void
setup_layout(void)
{
int c[4];
int i,n_mach;
int d[4];
#ifdef FIX_NODE_GEOM
int *geom = node_geometry;
#else
int *geom = NULL;
#endif
if(mynode()==0){
printf("LAYOUT = Hypercubes, options = ");
printf("QDP");
printf("\n");
}
/* Is there already a grid?
This could be a grid architecture with a preset dimension, or
a geometry could have been set by the -qmp-geom command line arg.
In either case we have a nonzero allocated number of dimensions.
*/
if(QMP_get_allocated_number_of_dimensions() == 0)
/* Set the geometry if requested */
set_qmp_layout_grid(geom, 4);
c[0] = nx;
c[1] = ny;
c[2] = nz;
c[3] = nt;
QDP_set_latsize(4, c);
QDP_create_layout();
sites_on_node = QDP_sites_on_node;
even_sites_on_node = QDP_subset_len(QDP_even);
odd_sites_on_node = QDP_subset_len(QDP_odd);
n_mach = QMP_get_logical_number_of_dimensions();
dim_mach = QMP_get_logical_dimensions();
/* Initialize I/O node function */
#ifdef FIX_IONODE_GEOM
init_io_node();
#endif
/* Report sublattice dimensions */
for(i = 0; i < 4; i++){
/* Any extra machine dimensions are assumed to be 1 */
if(i < n_mach)d[i] = c[i]/dim_mach[i];
else d[i] = c[i];
}
if( mynode()==0)
printf("ON EACH NODE %d x %d x %d x %d\n",d[0],d[1],d[2],d[3]);
#if 0
mpi_whoami(); /* Debug */
#endif
}
/*--------------------------------------------------------------------*/
static void setup_qmp_grid(){
int ndim = 4;
int len[4];
int ndim2, i;
const int *nsquares2;
len[0] = nx; len[1] = ny; len[2] = nz; len[3] = nt;
if(mynode()==0){
printf("qmp_grid,");
printf("\n");
}
ndim2 = QMP_get_allocated_number_of_dimensions();
nsquares2 = QMP_get_allocated_dimensions();
/* If the dimensions are not already allocated, use the
node_geometry request. Otherwise a hardware or command line
specification trumps the parameter input. */
#ifdef FIX_NODE_GEOM
if(ndim2 == 0){
ndim2 = 4;
nsquares2 = node_geometry;
}
else{
node0_printf("setup_qmp_grid: Preallocated machine geometry overrides request\n");
}
#endif
if(mynode()==0){
printf("Using machine geometry: ");
for(i=0; i<ndim; i++){
printf("%d ",nsquares2[i]);
if(i < ndim-1)printf("X ");
}
printf("\n");
}
/* In principle, we could now rotate coordinate axes */
/* Save this for a future upgrade */
set_qmp_layout_grid(nsquares2, ndim2);
ndim2 = QMP_get_logical_number_of_dimensions();
nsquares2 = QMP_get_logical_dimensions();
for(i=0; i<ndim; i++) {
if(i<ndim2) nsquares[i] = nsquares2[i];
else nsquares[i] = 1;
}
for(i=0; i<ndim; i++) {
if(len[i]%nsquares[i] != 0) {
node0_printf("LATTICE SIZE DOESN'T FIT GRID\n");
QMP_abort(0);
}
squaresize[i] = len[i]/nsquares[i];
}
}
/* Total problem size */
static int* getLattSize()
{
static int first = 1;
static int tot_size[4];
if (first == 1) {
const int* phys_size = QMP_get_logical_dimensions();
int i;
for(i=0; i < getNumDim(); ++i) {
tot_size[i] = getSubgridSize()[i]*phys_size[i];
}
first = 0;
}
return tot_size;
}
void
setup_layout(void)
{
int c[4];
int i,n_mach;
int d[4];
if(mynode()==0){
printf("LAYOUT = Hypercubes, options = ");
printf("QDP");
printf("\n");
}
c[0] = nx;
c[1] = ny;
c[2] = nz;
c[3] = nt;
QDP_set_latsize(4, c);
QDP_create_layout();
sites_on_node = QDP_sites_on_node;
even_sites_on_node = QDP_subset_len(QDP_even);
odd_sites_on_node = QDP_subset_len(QDP_odd);
n_mach = QMP_get_logical_number_of_dimensions();
dim_mach = QMP_get_logical_dimensions();
#ifdef FIX_IONODE_GEOM
/* Initialize I/O node function */
init_io_node();
#endif
/* Report sublattice dimensions */
for(i = 0; i < 4; i++){
/* Any extra machine dimensions are assumed to be 1 */
if(i < n_mach)d[i] = c[i]/dim_mach[i];
else d[i] = c[i];
}
if( mynode()==0)
printf("ON EACH NODE %d x %d x %d x %d\n",d[0],d[1],d[2],d[3]);
}
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[])
{
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;
}
void setup_layout( void )
{
const int *p_machine_dimensions = NULL;
int number_machine_dimensions = -1;
int i;
number_machine_dimensions = QMP_get_logical_number_of_dimensions();
printf( "number of QMP machine dimensions = %i\n", number_machine_dimensions );
p_machine_dimensions = QMP_get_logical_dimensions();
if( p_machine_dimensions == NULL )
{
printf( "p_machines_dimensions is NULL\n" );
terminate( 0 );
}
for( i = 0; i < number_machine_dimensions; i++ )
{
printf( "QMP machine dimension ( %i ) = %i\n", i, p_machine_dimensions[ i ] );
}
machine_nx = p_machine_dimensions[ 0 ];
machine_ny = p_machine_dimensions[ 1 ];
machine_nz = p_machine_dimensions[ 2 ];
machine_nt = p_machine_dimensions[ 3 ];
p_machine_dimensions = NULL;
machine_dimensions[ XUP ] = machine_nx;
machine_dimensions[ YUP ] = machine_ny;
machine_dimensions[ ZUP ] = machine_nz;
machine_dimensions[ TUP ] = machine_nt;
printf( "machine_nx = %i\n", machine_nx );
printf( "machine_ny = %i\n", machine_ny );
printf( "machine_nz = %i\n", machine_nz );
printf( "machine_nt = %i\n", machine_nt );
/* Each lattice dimension must be a mutliple of the corresponding machine dimension. */
if( ( nx % machine_nx ) != 0 )
{
printf( "nx = %i is not a multiple of machine_nx = %i\n", nx, machine_nx );
terminate( 0 );
}
if( ( ny % machine_ny ) != 0 )
{
printf( "ny = %i is not a multiple of machine_ny = %i\n", ny, machine_ny );
terminate( 0 );
}
if( ( nz % machine_nz ) != 0 )
{
printf( "nz = %i is not a multiple of machine_nz = %i\n", nz, machine_nz );
terminate( 0 );
}
if( ( nt % machine_nt ) != 0 )
{
printf( "nt = %i is not a multiple of machine_nt = %i\n", nt, machine_nt );
terminate( 0 );
}
sub_lattice_nx = nx / machine_nx;
sub_lattice_ny = ny / machine_ny;
sub_lattice_nz = nz / machine_nz;
sub_lattice_nt = nt / machine_nt;
sub_lattice_dimensions[ XUP ] = sub_lattice_nx;
sub_lattice_dimensions[ YUP ] = sub_lattice_ny;
sub_lattice_dimensions[ ZUP ] = sub_lattice_nz;
sub_lattice_dimensions[ TUP ] = sub_lattice_nt;
printf( "sub_lattice_nx = %i\n", sub_lattice_nx );
printf( "sub_lattice_ny = %i\n", sub_lattice_ny );
printf( "sub_lattice_nz = %i\n", sub_lattice_nz );
printf( "sub_lattice_nt = %i\n", sub_lattice_nt );
sites_on_node = sub_lattice_nx * sub_lattice_ny * sub_lattice_nz * sub_lattice_nt;
sub_lattice_volume = sites_on_node;
/* The number of sites per node must be even. */
if( mynode() == 0 )
{
if( sites_on_node % 2 != 0)
{
printf( "sites_on_node is not even\n" );
terminate(0);
}
}
even_sites_on_node = sites_on_node / 2;
odd_sites_on_node = sites_on_node / 2;
if( mynode()==0)
printf("ON EACH NODE %d x %d x %d x %d\n",sub_lattice_nx,sub_lattice_ny,
sub_lattice_nz,sub_lattice_nt);
if( mynode()==0 && sites_on_node%2 != 0)
printf("WATCH OUT FOR EVEN/ODD SITES ON NODE BUG!!!\n");
}
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;
}