exit data delete(bufPtr)
jl/gs.c:#pragma acc enter data create(bufPtr[0:vn*gs_dom_
size[dom]*gsh->r.buffer_size])
jl/gs.c:#pragma acc exit data delete(map_local0,map_local1,
flagged_primaries)
jl/gs.c:#pragma acc enter data pcopyin(map[0:*m_size],mapf2[0:2*mf_temp])
jl/gs_acc.c://#pragma acc data present(buf[0:l])
jl/gs_acc.c://#pragma acc host_data use_device(buf)
jl/gs_acc.c://#pragma acc data present(buf[0:l])
jl/gs_acc.c://#pragma acc host_data use_device(buf)
jl/gs_acc.c: //#pragma acc enter data copyin(t_mapf[0:t_m_nt*2],
mapf[0:m_nt*2],snd_mapf[0:snd_m_nt*2],rcv_mapf[0:rcv_m_nt*2],fp_mapf[0:fp_m_nt*2],
t_map[0:t_m_size],map[0:m_size],fp_map[0:fp_m_size],snd_
map[0:snd_m_size],rcv_map[0:rcv_m_size])
jl/gs_acc.c: //#pragma acc enter data copyin(t_mapf[0:t_m_nt*2],
mapf[0:m_nt*2],snd_mapf[0:snd_m_nt*2],rcv_mapf[0:rcv_m_nt*2],fp_mapf[0:fp_m_nt*2],
t_map[0:t_m_size],map[0:m_size],fp_map[0:fp_m_size],snd_
map[0:snd_m_size],rcv_map[0:rcv_m_size])
jl/gs_acc.c://#pragma acc enter data pcopyin(t_mapf[0:t_m_nt*2],
mapf[0:m_nt*2],snd_mapf[0:snd_m_nt*2],rcv_mapf[0:rcv_m_nt*2],fp_mapf[0:fp_m_nt*2],
t_map[0:t_m_size],map[0:m_size],fp_map[0:fp_m_size],snd_
map[0:snd_m_size],rcv_map[0:rcv_m_size])
jl/gs_acc.c://#pragma acc data present(u[0:uds],mapf[0:m_nt*
2],snd_mapf[0:snd_m_nt*2],rcv_mapf[0:rcv_m_nt*2],fp_mapf[0:
fp_m_nt*2],t_map[0:t_m_size],map[0:m_size],fp_map[0:fp_m_
size],snd_map[0:snd_m_size],rcv_map[0:rcv_m_size])
jl/gs_acc.c://#pragma acc data create(sbuf[0:bl],rbuf[0:bl]) if(bl!=0)
jl/gs_acc.c://#pragma acc parallel loop gang vector
present(u[0:uds],map[0:m_size],mapf[0:m_nt*2]) private(i,j,t) async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc parallel loop gang vector
present(u[0:uds],fp_map[0:fp_m_size],fp_mapf[0:fp_m_nt*2]) private(i,j)
async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc parallel loop gang vector
present(u[0:uds],fp_map[0:fp_m_size]) private(i,k)
jl/gs_acc.c://#pragma acc parallel loop gang vector
present(u[0:uds],snd_map[0:snd_m_size],snd_mapf[0:snd_m_nt*2],sbuf[0:bl])
private(i,j,t) async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc parallel loop gang vector
present(u[0:uds],snd_map[0:snd_m_size],sbuf[0:bl]) private(i,j,k)
jl/gs_acc.c://#pragma acc update host(sbuf[0:bl]) async(vn+2)
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc update device(rbuf[0:bl]) async(vn+2)
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c://#pragma acc parallel loop gang vector
present(u[0:uds],rcv_map[0:rcv_m_size],rcv_mapf[0:rcv_m_nt*2],rbuf[0:bl])
private(i,j,t) async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait
jl/gs_acc.c: //#pragma acc parallel loop gang vector
present(u[0:uds],rcv_map[0:rcv_m_size],rbuf[0:bl]) private(i,j,k)
jl/gs_acc.c://#pragma acc parallel loop gang vector
present(u[0:uds],t_map[0:t_m_size],t_mapf[0:t_m_nt*2]) private(i,j,t)
async(k+1)
jl/gs_acc.c://#pragma acc loop seq
jl/gs_acc.c://#pragma acc wait

>>
>>
>> > PowerGrid - Develop Iterative MR Image Reconstructions and Run Anywhere - Home
>>
>> /tmp/PowerGrid$ git grep -il "pragma acc"
>> PowerGrid/Gfft.hpp
>> PowerGrid/Gnufft.hpp
>> PowerGrid/ftCpu.hpp
>> PowerGrid/gridding.hpp
>> PowerGrid/griddingSupport.hpp
>>
>> From http://mrfil.github.io/PowerGrid/docs/Installation:
>>
>> We have experience with PGC++ 15.7 from NVIDIA/The Portland Group as
the version we have used most extensively. There is a free license
available as part of the OpenACC Toolkit for academic users.
>>
>> GCC 6.1 has OpenACC support but has not yet been tested by the
developers, we welcome reports of anyone trying to compile with it. We hope
to support it alongside PGI compilers in the near future.
>>
>> For those lucky enough to have access to Cray supercomputers, the Cray
compiler does support OpenACC, but we have not tried to build with it.
Because the Cray compilers are not available on desktops, workstations, or
non-Cray branded clusters, we have not dedicated resources to testing
PowerGrid on it.
>>
>> So these folks support OpenACC, but haven't bothered to try the GCC
implementation in the 1+ year that it's been available. How likely are
they to use Clang's?
>
>
> I cannot answer that. Perhaps they were waiting for GCC support to
mature?

Or maybe they aren't interested using in OpenACC compiler support outside
of PGI.

They said they are interested. I don't yet see sufficient evidence to
believe that interest is not genuine.

Thanks.

Joel

[Finkel_Hal_J]

Hi, Jeff,

First, I don’t think we should spam all of cfe-dev the source listings of every project using OpenACC – We actually do need to understand what’s out there (both open source and otherwise), for many different reasons, but this is not the place to collect that information. Of course, one of the reasons that we’d like OpenACC support in Clang is to make it easier to do this kind of analysis (current internal work is using GCC).

Second, we have a significant body of code using OpenACC, and moreover, the rate at which OpenACC code is being written is increasing. As I mentioned, this is because the compiler OpenACC support has now reached a point where using it makes sense. On many machines with GPUs, including large DOE machines, OpenACC is the recommended way to exploit them at this point. As OpenMP accelerator support matures, that may change. We’re (DOE is) investing heavily in OpenMP for the future. Nevertheless, OpenACC is a cross-vendor standard (), which is actively developed, supported by the Cray and PGI compilers, and also by GCC (as of GCC v7). The GCC support is new, and less mature than what is provided by the vendor compilers, but I expect that it will get better over time as well. I think that it makes sense for Clang to support OpenACC. We’d like to use Clang’s tooling capabilities on this code. Moreover, I’d like to do this in a way that enables maximum interoperability with OpenMP. Over time, this will make it easier to transition code from OpenACC to OpenMP when that’s desired. There’s a significant overlap in functionality between the two sets of directives. That should lower the implementation cost significantly. OpenACC is not a pure subset of OpenMP, however, and that’s also important. Thanks again, Hal

[Finkel_Hal_J]

That’s probably a good idea. Please share some details on this front. It is this last part that is potentially concerning. If you try it, however, and it sounds like you are, then we’ll know for sure soon enough. Obviously the most efficient way to write some piece of code, and the way to write it to maximize maintainability and ease of extension, may be different. To the extent that they’re the same, in terms of upstream functionality, we’ll learn something. You’re free to prototype things however you’d like Thanks again, Hal

[jdenny-ornl]

Will do. However, due to the holidays, it might take me a little time to put that in an email.

Agreed. My hope is that we can get the best of both if design 3 and design 2 prove easily composable.

Thanks.

Joel

[_C_Bergstrom]

Sorry the top post, but just to chime in here a bit…

As a company who has implemented OpenACC on top of clang ALREADY and supported customers for C (and a limited amount of C++) I see a lot of benefit here. Frankly, the amount of accelerated C/C++ code is growing and what does or doesn’t exist now doesn’t mean nearly as much as the industry trend towards C++. I’m biased in that I think OMP is just rubbish for offloading and needs a lot of work. I don’t think OpenACC is any perfect solution, but certainly has the benefit of being easier for users, but at the expense of needing smart and high quality tools.

2nd - there was no community usage requirement when Intel started working on getting OMP4 added to clang. It was allowed that it could incrementally be merged and reviewed. Adding this as a troll blocker just doesn’t seem to be fair. Please stop with the politics and judge this based on technical merit alone. OpenACC and OpenMP4+ have similar goals, but try to achieve them in different ways. All my biases aside please stick to technical reasons for why this should be blocked. (For example, if there isn’t anyone who has pledged to continue to support it long term… etc)

So basically if this is just some lame research project with no extended maintenance plan, make a fork and put your code there and not bother upstream. Otherwise, I think it’s a very welcome idea.

[andreybokhanko]

There is an *explicit* requirement in the "Contributing Extensions to
Clang" section of "Getting Involved with the Clang Project" document (
http://clang.llvm.org/get_involved.html#criteria):

"Evidence of a significant user community: This is based on a number of
factors, including an actual, existing user community, the perceived
likelihood that users would adopt such a feature if it were available, and
any "trickle-down" effects that come from, e.g., a library adopting the
feature and providing benefits to its users."

As an example, CilkPlus support was rejected partly due to concerns of not
significant user community present at the time.

I don't remember the exact situation with OpenMP, but perhaps the concern
hadn't been risen simply because usage interest (including from your
company, no?) was just too obvious for everyone?

Yours,
Andrey