void comm_init(int ndim, const int *dims, QudaCommsMap rank_from_coords, void *map_data)
{
if ( QMP_is_initialized() != QMP_TRUE ) {
errorQuda("QMP has not been initialized");
}
int grid_size = 1;
for (int i = 0; i < ndim; i++) {
grid_size *= dims[i];
}
if (grid_size != QMP_get_number_of_nodes()) {
errorQuda("Communication grid size declared via initCommsGridQuda() does not match"
" total number of QMP nodes (%d != %d)", grid_size, QMP_get_number_of_nodes());
}
Topology *topo = comm_create_topology(ndim, dims, rank_from_coords, map_data);
comm_set_default_topology(topo);
// determine which GPU this process will use (FIXME: adopt the scheme in comm_mpi.cpp)
int device_count;
cudaGetDeviceCount(&device_count);
if (device_count == 0) {
errorQuda("No CUDA devices found");
}
gpuid = (comm_rank() % device_count);
}
int main(int argc, char **argv) {
#ifdef QMP_COMMS
int ndim=4, dims[4];
QMP_thread_level_t tl;
QMP_init_msg_passing(&argc, &argv, QMP_THREAD_SINGLE, &tl);
dims[0] = dims[1] = dims[2] = 1;
dims[3] = QMP_get_number_of_nodes();
QMP_declare_logical_topology(dims, ndim);
#endif
SU3Test();
#ifdef QMP_COMMS
QMP_finalize_msg_passing();
#endif
return 0;
}
void
stupid_broadcast(void *send_buf, int count)
{
int node;
int num_nodes = QMP_get_number_of_nodes();
QMP_msgmem_t request_msg = QMP_declare_msgmem(send_buf, count);
QMP_msghandle_t request_mh;
// Send to each node
for(node=1; node < num_nodes; ++node)
{
if (QMP_get_node_number() == node)
{
request_mh = QMP_declare_receive_from(request_msg, 0, 0);
if (QMP_start(request_mh) != QMP_SUCCESS)
QMP_abort_string(1, "recvFromWait failed\n");
QMP_wait(request_mh);
QMP_free_msghandle(request_mh);
}
if (QMP_is_primary_node())
{
request_mh = QMP_declare_send_to(request_msg, node, 0);
if (QMP_start(request_mh) != QMP_SUCCESS)
QMP_abort_string(1, "sendToWait failed\n");
QMP_wait(request_mh);
QMP_free_msghandle(request_mh);
}
}
QMP_free_msgmem(request_msg);
}
int comm_size(void)
{
return QMP_get_number_of_nodes();
}
void initQuda(int dev)
{
static int initialized = 0;
if (initialized) {
return;
}
initialized = 1;
#if (CUDA_VERSION >= 4000) && defined(MULTI_GPU)
//check if CUDA_NIC_INTEROP is set to 1 in the enviroment
char* cni_str = getenv("CUDA_NIC_INTEROP");
if(cni_str == NULL){
errorQuda("Environment variable CUDA_NIC_INTEROP is not set\n");
}
int cni_int = atoi(cni_str);
if (cni_int != 1){
errorQuda("Environment variable CUDA_NIC_INTEROP is not set to 1\n");
}
#endif
int deviceCount;
cudaGetDeviceCount(&deviceCount);
if (deviceCount == 0) {
errorQuda("No devices supporting CUDA");
}
for(int i=0; i<deviceCount; i++) {
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, i);
printfQuda("QUDA: Found device %d: %s\n", i, deviceProp.name);
}
#ifdef QMP_COMMS
int ndim;
const int *dim;
if ( QMP_is_initialized() != QMP_TRUE ) {
errorQuda("QMP is not initialized");
}
num_QMP=QMP_get_number_of_nodes();
rank_QMP=QMP_get_node_number();
dev += rank_QMP % deviceCount;
ndim = QMP_get_logical_number_of_dimensions();
dim = QMP_get_logical_dimensions();
#elif defined(MPI_COMMS)
comm_init();
dev=comm_gpuid();
#else
if (dev < 0) dev = deviceCount - 1;
#endif
// Used for applying the gauge field boundary condition
if( commCoords(3) == 0 ) qudaPt0=true;
else qudaPt0=false;
if( commCoords(3) == commDim(3)-1 ) qudaPtNm1=true;
else qudaPtNm1=false;
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, dev);
if (deviceProp.major < 1) {
errorQuda("Device %d does not support CUDA", dev);
}
printfQuda("QUDA: Using device %d: %s\n", dev, deviceProp.name);
cudaSetDevice(dev);
#ifdef HAVE_NUMA
if(numa_config_set){
if(gpu_affinity[dev] >=0){
printfQuda("Numa setting to cpu node %d\n", gpu_affinity[dev]);
if(numa_run_on_node(gpu_affinity[dev]) != 0){
printfQuda("Warning: Setting numa to cpu node %d failed\n", gpu_affinity[dev]);
}
}
}
#endif
initCache();
quda::initBlas();
}
int qio_test(int output_volfmt, int output_serpar, int ildgstyle,
int input_volfmt, int input_serpar, int argc, char *argv[]){
float array_in[NARRAY], array_out[NARRAY];
float *field_in[NREAL], *subset_in[NREAL],
*field_out[NREAL], *subset_out[NREAL];
suN_matrix *field_su3_out[NMATRIX], *field_su3_in[NMATRIX];
QIO_Writer *outfile;
QIO_Reader *infile;
float diff_field = 0, diff_array = 0, diff_su3 = 0, diff_subset = 0;
QMP_thread_level_t provided;
int status;
int sites_on_node = 0;
int i,volume;
char filename[] = "binary_test";
int dim = 4;
int lower[4] = {1, 0, 0, 2};
int upper[4] = {2, 3, 3, 2};
char myname[] = "qio_test";
/* Start message passing */
QMP_init_msg_passing(&argc, &argv, QMP_THREAD_SINGLE, &provided);
this_node = mynode();
printf("%s(%d) QMP_init_msg_passing done\n",myname,this_node);
/* Lattice dimensions */
lattice_dim = 4;
lattice_size[0] = 8;
lattice_size[1] = 4;
lattice_size[2] = 4;
lattice_size[3] = 4;
volume = 1;
for(i = 0; i < lattice_dim; i++){
volume *= lattice_size[i];
}
/* Set the mapping of coordinates to nodes */
if(setup_layout(lattice_size, 4, QMP_get_number_of_nodes())!=0)
return 1;
printf("%s(%d) layout set for %d nodes\n",myname,this_node,
QMP_get_number_of_nodes());
sites_on_node = num_sites(this_node);
/* Build the layout structure */
layout.node_number = node_number;
layout.node_index = node_index;
layout.get_coords = get_coords;
layout.num_sites = num_sites;
layout.latsize = lattice_size;
layout.latdim = lattice_dim;
layout.volume = volume;
layout.sites_on_node = sites_on_node;
layout.this_node = this_node;
layout.number_of_nodes = QMP_get_number_of_nodes();
/* Open the test output file */
outfile = open_test_output(filename, output_volfmt, output_serpar,
ildgstyle, myname);
if(outfile == NULL)return 1;
/* If this is not the ILDG file test */
if(ildgstyle == QIO_ILDGNO){
/* Create the test output field */
status = vcreate_R(field_out, NREAL);
if(status)return status;
/* Set some values for the field */
vset_R(field_out, NREAL);
/* Write the real test field */
status = write_real_field(outfile, NREAL, field_out, myname);
if(status)return status;
/* Write a subset of the real test field */
status = write_real_field_subset(outfile, NREAL, field_out,
lower, upper, dim, myname);
if(status)return status;
/* Set some values for the global array */
for(i = 0; i < NARRAY; i++)
array_out[i] = i;
/* Write the real global array */
status = write_real_global(outfile, NARRAY, array_out, myname);
if(status)return status;
}
/* Create the test output su3 field */
status = vcreate_M(field_su3_out, NMATRIX);
if(status)return status;
/* Set some values for the su3 field */
vset_M(field_su3_out, NMATRIX);
/* Write the su3 test field */
status = write_su3_field(outfile, NMATRIX, field_su3_out, myname);
if(status)return status;
/* Close the file */
QIO_close_write(outfile);
printf("%s(%d): Closed file for writing\n",myname,this_node);
/* Set up a dummy input field */
status = vcreate_R(field_in, NREAL);
if(status)return status;
/* Set up a dummy input field for subset */
status = vcreate_R(subset_in, NREAL);
if(status)return status;
/* Set up a dummy input SU(N) field */
status = vcreate_M(field_su3_in, NMATRIX);
if(status)return status;
/* Open the test file for reading */
infile = open_test_input(filename, input_volfmt, input_serpar, myname);
if(infile == NULL)return 1;
if(ildgstyle == QIO_ILDGNO){
/* Peek at the field record */
status = peek_record_info(infile, myname);
if(status != QIO_SUCCESS)return status;
/* Skip the record */
#if(0)
/* Skip the field */
status = QIO_next_record(infile);
if(status != QIO_SUCCESS)return status;
#else
/* Read the field record */
printf("%s(%d) reading real field\n",myname,this_node); fflush(stdout);
status = read_real_field(infile, NREAL, field_in, myname);
if(status)return status;
#endif
/* Read the subset of the field */
printf("%s(%d) reading subset of real field\n",
myname,this_node); fflush(stdout);
status = read_real_field_subset(infile, NREAL, subset_in, myname);
if(status)return status;
/* Read the global array record */
printf("%s(%d) reading global field\n",myname,this_node); fflush(stdout);
status = read_real_global(infile, NARRAY, array_in, myname);
if(status)return status;
}
/* Read the su3 field record */
printf("%s(%d) reading su3 field\n",myname,this_node); fflush(stdout);
status = read_su3_field(infile, NMATRIX, field_su3_in, myname);
if(status)return status;
/* Close the file */
QIO_close_read(infile);
printf("%s(%d): Closed file for reading\n",myname,this_node);
if(ildgstyle == QIO_ILDGNO){
/* Compare the input and output fields */
diff_field = vcompare_R(field_out, field_in, NREAL);
if(this_node == 0){
printf("%s(%d): Comparison of in and out real fields |in - out|^2 = %e\n",
myname,this_node,diff_field);
}
/* Create the subset output field */
status = vcreate_R(subset_out, NREAL);
if(status)return status;
/* Copy the subset */
vsubset_R(subset_out, field_out, lower, upper, NREAL);
/* Compare the input and output subsets */
diff_subset = vcompare_R(subset_out, subset_in, NREAL);
if(this_node == 0){
printf("%s(%d): Comparison of subsets of in and out real fields |in - out|^2 = %e\n",
myname,this_node,diff_subset);
}
/* Compare the input and output global arrays */
diff_array = vcompare_r(array_out, array_in, NREAL);
if(this_node == 0){
printf("%s(%d): Comparison of in and out real global arrays |in - out|^2 = %e\n",
myname, this_node, diff_array);
}
}
/* Compare the input and output suN fields */
diff_su3 = vcompare_M(field_su3_out, field_su3_in, NMATRIX);
if(this_node == 0){
printf("%s(%d): Comparison of in and out suN fields |in - out|^2 = %e\n",
myname, this_node, diff_field);
}
/* Clean up */
if(ildgstyle == QIO_ILDGNO){
vdestroy_R(field_out, NREAL);
vdestroy_R(field_in, NREAL);
vdestroy_R(subset_in, NREAL);
vdestroy_R(subset_out, NREAL);
}
vdestroy_M(field_su3_in, NMATRIX);
vdestroy_M(field_su3_out, NMATRIX);
/* Shut down QMP */
QMP_finalize_msg_passing();
/* Report result */
if(diff_field + diff_subset + diff_su3 + diff_array > 0){
printf("%s(%d): Test failed\n",myname,this_node);
return 1;
}
printf("%s(%d): Test passed\n",myname,this_node);
return 0;
}
static void
eo_setup(QDP_Lattice *lat, void *args)
{
mLattice *S = args;
QDP_allocate_lattice_params(lat, sizeof (params));
params *p = QDP_get_lattice_params(lat);
p->S = S;
if (QMP_get_msg_passing_type() != QMP_SWITCH) {
int nd2 = QMP_get_allocated_number_of_dimensions();
const int *nsquares2 = QMP_get_allocated_dimensions();
int i;
for (i = 0; i < S->rank; i++) {
S->net[i] = (i < nd2) ? nsquares2[i] : 1;
}
} else { /* not QMP_GRID */
int squaresize[QLUA_MAX_LATTICE_RANK];
int extrafactors[QLUA_MAX_LATTICE_RANK];
int i;
for (i = 0; i < S->rank; i++) {
squaresize[i] = S->dim[i];
extrafactors[i] = 1;
S->net[i] = 1;
}
/* Figure out dimensions of rectangle */
int n = QMP_get_number_of_nodes(); /* nodes to factor */
int k = MAXPRIMES-1;
while (n > 1) {
/* figure out which prime to divide by starting with largest */
/* if no factor found, assume n is prime */
while ((k >= 0) && (n % prime[k] != 0)) --k;
int pfac = (k>=0) ? prime[k] : n;
/* figure out which direction to divide */
/* find largest divisible dimension of h-cubes */
/* if one direction with largest dimension has already been
divided, divide it again. Otherwise divide first direction
with largest dimension. */
int j = -1;
int i;
for (i = 0; i < S->rank; i++) {
if (squaresize[i] % pfac == 0) {
if ((j<0) ||
(extrafactors[j] * squaresize[i] >
extrafactors[i] * squaresize[j])) {
j = i;
} else if (extrafactors[j] * squaresize[i] ==
extrafactors[i] * squaresize[j]) {
if ((S->net[j] == 1) || (S->net[i] != 1))
j = i;
}
}
}
/* This can fail if we run out of prime factors in the dimensions */
/* then just choose largest dimension */
if (j < 0) {
int i;
for (i = 0; i < S->rank; i++) {
if ((j<0) ||
(extrafactors[j] * squaresize[i] >
extrafactors[i] * squaresize[j]) ) {
j = i;
} else if (extrafactors[j] * squaresize[i] ==
extrafactors[i] * squaresize[j]) {
if((S->net[j] == 1) || (S->net[i] != 1))
j = i;
}
}
n /= pfac;
extrafactors[j] *= pfac;
S->net[j] *= pfac;
} else {
n /= pfac;
squaresize[j] /= pfac;
S->net[j] *= pfac;
}
}
} /* not QMP_GRID */
int mc[QLUA_MAX_LATTICE_RANK];
int i;
S->node = QDP_this_node;
node2coord(mc, QDP_this_node, S);
for (i = 0; i < S->rank; i++) {
int x = mc[i];
mc[i] = x + 1;
if (mc[i] == S->net[i])
mc[i] = 0;
S->neighbor_up[i] = coord2node(mc, S);
mc[i] = x - 1;
if (mc[i] < 0)
mc[i] = S->net[i] - 1;
S->neighbor_down[i] = coord2node(mc, S);
mc[i] = x;
}
}
void init_qmp(int * argc, char ***argv) {
#if 0
printf("init_qmp(%d %p)\n",*argc,*argv);
for(int i = 0; i<*argc;i++){
printf("argv[%d](before)=%s\n",i,(*argv)[i]);
}
#endif
#if 0
spi_init();
#endif
QMP_thread_level_t prv;
#ifndef UNIFORM_SEED_NO_COMMS
QMP_status_t init_status = QMP_init_msg_passing(argc, argv, QMP_THREAD_SINGLE, &prv);
if (init_status) printf("QMP_init_msg_passing returned %d\n",init_status);
peRank = QMP_get_node_number();
peNum = QMP_get_number_of_nodes();
if(!peRank)printf("QMP_init_msg_passing returned %d\n",init_status);
if (init_status != QMP_SUCCESS) {
QMP_error("%s\n",QMP_error_string(init_status));
}
// check QMP thread level
// Added by Hantao
if(peRank == 0) {
switch(prv) {
case QMP_THREAD_SINGLE:
printf("QMP thread level = QMP_THREAD_SINGLE\n");
break;
case QMP_THREAD_FUNNELED:
printf("QMP thread level = QMP_THREAD_FUNNELED\n");
break;
case QMP_THREAD_SERIALIZED:
printf("QMP thread level = QMP_THREAD_SERIALIZED\n");
break;
case QMP_THREAD_MULTIPLE:
printf("QMP thread level = QMP_THREAD_MULTIPLE\n");
break;
default:
printf("QMP thread level = no idea what this is, boom!\n");
}
}
//Check to make sure that this machine is a GRID machine
//Exit if not GRID machine
QMP_ictype qmp_type = QMP_get_msg_passing_type();
//Get information about the allocated machine
peNum = QMP_get_number_of_nodes();
NDIM = QMP_get_allocated_number_of_dimensions();
peGrid = QMP_get_allocated_dimensions();
pePos = QMP_get_allocated_coordinates();
if(peRank==0){
for(int i = 0; i<*argc;i++){
printf("argv[%d])(after)=%s\n",i,(*argv)[i]);
}
}
#else
QMP_status_t init_status = QMP_SUCCESS;
peRank=0;
peNum=1;
NDIM=4;
#endif
//#if (TARGET == BGL) || (TARGET == BGP)
if (NDIM>5){
peNum = 1;
for(int i = 0;i<5;i++)
peNum *= peGrid[i];
peRank = peRank % peNum;
}
int if_print=1;
for(int i = 0;i<NDIM;i++)
if (pePos[i]>=2) if_print=0;
if (if_print){
printf("Rank=%d Num=%d NDIM=%d\n",peRank,peNum,NDIM);
printf("dim:");
for(int i = 0;i<NDIM;i++)
printf(" %d",peGrid[i]);
printf("\n");
printf("pos:");
for(int i = 0;i<NDIM;i++)
printf(" %d",pePos[i]);
printf("\n");
#if 0
int rc;
BGLPersonality pers;
rts_get_personality(&pers, sizeof(pers));
printf("from personality: %d %d %d %d\n",pers.xCoord,pers.yCoord,pers.zCoord,rts_get_processor_id());
#endif
}
// printf("from personality:\n");
#if 0
if ( (qmp_type!= QMP_GRID) && (qmp_type !=QMP_MESH) ) {
QMP_error("CPS on QMP only implemented for GRID or MESH, not (%d) machines\n",qmp_type);
}
#endif
// printf("QMP_declare_logical_topology(peGrid, NDIM)\n");
#ifndef UNIFORM_SEED_NO_COMMS
//Declare the logical topology (Redundant for GRID machines)
if (QMP_declare_logical_topology(peGrid, NDIM) != QMP_SUCCESS) {
QMP_error("Node %d: Failed to declare logical topology\n",peRank);
exit(-4);
}
#endif
initialized = true;
printf("Rank=%d init_qmp() done\n",peRank);
}
int
main (int argc, char** argv)
{
int i, nc;
QMP_status_t status;
int **smem, **rmem;
QMP_msgmem_t *recvmem;
QMP_msghandle_t *recvh;
QMP_msgmem_t *sendmem;
QMP_msghandle_t *sendh;
struct perf_argv pargv;
QMP_thread_level_t req, prv;
/**
* Simple point to point topology
*/
int dims[4] = {2,2,2,2};
int ndims = 1;
//if(QMP_get_node_number()==0)
//printf("starting init\n"); fflush(stdout);
req = QMP_THREAD_SINGLE;
status = QMP_init_msg_passing (&argc, &argv, req, &prv);
if (status != QMP_SUCCESS) {
fprintf (stderr, "QMP_init failed\n");
return -1;
}
if(QMP_get_node_number()==0)
printf("finished init\n"); fflush(stdout);
if (parse_options (argc, argv, &pargv) == -1) {
if(QMP_get_node_number()==0)
usage (argv[0]);
exit (1);
}
{
int maxdims = 4;
int k=0;
int nodes = QMP_get_number_of_nodes();
ndims = 0;
while( (nodes&1) == 0 ) {
if(ndims<maxdims) ndims++;
else {
dims[k] *= 2;
k++;
if(k>=maxdims) k = 0;
}
nodes /= 2;
}
if(nodes != 1) {
QMP_error("invalid number of nodes %i", QMP_get_number_of_nodes());
QMP_error(" must power of 2");
QMP_abort(1);
}
pargv.ndims = ndims;
}
status = QMP_declare_logical_topology (dims, ndims);
if (status != QMP_SUCCESS) {
fprintf (stderr, "Cannot declare logical grid\n");
return -1;
}
/* do a broadcast of parameter */
if (QMP_broadcast (&pargv, sizeof (pargv)) != QMP_SUCCESS) {
QMP_printf ("Broadcast parameter failed\n");
exit (1);
}
{
int k=1;
const int *lc = QMP_get_logical_coordinates();
for(i=0; i<ndims; i++) k += lc[i];
pargv.sender = k&1;
}
QMP_printf("%s options: num_channels[%d] verify[%d] option[%d] datasize[%d] numloops[%d] sender[%d] strided_send[%i] strided_recv[%i] strided_array_send[%i] ",
argv[0], pargv.num_channels, pargv.verify,
pargv.option, pargv.size, pargv.loops, pargv.sender,
strided_send, strided_recv, strided_array_send);
fflush(stdout);
/**
* Create memory
*/
nc = pargv.num_channels;
smem = (int **)malloc(nc*sizeof (int *));
rmem = (int **)malloc(nc*sizeof (int *));
sendmem = (QMP_msgmem_t *)malloc(ndims*nc*sizeof (QMP_msgmem_t));
recvmem = (QMP_msgmem_t *)malloc(ndims*nc*sizeof (QMP_msgmem_t));
sendh = (QMP_msghandle_t *)malloc(nc*sizeof (QMP_msghandle_t));
recvh = (QMP_msghandle_t *)malloc(nc*sizeof (QMP_msghandle_t));
QMP_barrier();
if(QMP_get_node_number()==0) printf("\n"); fflush(stdout);
if(pargv.option & TEST_SIMUL) {
int opts = pargv.option;
pargv.option = TEST_SIMUL;
if(QMP_get_node_number()==0)
QMP_printf("starting simultaneous sends"); fflush(stdout);
for(i=pargv.minsize; i<=pargv.maxsize; i*=pargv.facsize) {
pargv.size = i;
create_msgs(smem, rmem, sendmem, recvmem, sendh, recvh, ndims, nc, i, &pargv);
test_simultaneous_send (smem, rmem, sendh, recvh, &pargv);
check_mem(rmem, ndims, nc, i);
free_msgs(smem, rmem, sendmem, recvmem, sendh, recvh, ndims, nc);
}
if(QMP_get_node_number()==0)
QMP_printf("finished simultaneous sends\n"); fflush(stdout);
pargv.option = opts;
}
if(pargv.option & TEST_PINGPONG) {
int opts = pargv.option;
pargv.option = TEST_PINGPONG;
if(QMP_get_node_number()==0)
QMP_printf("starting ping pong sends"); fflush(stdout);
for(i=pargv.minsize; i<=pargv.maxsize; i*=pargv.facsize) {
pargv.size = i;
create_msgs(smem, rmem, sendmem, recvmem, sendh, recvh, ndims, nc, i, &pargv);
if(pargv.verify)
test_pingpong_verify(smem, rmem, sendh, recvh, &pargv);
else
test_pingpong(smem, rmem, sendh, recvh, &pargv);
check_mem(rmem, ndims, nc, i);
free_msgs(smem, rmem, sendmem, recvmem, sendh, recvh, ndims, nc);
}
if(QMP_get_node_number()==0)
QMP_printf("finished ping pong sends\n"); fflush(stdout);
pargv.option = opts;
}
if(pargv.option & TEST_ONEWAY) {
int opts = pargv.option;
pargv.option = TEST_ONEWAY;
if(QMP_get_node_number()==0)
QMP_printf("starting one way sends"); fflush(stdout);
for(i=pargv.minsize; i<=pargv.maxsize; i*=pargv.facsize) {
pargv.size = i;
create_msgs(smem, rmem, sendmem, recvmem, sendh, recvh, ndims, nc, i, &pargv);
test_oneway (smem, rmem, sendh, recvh, &pargv);
if(!pargv.sender) check_mem(rmem, ndims, nc, i);
free_msgs(smem, rmem, sendmem, recvmem, sendh, recvh, ndims, nc);
}
if(QMP_get_node_number()==0)
QMP_printf("finished one way sends"); fflush(stdout);
pargv.option = opts;
}
/**
* Free memory
*/
free (smem);
free (rmem);
free (sendh);
free (recvh);
free (sendmem);
free (recvmem);
QMP_finalize_msg_passing ();
return 0;
}