void SU3Test(int argc, char **argv) {
for (int i =1;i < argc; i++){
if(process_command_line_option(argc, argv, &i) == 0){
continue;
}
fprintf(stderr, "ERROR: Invalid option:%s\n", argv[i]);
usage(argv);
}
init();
if (strcmp(latfile,"")) { // load in the command line supplied gauge field
read_gauge_field(latfile, gauge, param.cpu_prec, param.X, argc, argv);
construct_gauge_field((void**)gauge, 2, param.cpu_prec, ¶m);
} else { // generate a random SU(3) field
printf("Randomizing fields...");
construct_gauge_field((void**)gauge, 1, param.cpu_prec, ¶m);
printf("done.\n");
}
loadGaugeQuda(gauge, ¶m);
saveGaugeQuda(new_gauge, ¶m);
check_gauge(gauge, new_gauge, 1e-3, param.cpu_prec);
end();
}
void SU3Test(int argc, char **argv) {
init();
char *latfile = "";//"16_64.lat";
if (strcmp(latfile,"")) { // load in the command line supplied gauge field
read_gauge_field(latfile, gauge, param.cpu_prec, param.X, argc, argv);
construct_gauge_field((void**)gauge, 2, param.cpu_prec, ¶m);
} else { // generate a random SU(3) field
printf("Randomizing fields...");
construct_gauge_field((void**)gauge, 1, param.cpu_prec, ¶m);
printf("done.\n");
}
loadGaugeQuda(gauge, ¶m);
saveGaugeQuda(new_gauge, ¶m);
check_gauge(gauge, new_gauge, 1e-3, param.cpu_prec);
end();
}
void SU3Test() {
init();
printf("Randomizing fields...");
construct_gauge_field((void**)gauge, 1, param.cpu_prec, ¶m);
printf("done.\n");
loadGaugeQuda(gauge, ¶m);
saveGaugeQuda(new_gauge, ¶m);
check_gauge(gauge, new_gauge, 1e-3, param.cpu_prec);
end();
}
void init() {
gauge_param = newQudaGaugeParam();
inv_param = newQudaInvertParam();
gauge_param.X[0] = 12;
gauge_param.X[1] = 12;
gauge_param.X[2] = 12;
gauge_param.X[3] = 12;
setDims(gauge_param.X, Ls);
gauge_param.anisotropy = 2.3;
gauge_param.type = QUDA_WILSON_LINKS;
gauge_param.gauge_order = QUDA_QDP_GAUGE_ORDER;
gauge_param.t_boundary = QUDA_ANTI_PERIODIC_T;
gauge_param.cpu_prec = cpu_prec;
gauge_param.cuda_prec = cuda_prec;
gauge_param.reconstruct = QUDA_RECONSTRUCT_12;
gauge_param.reconstruct_sloppy = gauge_param.reconstruct;
gauge_param.cuda_prec_sloppy = gauge_param.cuda_prec;
gauge_param.gauge_fix = QUDA_GAUGE_FIXED_NO;
gauge_param.type = QUDA_WILSON_LINKS;
inv_param.inv_type = QUDA_CG_INVERTER;
inv_param.mass = 0.01;
inv_param.m5 = -1.5;
kappa5 = 0.5/(5 + inv_param.m5);
inv_param.Ls = Ls;
inv_param.matpc_type = QUDA_MATPC_EVEN_EVEN;
inv_param.dagger = dagger;
inv_param.cpu_prec = cpu_prec;
inv_param.cuda_prec = cuda_prec;
gauge_param.ga_pad = 0;
inv_param.sp_pad = 0;
inv_param.cl_pad = 0;
inv_param.gamma_basis = QUDA_DEGRAND_ROSSI_GAMMA_BASIS;
inv_param.dirac_order = QUDA_DIRAC_ORDER;
if (test_type == 2) {
inv_param.solution_type = QUDA_MAT_SOLUTION;
} else {
inv_param.solution_type = QUDA_MATPC_SOLUTION;
}
inv_param.dslash_type = QUDA_DOMAIN_WALL_DSLASH;
inv_param.verbosity = QUDA_VERBOSE;
// construct input fields
for (int dir = 0; dir < 4; dir++) hostGauge[dir] = malloc(V*gaugeSiteSize*gauge_param.cpu_prec);
ColorSpinorParam csParam;
csParam.fieldLocation = QUDA_CPU_FIELD_LOCATION;
csParam.nColor = 3;
csParam.nSpin = 4;
csParam.nDim = 5;
for (int d=0; d<4; d++) csParam.x[d] = gauge_param.X[d];
csParam.x[4] = Ls;
csParam.precision = inv_param.cpu_prec;
csParam.pad = 0;
if (test_type < 2) {
csParam.siteSubset = QUDA_PARITY_SITE_SUBSET;
csParam.x[0] /= 2;
} else {
csParam.siteSubset = QUDA_FULL_SITE_SUBSET;
}
csParam.siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
csParam.fieldOrder = QUDA_SPACE_SPIN_COLOR_FIELD_ORDER;
csParam.gammaBasis = inv_param.gamma_basis;
csParam.create = QUDA_ZERO_FIELD_CREATE;
spinor = new cpuColorSpinorField(csParam);
spinorOut = new cpuColorSpinorField(csParam);
spinorRef = new cpuColorSpinorField(csParam);
csParam.siteSubset = QUDA_FULL_SITE_SUBSET;
csParam.x[0] = gauge_param.X[0];
printfQuda("Randomizing fields... ");
construct_gauge_field(hostGauge, 1, gauge_param.cpu_prec, &gauge_param);
spinor->Source(QUDA_RANDOM_SOURCE);
printfQuda("done.\n"); fflush(stdout);
int dev = 0;
initQuda(dev);
printfQuda("Sending gauge field to GPU\n");
loadGaugeQuda(hostGauge, &gauge_param);
if (!transfer) {
csParam.fieldLocation = QUDA_CUDA_FIELD_LOCATION;
csParam.gammaBasis = QUDA_UKQCD_GAMMA_BASIS;
csParam.pad = inv_param.sp_pad;
csParam.precision = inv_param.cuda_prec;
if (csParam.precision == QUDA_DOUBLE_PRECISION ) {
csParam.fieldOrder = QUDA_FLOAT2_FIELD_ORDER;
} else {
/* Single and half */
csParam.fieldOrder = QUDA_FLOAT4_FIELD_ORDER;
}
if (test_type < 2) {
csParam.siteSubset = QUDA_PARITY_SITE_SUBSET;
csParam.x[0] /= 2;
}
printfQuda("Creating cudaSpinor\n");
cudaSpinor = new cudaColorSpinorField(csParam);
printfQuda("Creating cudaSpinorOut\n");
cudaSpinorOut = new cudaColorSpinorField(csParam);
if (test_type == 2) csParam.x[0] /= 2;
csParam.siteSubset = QUDA_PARITY_SITE_SUBSET;
tmp = new cudaColorSpinorField(csParam);
printfQuda("Sending spinor field to GPU\n");
*cudaSpinor = *spinor;
std::cout << "Source: CPU = " << norm2(*spinor) << ", CUDA = " <<
norm2(*cudaSpinor) << std::endl;
bool pc = (test_type != 2);
DiracParam diracParam;
setDiracParam(diracParam, &inv_param, pc);
diracParam.verbose = QUDA_DEBUG_VERBOSE;
diracParam.tmp1 = tmp;
diracParam.tmp2 = tmp2;
dirac = Dirac::create(diracParam);
} else {
std::cout << "Source: CPU = " << norm2(*spinor) << std::endl;
}
}
void init(int argc, char **argv) {
kernelPackT = false; // Set true for kernel T face packing
cuda_prec= prec;
gauge_param = newQudaGaugeParam();
inv_param = newQudaInvertParam();
gauge_param.X[0] = xdim;
gauge_param.X[1] = ydim;
gauge_param.X[2] = zdim;
gauge_param.X[3] = tdim;
setDims(gauge_param.X);
gauge_param.anisotropy = 1.0;
gauge_param.type = QUDA_WILSON_LINKS;
gauge_param.gauge_order = QUDA_QDP_GAUGE_ORDER;
gauge_param.t_boundary = QUDA_PERIODIC_T;
gauge_param.cpu_prec = cpu_prec;
gauge_param.cuda_prec = cuda_prec;
gauge_param.reconstruct = link_recon;
gauge_param.reconstruct_sloppy = link_recon;
gauge_param.cuda_prec_sloppy = cuda_prec;
gauge_param.gauge_fix = QUDA_GAUGE_FIXED_NO;
inv_param.kappa = 0.1;
if (dslash_type == QUDA_TWISTED_MASS_DSLASH) {
inv_param.mu = 0.01;
inv_param.twist_flavor = QUDA_TWIST_MINUS;
}
inv_param.matpc_type = QUDA_MATPC_EVEN_EVEN;
inv_param.dagger = dagger;
inv_param.cpu_prec = cpu_prec;
if (inv_param.cpu_prec != gauge_param.cpu_prec)
errorQuda("Gauge and spinor cpu precisions must match");
inv_param.cuda_prec = cuda_prec;
#ifndef MULTI_GPU // free parameter for single GPU
gauge_param.ga_pad = 0;
#else // must be this one c/b face for multi gpu
int x_face_size = gauge_param.X[1]*gauge_param.X[2]*gauge_param.X[3]/2;
int y_face_size = gauge_param.X[0]*gauge_param.X[2]*gauge_param.X[3]/2;
int z_face_size = gauge_param.X[0]*gauge_param.X[1]*gauge_param.X[3]/2;
int t_face_size = gauge_param.X[0]*gauge_param.X[1]*gauge_param.X[2]/2;
int pad_size =MAX(x_face_size, y_face_size);
pad_size = MAX(pad_size, z_face_size);
pad_size = MAX(pad_size, t_face_size);
gauge_param.ga_pad = pad_size;
#endif
inv_param.sp_pad = 0;
inv_param.cl_pad = 0;
//inv_param.sp_pad = 24*24*24;
//inv_param.cl_pad = 24*24*24;
inv_param.gamma_basis = QUDA_DEGRAND_ROSSI_GAMMA_BASIS; // test code only supports DeGrand-Rossi Basis
inv_param.dirac_order = QUDA_DIRAC_ORDER;
if (test_type == 2) {
inv_param.solution_type = QUDA_MAT_SOLUTION;
} else {
inv_param.solution_type = QUDA_MATPC_SOLUTION;
}
inv_param.dslash_type = dslash_type;
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
inv_param.clover_cpu_prec = cpu_prec;
inv_param.clover_cuda_prec = cuda_prec;
inv_param.clover_cuda_prec_sloppy = inv_param.clover_cuda_prec;
inv_param.clover_order = QUDA_PACKED_CLOVER_ORDER;
//if (test_type > 0) {
hostClover = malloc(V*cloverSiteSize*inv_param.clover_cpu_prec);
hostCloverInv = hostClover; // fake it
/*} else {
hostClover = NULL;
hostCloverInv = malloc(V*cloverSiteSize*inv_param.clover_cpu_prec);
}*/
} else if (dslash_type == QUDA_TWISTED_MASS_DSLASH) {
}
//inv_param.verbosity = QUDA_VERBOSE;
// construct input fields
for (int dir = 0; dir < 4; dir++) hostGauge[dir] = malloc(V*gaugeSiteSize*gauge_param.cpu_prec);
ColorSpinorParam csParam;
csParam.fieldLocation = QUDA_CPU_FIELD_LOCATION;
csParam.nColor = 3;
csParam.nSpin = 4;
if (dslash_type == QUDA_TWISTED_MASS_DSLASH) {
csParam.twistFlavor = inv_param.twist_flavor;
}
csParam.nDim = 4;
for (int d=0; d<4; d++) csParam.x[d] = gauge_param.X[d];
csParam.precision = inv_param.cpu_prec;
csParam.pad = 0;
if (test_type < 2) {
csParam.siteSubset = QUDA_PARITY_SITE_SUBSET;
csParam.x[0] /= 2;
} else {
csParam.siteSubset = QUDA_FULL_SITE_SUBSET;
}
csParam.siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
csParam.fieldOrder = QUDA_SPACE_SPIN_COLOR_FIELD_ORDER;
csParam.gammaBasis = inv_param.gamma_basis;
csParam.create = QUDA_ZERO_FIELD_CREATE;
//csParam.verbose = QUDA_DEBUG_VERBOSE;
spinor = new cpuColorSpinorField(csParam);
spinorOut = new cpuColorSpinorField(csParam);
spinorRef = new cpuColorSpinorField(csParam);
csParam.siteSubset = QUDA_FULL_SITE_SUBSET;
csParam.x[0] = gauge_param.X[0];
printfQuda("Randomizing fields... ");
if (strcmp(latfile,"")) { // load in the command line supplied gauge field
read_gauge_field(latfile, hostGauge, gauge_param.cpu_prec, gauge_param.X, argc, argv);
construct_gauge_field(hostGauge, 2, gauge_param.cpu_prec, &gauge_param);
} else { // else generate a random SU(3) field
construct_gauge_field(hostGauge, 1, gauge_param.cpu_prec, &gauge_param);
}
spinor->Source(QUDA_RANDOM_SOURCE);
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
double norm = 0.0; // clover components are random numbers in the range (-norm, norm)
double diag = 1.0; // constant added to the diagonal
if (test_type == 2) {
construct_clover_field(hostClover, norm, diag, inv_param.clover_cpu_prec);
} else {
construct_clover_field(hostCloverInv, norm, diag, inv_param.clover_cpu_prec);
}
}
printfQuda("done.\n"); fflush(stdout);
initQuda(device);
printfQuda("Sending gauge field to GPU\n");
loadGaugeQuda(hostGauge, &gauge_param);
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
printfQuda("Sending clover field to GPU\n");
loadCloverQuda(hostClover, hostCloverInv, &inv_param);
//clover = cudaCloverPrecise;
}
if (!transfer) {
csParam.fieldLocation = QUDA_CUDA_FIELD_LOCATION;
csParam.gammaBasis = QUDA_UKQCD_GAMMA_BASIS;
csParam.pad = inv_param.sp_pad;
csParam.precision = inv_param.cuda_prec;
if (csParam.precision == QUDA_DOUBLE_PRECISION ) {
csParam.fieldOrder = QUDA_FLOAT2_FIELD_ORDER;
} else {
/* Single and half */
csParam.fieldOrder = QUDA_FLOAT4_FIELD_ORDER;
}
if (test_type < 2) {
csParam.siteSubset = QUDA_PARITY_SITE_SUBSET;
csParam.x[0] /= 2;
}
printfQuda("Creating cudaSpinor\n");
cudaSpinor = new cudaColorSpinorField(csParam);
printfQuda("Creating cudaSpinorOut\n");
cudaSpinorOut = new cudaColorSpinorField(csParam);
if (test_type == 2) csParam.x[0] /= 2;
csParam.siteSubset = QUDA_PARITY_SITE_SUBSET;
tmp1 = new cudaColorSpinorField(csParam);
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH ||
dslash_type == QUDA_TWISTED_MASS_DSLASH) {
tmp2 = new cudaColorSpinorField(csParam);
}
printfQuda("Sending spinor field to GPU\n");
*cudaSpinor = *spinor;
std::cout << "Source: CPU = " << norm2(*spinor) << ", CUDA = " <<
norm2(*cudaSpinor) << std::endl;
bool pc = (test_type != 2);
DiracParam diracParam;
setDiracParam(diracParam, &inv_param, pc);
diracParam.verbose = QUDA_VERBOSE;
diracParam.tmp1 = tmp1;
diracParam.tmp2 = tmp2;
dirac = Dirac::create(diracParam);
} else {
std::cout << "Source: CPU = " << norm2(*spinor) << std::endl;
}
}
int main(int argc, char **argv)
{
// set QUDA parameters
int device = 0; // CUDA device number
QudaPrecision cpu_prec = QUDA_DOUBLE_PRECISION;
QudaPrecision cuda_prec = QUDA_SINGLE_PRECISION;
QudaPrecision cuda_prec_sloppy = QUDA_HALF_PRECISION;
QudaGaugeParam gauge_param = newQudaGaugeParam();
QudaInvertParam inv_param = newQudaInvertParam();
gauge_param.X[0] = 16;
gauge_param.X[1] = 16;
gauge_param.X[2] = 16;
gauge_param.X[3] = 16;
inv_param.Ls = 16;
gauge_param.anisotropy = 1.0;
gauge_param.type = QUDA_WILSON_LINKS;
gauge_param.gauge_order = QUDA_QDP_GAUGE_ORDER;
gauge_param.t_boundary = QUDA_ANTI_PERIODIC_T;
gauge_param.cpu_prec = cpu_prec;
gauge_param.cuda_prec = cuda_prec;
gauge_param.reconstruct = QUDA_RECONSTRUCT_12;
gauge_param.cuda_prec_sloppy = cuda_prec_sloppy;
gauge_param.reconstruct_sloppy = QUDA_RECONSTRUCT_12;
gauge_param.gauge_fix = QUDA_GAUGE_FIXED_NO;
inv_param.dslash_type = QUDA_DOMAIN_WALL_DSLASH;
inv_param.inv_type = QUDA_CG_INVERTER;
inv_param.mass = 0.01;
inv_param.m5 = -1.5;
double kappa5 = 0.5/(5 + inv_param.m5);
inv_param.tol = 5e-8;
inv_param.maxiter = 1000;
inv_param.reliable_delta = 0.1;
inv_param.solution_type = QUDA_MAT_SOLUTION;
inv_param.solve_type = QUDA_NORMEQ_PC_SOLVE;
inv_param.matpc_type = QUDA_MATPC_EVEN_EVEN;
inv_param.dagger = QUDA_DAG_NO;
inv_param.mass_normalization = QUDA_KAPPA_NORMALIZATION;
inv_param.cpu_prec = cpu_prec;
inv_param.cuda_prec = cuda_prec;
inv_param.cuda_prec_sloppy = cuda_prec_sloppy;
inv_param.prec_precondition = cuda_prec_sloppy;
inv_param.preserve_source = QUDA_PRESERVE_SOURCE_NO;
inv_param.gamma_basis = QUDA_DEGRAND_ROSSI_GAMMA_BASIS;
inv_param.dirac_order = QUDA_DIRAC_ORDER;
inv_param.dirac_tune = QUDA_TUNE_YES;
inv_param.preserve_dirac = QUDA_PRESERVE_DIRAC_YES;
gauge_param.ga_pad = 0; // 24*24*24;
inv_param.sp_pad = 0; // 24*24*24;
inv_param.cl_pad = 0; // 24*24*24;
inv_param.verbosity = QUDA_VERBOSE;
// Everything between here and the call to initQuda() is application-specific.
// set parameters for the reference Dslash, and prepare fields to be loaded
setDims(gauge_param.X, inv_param.Ls);
size_t gSize = (gauge_param.cpu_prec == QUDA_DOUBLE_PRECISION) ? sizeof(double) : sizeof(float);
size_t sSize = (inv_param.cpu_prec == QUDA_DOUBLE_PRECISION) ? sizeof(double) : sizeof(float);
void *gauge[4];
for (int dir = 0; dir < 4; dir++) {
gauge[dir] = malloc(V*gaugeSiteSize*gSize);
}
construct_gauge_field(gauge, 1, gauge_param.cpu_prec, &gauge_param);
void *spinorIn = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
void *spinorOut = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
void *spinorCheck = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
// create a point source at 0
if (inv_param.cpu_prec == QUDA_SINGLE_PRECISION) *((float*)spinorIn) = 1.0;
else *((double*)spinorIn) = 1.0;
// start the timer
double time0 = -((double)clock());
// initialize the QUDA library
initQuda(device);
// load the gauge field
loadGaugeQuda((void*)gauge, &gauge_param);
// perform the inversion
invertQuda(spinorOut, spinorIn, &inv_param);
// stop the timer
time0 += clock();
time0 /= CLOCKS_PER_SEC;
printf("Device memory used:\n Spinor: %f GiB\n Gauge: %f GiB\n",
inv_param.spinorGiB, gauge_param.gaugeGiB);
printf("\nDone: %i iter / %g secs = %g Gflops, total time = %g secs\n",
inv_param.iter, inv_param.secs, inv_param.gflops/inv_param.secs, time0);
if (inv_param.solution_type == QUDA_MAT_SOLUTION) {
mat(spinorCheck, gauge, spinorOut, kappa5, 0, inv_param.cpu_prec,
gauge_param.cpu_prec, inv_param.mass);
if (inv_param.mass_normalization == QUDA_MASS_NORMALIZATION)
ax(0.5/kappa5, spinorCheck, V*spinorSiteSize, inv_param.cpu_prec);
} else if(inv_param.solution_type == QUDA_MATPC_SOLUTION) {
matpc(spinorCheck, gauge, spinorOut, kappa5, inv_param.matpc_type, 0,
inv_param.cpu_prec, gauge_param.cpu_prec, inv_param.mass);
if (inv_param.mass_normalization == QUDA_MASS_NORMALIZATION)
ax(0.25/(kappa5*kappa5), spinorCheck, V*spinorSiteSize, inv_param.cpu_prec);
}
mxpy(spinorIn, spinorCheck, V*spinorSiteSize, inv_param.cpu_prec);
double nrm2 = norm_2(spinorCheck, V*spinorSiteSize, inv_param.cpu_prec);
double src2 = norm_2(spinorIn, V*spinorSiteSize, inv_param.cpu_prec);
printf("Relative residual: requested = %g, actual = %g\n", inv_param.tol, sqrt(nrm2/src2));
// finalize the QUDA library
endQuda();
return 0;
}
int main(int argc, char **argv)
{
for (int i = 1; i < argc; i++){
if(process_command_line_option(argc, argv, &i) == 0){
continue;
}
printfQuda("ERROR: Invalid option:%s\n", argv[i]);
usage(argv);
}
if (prec_sloppy == QUDA_INVALID_PRECISION){
prec_sloppy = prec;
}
if (link_recon_sloppy == QUDA_RECONSTRUCT_INVALID){
link_recon_sloppy = link_recon;
}
// initialize QMP/MPI, QUDA comms grid and RNG (test_util.cpp)
initComms(argc, argv, gridsize_from_cmdline);
display_test_info();
// *** QUDA parameters begin here.
if (dslash_type != QUDA_WILSON_DSLASH &&
dslash_type != QUDA_CLOVER_WILSON_DSLASH &&
dslash_type != QUDA_TWISTED_MASS_DSLASH &&
dslash_type != QUDA_DOMAIN_WALL_4D_DSLASH &&
dslash_type != QUDA_MOBIUS_DWF_DSLASH &&
dslash_type != QUDA_TWISTED_CLOVER_DSLASH &&
dslash_type != QUDA_DOMAIN_WALL_DSLASH) {
printfQuda("dslash_type %d not supported\n", dslash_type);
exit(0);
}
QudaPrecision cpu_prec = QUDA_DOUBLE_PRECISION;
QudaPrecision cuda_prec = prec;
QudaPrecision cuda_prec_sloppy = prec_sloppy;
QudaPrecision cuda_prec_precondition = QUDA_HALF_PRECISION;
QudaGaugeParam gauge_param = newQudaGaugeParam();
QudaInvertParam inv_param = newQudaInvertParam();
double kappa5;
gauge_param.X[0] = xdim;
gauge_param.X[1] = ydim;
gauge_param.X[2] = zdim;
gauge_param.X[3] = tdim;
inv_param.Ls = 1;
gauge_param.anisotropy = 1.0;
gauge_param.type = QUDA_WILSON_LINKS;
gauge_param.gauge_order = QUDA_QDP_GAUGE_ORDER;
gauge_param.t_boundary = QUDA_ANTI_PERIODIC_T;
gauge_param.cpu_prec = cpu_prec;
gauge_param.cuda_prec = cuda_prec;
gauge_param.reconstruct = link_recon;
gauge_param.cuda_prec_sloppy = cuda_prec_sloppy;
gauge_param.reconstruct_sloppy = link_recon_sloppy;
gauge_param.cuda_prec_precondition = cuda_prec_precondition;
gauge_param.reconstruct_precondition = link_recon_sloppy;
gauge_param.gauge_fix = QUDA_GAUGE_FIXED_NO;
inv_param.dslash_type = dslash_type;
inv_param.mass = mass;
inv_param.kappa = 1.0 / (2.0 * (1 + 3/gauge_param.anisotropy + mass));
if (dslash_type == QUDA_TWISTED_MASS_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) {
inv_param.mu = 0.12;
inv_param.epsilon = 0.1385;
inv_param.twist_flavor = twist_flavor;
inv_param.Ls = (inv_param.twist_flavor == QUDA_TWIST_NONDEG_DOUBLET) ? 2 : 1;
} else if (dslash_type == QUDA_DOMAIN_WALL_DSLASH ||
dslash_type == QUDA_DOMAIN_WALL_4D_DSLASH) {
inv_param.m5 = -1.8;
kappa5 = 0.5/(5 + inv_param.m5);
inv_param.Ls = Lsdim;
} else if (dslash_type == QUDA_MOBIUS_DWF_DSLASH) {
inv_param.m5 = -1.8;
kappa5 = 0.5/(5 + inv_param.m5);
inv_param.Ls = Lsdim;
for(int k = 0; k < Lsdim; k++)
{
// b5[k], c[k] values are chosen for arbitrary values,
// but the difference of them are same as 1.0
inv_param.b_5[k] = 1.452;
inv_param.c_5[k] = 0.452;
}
}
// offsets used only by multi-shift solver
inv_param.num_offset = 4;
double offset[4] = {0.01, 0.02, 0.03, 0.04};
for (int i=0; i<inv_param.num_offset; i++) inv_param.offset[i] = offset[i];
inv_param.inv_type = inv_type;
if (inv_param.dslash_type == QUDA_TWISTED_MASS_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) {
inv_param.solution_type = QUDA_MAT_SOLUTION;
} else {
inv_param.solution_type = multishift ? QUDA_MATPCDAG_MATPC_SOLUTION : QUDA_MATPC_SOLUTION;
}
inv_param.matpc_type = matpc_type;
inv_param.dagger = QUDA_DAG_NO;
inv_param.mass_normalization = normalization;
inv_param.solver_normalization = QUDA_DEFAULT_NORMALIZATION;
if (dslash_type == QUDA_DOMAIN_WALL_DSLASH ||
dslash_type == QUDA_DOMAIN_WALL_4D_DSLASH ||
dslash_type == QUDA_MOBIUS_DWF_DSLASH ||
dslash_type == QUDA_TWISTED_MASS_DSLASH ||
dslash_type == QUDA_TWISTED_CLOVER_DSLASH ||
multishift || inv_type == QUDA_CG_INVERTER) {
inv_param.solve_type = QUDA_NORMOP_PC_SOLVE;
} else {
inv_param.solve_type = QUDA_DIRECT_PC_SOLVE;
}
inv_param.pipeline = 0;
inv_param.Nsteps = 2;
inv_param.gcrNkrylov = 10;
inv_param.tol = 1e-7;
inv_param.tol_restart = 1e-3; //now theoretical background for this parameter...
#if __COMPUTE_CAPABILITY__ >= 200
// require both L2 relative and heavy quark residual to determine convergence
inv_param.residual_type = static_cast<QudaResidualType>(QUDA_L2_RELATIVE_RESIDUAL | QUDA_HEAVY_QUARK_RESIDUAL);
inv_param.tol_hq = 1e-3; // specify a tolerance for the residual for heavy quark residual
#else
// Pre Fermi architecture only supports L2 relative residual norm
inv_param.residual_type = QUDA_L2_RELATIVE_RESIDUAL;
#endif
// these can be set individually
for (int i=0; i<inv_param.num_offset; i++) {
inv_param.tol_offset[i] = inv_param.tol;
inv_param.tol_hq_offset[i] = inv_param.tol_hq;
}
inv_param.maxiter = 10000;
inv_param.reliable_delta = 1e-1;
inv_param.use_sloppy_partial_accumulator = 0;
inv_param.max_res_increase = 1;
// domain decomposition preconditioner parameters
inv_param.inv_type_precondition = precon_type;
inv_param.schwarz_type = QUDA_ADDITIVE_SCHWARZ;
inv_param.precondition_cycle = 1;
inv_param.tol_precondition = 1e-1;
inv_param.maxiter_precondition = 10;
inv_param.verbosity_precondition = QUDA_SILENT;
inv_param.cuda_prec_precondition = cuda_prec_precondition;
inv_param.omega = 1.0;
inv_param.cpu_prec = cpu_prec;
inv_param.cuda_prec = cuda_prec;
inv_param.cuda_prec_sloppy = cuda_prec_sloppy;
inv_param.preserve_source = QUDA_PRESERVE_SOURCE_NO;
inv_param.gamma_basis = QUDA_DEGRAND_ROSSI_GAMMA_BASIS;
inv_param.dirac_order = QUDA_DIRAC_ORDER;
inv_param.input_location = QUDA_CPU_FIELD_LOCATION;
inv_param.output_location = QUDA_CPU_FIELD_LOCATION;
inv_param.tune = tune ? QUDA_TUNE_YES : QUDA_TUNE_NO;
gauge_param.ga_pad = 0; // 24*24*24/2;
inv_param.sp_pad = 0; // 24*24*24/2;
inv_param.cl_pad = 0; // 24*24*24/2;
// For multi-GPU, ga_pad must be large enough to store a time-slice
#ifdef MULTI_GPU
int x_face_size = gauge_param.X[1]*gauge_param.X[2]*gauge_param.X[3]/2;
int y_face_size = gauge_param.X[0]*gauge_param.X[2]*gauge_param.X[3]/2;
int z_face_size = gauge_param.X[0]*gauge_param.X[1]*gauge_param.X[3]/2;
int t_face_size = gauge_param.X[0]*gauge_param.X[1]*gauge_param.X[2]/2;
int pad_size =MAX(x_face_size, y_face_size);
pad_size = MAX(pad_size, z_face_size);
pad_size = MAX(pad_size, t_face_size);
gauge_param.ga_pad = pad_size;
#endif
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) {
inv_param.clover_cpu_prec = cpu_prec;
inv_param.clover_cuda_prec = cuda_prec;
inv_param.clover_cuda_prec_sloppy = cuda_prec_sloppy;
inv_param.clover_cuda_prec_precondition = cuda_prec_precondition;
inv_param.clover_order = QUDA_PACKED_CLOVER_ORDER;
inv_param.clover_coeff = 1.5*inv_param.kappa;
}
inv_param.verbosity = QUDA_VERBOSE;
// *** Everything between here and the call to initQuda() is
// *** application-specific.
// set parameters for the reference Dslash, and prepare fields to be loaded
if (dslash_type == QUDA_DOMAIN_WALL_DSLASH ||
dslash_type == QUDA_DOMAIN_WALL_4D_DSLASH ||
dslash_type == QUDA_MOBIUS_DWF_DSLASH) {
dw_setDims(gauge_param.X, inv_param.Ls);
} else {
setDims(gauge_param.X);
}
setSpinorSiteSize(24);
size_t gSize = (gauge_param.cpu_prec == QUDA_DOUBLE_PRECISION) ? sizeof(double) : sizeof(float);
size_t sSize = (inv_param.cpu_prec == QUDA_DOUBLE_PRECISION) ? sizeof(double) : sizeof(float);
void *gauge[4], *clover_inv=0, *clover=0;
for (int dir = 0; dir < 4; dir++) {
gauge[dir] = malloc(V*gaugeSiteSize*gSize);
}
if (strcmp(latfile,"")) { // load in the command line supplied gauge field
read_gauge_field(latfile, gauge, gauge_param.cpu_prec, gauge_param.X, argc, argv);
construct_gauge_field(gauge, 2, gauge_param.cpu_prec, &gauge_param);
} else { // else generate a random SU(3) field
construct_gauge_field(gauge, 1, gauge_param.cpu_prec, &gauge_param);
}
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
double norm = 0.0; // clover components are random numbers in the range (-norm, norm)
double diag = 1.0; // constant added to the diagonal
size_t cSize = (inv_param.clover_cpu_prec == QUDA_DOUBLE_PRECISION) ? sizeof(double) : sizeof(float);
clover_inv = malloc(V*cloverSiteSize*cSize);
construct_clover_field(clover_inv, norm, diag, inv_param.clover_cpu_prec);
// The uninverted clover term is only needed when solving the unpreconditioned
// system or when using "asymmetric" even/odd preconditioning.
int preconditioned = (inv_param.solve_type == QUDA_DIRECT_PC_SOLVE ||
inv_param.solve_type == QUDA_NORMOP_PC_SOLVE);
int asymmetric = preconditioned &&
(inv_param.matpc_type == QUDA_MATPC_EVEN_EVEN_ASYMMETRIC ||
inv_param.matpc_type == QUDA_MATPC_ODD_ODD_ASYMMETRIC);
if (!preconditioned) {
clover = clover_inv;
clover_inv = NULL;
} else if (asymmetric) { // fake it by using the same random matrix
clover = clover_inv; // for both clover and clover_inv
} else {
clover = NULL;
}
}
void *spinorIn = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
void *spinorCheck = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
void *spinorOut = NULL, **spinorOutMulti = NULL;
if (multishift) {
spinorOutMulti = (void**)malloc(inv_param.num_offset*sizeof(void *));
for (int i=0; i<inv_param.num_offset; i++) {
spinorOutMulti[i] = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
}
} else {
spinorOut = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
}
memset(spinorIn, 0, inv_param.Ls*V*spinorSiteSize*sSize);
memset(spinorCheck, 0, inv_param.Ls*V*spinorSiteSize*sSize);
if (multishift) {
for (int i=0; i<inv_param.num_offset; i++) memset(spinorOutMulti[i], 0, inv_param.Ls*V*spinorSiteSize*sSize);
} else {
memset(spinorOut, 0, inv_param.Ls*V*spinorSiteSize*sSize);
}
// create a point source at 0 (in each subvolume... FIXME)
// create a point source at 0 (in each subvolume... FIXME)
if (inv_param.cpu_prec == QUDA_SINGLE_PRECISION) {
//((float*)spinorIn)[0] = 1.0;
for (int i=0; i<inv_param.Ls*V*spinorSiteSize; i++) ((float*)spinorIn)[i] = rand() / (float)RAND_MAX;
} else {
//((double*)spinorIn)[0] = 1.0;
for (int i=0; i<inv_param.Ls*V*spinorSiteSize; i++) ((double*)spinorIn)[i] = rand() / (double)RAND_MAX;
}
// start the timer
double time0 = -((double)clock());
// initialize the QUDA library
initQuda(device);
// load the gauge field
loadGaugeQuda((void*)gauge, &gauge_param);
// load the clover term, if desired
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH) loadCloverQuda(clover, clover_inv, &inv_param);
if (dslash_type == QUDA_TWISTED_CLOVER_DSLASH) loadCloverQuda(NULL, NULL, &inv_param);
// perform the inversion
if (multishift) {
invertMultiShiftQuda(spinorOutMulti, spinorIn, &inv_param);
} else {
invertQuda(spinorOut, spinorIn, &inv_param);
}
// stop the timer
time0 += clock();
time0 /= CLOCKS_PER_SEC;
printfQuda("Device memory used:\n Spinor: %f GiB\n Gauge: %f GiB\n",
inv_param.spinorGiB, gauge_param.gaugeGiB);
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) {
printfQuda(" Clover: %f GiB\n", inv_param.cloverGiB);
}
printfQuda("\nDone: %i iter / %g secs = %g Gflops, total time = %g secs\n",
inv_param.iter, inv_param.secs, inv_param.gflops/inv_param.secs, time0);
if (multishift) {
if (inv_param.mass_normalization == QUDA_MASS_NORMALIZATION) {
errorQuda("Mass normalization not supported for multi-shift solver in invert_test");
}
void *spinorTmp = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
printfQuda("Host residuum checks: \n");
for(int i=0; i < inv_param.num_offset; i++) {
ax(0, spinorCheck, V*spinorSiteSize, inv_param.cpu_prec);
if (dslash_type == QUDA_TWISTED_MASS_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) {
if (inv_param.twist_flavor != QUDA_TWIST_MINUS && inv_param.twist_flavor != QUDA_TWIST_PLUS)
errorQuda("Twisted mass solution type not supported");
tm_matpc(spinorTmp, gauge, spinorOutMulti[i], inv_param.kappa, inv_param.mu, inv_param.twist_flavor,
inv_param.matpc_type, 0, inv_param.cpu_prec, gauge_param);
tm_matpc(spinorCheck, gauge, spinorTmp, inv_param.kappa, inv_param.mu, inv_param.twist_flavor,
inv_param.matpc_type, 1, inv_param.cpu_prec, gauge_param);
} else if (dslash_type == QUDA_WILSON_DSLASH || dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
wil_matpc(spinorTmp, gauge, spinorOutMulti[i], inv_param.kappa, inv_param.matpc_type, 0,
inv_param.cpu_prec, gauge_param);
wil_matpc(spinorCheck, gauge, spinorTmp, inv_param.kappa, inv_param.matpc_type, 1,
inv_param.cpu_prec, gauge_param);
} else {
printfQuda("Domain wall not supported for multi-shift\n");
exit(-1);
}
axpy(inv_param.offset[i], spinorOutMulti[i], spinorCheck, Vh*spinorSiteSize, inv_param.cpu_prec);
mxpy(spinorIn, spinorCheck, Vh*spinorSiteSize, inv_param.cpu_prec);
double nrm2 = norm_2(spinorCheck, Vh*spinorSiteSize, inv_param.cpu_prec);
double src2 = norm_2(spinorIn, Vh*spinorSiteSize, inv_param.cpu_prec);
double l2r = sqrt(nrm2 / src2);
printfQuda("Shift %d residuals: (L2 relative) tol %g, QUDA = %g, host = %g; (heavy-quark) tol %g, QUDA = %g\n",
i, inv_param.tol_offset[i], inv_param.true_res_offset[i], l2r,
inv_param.tol_hq_offset[i], inv_param.true_res_hq_offset[i]);
}
free(spinorTmp);
} else {
if (inv_param.solution_type == QUDA_MAT_SOLUTION) {
if (dslash_type == QUDA_TWISTED_MASS_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) {
if(inv_param.twist_flavor == QUDA_TWIST_PLUS || inv_param.twist_flavor == QUDA_TWIST_MINUS)
tm_mat(spinorCheck, gauge, spinorOut, inv_param.kappa, inv_param.mu, inv_param.twist_flavor, 0, inv_param.cpu_prec, gauge_param);
else
{
int tm_offset = V*spinorSiteSize; //12*spinorRef->Volume();
void *evenOut = spinorCheck;
void *oddOut = cpu_prec == sizeof(double) ? (void*)((double*)evenOut + tm_offset): (void*)((float*)evenOut + tm_offset);
void *evenIn = spinorOut;
void *oddIn = cpu_prec == sizeof(double) ? (void*)((double*)evenIn + tm_offset): (void*)((float*)evenIn + tm_offset);
tm_ndeg_mat(evenOut, oddOut, gauge, evenIn, oddIn, inv_param.kappa, inv_param.mu, inv_param.epsilon, 0, inv_param.cpu_prec, gauge_param);
}
} else if (dslash_type == QUDA_WILSON_DSLASH || dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
wil_mat(spinorCheck, gauge, spinorOut, inv_param.kappa, 0, inv_param.cpu_prec, gauge_param);
} else if (dslash_type == QUDA_DOMAIN_WALL_DSLASH) {
dw_mat(spinorCheck, gauge, spinorOut, kappa5, inv_param.dagger, inv_param.cpu_prec, gauge_param, inv_param.mass);
// } else if (dslash_type == QUDA_DOMAIN_WALL_4D_DSLASH) {
// dw_4d_mat(spinorCheck, gauge, spinorOut, kappa5, inv_param.dagger, inv_param.cpu_prec, gauge_param, inv_param.mass);
// } else if (dslash_type == QUDA_MOBIUS_DWF_DSLASH) {
// mdw_mat(spinorCheck, gauge, spinorOut, kappa5, inv_param.dagger, inv_param.cpu_prec, gauge_param, inv_param.mass);
} else {
printfQuda("Unsupported dslash_type\n");
exit(-1);
}
if (inv_param.mass_normalization == QUDA_MASS_NORMALIZATION) {
if (dslash_type == QUDA_DOMAIN_WALL_DSLASH ||
dslash_type == QUDA_DOMAIN_WALL_4D_DSLASH ||
dslash_type == QUDA_MOBIUS_DWF_DSLASH) {
ax(0.5/kappa5, spinorCheck, V*spinorSiteSize*inv_param.Ls, inv_param.cpu_prec);
} else if (dslash_type == QUDA_TWISTED_MASS_DSLASH && twist_flavor == QUDA_TWIST_NONDEG_DOUBLET) {
ax(0.5/inv_param.kappa, spinorCheck, 2*V*spinorSiteSize, inv_param.cpu_prec);
} else {
ax(0.5/inv_param.kappa, spinorCheck, V*spinorSiteSize, inv_param.cpu_prec);
}
}
} else if(inv_param.solution_type == QUDA_MATPC_SOLUTION) {
if (dslash_type == QUDA_TWISTED_MASS_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) {
if (inv_param.twist_flavor != QUDA_TWIST_MINUS && inv_param.twist_flavor != QUDA_TWIST_PLUS)
errorQuda("Twisted mass solution type not supported");
tm_matpc(spinorCheck, gauge, spinorOut, inv_param.kappa, inv_param.mu, inv_param.twist_flavor,
inv_param.matpc_type, 0, inv_param.cpu_prec, gauge_param);
} else if (dslash_type == QUDA_WILSON_DSLASH || dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
wil_matpc(spinorCheck, gauge, spinorOut, inv_param.kappa, inv_param.matpc_type, 0,
inv_param.cpu_prec, gauge_param);
} else if (dslash_type == QUDA_DOMAIN_WALL_DSLASH) {
dw_matpc(spinorCheck, gauge, spinorOut, kappa5, inv_param.matpc_type, 0, inv_param.cpu_prec, gauge_param, inv_param.mass);
} else if (dslash_type == QUDA_DOMAIN_WALL_4D_DSLASH) {
dw_4d_matpc(spinorCheck, gauge, spinorOut, kappa5, inv_param.matpc_type, 0, inv_param.cpu_prec, gauge_param, inv_param.mass);
} else if (dslash_type == QUDA_MOBIUS_DWF_DSLASH) {
double *kappa_b, *kappa_c;
kappa_b = (double*)malloc(Lsdim*sizeof(double));
kappa_c = (double*)malloc(Lsdim*sizeof(double));
for(int xs = 0 ; xs < Lsdim ; xs++)
{
kappa_b[xs] = 1.0/(2*(inv_param.b_5[xs]*(4.0 + inv_param.m5) + 1.0));
kappa_c[xs] = 1.0/(2*(inv_param.c_5[xs]*(4.0 + inv_param.m5) - 1.0));
}
mdw_matpc(spinorCheck, gauge, spinorOut, kappa_b, kappa_c, inv_param.matpc_type, 0, inv_param.cpu_prec, gauge_param, inv_param.mass, inv_param.b_5, inv_param.c_5);
free(kappa_b);
free(kappa_c);
} else {
printfQuda("Unsupported dslash_type\n");
exit(-1);
}
if (inv_param.mass_normalization == QUDA_MASS_NORMALIZATION) {
if (dslash_type == QUDA_DOMAIN_WALL_DSLASH ||
dslash_type == QUDA_DOMAIN_WALL_4D_DSLASH ||
dslash_type == QUDA_MOBIUS_DWF_DSLASH) {
ax(0.25/(kappa5*kappa5), spinorCheck, V*spinorSiteSize*inv_param.Ls, inv_param.cpu_prec);
} else {
ax(0.25/(inv_param.kappa*inv_param.kappa), spinorCheck, Vh*spinorSiteSize, inv_param.cpu_prec);
}
}
} else if (inv_param.solution_type == QUDA_MATPCDAG_MATPC_SOLUTION) {
void *spinorTmp = malloc(V*spinorSiteSize*sSize*inv_param.Ls);
ax(0, spinorCheck, V*spinorSiteSize, inv_param.cpu_prec);
if (dslash_type == QUDA_TWISTED_MASS_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) {
if (inv_param.twist_flavor != QUDA_TWIST_MINUS && inv_param.twist_flavor != QUDA_TWIST_PLUS)
errorQuda("Twisted mass solution type not supported");
tm_matpc(spinorTmp, gauge, spinorOut, inv_param.kappa, inv_param.mu, inv_param.twist_flavor,
inv_param.matpc_type, 0, inv_param.cpu_prec, gauge_param);
tm_matpc(spinorCheck, gauge, spinorTmp, inv_param.kappa, inv_param.mu, inv_param.twist_flavor,
inv_param.matpc_type, 1, inv_param.cpu_prec, gauge_param);
} else if (dslash_type == QUDA_WILSON_DSLASH || dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
wil_matpc(spinorTmp, gauge, spinorOut, inv_param.kappa, inv_param.matpc_type, 0,
inv_param.cpu_prec, gauge_param);
wil_matpc(spinorCheck, gauge, spinorTmp, inv_param.kappa, inv_param.matpc_type, 1,
inv_param.cpu_prec, gauge_param);
} else {
printfQuda("Unsupported dslash_type\n");
exit(-1);
}
if (inv_param.mass_normalization == QUDA_MASS_NORMALIZATION) {
errorQuda("Mass normalization not implemented");
}
free(spinorTmp);
}
int vol = inv_param.solution_type == QUDA_MAT_SOLUTION ? V : Vh;
mxpy(spinorIn, spinorCheck, vol*spinorSiteSize*inv_param.Ls, inv_param.cpu_prec);
double nrm2 = norm_2(spinorCheck, vol*spinorSiteSize*inv_param.Ls, inv_param.cpu_prec);
double src2 = norm_2(spinorIn, vol*spinorSiteSize*inv_param.Ls, inv_param.cpu_prec);
double l2r = sqrt(nrm2 / src2);
printfQuda("Residuals: (L2 relative) tol %g, QUDA = %g, host = %g; (heavy-quark) tol %g, QUDA = %g\n",
inv_param.tol, inv_param.true_res, l2r, inv_param.tol_hq, inv_param.true_res_hq);
}
freeGaugeQuda();
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH || dslash_type == QUDA_TWISTED_CLOVER_DSLASH) freeCloverQuda();
// finalize the QUDA library
endQuda();
finalizeComms();
return 0;
}
