CFD/hypre
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

add #ifdef guards to MPI-3 functions hypre_MPI_Comm_split_type, hypre_MPI_Info_create, and hypre_MPI_Info_free

Browse Source
This commit is contained in:
Ruipeng Li 2018-09-12 14:42:55 -07:00
parent d356d7abc1
commit 1430853bda
4 changed files with 167 additions and 161 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/utilities/_hypre_utilities.h
View File

@ -372,9 +372,11 @@ HYPRE_Int hypre_MPI_Type_commit( hypre_MPI_Datatype *datatype );
HYPRE_Int hypre_MPI_Type_free( hypre_MPI_Datatype *datatype );
HYPRE_Int hypre_MPI_Op_free( hypre_MPI_Op *op );
HYPRE_Int hypre_MPI_Op_create( hypre_MPI_User_function *function , hypre_int commute , hypre_MPI_Op *op );
#if defined(HYPRE_USING_CUDA) || defined(HYPRE_USING_DEVICE_OPENMP)
HYPRE_Int hypre_MPI_Comm_split_type(hypre_MPI_Comm comm, HYPRE_Int split_type, HYPRE_Int key, hypre_MPI_Info info, hypre_MPI_Comm *newcomm);
HYPRE_Int hypre_MPI_Info_create(hypre_MPI_Info *info);
HYPRE_Int hypre_MPI_Info_free( hypre_MPI_Info *info );
#endif
#ifdef __cplusplus
}

317
src/utilities/gpuMem.c
View File

@ -24,146 +24,145 @@ hypre_int ggc(hypre_int id);
/* Global struct that holds device,library handles etc */
struct hypre__global_struct hypre__global_handle = { .initd=0, .device=0, .device_count=1, .memoryHWM=0};
/* Initialize GPU branch of Hypre AMG */
/* use_device =-1 */
/* Application passes device number it is using or -1 to let Hypre decide on which device to use */
void hypre_GPUInit(hypre_int use_device)
{
char pciBusId[80];
hypre_int myid;
hypre_int nDevices;
hypre_int device;
char pciBusId[80];
hypre_int myid;
hypre_int nDevices;
hypre_int device;
#if defined(TRACK_MEMORY_ALLOCATIONS)
hypre_printf("\n\n\n WARNING :: TRACK_MEMORY_ALLOCATIONS IS ON \n\n");
hypre_printf("\n\n\n WARNING :: TRACK_MEMORY_ALLOCATIONS IS ON \n\n");
#endif /* TRACK_MEMORY_ALLOCATIONS */
if (!HYPRE_GPU_HANDLE)
{
HYPRE_GPU_HANDLE=1;
HYPRE_DEVICE=0;
hypre_CheckErrorDevice(cudaGetDeviceCount(&nDevices));
if (!HYPRE_GPU_HANDLE)
{
HYPRE_GPU_HANDLE=1;
HYPRE_DEVICE=0;
hypre_CheckErrorDevice(cudaGetDeviceCount(&nDevices));
/* XXX */
nDevices = 1; /* DO NOT COMMENT ME OUT AGAIN! nDevices does NOT WORK !!!! */
HYPRE_DEVICE_COUNT=nDevices;
/* TODO cannot use nDevices to check if mpibind is used, need to rewrite
* E.g., NP=5 on 2 nodes, nDevices=1,1,1,1,4 */
/* XXX */
nDevices = 1; /* DO NOT COMMENT ME OUT AGAIN! nDevices does NOT WORK !!!! */
HYPRE_DEVICE_COUNT=nDevices;
if (use_device<0)
{
if (nDevices<4)
/* TODO cannot use nDevices to check if mpibind is used, need to rewrite
* E.g., NP=5 on 2 nodes, nDevices=1,1,1,1,4 */
if (use_device<0)
{
/* with mpibind each process will only see 1 GPU */
HYPRE_DEVICE=0;
hypre_CheckErrorDevice(cudaSetDevice(HYPRE_DEVICE));
cudaDeviceGetPCIBusId ( pciBusId, 80, HYPRE_DEVICE);
//hypre_printf("num Devices %d\n", nDevices);
}
else if (nDevices==4)
{
// THIS IS A HACK THAT WORKS ONLY AT LLNL
/* No mpibind or it is a single rank run */
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &myid );
//affs(myid);
hypre_MPI_Comm node_comm;
hypre_MPI_Info info;
hypre_MPI_Info_create(&info);
hypre_MPI_Comm_split_type(hypre_MPI_COMM_WORLD, hypre_MPI_COMM_TYPE_SHARED, myid, info, &node_comm);
hypre_int round_robin=1;
hypre_int myNodeid, NodeSize;
hypre_MPI_Comm_rank(node_comm, &myNodeid);
hypre_MPI_Comm_size(node_comm, &NodeSize);
if (round_robin)
if (nDevices<4)
{
/* Round robin allocation of GPUs. Does not account for affinities */
HYPRE_DEVICE=myNodeid%nDevices;
/* with mpibind each process will only see 1 GPU */
HYPRE_DEVICE=0;
hypre_CheckErrorDevice(cudaSetDevice(HYPRE_DEVICE));
cudaDeviceGetPCIBusId ( pciBusId, 80, HYPRE_DEVICE);
hypre_printf("WARNING:: Code running without mpibind\n");
hypre_printf("Global ID = %d , Node ID %d running on device %d of %d \n",myid,myNodeid,HYPRE_DEVICE,nDevices);
//hypre_printf("num Devices %d\n", nDevices);
}
else if (nDevices==4)
{
// THIS IS A HACK THAT WORKS ONLY AT LLNL
/* No mpibind or it is a single rank run */
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &myid );
//affs(myid);
hypre_MPI_Comm node_comm;
hypre_MPI_Info info;
hypre_MPI_Info_create(&info);
hypre_MPI_Comm_split_type(hypre_MPI_COMM_WORLD, hypre_MPI_COMM_TYPE_SHARED, myid, info, &node_comm);
hypre_int round_robin=1;
hypre_int myNodeid, NodeSize;
hypre_MPI_Comm_rank(node_comm, &myNodeid);
hypre_MPI_Comm_size(node_comm, &NodeSize);
if (round_robin)
{
/* Round robin allocation of GPUs. Does not account for affinities */
HYPRE_DEVICE=myNodeid%nDevices;
hypre_CheckErrorDevice(cudaSetDevice(HYPRE_DEVICE));
cudaDeviceGetPCIBusId ( pciBusId, 80, HYPRE_DEVICE);
hypre_printf("WARNING:: Code running without mpibind\n");
hypre_printf("Global ID = %d , Node ID %d running on device %d of %d \n",myid,myNodeid,HYPRE_DEVICE,nDevices);
}
else
{
/* Try to set the GPU based on process binding */
/* works correcly for all cases */
hypre_MPI_Comm numa_comm;
hypre_MPI_Comm_split(node_comm,getnuma(),myNodeid,&numa_comm);
hypre_int myNumaId,NumaSize;
hypre_MPI_Comm_rank(numa_comm, &myNumaId);
hypre_MPI_Comm_size(numa_comm, &NumaSize);
hypre_int domain_devices=nDevices/2; /* Again hardwired for 2 NUMA domains */
HYPRE_DEVICE = getnuma()*2+myNumaId%domain_devices;
hypre_CheckErrorDevice(cudaSetDevice(HYPRE_DEVICE));
hypre_printf("WARNING:: Code running without mpibind\n");
hypre_printf("NUMA %d GID %d , NodeID %d NumaID %d running on device %d (RR=%d) of %d \n",getnuma(),myid,myNodeid,myNumaId,HYPRE_DEVICE,myNodeid%nDevices,nDevices);
}
hypre_MPI_Info_free(&info);
}
else
{
/* Try to set the GPU based on process binding */
/* works correcly for all cases */
hypre_MPI_Comm numa_comm;
hypre_MPI_Comm_split(node_comm,getnuma(),myNodeid,&numa_comm);
hypre_int myNumaId,NumaSize;
hypre_MPI_Comm_rank(numa_comm, &myNumaId);
hypre_MPI_Comm_size(numa_comm, &NumaSize);
hypre_int domain_devices=nDevices/2; /* Again hardwired for 2 NUMA domains */
HYPRE_DEVICE = getnuma()*2+myNumaId%domain_devices;
hypre_CheckErrorDevice(cudaSetDevice(HYPRE_DEVICE));
hypre_printf("WARNING:: Code running without mpibind\n");
hypre_printf("NUMA %d GID %d , NodeID %d NumaID %d running on device %d (RR=%d) of %d \n",getnuma(),myid,myNodeid,myNumaId,HYPRE_DEVICE,myNodeid%nDevices,nDevices);
/* No device found */
hypre_fprintf(stderr,"ERROR:: NO GPUS found \n");
exit(2);
}
hypre_MPI_Info_free(&info);
}
else
{
/* No device found */
hypre_fprintf(stderr,"ERROR:: NO GPUS found \n");
exit(2);
HYPRE_DEVICE = use_device;
hypre_CheckErrorDevice(cudaSetDevice(HYPRE_DEVICE));
}
}
else
{
HYPRE_DEVICE = use_device;
hypre_CheckErrorDevice(cudaSetDevice(HYPRE_DEVICE));
}
#if defined(HYPRE_USING_OPENMP_OFFLOAD) || defined(HYPRE_USING_MAPPED_OPENMP_OFFLOAD)
omp_set_default_device(HYPRE_DEVICE);
printf("Set OMP Default device to %d \n",HYPRE_DEVICE);
omp_set_default_device(HYPRE_DEVICE);
printf("Set OMP Default device to %d \n",HYPRE_DEVICE);
#endif
/* Create NVTX domain for all the nvtx calls in HYPRE */
HYPRE_DOMAIN=nvtxDomainCreateA("Hypre");
/* Initialize streams */
hypre_int jj;
for(jj=0;jj<MAX_HGS_ELEMENTS;jj++)
hypre_CheckErrorDevice(cudaStreamCreateWithFlags(&(HYPRE_STREAM(jj)),cudaStreamNonBlocking));
hypre_CheckErrorDevice(cudaStreamCreateWithFlags(&(HYPRE_STREAM(jj)),cudaStreamNonBlocking));
/* Initialize the library handles and streams */
cusparseErrchk(cusparseCreate(&(HYPRE_CUSPARSE_HANDLE)));
cusparseErrchk(cusparseSetStream(HYPRE_CUSPARSE_HANDLE,HYPRE_STREAM(4)));
//cusparseErrchk(cusparseSetStream(HYPRE_CUSPARSE_HANDLE,0)); // Cusparse MxV happens in default stream
cusparseErrchk(cusparseCreateMatDescr(&(HYPRE_CUSPARSE_MAT_DESCR)));
cusparseErrchk(cusparseSetMatType(HYPRE_CUSPARSE_MAT_DESCR,CUSPARSE_MATRIX_TYPE_GENERAL));
cusparseErrchk(cusparseSetMatIndexBase(HYPRE_CUSPARSE_MAT_DESCR,CUSPARSE_INDEX_BASE_ZERO));
cublasErrchk(cublasCreate(&(HYPRE_CUBLAS_HANDLE)));
cublasErrchk(cublasSetStream(HYPRE_CUBLAS_HANDLE,HYPRE_STREAM(4)));
if (!checkDeviceProps()) hypre_printf("WARNING:: Concurrent memory access not allowed\n");
cusparseErrchk(cusparseCreate(&(HYPRE_CUSPARSE_HANDLE)));
cusparseErrchk(cusparseSetStream(HYPRE_CUSPARSE_HANDLE,HYPRE_STREAM(4)));
//cusparseErrchk(cusparseSetStream(HYPRE_CUSPARSE_HANDLE,0)); // Cusparse MxV happens in default stream
cusparseErrchk(cusparseCreateMatDescr(&(HYPRE_CUSPARSE_MAT_DESCR)));
cusparseErrchk(cusparseSetMatType(HYPRE_CUSPARSE_MAT_DESCR,CUSPARSE_MATRIX_TYPE_GENERAL));
cusparseErrchk(cusparseSetMatIndexBase(HYPRE_CUSPARSE_MAT_DESCR,CUSPARSE_INDEX_BASE_ZERO));
/* Check if the arch flags used for compiling the cuda kernels match the device */
cublasErrchk(cublasCreate(&(HYPRE_CUBLAS_HANDLE)));
cublasErrchk(cublasSetStream(HYPRE_CUBLAS_HANDLE,HYPRE_STREAM(4)));
if (!checkDeviceProps()) hypre_printf("WARNING:: Concurrent memory access not allowed\n");
/* Check if the arch flags used for compiling the cuda kernels match the device */
#if defined(HYPRE_USING_GPU)
CudaCompileFlagCheck();
CudaCompileFlagCheck();
#endif
}
}
}
void hypre_GPUFinalize()
{
cusparseErrchk(cusparseDestroy(HYPRE_CUSPARSE_HANDLE));
cublasErrchk(cublasDestroy(HYPRE_CUBLAS_HANDLE));
cusparseErrchk(cusparseDestroy(HYPRE_CUSPARSE_HANDLE));
cublasErrchk(cublasDestroy(HYPRE_CUBLAS_HANDLE));
#if defined(HYPRE_MEASURE_GPU_HWM)
hypre_printf("GPU Memory High Water Mark(per MPI_RANK) %f MB \n",(HYPRE_Real)HYPRE_GPU_HWM/1024/1024);
#endif
/* Destroy streams */
hypre_int jj;
for(jj=0;jj<MAX_HGS_ELEMENTS;jj++)
hypre_CheckErrorDevice(cudaStreamDestroy(HYPRE_STREAM(jj)));
hypre_printf("GPU Memory High Water Mark(per MPI_RANK) %f MB \n",(HYPRE_Real)HYPRE_GPU_HWM/1024/1024);
#endif
/* Destroy streams */
hypre_int jj;
for(jj=0;jj<MAX_HGS_ELEMENTS;jj++)
{
hypre_CheckErrorDevice(cudaStreamDestroy(HYPRE_STREAM(jj)));
}
}
void MemAdviseReadOnly(const void* ptr, hypre_int device){
@ -349,31 +348,31 @@ void affs(hypre_int myid){
hypre_int cpus[NCPUS],i;
hypre_int retval=sched_getaffinity(0, size,mask);
if (!retval){
for(i=0;i<NCPUS;i++){
if (CPU_ISSET(i,mask))
cpus[i]=1;
else
cpus[i]=0;
}
printf("Node(%d)::",myid);
for(i=0;i<160;i++)printf("%d",cpus[i]);
printf("\n");
for(i=0;i<NCPUS;i++){
if (CPU_ISSET(i,mask))
cpus[i]=1;
else
cpus[i]=0;
}
printf("Node(%d)::",myid);
for(i=0;i<160;i++)printf("%d",cpus[i]);
printf("\n");
} else {
fprintf(stderr,"sched_affinity failed\n");
switch(errno){
case EFAULT:
printf("INVALID MEMORY ADDRESS\n");
break;
case EINVAL:
printf("EINVAL:: NO VALID CPUS\n");
break;
default:
printf("%d something else\n",errno);
}
fprintf(stderr,"sched_affinity failed\n");
switch(errno){
case EFAULT:
printf("INVALID MEMORY ADDRESS\n");
break;
case EINVAL:
printf("EINVAL:: NO VALID CPUS\n");
break;
default:
printf("%d something else\n",errno);
}
}
CPU_FREE(mask);
}
hypre_int getcore(){
const hypre_int NCPUS=160;
@ -384,8 +383,8 @@ hypre_int getcore(){
if (!retval){
for(i=0;i<NCPUS;i+=20){
if (CPU_ISSET(i,mask)) {
CPU_FREE(mask);
return i;
CPU_FREE(mask);
return i;
}
}
} else {
@ -403,7 +402,7 @@ hypre_int getcore(){
}
return 0;
CPU_FREE(mask);
}
hypre_int getnuma(){
const hypre_int NCPUS=160;
@ -471,47 +470,47 @@ size_t mempush(const void *ptr, size_t size, hypre_int action){
nc++;
return size;
} else {
// Purge an address
if (action==1){
found=memfind(head,ptr);
if (found){
memdel(&head, found);
nc--;
return 0;
} else {
// Purge an address
if (action==1){
found=memfind(head,ptr);
if (found){
memdel(&head, found);
nc--;
return 0;
} else {
#ifdef FULL_WARN
fprintf(stderr,"ERROR :: Pointer for deletion not found in linked list %p\n",ptr);
fprintf(stderr,"ERROR :: Pointer for deletion not found in linked list %p\n",ptr);
#endif
return 0;
}
} // End purge
// Insertion
if (size>0){
found=memfind(head,ptr);
if (found){
#ifdef FULL_WARN
fprintf(stderr,"ERROR :: Pointer for insertion already in use in linked list %p\n",ptr);
//printlist(head,nc);
#endif
return 0;
} else {
nc++;
meminsert(&head,ptr,size);
return 0;
}
}
return 0;
}
} // End purge
// Getting allocation size
found=memfind(head,ptr);
if (found){
return found->size;
} else{
// Insertion
if (size>0){
found=memfind(head,ptr);
if (found){
#ifdef FULL_WARN
fprintf(stderr,"ERROR :: Pointer for size check NOT found in linked list\n");
fprintf(stderr,"ERROR :: Pointer for insertion already in use in linked list %p\n",ptr);
//printlist(head,nc);
#endif
return 0;
}
return 0;
} else {
nc++;
meminsert(&head,ptr,size);
return 0;
}
}
// Getting allocation size
found=memfind(head,ptr);
if (found){
return found->size;
} else{
#ifdef FULL_WARN
fprintf(stderr,"ERROR :: Pointer for size check NOT found in linked list\n");
#endif
return 0;
}
}
}
@ -569,7 +568,7 @@ void printlist(node *head,hypre_int nc){
/* num: number of bytes */
HYPRE_Int HYPRE_OMPOffload(HYPRE_Int device, void *ptr, size_t num,
const char *type1, const char *type2) {
const char *type1, const char *type2) {
hypre_omp45_offload(device, ptr, char, 0, num, type1, type2);
return 0;

7
src/utilities/mpistubs.c
View File

@ -651,7 +651,7 @@ hypre_MPI_Op_free( hypre_MPI_Op *op )
return(0);
}
#if defined(HYPRE_USING_CUDA) || defined(HYPRE_USING_DEVICE_OPENMP)
HYPRE_Int hypre_MPI_Comm_split_type( hypre_MPI_Comm comm, HYPRE_Int split_type, HYPRE_Int key, hypre_MPI_Info info, hypre_MPI_Comm *newcomm )
{
return (0);
@ -666,6 +666,7 @@ HYPRE_Int hypre_MPI_Info_free( hypre_MPI_Info *info )
{
return (0);
}
#endif
/******************************************************************************
* MPI stubs to do casting of HYPRE_Int and hypre_int correctly
@ -1307,7 +1308,8 @@ hypre_MPI_Op_create( hypre_MPI_User_function *function, hypre_int commute, hypre
return (HYPRE_Int) MPI_Op_create(function, commute, op);
}
HYPRE_Int
#if defined(HYPRE_USING_CUDA) || defined(HYPRE_USING_DEVICE_OPENMP)
HYPRE_Int
hypre_MPI_Comm_split_type( hypre_MPI_Comm comm, HYPRE_Int split_type, HYPRE_Int key, hypre_MPI_Info info, hypre_MPI_Comm *newcomm )
{
return (HYPRE_Int) MPI_Comm_split_type(comm, split_type, key, info, newcomm );
@ -1324,5 +1326,6 @@ hypre_MPI_Info_free( hypre_MPI_Info *info )
{
return (HYPRE_Int) MPI_Info_free(info);
}
#endif
#endif

2
src/utilities/mpistubs.h
View File

@ -300,9 +300,11 @@ HYPRE_Int hypre_MPI_Type_commit( hypre_MPI_Datatype *datatype );
HYPRE_Int hypre_MPI_Type_free( hypre_MPI_Datatype *datatype );
HYPRE_Int hypre_MPI_Op_free( hypre_MPI_Op *op );
HYPRE_Int hypre_MPI_Op_create( hypre_MPI_User_function *function , hypre_int commute , hypre_MPI_Op *op );
#if defined(HYPRE_USING_CUDA) || defined(HYPRE_USING_DEVICE_OPENMP)
HYPRE_Int hypre_MPI_Comm_split_type(hypre_MPI_Comm comm, HYPRE_Int split_type, HYPRE_Int key, hypre_MPI_Info info, hypre_MPI_Comm *newcomm);
HYPRE_Int hypre_MPI_Info_create(hypre_MPI_Info *info);
HYPRE_Int hypre_MPI_Info_free( hypre_MPI_Info *info );
#endif
#ifdef __cplusplus
}