Skip to content

Use host-pinned, mapped memory #8

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
gzagaris wants to merge 4 commits into
base: main
Choose a base branch
from
Draft

Use host-pinned, mapped memory #8

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
a88644c
ENH: add NVTX annotations to UpdateFringes
gzagaris Sep 24, 2021
038edfe
ENH: comment out Tecplot I/O for mesh partitions
gzagaris Sep 26, 2021
15fb777
ENH: methods to alloc/dealloc host-pinned memory
gzagaris Sep 26, 2021
b6bba0e
ENH: use host-pinned mapped memory
gzagaris Sep 26, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
53 changes: 29 additions & 24 deletions src/LocalMesh.C
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,10 +22,10 @@ LocalMesh::~LocalMesh()
FVSAND_FREE_DEVICE(device2host_d);
FVSAND_FREE_DEVICE(host2device_d);
//
if (qbuf) delete [] qbuf;
if (qbuf2) delete [] qbuf2;
FVSAND_FREE_DEVICE(qbuf_d);
FVSAND_FREE_DEVICE(qbuf_d2);
gpu::deallocate_host_pinned(&qbuf);
gpu::deallocate_host_pinned(&qbuf2);
gpu::deallocate_host_pinned(&qbuf_d);
gpu::deallocate_host_pinned(&qbuf_d2);
//
FVSAND_FREE_DEVICE(x_d);
FVSAND_FREE_DEVICE(flovar_d);
Expand All @@ -41,7 +41,7 @@ LocalMesh::~LocalMesh()
FVSAND_FREE_DEVICE(center_d);
FVSAND_FREE_DEVICE(normals_d);
FVSAND_FREE_DEVICE(volume_d);
FVSAND_FREE_DEVICE(res_d);
gpu::deallocate_host_pinned(&res_d);
//
FVSAND_FREE_DEVICE(rmatall_d);
FVSAND_FREE_DEVICE(Dall_d);
Expand All @@ -60,7 +60,7 @@ LocalMesh::~LocalMesh()
FVSAND_FREE_DEVICE(q);
FVSAND_FREE_DEVICE(qn);
FVSAND_FREE_DEVICE(qnn);
FVSAND_FREE_DEVICE(dq_d);
gpu::deallocate_host_pinned(&dq_d);
FVSAND_FREE_DEVICE(dqupdate_d);
}

Expand Down Expand Up @@ -256,8 +256,8 @@ void LocalMesh::InitSolution(double *flovar, int nfields)
q=gpu::allocate_on_device<double>(sizeof(double)*(ncells+nhalo)*nfields);
qn=gpu::allocate_on_device<double>(sizeof(double)*(ncells+nhalo)*nfields);
qnn=gpu::allocate_on_device<double>(sizeof(double)*(ncells+nhalo)*nfields);
res_d=gpu::allocate_on_device<double>(sizeof(double)*(ncells+nhalo)*nfields);
dq_d=gpu::allocate_on_device<double>(sizeof(double)*(ncells+nhalo)*nfields);
res_d=gpu::allocate_host_pinned<double>(sizeof(double)*(ncells+nhalo)*nfields);
Copy link
Copy Markdown
Owner

@jsitaraman jsitaraman Sep 27, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why allocate res_d and dq_d as host pinned ? I get that for the qbuf, but won't this be inefficient if we host pin the residual and dq, the main arrays that get filled in by GPU kernels ?

Copy link
Copy Markdown
Collaborator Author

@gzagaris gzagaris Sep 27, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I believe I made this change b/c both res_d and dq_d are passed to the UpdateFringes. Yes, I agree this is definitely not ideal -- I am still fiddling with it.

dq_d=gpu::allocate_host_pinned<double>(sizeof(double)*(ncells+nhalo)*nfields);
dqupdate_d=gpu::allocate_on_device<double>(sizeof(double)*(ncells+nhalo)*nfields);
);

Expand Down Expand Up @@ -291,10 +291,10 @@ void LocalMesh::InitSolution(double *flovar, int nfields)

int dsize=device2host.size();
dsize=(dsize < host2device.size())?host2device.size():dsize;
qbuf=new double [dsize];
qbuf2=new double [dsize];
qbuf_d=gpu::allocate_on_device<double>(sizeof(double)*dsize);
qbuf_d2=gpu::allocate_on_device<double>(sizeof(double)*dsize);
qbuf=gpu::allocate_host_pinned<double>(sizeof(double)*dsize);
qbuf2=gpu::allocate_host_pinned<double>(sizeof(double)*dsize);
qbuf_d=gpu::allocate_host_pinned<double>(sizeof(double)*dsize);
qbuf_d2=gpu::allocate_host_pinned<double>(sizeof(double)*dsize);

ireq=new MPI_Request [sndmap.size()+rcvmap.size()];
istatus=new MPI_Status [sndmap.size()+rcvmap.size()];
Expand Down Expand Up @@ -337,27 +337,32 @@ void LocalMesh::UpdateFringes(double *qh, double *qd)
// doesn't work CUDA-Aware -- debug this
void LocalMesh::UpdateFringes(double *qd)
{
FVSAND_NVTX_FUNCTION("UpdateFringes");

nthreads=device2host.size();
if(nthreads == 0) return;
FVSAND_GPU_KERNEL_LAUNCH( updateHost, nthreads,
qbuf_d,qd,device2host_d,nthreads);

FVSAND_NVTX_SECTION( "pack",
FVSAND_GPU_KERNEL_LAUNCH( updateHost, nthreads,
qbuf_d,qd,device2host_d,nthreads);
);

// separate sends and receives so that we can overlap comm and calculation
// in the residual and iteration loops.
// TODO (george) use qbuf2_d and qbuf_d instead of qbuf2 and qbuf for cuda-aware
int reqcount=0;
pc.postRecvs_direct(qbuf2,nfields_d,rcvmap,ireq,mycomm,&reqcount);
// TODO (george) with cuda-aware this pull is not required
// but it doesn't work now
gpu::pull_from_device<double>(qbuf,qbuf_d,sizeof(double)*device2host.size());
pc.postSends_direct(qbuf,nfields_d,sndmap,ireq,mycomm,&reqcount);
pc.finish_comm(reqcount,ireq,istatus);
// same as above
// not doing cuda-aware now
gpu::copy_to_device(qbuf_d2,qbuf2,sizeof(double)*host2device.size());

gpu::synchronize();
pc.postSends_direct(qbuf_d,nfields_d,sndmap,ireq,mycomm,&reqcount);
pc.finish_comm(reqcount,ireq,istatus);

nthreads=host2device.size();
FVSAND_GPU_KERNEL_LAUNCH( updateDevice, nthreads,
qd,qbuf_d2,host2device_d,nthreads);
FVSAND_NVTX_SECTION( "unpack",
FVSAND_GPU_KERNEL_LAUNCH( updateDevice, nthreads,
qd,qbuf2,host2device_d,nthreads );
);

}


Expand Down
2 changes: 1 addition & 1 deletion src/fvsand.C
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -143,7 +143,7 @@ int main(int argc, char *argv[])
printf("# ----------------------------------\n");
}

lm->WriteMesh(myid);
// lm->WriteMesh(myid);

MPI_Finalize();

Expand Down
15 changes: 15 additions & 0 deletions src/fvsand_cuda.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,6 +47,14 @@ inline T* allocate_on_device(const size_t size)
return dptr;
}

template <typename T>
inline T* allocate_host_pinned(const size_t size)
{
T* dptr = nullptr;
FVSAND_CUDA_CHECK_ERROR(cudaHostAlloc((void**)(&dptr), size,cudaHostAllocMapped));
return dptr;
}

template <typename T>
inline void copy_to_device(T* dptr, const T* hptr, const size_t size)
{
Expand Down Expand Up @@ -74,6 +82,13 @@ inline void deallocate_device(T** dptr)
*dptr = nullptr;
}

template <typename T>
inline void deallocate_host_pinned(T** ptr)
{
FVSAND_CUDA_CHECK_ERROR(cudaFreeHost(static_cast<void*>(*ptr)));
*ptr = nullptr;
}

template <typename T>
inline void memset_on_device(T* dptr, T val, const size_t sz)
{
Expand Down
17 changes: 17 additions & 0 deletions src/fvsand_hip.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,15 @@ inline T* allocate_on_device(const size_t size)
return dptr;
}

template <typename T>
inline T* allocate_host_pinned(const size_t size)
{
T* dptr = nullptr;
FVSAND_CUDA_CHECK_ERROR(hipHostMalloc((void**)(&dptr), size));
return dptr;
}


template <typename T>
inline void copy_to_device(T* dptr, const T* hptr, const size_t size)
{
Expand Down Expand Up @@ -75,6 +84,14 @@ inline void deallocate_device(T** dptr)
*dptr = nullptr;
}

template < typename T>
inline void deallocate_host_pinned(T** ptr)
{
FVSAND_CUDA_CHECK_ERROR(hipHostFree(static_cast<void*>(*ptr)));
*ptr = nullptr;
}


template <typename T>
inline void memset_on_device(T* dptr, T val, const size_t sz)
{
Expand Down
13 changes: 13 additions & 0 deletions src/fvsand_nogpu.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,12 @@ inline T* allocate_on_device(const size_t size)
return static_cast<T*>(std::malloc(size));
}

template <typename T>
inline T* allocate_host_pinned(const size_t size)
{
return static_cast<T*>(std::malloc(size));
}

template <typename T>
inline void copy_to_device(T* dptr, const T* hptr, const size_t size)
{
Expand Down Expand Up @@ -54,6 +60,13 @@ inline void deallocate_device(T** dptr)
*dptr = nullptr;
}

template < typename T>
inline void deallocate_host_pinned(T** ptr)
{
std::free(*ptr);
*ptr = nullptr;
}

template <typename T>
inline void memset_on_device(T* dptr, T val, const size_t sz)
{
Expand Down