// if even_odd_flag set
void M_full_quda(spinor * const Even_new, spinor * const Odd_new, spinor * const Even, spinor * const Odd) {
inv_param.kappa = g_kappa;
inv_param.mu = fabs(g_mu);
inv_param.epsilon = 0.0;
// IMPORTANT: use opposite TM flavor since gamma5 -> -gamma5 (until LXLYLZT prob. resolved)
inv_param.twist_flavor = (g_mu < 0.0 ? QUDA_TWIST_PLUS : QUDA_TWIST_MINUS);
inv_param.Ls = (inv_param.twist_flavor == QUDA_TWIST_NONDEG_DOUBLET ||
inv_param.twist_flavor == QUDA_TWIST_DEG_DOUBLET ) ? 2 : 1;
void *spinorIn = (void*)g_spinor_field[DUM_DERI]; // source
void *spinorOut = (void*)g_spinor_field[DUM_DERI+1]; // solution
// reorder spinor
convert_eo_to_lexic( spinorIn, Even, Odd );
reorder_spinor_toQuda( (double*)spinorIn, inv_param.cpu_prec, 0, NULL );
// multiply
inv_param.solution_type = QUDA_MAT_SOLUTION;
MatQuda( spinorOut, spinorIn, &inv_param);
// reorder spinor
reorder_spinor_fromQuda( (double*)spinorOut, inv_param.cpu_prec, 0, NULL );
convert_lexic_to_eo( Even_new, Odd_new, spinorOut );
}
// execute kernel
double dslashCUDA() {
printfQuda("Executing %d kernel loops...\n", loops);
fflush(stdout);
if (test_type < 2)
dirac->Tune(*cudaSpinorOut, *cudaSpinor, *tmp);
else
dirac->Tune(cudaSpinorOut->Even(), cudaSpinor->Even(), *tmp);
cudaEvent_t start, end;
cudaEventCreate(&start);
cudaEventRecord(start, 0);
cudaEventSynchronize(start);
for (int i = 0; i < loops; i++) {
switch (test_type) {
case 0:
if (transfer) {
dslashQuda(spinorOut->V(), spinor->V(), &inv_param, parity);
} else {
dirac->Dslash(*cudaSpinorOut, *cudaSpinor, parity);
}
break;
case 1:
case 2:
if (transfer) {
MatQuda(spinorOut->V(), spinor->V(), &inv_param);
} else {
dirac->M(*cudaSpinorOut, *cudaSpinor);
}
break;
}
}
cudaEventCreate(&end);
cudaEventRecord(end, 0);
cudaEventSynchronize(end);
float runTime;
cudaEventElapsedTime(&runTime, start, end);
cudaEventDestroy(start);
cudaEventDestroy(end);
double secs = runTime / 1000; //stopwatchReadSeconds();
// check for errors
cudaError_t stat = cudaGetLastError();
if (stat != cudaSuccess)
printf("with ERROR: %s\n", cudaGetErrorString(stat));
printf("done.\n\n");
return secs;
}
// execute kernel
double dslashCUDA(int niter) {
cudaEvent_t start, end;
cudaEventCreate(&start);
cudaEventCreate(&end);
cudaEventRecord(start, 0);
for (int i = 0; i < niter; i++) {
switch (test_type) {
case 0:
if (transfer) {
dslashQuda(spinorOut->V(), spinor->V(), &inv_param, parity);
} else {
//inv_param.input_location = QUDA_CUDA_FIELD_LOCATION;
//inv_param.output_location = QUDA_CUDA_FIELD_LOCATION;
//dslashQuda(cudaSpinorOut->V(), cudaSpinor->V(), &inv_param, parity);
dirac->Dslash(*cudaSpinorOut, *cudaSpinor, parity);
}
break;
case 1:
case 2:
if (transfer) {
MatQuda(spinorOut->V(), spinor->V(), &inv_param);
} else {
dirac->M(*cudaSpinorOut, *cudaSpinor);
}
break;
case 3:
case 4:
if (transfer) {
MatDagMatQuda(spinorOut->V(), spinor->V(), &inv_param);
} else {
dirac->MdagM(*cudaSpinorOut, *cudaSpinor);
}
break;
}
}
cudaEventRecord(end, 0);
cudaEventSynchronize(end);
float runTime;
cudaEventElapsedTime(&runTime, start, end);
cudaEventDestroy(start);
cudaEventDestroy(end);
double secs = runTime / 1000; //stopwatchReadSeconds();
// check for errors
cudaError_t stat = cudaGetLastError();
if (stat != cudaSuccess)
printfQuda("with ERROR: %s\n", cudaGetErrorString(stat));
return secs;
}
// no even-odd
void D_psi_quda(spinor * const P, spinor * const Q) {
inv_param.kappa = g_kappa;
inv_param.mu = fabs(g_mu);
inv_param.epsilon = 0.0;
// IMPORTANT: use opposite TM flavor since gamma5 -> -gamma5 (until LXLYLZT prob. resolved)
inv_param.twist_flavor = (g_mu < 0.0 ? QUDA_TWIST_PLUS : QUDA_TWIST_MINUS);
inv_param.Ls = (inv_param.twist_flavor == QUDA_TWIST_NONDEG_DOUBLET ||
inv_param.twist_flavor == QUDA_TWIST_DEG_DOUBLET ) ? 2 : 1;
void *spinorIn = (void*)Q;
void *spinorOut = (void*)P;
// reorder spinor
reorder_spinor_toQuda( (double*)spinorIn, inv_param.cpu_prec, 0, NULL );
// multiply
inv_param.solution_type = QUDA_MAT_SOLUTION;
MatQuda( spinorOut, spinorIn, &inv_param);
// reorder spinor
reorder_spinor_fromQuda( (double*)spinorIn, inv_param.cpu_prec, 0, NULL );
reorder_spinor_fromQuda( (double*)spinorOut, inv_param.cpu_prec, 0, NULL );
}
