void pgp_memzero(Integer g_p, Integer intsize)
{
void *gp_ptr, *size_array;
Integer handle, i, me, nelems, ndim;
Integer lo[GP_MAX_DIM], hi[GP_MAX_DIM], ld[GP_MAX_DIM-1];
Integer j;
pnga_sync();
handle = g_p + GP_OFFSET;
me = pnga_nodeid();
ndim = GP[handle].ndim;
/* Determine number of elements held locally */
pgp_distribution(g_p, me, lo, hi);
nelems = 1;
for (i=0; i<ndim; i++) {
nelems *= (hi[i]-lo[i]+1);
}
/* Get pointers to local data elements and their sizes */
pnga_access_ptr(GP[handle].g_ptr_array,lo,hi,&gp_ptr,ld);
pnga_access_ptr(GP[handle].g_size_array,lo,hi,&size_array,ld);
/* Zero bits in data elements */
/*bjp
printf("p[%d] nelems: %d ld[0]: %d\n",me,nelems,ld[0]);
*/
for (i=0; i<nelems; i++) {
/*bjp
printf("p[%d] gp_ptr[%d].addr: %p size_array[%d]: %d\n",me,
i,(void*)((armci_meminfo_t*)gp_ptr)[i].addr,i,
(int)((int*)size_array)[i]);
*/
if (intsize == 4) {
memset((void*)((armci_meminfo_t*)gp_ptr)[i].addr, 0,
(size_t)((int*)size_array)[i]);
#if 0
for (j=0; j<((int*)size_array)[i]; j++) {
if (((char*)((armci_meminfo_t*)gp_ptr)[i].addr)[j] != 0) {
printf("p[%d] mismatch for i: %d j: %d\n",me,i,j);
}
}
#endif
} else {
memset((void*)((armci_meminfo_t*)gp_ptr)[i].addr, 0,
(size_t)((int64_t*)size_array)[i]);
#if 0
for (j=0; j<((int64_t*)size_array)[i]; j++) {
if (((char*)((armci_meminfo_t*)gp_ptr)[i].addr)[j] != 0) {
printf("p[%d] mismatch for i: %d j: %d\n",me,i,j);
}
}
#endif
}
}
pnga_release_update(GP[handle].g_ptr_array,lo,hi);
pnga_release_update(GP[handle].g_size_array,lo,hi);
pnga_sync();
}
void* pgp_free_local_element(Integer g_p, Integer *subscript)
{
armci_meminfo_t *gp_ptr;
void *ptr;
Integer handle, ld[GP_MAX_DIM-1], i;
GP_Int buf;
handle = g_p + GP_OFFSET;
/* check to make sure that element is located in local block of GP array */
for (i=0; i<GP[handle].ndim; i++) {
if (subscript[i]<GP[handle].lo[i] || subscript[i]>GP[handle].hi[i]) {
pnga_error("gp_free_local_element: subscript out of bounds", i);
}
}
pnga_access_ptr(GP[handle].g_ptr_array,subscript,subscript,&gp_ptr,ld);
ptr = (*gp_ptr).addr;
memset((void*)gp_ptr,0,sizeof(armci_meminfo_t));
pnga_release_update(GP[handle].g_ptr_array, subscript, subscript);
/* set corresponding element of size array to zero */
buf = 0;
for (i=0; i<GP[handle].ndim-1; i++) {
ld[i] = 1;
}
pnga_put(GP[handle].g_size_array, subscript, subscript, &buf, ld);
return ptr;
}
/**
* Get the next sub-block from local portion of global array
* @param hdl handle for iterator
* @param plo indices for lower corner of block
* @param phi indices for upper corner of block
* @param ptr pointer to local buffer
* @param ld array of strides for local block
* @return returns false if there is no new block, true otherwise
*/
int pnga_local_iterator_next(_iterator_hdl *hdl, Integer plo[],
Integer phi[], char **ptr, Integer ld[])
{
Integer i;
Integer handle = GA_OFFSET + hdl->g_a;
Integer grp = GA[handle].p_handle;
Integer elemsize = GA[handle].elemsize;
int ndim;
int me = pnga_pgroup_nodeid(grp);
ndim = GA[handle].ndim;
if (GA[handle].distr_type == REGULAR) {
Integer nelems;
/* no blocks left, so return */
if (hdl->count>0) return 0;
/* Find visible portion of patch held by this processor and
* return the result in plo and phi. Return pointer to local
* data as well
*/
pnga_distribution(hdl->g_a, me, plo, phi);
/* Check to see if this process has any data. Return 0 if
* it does not */
for (i=0; i<ndim; i++) {
if (phi[i]<plo[i]) return 0;
}
pnga_access_ptr(hdl->g_a,plo,phi,ptr,ld);
hdl->count++;
} else if (GA[handle].distr_type == BLOCK_CYCLIC) {
/* Simple block-cyclic distribution */
if (hdl->count >= pnga_total_blocks(hdl->g_a)) return 0;
pnga_distribution(hdl->g_a,hdl->count,plo,phi);
pnga_access_block_ptr(hdl->g_a,hdl->count,ptr,ld);
hdl->count += pnga_pgroup_nnodes(grp);
} else if (GA[handle].distr_type == SCALAPACK ||
GA[handle].distr_type == TILED) {
/* Scalapack-type data distribution */
if (hdl->index[ndim-1] >= hdl->blk_num[ndim-1]) return 0;
/* Find coordinates of bounding block */
for (i=0; i<ndim; i++) {
plo[i] = hdl->index[i]*hdl->blk_size[i]+1;
phi[i] = (hdl->index[i]+1)*hdl->blk_size[i];
if (phi[i] > hdl->blk_dim[i]) phi[i] = hdl->blk_dim[i];
}
pnga_access_block_grid_ptr(hdl->g_a,hdl->index,ptr,ld);
hdl->index[0] += hdl->blk_inc[0];
for (i=0; i<ndim; i++) {
if (hdl->index[i] >= hdl->blk_num[i] && i<ndim-1) {
hdl->index[i] = hdl->proc_index[i];
hdl->index[i+1] += hdl->blk_inc[i+1];
}
}
}
return 1;
}
void pgp_assign_local_element(Integer g_p, Integer *subscript, void *ptr,
Integer size, Integer intsize)
{
void *gp_ptr;
Integer handle, ld[GP_MAX_DIM-1], i;
handle = g_p + GP_OFFSET;
/* check to make sure that element is located in local block of GP array */
for (i=0; i<GP[handle].ndim; i++) {
if (subscript[i]<GP[handle].lo[i] || subscript[i]>GP[handle].hi[i]) {
/*bjp
printf("p[%d] subscript[%d]: %d\n",pnga_nodeid(),i,subscript[i]);
printf("p[%d] lo[%d]: %d hi[%d]: %d\n",pnga_nodeid(),i,GP[handle].lo[i],i,
GP[handle].hi[i]);
*/
/*
printf("p[%d] subscript[%d]: %d lo[%d]: %d hi[%d]: %d\n",pnga_nodeid(),
i, subscript[i], i, GP[handle].lo[i], i, GP[handle].hi[i]);
*/
pnga_error("gp_assign_local_element: subscript out of bounds", i);
}
}
pnga_access_ptr(GP[handle].g_size_array,subscript,subscript,&gp_ptr,ld);
if (intsize == 4) {
*((int*)gp_ptr) = (int)size;
} else {
*((int64_t*)gp_ptr) = (int64_t)size;
}
/*bjp
printf("p[%ld] (internal) size %d at location [%ld:%ld]\n",
(long)pnga_nodeid(), *((int*)gp_ptr),
(long)subscript[0],(long)subscript[1]);
*/
pnga_release_update(GP[handle].g_size_array, subscript, subscript);
pnga_access_ptr(GP[handle].g_ptr_array,subscript,subscript,&gp_ptr,ld);
*((armci_meminfo_t*)gp_ptr) =
*((armci_meminfo_t*)(((char*)ptr)-sizeof(armci_meminfo_t)));
pnga_release_update(GP[handle].g_ptr_array, subscript, subscript);
}
void pnga_select_elem(Integer g_a, char* op, void* val, Integer *subscript)
{
Integer ndim, type, me, elems, ind=0, i;
Integer lo[MAXDIM],hi[MAXDIM],dims[MAXDIM],ld[MAXDIM-1];
elem_info_t info;
Integer num_blocks;
int participate=0;
int local_sync_begin;
local_sync_begin = _ga_sync_begin;
_ga_sync_begin = 1; _ga_sync_end=1; /*remove any previous masking*/
if(local_sync_begin)pnga_sync();
me = pnga_nodeid();
pnga_check_handle(g_a, "ga_select_elem");
GA_PUSH_NAME("ga_elem_op");
if (strncmp(op,"min",3) == 0);
else if (strncmp(op,"max",3) == 0);
else pnga_error("operator not recognized",0);
pnga_inquire(g_a, &type, &ndim, dims);
num_blocks = pnga_total_blocks(g_a);
if (num_blocks < 0) {
pnga_distribution(g_a, me, lo, hi);
if ( lo[0]> 0 ){ /* base index is 1: we get 0 if no elements stored on p */
/******************* calculate local result ************************/
void *ptr;
pnga_access_ptr(g_a, lo, hi, &ptr, ld);
GET_ELEMS(ndim,lo,hi,ld,&elems);
participate =1;
/* select local element */
snga_select_elem(type, op, ptr, elems, &info, &ind);
/* release access to the data */
pnga_release(g_a, lo, hi);
/* determine element subscript in the ndim-array */
for(i = 0; i < ndim; i++){
int elems = (int)( hi[i]-lo[i]+1);
info.subscr[i] = ind%elems + lo[i] ;
ind /= elems;
}
}
} else {
void *ptr;
Integer j, offset, jtot, upper;
Integer nproc = pnga_nnodes();
pnga_access_block_segment_ptr(g_a, me, &ptr, &elems);
if (elems > 0) {
participate =1;
/* select local element */
snga_select_elem(type, op, ptr, elems, &info, &ind);
/* release access to the data */
pnga_release_block_segment(g_a, me);
/* convert local index back into a global array index */
if (!pnga_uses_proc_grid(g_a)) {
offset = 0;
for (i=me; i<num_blocks; i += nproc) {
pnga_distribution(g_a, i, lo, hi);
jtot = 1;
for (j=0; j<ndim; j++) {
jtot *= (hi[j]-lo[j]+1);
}
upper = offset + jtot;
if (ind >= offset && ind < upper) {
break;
} else {
offset += jtot;
}
}
/* determine element subscript in the ndim-array */
ind -= offset;
for(i = 0; i < ndim; i++){
int elems = (int)( hi[i]-lo[i]+1);
info.subscr[i] = ind%elems + lo[i] ;
ind /= elems;
}
} else {
Integer stride[MAXDIM], index[MAXDIM];
Integer blocks[MAXDIM], block_dims[MAXDIM];
Integer proc_index[MAXDIM], topology[MAXDIM];
Integer l_index[MAXDIM];
Integer min, max;
pnga_get_proc_index(g_a, me, proc_index);
pnga_get_block_info(g_a, blocks, block_dims);
pnga_get_proc_grid(g_a, topology);
/* figure out strides for locally held block of data */
for (i=0; i<ndim; i++) {
stride[i] = 0;
for (j=proc_index[i]; j<blocks[i]; j += topology[i]) {
min = j*block_dims[i] + 1;
max = (j+1)*block_dims[i];
if (max > dims[i])
max = dims[i];
stride[i] += (max - min + 1);
}
}
/* use strides to figure out local index */
l_index[0] = ind%stride[0];
for (i=1; i<ndim; i++) {
ind = (ind-l_index[i-1])/stride[i-1];
l_index[i] = ind%stride[i];
}
/* figure out block index for block holding data element */
for (i=0; i<ndim; i++) {
index[i] = l_index[i]/block_dims[i];
}
for (i=0; i<ndim; i++) {
lo[i] = (topology[i]*index[i] + proc_index[i])*block_dims[i];
info.subscr[i] = l_index[i]%block_dims[i] + lo[i];
}
}
}
}
/* calculate global result */
if(type==C_INT){
int size = sizeof(double) + sizeof(Integer)*(int)ndim;
armci_msg_sel(&info,size,op,ARMCI_INT,participate);
*(int*)val = (int)info.v.ival;
}else if(type==C_LONG){
int size = sizeof(double) + sizeof(Integer)*(int)ndim;
armci_msg_sel(&info,size,op,ARMCI_LONG,participate);
*(long*)val = info.v.lval;
}else if(type==C_LONGLONG){
int size = sizeof(double) + sizeof(Integer)*(int)ndim;
armci_msg_sel(&info,size,op,ARMCI_LONG_LONG,participate);
*(long long*)val = info.v.llval;
}else if(type==C_DBL){
int size = sizeof(double) + sizeof(Integer)*(int)ndim;
armci_msg_sel(&info,size,op,ARMCI_DOUBLE,participate);
*(DoublePrecision*)val = info.v.dval;
}else if(type==C_FLOAT){
int size = sizeof(double) + sizeof(Integer)*ndim;
armci_msg_sel(&info,size,op,ARMCI_FLOAT,participate);
*(float*)val = info.v.fval;
}else if(type==C_SCPL){
int size = sizeof(info); /* for simplicity we send entire info */
armci_msg_sel(&info,size,op,ARMCI_FLOAT,participate);
*(SingleComplex*)val = info.extra2;
}else{
int size = sizeof(info); /* for simplicity we send entire info */
armci_msg_sel(&info,size,op,ARMCI_DOUBLE,participate);
*(DoubleComplex*)val = info.extra;
}
for(i = 0; i < ndim; i++) subscript[i]= info.subscr[i];
GA_POP_NAME;
}
