// This is not needed as it can be done transitively:
// ie lookup the QDP index and then lookup the coord with that
static void mySiteCoords3D(int gcoords[], int node, int linearsite)
{
int mu;
int subgrid_cb_nrow[4];
int tmp_coord[4];
int cb3,cbb3;
int my_node = QMP_get_node_number();
for(mu=0; mu < 4; mu++) {
subgrid_cb_nrow[mu] = getLattSize()[mu];
}
subgrid_cb_nrow[0] /=2; /* Checkerboarding */
/* Single processor -- all coords 0 */
for(mu=0; mu < 4; mu++) {
gcoords[mu] = 0;
}
cb3=linearsite/total_vol_3d_cb;
crtesn3d(linearsite % total_vol_3d_cb, subgrid_cb_nrow, tmp_coord);
// Add on position within the node
// NOTE: the cb for the x-coord is not yet determined
gcoords[0] += 2*tmp_coord[0];
for(mu=1; mu < 4; ++mu) {
gcoords[mu] += tmp_coord[mu];
}
cbb3 = cb3;
for(mu=1; mu < 3; ++mu) {
cbb3 += gcoords[mu];
}
gcoords[0] += (cbb3 & 1);
}
int
main (int argc, char** argv)
{
QMP_status_t status;
int this_node;
QMP_thread_level_t req, prv;
/* Start QMP */
req = QMP_THREAD_SINGLE;
status = QMP_init_msg_passing (&argc, &argv, req, &prv);
if (status != QMP_SUCCESS) {
QMP_error ("QMP_init failed: %s\n", QMP_error_string(status));
QMP_abort(1);
}
/* Get my logical node number */
this_node = QMP_get_node_number();
/* Print the result */
printf("%04d",this_node);
/* Quit */
QMP_finalize_msg_passing ();
return 0;
}
static void
test_simultaneous_send (int** smem,
int** rmem,
QMP_msghandle_t* sendh,
QMP_msghandle_t* recvh,
struct perf_argv* pargv)
{
double it, ft, dt, bwval;
int i, j;
QMP_status_t err;
int nc, ndims;
int nloops;
int dsize;
nc = pargv->num_channels;
nloops = pargv->loops;
dsize = pargv->size;
ndims = pargv->ndims;
/* do a test for nloops */
QMP_barrier();
it = get_current_time ();
for (i = 0; i < nloops; i++) {
for (j = 0; j < nc; j++) {
/* receive operation */
if ((err = QMP_start (recvh[j])) != QMP_SUCCESS)
QMP_printf ("Start receiving failed: %s\n", QMP_error_string(err));
}
for (j = 0; j < nc; j++) {
/* Send operation */
if ((err = QMP_start (sendh[j])) != QMP_SUCCESS)
QMP_printf ("Start sending failed: %s\n", QMP_error_string(err));
}
for (j = 0; j < nc; j++) {
if (QMP_wait (sendh[j]) != QMP_SUCCESS)
QMP_printf ("Error in sending %d\n", j );
}
for (j = 0; j < nc; j++) {
if (QMP_wait (recvh[j]) != QMP_SUCCESS)
QMP_printf ("Error in receiving %d\n", j);
}
}
ft = get_current_time ();
if(QMP_get_node_number()==0) {
dt = (ft - it); /* in milli seconds */
bwval = dsize/(double)1000.0 * 4.0 * nloops * nc*ndims/dt;
QMP_printf ("Simultaneous send B/W for datasize %d is %g (MB/s)", dsize * 4, bwval);
QMP_printf ("Time difference is %lf micro seconds", dt*1000.0/nloops);
}
QMP_barrier();
}
int DML_clear_to_send(char *scratch_buf, size_t size,
int my_io_node, int new_node)
{
if (QMP_get_msg_passing_type() != QMP_GRID)
{
int this_node = QMP_get_node_number();
if(this_node == my_io_node && new_node != my_io_node)
DML_send_bytes(scratch_buf,size,new_node); /* junk message */
if(this_node == new_node && new_node != my_io_node)
DML_get_bytes(scratch_buf,size,my_io_node);
}
return 0;
}
int DML_route_bytes(char *buf, size_t size, int fromnode, int tonode)
{
if (QMP_get_msg_passing_type() == QMP_GRID)
{
DML_grid_route(buf, size, fromnode, tonode);
}
else
{
int this_node = QMP_get_node_number();
if (this_node == tonode)
DML_get_bytes(buf,size,fromnode);
if (this_node == fromnode)
DML_send_bytes(buf,size,tonode);
}
return 0;
}
unsigned int test_checksum(Float * float_p, int len) {
static int gcsum_count = 1;
static int peRank = 0;
if (gcsum_count==1){
peRank = QMP_get_node_number();
// fprintf(stderr,"peRank=%d\n",peRank);
}
unsigned int locsum = local_checksum(float_p, len);
unsigned int bosssum;
// if(UniqueID() == 0) bosssum = locsum;
if(CoorX()==0 &&
CoorY()==0 &&
CoorZ()==0 &&
CoorT()==0 ) bosssum = locsum;
else bosssum = 0;
glb_sum_internal2(&bosssum,4,0); // 0 for XOR
if(bosssum != locsum) {
fprintf( stderr,
"GCheckSum %d : Node %d : Oops I did it again: me (%d,%d,%d,%d,%d) %u != boss %u\n",
gcsum_count,
UniqueID(),
CoorX(),
CoorY(),
CoorZ(),
CoorT(),
CoorS(),
locsum, bosssum );
exit(-30);
}
#if 0
if(gcsum_count>2 && peRank==0){
fprintf(stderr,"peRank=%d exiting\n",peRank);
exit(-24);
}
#endif
gcsum_count++;
return bosssum;
}
// This is not needed as it can be done transitively:
// ie lookup the QDP index and then lookup the coord with that
static void mySiteCoords4D(int gcoords[], int node, int linearsite)
{
int mu;
int subgrid_cb_nrow[4];
int tmp_coord[4];
int cb,cbb;
int* log_coords=QMP_get_logical_coordinates_from(node);
int my_node = QMP_get_node_number();
for(mu=0; mu < 4; mu++) {
subgrid_cb_nrow[mu] = getSubgridSize()[mu];
}
subgrid_cb_nrow[0] /=2; /* Checkerboarding */
for(mu=0; mu < 4; mu++) {
gcoords[mu] = log_coords[mu]*getSubgridSize()[mu];
}
cb=linearsite/subgrid_vol_cb;
crtesn4d(linearsite % subgrid_vol_cb, subgrid_cb_nrow, tmp_coord);
// Add on position within the node
// NOTE: the cb for the x-coord is not yet determined
gcoords[0] += 2*tmp_coord[0];
for(mu=1; mu < 4; ++mu) {
gcoords[mu] += tmp_coord[mu];
}
cbb = cb;
for(mu=1; mu < 4; ++mu) {
cbb += gcoords[mu];
}
gcoords[0] += (cbb & 1);
}
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);
}
/**
* Test oneway blast send
*/
static void
test_oneway (int** smem,
int** rmem,
QMP_msghandle_t* sendh,
QMP_msghandle_t* recvh,
struct perf_argv* pargv)
{
double it, ft, dt, bwval;
int i, j;
QMP_status_t err;
int nc, ndims;
int nloops;
int dsize;
QMP_bool_t sender;
nc = pargv->num_channels;
nloops = pargv->loops;
dsize = pargv->size;
sender = pargv->sender;
ndims = pargv->ndims;
QMP_barrier();
if (sender) {
it = get_current_time ();
for (i = 0; i < nloops; i++) {
for (j = 0; j < nc; j++) {
/* Send operation */
if ((err = QMP_start (sendh[j])) != QMP_SUCCESS)
QMP_printf ("Start sending failed: %s\n", QMP_error_string(err));
}
for (j = 0; j < nc; j++) {
if (QMP_wait (sendh[j]) != QMP_SUCCESS)
QMP_printf ("Error in sending %d\n", j );
}
}
ft = get_current_time (); /* In milli seconds */
dt = (ft - it); /* actual send time milli seconds */
bwval = dsize/(double)1000.0 * 4 * nloops*nc*ndims/dt;
if(QMP_get_node_number()==0) {
QMP_printf ("Oneway Bandwidth for datasize %d is %g (MB/s)",
dsize * 4, bwval);
QMP_printf ("time is %lf micro seconds", dt*1000.0/nloops);
fflush(stdout);
}
QMP_barrier();
}
else {
it = get_current_time ();
for (i = 0; i < nloops; i++) {
for (j = 0; j < nc; j++) {
/* receive operation */
if ((err = QMP_start (recvh[j])) != QMP_SUCCESS)
QMP_printf ("Start receiving failed: %s\n", QMP_error_string(err));
}
for (j = 0; j < nc; j++) {
if (QMP_wait (recvh[j]) != QMP_SUCCESS)
QMP_printf ("Error in receiving %d\n", j);
}
}
ft = get_current_time (); /* In milli seconds */
dt = (ft - it); /* actual send time milli seconds */
bwval = dsize/(double)1000.0 * 4 * nloops*nc*ndims/dt;
QMP_barrier();
if(QMP_get_node_number()==1) {
QMP_printf ("Oneway Bandwidth for datasize %d is %g (MB/s)",
dsize * 4, bwval);
QMP_printf ("time is %lf micro seconds", dt*1000.0/nloops);
fflush(stdout);
}
}
QMP_barrier();
}
int
main(int argc, char *argv[])
{
const char *msg;
int status = 1;
int mu, i;
struct QOP_CLOVER_State *clover_state;
QDP_Int *I_seed;
int i_seed;
QDP_RandomState *state;
QLA_Real plaq;
QLA_Real n[NELEMS(F)];
struct QOP_CLOVER_Gauge *c_g;
struct QOP_CLOVER_Fermion *c_f[NELEMS(F)];
double kappa;
double c_sw;
double in_eps;
int in_iter;
int log_flag;
double out_eps;
int out_iter;
int cg_status;
double run_time;
long long flops, sent, received;
/* start QDP */
QDP_initialize(&argc, &argv);
if (argc != 1 + NDIM + 6) {
printf0("ERROR: usage: %s Lx ... seed kappa c_sw iter eps log?\n",
argv[0]);
goto end;
}
for (mu = 0; mu < NDIM; mu++) {
lattice[mu] = atoi(argv[1 + mu]);
}
i_seed = atoi(argv[1 + NDIM]);
kappa = atof(argv[2 + NDIM]);
c_sw = atof(argv[3 + NDIM]);
in_iter = atoi(argv[4 + NDIM]);
in_eps = atof(argv[5 + NDIM]);
log_flag = atoi(argv[6 + NDIM]) == 0? 0: QOP_CLOVER_LOG_EVERYTHING;
/* set lattice size and create layout */
QDP_set_latsize(NDIM, lattice);
QDP_create_layout();
primary = QMP_is_primary_node();
self = QMP_get_node_number();
get_vector(network, 1, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_dimensions());
get_vector(node, 0, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_coordinates());
printf0("network: ");
for (i = 0; i < NDIM; i++)
printf0(" %d", network[i]);
printf0("\n");
printf0("node: ");
for (i = 0; i < NDIM; i++)
printf0(" %d", node[i]);
printf0("\n");
printf0("kappa: %20.15f\n", kappa);
printf0("c_sw: %20.15f\n", c_sw);
printf0("in_iter: %d\n", in_iter);
printf0("in_eps: %15.2e\n", in_eps);
/* allocate the gauge field */
create_Mvector(U, NELEMS(U));
create_Mvector(C, NELEMS(C));
create_Dvector(F, NELEMS(F));
I_seed = QDP_create_I();
QDP_I_eq_funci(I_seed, icoord, QDP_all);
state = QDP_create_S();
QDP_S_eq_seed_i_I(state, i_seed, I_seed, QDP_all);
for (mu = 0; mu < NELEMS(U); mu++) {
QDP_M_eq_gaussian_S(U[mu], state, QDP_all);
}
for (i = 0; i < NELEMS(F); i++) {
QDP_D_eq_gaussian_S(F[i], state, QDP_all);
}
/* build the clovers */
clover(C, U);
/* initialize CLOVER */
if (QOP_CLOVER_init(&clover_state, lattice, network, node, primary,
sublattice, NULL)) {
printf0("CLOVER_init() failed\n");
goto end;
}
if (QOP_CLOVER_import_fermion(&c_f[0], clover_state, f_reader, F[0])) {
printf0("CLOVER_import_fermion(0) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[1], clover_state)) {
printf0("CLOVER_allocate_fermion(1) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[2], clover_state)) {
printf0("CLOVER_allocate_fermion(2) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[3], clover_state)) {
printf0("CLOVER_allocate_fermion(3) failed\n");
goto end;
}
if (QOP_CLOVER_import_gauge(&c_g, clover_state, kappa, c_sw,
u_reader, c_reader, NULL)) {
printf("CLOVER_import_gauge() failed\n");
goto end;
}
QOP_CLOVER_D_operator(c_f[2], c_g, c_f[0]);
cg_status = QOP_CLOVER_D_CG(c_f[3], &out_iter, &out_eps,
c_f[2], c_g, c_f[2], in_iter, in_eps,
log_flag);
msg = QOP_CLOVER_error(clover_state);
QOP_CLOVER_performance(&run_time, &flops, &sent, &received, clover_state);
QOP_CLOVER_export_fermion(f_writer, F[3], c_f[3]);
printf0("CG status: %d\n", cg_status);
printf0("CG error message: %s\n", msg? msg: "<NONE>");
printf0("CG iter: %d\n", out_iter);
printf0("CG eps: %20.10e\n", out_eps);
printf0("CG performance: runtime %e sec\n", run_time);
printf0("CG performance: flops %.3e MFlop/s (%lld)\n",
flops * 1e-6 / run_time, flops);
printf0("CG performance: snd %.3e MB/s (%lld)\n",
sent * 1e-6 / run_time, sent);
printf0("CG performance: rcv %.3e MB (%lld)/s\n",
received * 1e-6 / run_time, received);
/* free CLOVER */
QOP_CLOVER_free_gauge(&c_g);
for (i = 0; i < NELEMS(c_f); i++)
QOP_CLOVER_free_fermion(&c_f[i]);
QOP_CLOVER_fini(&clover_state);
/* Compute plaquette */
plaq = plaquette(U);
/* field norms */
for (i = 0; i < NELEMS(F); i++)
QDP_r_eq_norm2_D(&n[i], F[i], QDP_all);
/* Display the values */
printf0("plaquette = %g\n",
plaq / (QDP_volume() * QDP_Nc * NDIM * (NDIM - 1) / 2 ));
for (i = 0; i < NELEMS(F); i++)
printf0(" |f|^2 [%d] = %20.10e\n", i, (double)(n[i]));
/* Compute and display <f[1] f[0]> */
show_dot("1|orig", F[1], F[0]);
/* Compute and display <f[1] f[3]> */
show_dot("1|solv", F[1], F[3]);
QDP_destroy_S(state);
QDP_destroy_I(I_seed);
destroy_Mvector(U, NELEMS(U));
destroy_Mvector(C, NELEMS(C));
destroy_Dvector(F, NELEMS(F));
status = 0;
end:
/* shutdown QDP */
printf0("end\n");
QDP_finalize();
return status;
}
int
main(int argc, char *argv[])
{
struct QOP_MDWF_State *mdwf_state = NULL;
QMP_thread_level_t qt = QMP_THREAD_SINGLE;
int status = 1;
int vx[4];
int i;
if (QMP_init_msg_passing(&argc, &argv, qt, &qt) != QMP_SUCCESS) {
fprintf(stderr, "QMP_init() failed\n");
return 1;
}
for (i = 0; i < NELEM(b5); i++) {
b5[i] = 0.1 * i * (NELEM(b5) - i);
c5[i] = 0.1 * i * i * (NELEM(b5) - i);
}
self = QMP_get_node_number();
primary = QMP_is_primary_node();
for (i = 0; i < argc; i++)
zprint("arg[%d]=%s", i, argv[i]);
if (argc != 14) {
zprint("14 arguments expected, found %d", argc);
QMP_finalize_msg_passing();
return 1;
}
for (i = 0; i < 4; i++) {
mynetwork[i] = atoi(argv[i+1]);
mylocal[i] = atoi(argv[i+5]);
vx[i] = atoi(argv[i+10]);
mylattice[i] = mylocal[i] * mynetwork[i];
}
mylocal[4] = mylattice[4] = atoi(argv[9]);
zshowv4("network", mynetwork);
zshowv5("local lattice", mylocal);
zshowv5("lattice", mylattice);
getv(mynode, 0, 4, vx);
xshowv("node", mynode);
if (QOP_MDWF_init(&mdwf_state,
mylattice, mynetwork, mynode, primary,
getsub, NULL)) {
zprint("MDWF_init() failed");
goto end;
}
zprint("MDWF_init() done");
dump_state(mdwf_state);
QOP_MDWF_fini(&mdwf_state);
status = 0;
end:
QMP_finalize_msg_passing();
return status;
}
void PT::mat(int n, matrix **mout, matrix **min, const int *dir){
int wire[MAX_DIR];
int i;
QMP_msgmem_t msg_mem_p[2*MAX_DIR];
QMP_msghandle_t msg_handle_p[2*MAX_DIR];
QMP_msghandle_t multiple;
static double setup=0.,qmp=0.,localt=0.,nonlocal=0.;
static int call_num = 0;
call_num++;
// char *fname="pt_mat()";
// VRB.Func("",fname);
// if (call_num%100==1) printf("PT:mat()\n");
for(i=0;i<n;i++) wire[i] = dir[i];
#ifdef PROFILE
Float dtime2 = - dclock();
#endif
double dtime = -dclock();
int non_local_dir=0;
for(i=0;i<n;i++)
if (!local[wire[i]/2]) {
//Calculate the address for transfer in a particular direction
Float * addr = ((Float *)min[i]+GAUGE_LEN*offset[wire[i]]);
msg_mem_p[2*non_local_dir] = QMP_declare_msgmem((void *)rcv_buf[wire[i]], 3*non_local_chi[wire[i]]*VECT_LEN*sizeof(IFloat));
msg_mem_p[2*non_local_dir+1] = QMP_declare_strided_msgmem((void *)addr, (size_t)(3*blklen[wire[i]]), numblk[wire[i]], (ptrdiff_t)(3*stride[wire[i]]+3*blklen[wire[i]]));
msg_handle_p[2*non_local_dir] = QMP_declare_receive_relative(msg_mem_p[2*non_local_dir], wire[i]/2, 1-2*(wire[i]%2), 0);
msg_handle_p[2*non_local_dir+1] = QMP_declare_send_relative(msg_mem_p[2*non_local_dir+1], wire[i]/2, 2*(wire[i]%2)-1, 0);
non_local_dir++;
}
if (call_num==1 && !QMP_get_node_number())
printf("non_local_dir=%d\n",non_local_dir);
if(non_local_dir) {
multiple = QMP_declare_multiple(msg_handle_p, 2*non_local_dir);
QMP_start(multiple);
}
dtime += dclock();
setup +=dtime;
dtime = -dclock();
int if_print = 0;
// if ( (call_num%10000==1) && (!QMP_get_node_number()) ) if_print=1;
#define USE_TEST2
#ifdef USE_TEST2
//assume nt > n!
static char *cname="mat()";
#pragma omp parallel default(shared)
{
int iam,nt,ipoints,istart,offset;
iam = omp_get_thread_num();
nt = omp_get_num_threads();
int nt_dir = nt/n;
int n_t = iam/nt_dir;
int i_t = iam%nt_dir;
if (n_t >= n ){ n_t = n-1;
i_t = iam - (n-1)*nt_dir;
nt_dir = nt -(n-1)*nt_dir;
}
int w_t = wire[n_t];
ipoints = (local_chi[w_t]/2)/nt_dir;
offset = ipoints*i_t;
if (i_t == (nt_dir-1)) ipoints = (local_chi[w_t]/2)-offset;
if ( if_print )
printf("thread %d of %d nt_dir n_t i_t ipoints offset= %d %d %d %d %d\n",iam,nt,nt_dir,n_t,i_t,ipoints,offset);
//Interleaving of local computation of matrix multiplication
partrans_cmm_agg((uc_l[w_t]+offset*2),min[n_t],mout[n_t],ipoints);
if ( if_print )
printf("thread %d of %d done\n",iam,nt);
}
#else
{
//Interleaving of local computation of matrix multiplication
#pragma omp parallel for default(shared)
for(i=0;i<n;i++){
partrans_cmm_agg(uc_l[wire[i]],min[i],mout[i],local_chi[wire[i]]/2);
}
}
#endif
dtime += dclock();
localt +=dtime;
dtime = -dclock();
//#pragma omp barrier
//#pragma omp master
{
if(non_local_dir) {
QMP_status_t qmp_complete_status = QMP_wait(multiple);
if (qmp_complete_status != QMP_SUCCESS)
QMP_error("Send failed in vec_cb_norm: %s\n", QMP_error_string(qmp_complete_status));
QMP_free_msghandle(multiple);
for(int i = 0; i < 2*non_local_dir; i++)
QMP_free_msgmem(msg_mem_p[i]);
// Free(msg_handle_p);
// Free(msg_mem_p);
}
} //#pragma omp master {
dtime += dclock();
qmp +=dtime;
dtime = -dclock();
//Do non-local computations
#ifdef USE_TEST2
//assume nt > n!
#pragma omp parallel default(shared)
{
int iam,nt,ipoints,istart,offset;
iam = omp_get_thread_num();
nt = omp_get_num_threads();
int nt_dir = nt/n;
int n_t = iam/nt_dir;
int i_t = iam%nt_dir;
if (n_t >= n ){ n_t = n-1;
i_t = iam - (n-1)*nt_dir;
nt_dir = nt -(n-1)*nt_dir;
}
int w_t = wire[n_t];
ipoints = (non_local_chi[w_t]/2)/nt_dir;
offset = ipoints*i_t;
if (i_t == (nt_dir-1)) ipoints = (non_local_chi[w_t]/2)-offset;
if ( if_print )
printf("thread %d of %d nt_dir n_t i_t ipoints offset= %d %d %d %d %d\n",iam,nt,nt_dir,n_t,i_t,ipoints,offset);
//Non-local computation
if (ipoints>0)
partrans_cmm_agg((uc_nl[w_t]+offset*2),(matrix *)rcv_buf[w_t],mout[n_t],ipoints);
if ( if_print )
printf("thread %d of %d done\n",iam,nt);
}
#else
{
#pragma omp parallel for
for(i=0;i<n;i++)
if (!local[wire[i]/2]) {
#ifdef USE_OMP
if (call_num%10000==1 && !QMP_get_node_number() )
printf("thread %d of %d i=%d\n",omp_get_thread_num(),omp_get_num_threads(),i);
#endif
partrans_cmm_agg(uc_nl[wire[i]],(matrix *)rcv_buf[wire[i]],mout[i],non_local_chi[wire[i]]/2);
}
}//#pragma omp parallel
#endif
dtime += dclock();
nonlocal +=dtime;
if (call_num%100==0){
static char *cname="mat()";
if (!QMP_get_node_number() ) {
print_flops("mat():local*100",0,localt);
print_flops("mat():nonlocal*100",0,nonlocal);
print_flops("mat():qmp*100",0,qmp);
print_flops("mat():setup*100",0,setup);
}
localt=nonlocal=qmp=setup=0.;
}
#ifdef PROFILE
dtime2 +=dclock();
print_flops("",fname,198*vol*n,dtime2);
#endif
// ParTrans::PTflops +=198*n*vol;
}
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;
}
void make_shift_tables(int bound[2][4][4], halfspinor_array* chi1,
halfspinor_array* chi2,
halfspinor_array* recv_bufs[2][4],
halfspinor_array* send_bufs[2][4],
void (*QDP_getSiteCoords)(int coord[], int node, int linearsite),
int (*QDP_getLinearSiteIndex)(const int coord[]),
int (*QDP_getNodeNumber)(const int coord[]))
{
volatile int dir,i;
const int my_node = QMP_get_node_number();
int coord[4];
int gcoord[4];
int gcoord2[4];
int linear;
int **shift_table;
int x,y,z,t;
int *subgrid_size = getSubgridSize();
int mu;
int offset;
int cb;
const int *node_coord = QMP_get_logical_coordinates();
int p;
int site, index;
InvTab4 *xinvtab;
InvTab4 *invtab;
int qdp_index;
int my_index;
int num;
int offsite_found;
/* Setup the subgrid volume for ever after */
subgrid_vol = 1;
for(i=0; i < getNumDim(); ++i) {
subgrid_vol *= getSubgridSize()[i];
}
/* Get the checkerboard size for ever after */
subgrid_vol_cb = subgrid_vol / 2;
/* Now I want to build the site table */
/* I want it cache line aligned? */
xsite_table = (int *)malloc(sizeof(int)*subgrid_vol+63L);
if(xsite_table == 0x0 ) {
QMP_error("Couldnt allocate site table");
QMP_abort(1);
}
site_table = (int *)((((ptrdiff_t)(xsite_table))+63L)&(-64L));
xinvtab = (InvTab4 *)malloc(sizeof(InvTab4)*subgrid_vol + 63L);
if(xinvtab == 0x0 ) {
QMP_error("Couldnt allocate site table");
QMP_abort(1);
}
invtab = (InvTab4 *)((((ptrdiff_t)(xinvtab))+63L)&(-64L));
/* Inversity of functions check:
Check that myLinearSiteIndex3D is in fact the inverse
of mySiteCoords3D, and that QDP_getSiteCoords is the
inverse of QDP_linearSiteIndex()
*/
for(p=0; p < 2; p++) {
for(site=0; site < subgrid_vol_cb; site++) {
/* Linear site index */
my_index = site + subgrid_vol_cb*p;
QDP_getSiteCoords(gcoord, my_node, my_index);
linear=QDP_getLinearSiteIndex(gcoord);
if( linear != my_index ) {
printf("P%d cb=%d site=%d : QDP_getSiteCoords not inverse of QDP_getLinearSiteIndex(): my_index=%d linear=%d\n", my_node, p,site, my_index,linear);
}
mySiteCoords4D(gcoord, my_node, my_index);
linear=myLinearSiteIndex4D(gcoord);
if( linear != my_index ) {
printf("P%d cb=%d site=%d : mySiteCoords3D not inverse of myLinearSiteIndex3D(): my_index=%d linear=%d\n", my_node, p,site, my_index,linear);
}
}
}
/* Loop through sites - you can choose your path below */
/* This is a checkerboarded order which is identical hopefully
to QDP++'s rb2 subset when QDP++ is in a CB2 layout */
for(p=0; p < 2; p++) {
for(t=0; t < subgrid_size[3]; t++) {
for(z=0; z < subgrid_size[2]; z++) {
for(y=0; y < subgrid_size[1]; y++) {
for(x=0; x < subgrid_size[0]/2; x++) {
coord[0] = 2*x + p;
coord[1] = y;
coord[2] = z;
coord[3] = t;
/* Make global */
for(i=0; i < 4; i++) {
coord[i] += subgrid_size[i]*node_coord[i];
}
/* Index of coordinate -- NB this is not lexicographic
but takes into account checkerboarding in QDP++ */
qdp_index = QDP_getLinearSiteIndex(coord);
/* Index of coordinate in my layout. -- NB this is not lexicographic
but takes into account my 3D checkerbaording */
my_index = myLinearSiteIndex4D(coord);
site_table[my_index] = qdp_index;
cb=parity(coord);
linear = my_index%subgrid_vol_cb;
invtab[qdp_index].cb=cb;
invtab[qdp_index].linearcb=linear;
}
}
}
}
}
/* Site table transitivity check:
for each site, convert to index in cb3d, convert to qdp index
convert qdp_index to coordinate
convert coordinate to back index in cb3d
Check that your cb3d at the end is the same as you
started with */
for(p=0; p < 2; p++) {
for(site=0; site < subgrid_vol_cb; site++) {
/* My local index */
my_index = site + subgrid_vol_cb*p;
/* Convert to QDP index */
qdp_index = site_table[ my_index ];
/* Switch QDP index to coordinates */
QDP_getSiteCoords(gcoord, my_node,qdp_index);
/* Convert back to cb3d index */
linear = myLinearSiteIndex4D(gcoord);
/* Check new cb,cbsite index matches the old cb index */
if (linear != my_index) {
printf("P%d The Circle is broken. My index=%d qdp_index=%d coords=%d,%d,%d,%d linear(=my_index?)=%d\n", my_node, my_index, qdp_index, gcoord[0],gcoord[1],gcoord[2],gcoord[3],linear);
}
}
}
/* Consistency check 2: Test mySiteCoords 3D
for all 3d cb,cb3index convert to
cb3d linear index (my_index)
convert to qdp_index (lookup in site table)
Now convert qdp_index and my_index both to
coordinates. They should produce the same coordinates
*/
for(p=0; p < 2; p++) {
for(site=0; site < subgrid_vol_cb; site++) {
/* My local index */
my_index = site + subgrid_vol_cb*p;
mySiteCoords4D(gcoord, my_node, my_index);
qdp_index = site_table[ my_index ];
QDP_getSiteCoords(gcoord2, my_node,qdp_index);
for(mu=0 ; mu < 4; mu++) {
if( gcoord2[mu] != gcoord[mu] ) {
printf("P%d: my_index=%d qdp_index=%d mySiteCoords=(%d,%d,%d,%d) QDPsiteCoords=(%d,%d,%d,%d)\n", my_node, my_index, qdp_index, gcoord[0], gcoord[1], gcoord[2], gcoord[3], gcoord2[0], gcoord2[1], gcoord2[2], gcoord2[3]);
continue;
}
}
}
}
/* Allocate the shift table */
/* The structure is as follows: There are 4 shift tables in order:
[ Table 1 | Table 2 | Table 3 | Table 4 ]
Table 1: decomp_scatter_index[mu][site]
Table 2: decomp_hvv_scatter_index[mu][site]
Table 3: recons_mvv_gather_index[mu][site]
Table 4: recons_gather_index[mu][site]
*/
/* This 4 is for the 4 tables: Table 1-4*/
if ((shift_table = (int **)malloc(4*sizeof(int*))) == 0 ) {
QMP_error("init_wnxtsu3dslash: could not initialize shift_table");
QMP_abort(1);
}
for(i=0; i < 4; i++) {
/* This 4 is for the 4 comms dierctions: */
if ((shift_table[i] = (int *)malloc(4*subgrid_vol*sizeof(int))) == 0) {
QMP_error("init_wnxtsu3dslash: could not initialize shift_table");
QMP_abort(1);
}
}
/* Initialize the boundary counters */
for(cb=0; cb < 2; cb++) {
for(dir=0; dir < 4; dir++) {
bound[cb][0][dir] = 0;
bound[cb][1][dir] = 0;
bound[cb][2][dir] = 0;
bound[cb][3][dir] = 0;
}
}
for(cb=0; cb < 2; cb++) {
for(site=0; site < subgrid_vol_cb; ++site) {
index = cb*subgrid_vol_cb + site;
/* Fetch site from site table */
qdp_index = site_table[index];
/* Get its coords */
QDP_getSiteCoords(coord, my_node, qdp_index);
/* Loop over directions building up shift tables */
for(dir=0; dir < 4; dir++) {
int fcoord[4], bcoord[4];
int fnode, bnode;
int blinear, flinear;
/* Backwards displacement*/
offs(bcoord, coord, dir, -1);
bnode = QDP_getNodeNumber(bcoord);
blinear = QDP_getLinearSiteIndex(bcoord);
/* Forward displacement */
offs(fcoord, coord, dir, +1);
fnode = QDP_getNodeNumber(fcoord);
flinear = QDP_getLinearSiteIndex(fcoord);
/* Scatter: decomp_{plus,minus} */
/* Operation: a^F(shift(x,type=0),dir) <- decomp(psi(x),dir) */
/* Send backwards - also called a receive from forward */
if (bnode != my_node) {
/* Offnode */
/* Append to Tail 1, increase boundary count */
/* This is the correct code */
shift_table[DECOMP_SCATTER][dir+4*index]
= subgrid_vol_cb + bound[1-cb][DECOMP_SCATTER][dir];
bound[1-cb][DECOMP_SCATTER][dir]++;
}
else {
/* On node. Note the linear part of its (cb3, linear) bit,
using a reverse lookup */
shift_table[DECOMP_SCATTER][dir+4*index] =
invtab[blinear].linearcb;
}
/* Scatter: decomp_hvv_{plus,minus} */
/* Operation: a^B(shift(x,type=1),dir) <- U^dag(x,dir)*decomp(psi(x),dir) */
/* Send forwards - also called a receive from backward */
if (fnode != my_node) {
/* Offnode */
/* Append to Tail 1, increase boundary count */
shift_table[DECOMP_HVV_SCATTER][dir+4*index]
= subgrid_vol_cb + bound[1-cb][DECOMP_HVV_SCATTER][dir];
bound[1-cb][DECOMP_HVV_SCATTER][dir]++;
}
else {
/* On node. Note the linear part of its (cb3, linear) bit,
using a reverse lookup */
shift_table[DECOMP_HVV_SCATTER][dir+4*index] /* Onnode */
= invtab[flinear].linearcb ;
}
/* Gather: mvv_recons_{plus,minus} */
/* Operation: chi(x) <- \sum_dir U(x,dir)*a^F(shift(x,type=2),dir) */
/* Receive from forward */
if (fnode != my_node) {
/* Offnode */
/* Append to Tail 2, increase boundary count */
shift_table[RECONS_MVV_GATHER][dir+4*index] =
2*subgrid_vol_cb + (bound[cb][RECONS_MVV_GATHER][dir]);
bound[cb][RECONS_MVV_GATHER][dir]++;
}
else {
/* On node. Note the linear part of its (cb3, linear) bit,
using a reverse lookup. Note this is a recons post shift,
so the linear coordinate to invert is mine rather than the neighbours */
shift_table[RECONS_MVV_GATHER][dir+4*index] =
invtab[qdp_index].linearcb ;
}
/* Gather: recons_{plus,minus} */
/* Operation: chi(x) += \sum_dir recons(a^B(shift(x,type=3),dir),dir) */
/* Receive from backward */
if (bnode != my_node) {
shift_table[RECONS_GATHER][dir+4*index] =
2*subgrid_vol_cb + bound[cb][RECONS_GATHER][dir];
bound[cb][RECONS_GATHER][dir]++;
}
else {
/* On node. Note the linear part of its (cb3, linear) bit,
using a reverse lookup. Note this is a recons post shift,
so the linear coordinate to invert is mine rather than the neighbours */
shift_table[RECONS_GATHER][dir+4*index] =
invtab[qdp_index].linearcb ;
}
}
}
}
/* Sanity check - make sure the sending and receiving counters match */
for(cb=0; cb < 2; cb++) {
for(dir=0; dir < 4; dir++) {
/* Sanity 1: Must have same number of boundary sites on each cb for
a given operation */
for(i = 0; i < 4; i++) {
if (bound[1-cb][i][dir] != bound[cb][i][dir]) {
QMP_error("SSE Wilson dslash - make_shift_tables: type 0 diff. cb send/recv counts do not match: %d %d",
bound[1-cb][i][dir],bound[cb][i][dir]);
QMP_abort(1);
}
}
}
}
/* Now I want to make the offset table into the half spinor temporaries */
/* The half spinor temporaries will look like this:
dir=0 [ Body Half Spinors ][ Tail 1 Half Spinors ][ Tail 2 Half Spinors ]
dir=1 [ Body Half Spinors ][ Tail 1 Half Spinors ][ Tail 2 Half Spinors ]
...
And each of these blocks of half spinors will be sized to vol_cb
sites (ie half volume only). The shift_table() for a given site and
direction indexes into one of these lines. So the offset table essentially
delineates which line one picks, by adding an offset of
3*subgrid_vol_cb*dir
To the shift. The result from offset table, can be used directly as a
pointer displacement on the temporaries.
Perhaps the best way to condsider this is to consider a value
of shift_table[type][dir/site] that lands in the body. The
shift table merely gives me a site index. But the data needs
to be different for each direction for that site index. Hence
we need to replicate the body, for each dir. The 3xsubgrid_vol_cb
is just there to take care of the buffers.
Or another way to think of it is that there is a 'body element' index
specified by the shift table lookup, and that dir is just the slowest
varying index.
*/
/* 4 dims, 4 types, rest of the magic is to align the thingie */
xoffset_table = (halfspinor_array **)malloc(4*4*subgrid_vol*sizeof(halfspinor_array*)+63L);
if( xoffset_table == 0 ) {
QMP_error("init_wnxtsu3dslash: could not initialize offset_table[i]");
QMP_abort(1);
}
/* This is the bit what aligns straight from AMD Manual */
offset_table = (halfspinor_array**)((((ptrdiff_t)(xoffset_table)) + 63L) & (-64L));
/* Walk through the shift_table and remap the offsets into actual
pointers */
/* DECOMP_SCATTER */
num=0;
for(dir =0; dir < Nd; dir++) {
/* Loop through all the sites. Remap the offsets either to local
arrays or pointers */
offsite_found=0;
for(site=0; site < subgrid_vol; site++) {
offset = shift_table[DECOMP_SCATTER][dir+4*site];
if( offset >= subgrid_vol_cb ) {
/* Found an offsite guy. It's address must be to the send back buffer */
/* send to back index = recv from forward index = 0 */
offsite_found++;
offset_table[ dir + 4*(site + subgrid_vol*DECOMP_SCATTER) ] =
send_bufs[0][num]+(offset - subgrid_vol_cb);
}
else {
/* Guy is onsite: This is DECOMP_SCATTER so offset to chi1 */
offset_table[ dir + 4*(site + subgrid_vol*DECOMP_SCATTER) ] =
chi1+shift_table[DECOMP_SCATTER][dir+4*site]+subgrid_vol_cb*dir;
}
}
if( offsite_found > 0 ) {
/* If we found an offsite guy, next direction has to
go into the next dir part of the send bufs */
num++;
}
}
/* DECOMP_HVV_SCATTER */
/* Restart num-s */
num=0;
for(dir =0; dir <Nd; dir++) {
offsite_found=0;
for(site=0; site < subgrid_vol; site++) {
offset = shift_table[DECOMP_HVV_SCATTER][dir+4*site];
if( offset >= subgrid_vol_cb ) {
/* Found an offsite guy. It's address must be to the send forw buffer */
/* send to forward / receive from backward index = 1 */
offsite_found++;
offset_table[ dir + 4*(site + subgrid_vol*DECOMP_HVV_SCATTER) ] =
send_bufs[1][num]+(offset - subgrid_vol_cb);
}
else {
/* Guy is onsite. This is DECOMP_HVV_SCATTER so offset to chi2 */
offset_table[ dir + 4*(site + subgrid_vol*DECOMP_HVV_SCATTER) ] =
chi2+shift_table[DECOMP_HVV_SCATTER][dir+4*site ]+subgrid_vol_cb*dir;
}
}
if( offsite_found > 0 ) {
num++;
}
}
/* RECONS_MVV_GATHER */
num=0;
for(dir =0; dir <Nd; dir++) {
offsite_found=0;
for(site=0; site < subgrid_vol; site++) {
offset = shift_table[RECONS_MVV_GATHER][dir+4*site];
if( offset >= 2*subgrid_vol_cb ) {
/* Found an offsite guy. It's address must be to the recv from front buffer */
/* recv_from front index = send to back index = 0 */
offsite_found++;
offset_table[ dir + 4*(site + subgrid_vol*RECONS_MVV_GATHER) ] =
recv_bufs[0][num]+(offset - 2*subgrid_vol_cb);
}
else {
/* Guy is onsite */
/* This is RECONS_MVV_GATHER so offset with respect to chi1 */
offset_table[ dir + 4*(site + subgrid_vol*RECONS_MVV_GATHER) ] =
chi1+shift_table[RECONS_MVV_GATHER][dir+4*site ]+subgrid_vol_cb*dir;
}
}
if( offsite_found > 0 ) {
num++;
}
}
/* RECONS_GATHER */
num=0;
for(dir =0; dir <Nd; dir++) {
offsite_found=0;
for(site=0; site < subgrid_vol; site++) {
offset = shift_table[RECONS_GATHER][dir+4*site];
if( offset >= 2*subgrid_vol_cb ) {
/* Found an offsite guy. It's address must be to the recv from back buffer */
/* receive from back = send to forward index = 1*/
offsite_found++;
offset_table[ dir + 4*(site + subgrid_vol*RECONS_GATHER) ] =
recv_bufs[1][num]+(offset - 2*subgrid_vol_cb);
}
else {
/* Guy is onsite */
/* This is RECONS_GATHER so offset with respect to chi2 */
offset_table[ dir + 4*(site + subgrid_vol*RECONS_GATHER ) ] =
chi2+shift_table[RECONS_GATHER][dir+4*site ]+subgrid_vol_cb*dir;
}
}
if( offsite_found > 0 ) {
num++;
}
}
/* Free shift table - it is no longer needed. We deal solely with offsets */
for(i=0; i < 4; i++) {
free( (shift_table)[i] );
}
free( shift_table );
free( xinvtab );
}
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
main(int argc, char *argv[])
{
struct QOP_MDWF_State *mdwf_state = NULL;
struct QOP_MDWF_Parameters *mdwf_params = NULL;
QMP_thread_level_t qt = QMP_THREAD_SINGLE;
int status = 1;
int i;
if (QMP_init_msg_passing(&argc, &argv, qt, &qt) != QMP_SUCCESS) {
fprintf(stderr, "QMP_init() failed\n");
return 1;
}
for (i = 0; i < NELEM(b5); i++) {
b5[i] = 0.1 * i * (NELEM(b5) - i);
c5[i] = 0.1 * i * i * (NELEM(b5) - i);
}
self = QMP_get_node_number();
primary = QMP_is_primary_node();
if (argc != 7) {
zprint("7 arguments expected, found %d", argc);
zprint("usage: localheat Lx Ly Lz Lt Ls time");
QMP_finalize_msg_passing();
return 1;
}
for (i = 0; i < 4; i++) {
mynetwork[i] = 1;
mylocal[i] = atoi(argv[i+1]);
mylattice[i] = mylocal[i] * mynetwork[i];
}
mylocal[4] = mylattice[4] = atoi(argv[5]);
total_sec = atoi(argv[6]);
zshowv4("network", mynetwork);
zshowv5("local lattice", mylocal);
zshowv5("lattice", mylattice);
zprint("total requested runtime %.0f sec", total_sec);
#if 0
if (QMP_declare_logical_topology(mynetwork, 4) != QMP_SUCCESS) {
zprint("declare_logical_top failed");
goto end;
}
getv(mynode, 0, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_coordinates());
#else
{ int i;
for (i = 0; i < 4; i++)
mynode[i] = 0;
}
#endif
if (QOP_MDWF_init(&mdwf_state,
mylattice, mynetwork, mynode, primary,
getsub, NULL)) {
zprint("MDWF_init() failed");
goto end;
}
zprint("MDWF_init() done");
if (QOP_MDWF_set_generic(&mdwf_params, mdwf_state, b5, c5, 0.123, 0.05)) {
zprint("MDW_set_generic() failed");
goto end;
}
zprint("MDWF_set_generic() done");
if (do_run(mdwf_state, mdwf_params)) {
zprint("float test failed");
goto end;
}
QOP_MDWF_fini(&mdwf_state);
zprint("Heater test finished");
status = 0;
end:
QMP_finalize_msg_passing();
return status;
}
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 comm_rank(void)
{
return QMP_get_node_number();
}
int
main(int argc, char *argv[])
{
int status = 1;
int mu, i;
struct QOP_CLOVER_State *clover_state;
QDP_Int *I_seed;
int i_seed;
QDP_RandomState *state;
QLA_Real plaq;
QLA_Real n[NELEMS(F)];
struct QOP_CLOVER_Gauge *c_g;
struct QOP_CLOVER_Fermion *c_f[NELEMS(F)];
double kappa;
double c_sw;
/* start QDP */
QDP_initialize(&argc, &argv);
if (argc != 1 + NDIM + 3) {
printf0("ERROR: usage: %s Lx ... seed kappa c_sw\n", argv[0]);
goto end;
}
for (mu = 0; mu < NDIM; mu++) {
lattice[mu] = atoi(argv[1 + mu]);
}
i_seed = atoi(argv[1 + NDIM]);
kappa = atof(argv[2 + NDIM]);
c_sw = atof(argv[3 + NDIM]);
/* set lattice size and create layout */
QDP_set_latsize(NDIM, lattice);
QDP_create_layout();
primary = QMP_is_primary_node();
self = QMP_get_node_number();
get_vector(network, 1, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_dimensions());
get_vector(node, 0, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_coordinates());
printf0("network: ");
for (i = 0; i < NDIM; i++)
printf0(" %d", network[i]);
printf0("\n");
printf0("node: ");
for (i = 0; i < NDIM; i++)
printf0(" %d", node[i]);
printf0("\n");
printf0("kappa: %20.15f\n", kappa);
printf0("c_sw: %20.15f\n", c_sw);
/* allocate the gauge field */
create_Mvector(U, NELEMS(U));
create_Mvector(C, NELEMS(C));
create_Dvector(F, NELEMS(F));
I_seed = QDP_create_I();
QDP_I_eq_funci(I_seed, icoord, QDP_all);
state = QDP_create_S();
QDP_S_eq_seed_i_I(state, i_seed, I_seed, QDP_all);
for (mu = 0; mu < NELEMS(U); mu++) {
QDP_M_eq_gaussian_S(U[mu], state, QDP_all);
}
for (i = 0; i < NELEMS(F); i++) {
QDP_D_eq_gaussian_S(F[i], state, QDP_all);
}
/* build the clovers */
clover(C, U);
/* initialize CLOVER */
if (QOP_CLOVER_init(&clover_state, lattice, network, node, primary,
sublattice, NULL)) {
printf0("CLOVER_init() failed\n");
goto end;
}
if (QOP_CLOVER_import_fermion(&c_f[0], clover_state, f_reader, F[0])) {
printf0("CLOVER_import_fermion(0) failed\n");
goto end;
}
if (QOP_CLOVER_import_fermion(&c_f[1], clover_state, f_reader, F[1])) {
printf0("CLOVER_import_fermion(1) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[2], clover_state)) {
printf0("CLOVER_allocate_fermion(2) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[3], clover_state)) {
printf0("CLOVER_allocate_fermion(3) failed\n");
goto end;
}
if (QOP_CLOVER_import_gauge(&c_g, clover_state, kappa, c_sw,
u_reader, c_reader, NULL)) {
printf("CLOVER_import_gauge() failed\n");
goto end;
}
QOP_CLOVER_D_operator(c_f[2], c_g, c_f[0]);
QOP_CLOVER_export_fermion(f_writer, F[2], c_f[2]);
QOP_CLOVER_D_operator_conjugated(c_f[3], c_g, c_f[1]);
QOP_CLOVER_export_fermion(f_writer, F[3], c_f[3]);
/* free CLOVER */
QOP_CLOVER_free_gauge(&c_g);
for (i = 0; i < NELEMS(c_f); i++)
QOP_CLOVER_free_fermion(&c_f[i]);
QOP_CLOVER_fini(&clover_state);
/* Compute plaquette */
plaq = plaquette(U);
/* field norms */
for (i = 0; i < NELEMS(F); i++)
QDP_r_eq_norm2_D(&n[i], F[i], QDP_all);
/* Display the values */
printf0("plaquette = %g\n",
plaq / (QDP_volume() * QDP_Nc * NDIM * (NDIM - 1) / 2 ));
for (i = 0; i < NELEMS(F); i++)
printf0(" |f|^2 [%d] = %20.10e\n", i, (double)(n[i]));
/* Compute and display <f[1] f[2]> */
show_dot("1|D0", F[1], F[2]);
/* Compute and display <f[3] f[0]> */
show_dot("X1|0", F[3], F[0]);
QDP_destroy_S(state);
QDP_destroy_I(I_seed);
destroy_Mvector(U, NELEMS(U));
destroy_Mvector(C, NELEMS(C));
destroy_Dvector(F, NELEMS(F));
status = 0;
end:
/* shutdown QDP */
printf0("end\n");
QDP_finalize();
return status;
}