int gpuCgSpiritGadget::process(GadgetContainerMessage<ISMRMRD::ImageHeader> *m1, GadgetContainerMessage<GenericReconJob> *m2)
{
// Is this data for this gadget's set/slice?
//
if( m1->getObjectPtr()->set != set_number_ || m1->getObjectPtr()->slice != slice_number_ ) {
// No, pass it downstream...
return this->next()->putq(m1);
}
//GDEBUG("gpuCgSpiritGadget::process\n");
boost::shared_ptr<GPUTimer> process_timer;
if( output_timing_ )
process_timer = boost::shared_ptr<GPUTimer>( new GPUTimer("gpuCgSpiritGadget::process()") );
if (!is_configured_) {
GDEBUG("Data received before configuration was completed\n");
return GADGET_FAIL;
}
GenericReconJob* j = m2->getObjectPtr();
// Some basic validation of the incoming Spirit job
if (!j->csm_host_.get() || !j->dat_host_.get() || !j->tra_host_.get() || !j->dcw_host_.get() || !j->reg_host_.get()) {
GDEBUG("Received an incomplete Spirit job\n");
return GADGET_FAIL;
}
unsigned int samples = j->dat_host_->get_size(0);
unsigned int channels = j->dat_host_->get_size(1);
unsigned int rotations = samples / j->tra_host_->get_number_of_elements();
unsigned int frames = j->tra_host_->get_size(1)*rotations;
if( samples%j->tra_host_->get_number_of_elements() ) {
GDEBUG("Mismatch between number of samples (%d) and number of k-space coordinates (%d).\nThe first should be a multiplum of the latter.\n",
samples, j->tra_host_->get_number_of_elements());
return GADGET_FAIL;
}
boost::shared_ptr< cuNDArray<floatd2> > traj(new cuNDArray<floatd2> (j->tra_host_.get()));
boost::shared_ptr< cuNDArray<float> > dcw(new cuNDArray<float> (j->dcw_host_.get()));
sqrt_inplace(dcw.get()); //Take square root to use for weighting
boost::shared_ptr< cuNDArray<float_complext> > csm(new cuNDArray<float_complext> (j->csm_host_.get()));
boost::shared_ptr< cuNDArray<float_complext> > device_samples(new cuNDArray<float_complext> (j->dat_host_.get()));
cudaDeviceProp deviceProp;
if( cudaGetDeviceProperties( &deviceProp, device_number_ ) != cudaSuccess) {
GDEBUG( "Error: unable to query device properties.\n" );
return GADGET_FAIL;
}
unsigned int warp_size = deviceProp.warpSize;
matrix_size_ = uint64d2( j->reg_host_->get_size(0), j->reg_host_->get_size(1) );
matrix_size_os_ =
uint64d2(((static_cast<unsigned int>(std::ceil(matrix_size_[0]*oversampling_factor_))+warp_size-1)/warp_size)*warp_size,
((static_cast<unsigned int>(std::ceil(matrix_size_[1]*oversampling_factor_))+warp_size-1)/warp_size)*warp_size);
if( !matrix_size_reported_ ) {
GDEBUG("Matrix size : [%d,%d] \n", matrix_size_[0], matrix_size_[1]);
GDEBUG("Matrix size OS : [%d,%d] \n", matrix_size_os_[0], matrix_size_os_[1]);
matrix_size_reported_ = true;
}
std::vector<size_t> image_dims = to_std_vector(matrix_size_);
image_dims.push_back(frames);
image_dims.push_back(channels);
GDEBUG("Number of coils: %d %d \n",channels,image_dims.size());
E_->set_domain_dimensions(&image_dims);
E_->set_codomain_dimensions(device_samples->get_dimensions().get());
E_->set_dcw(dcw);
E_->setup( matrix_size_, matrix_size_os_, static_cast<float>(kernel_width_) );
E_->preprocess(traj.get());
boost::shared_ptr< cuNDArray<float_complext> > csm_device( new cuNDArray<float_complext>( csm.get() ));
S_->set_calibration_kernels(csm_device);
S_->set_domain_dimensions(&image_dims);
S_->set_codomain_dimensions(&image_dims);
/*
boost::shared_ptr< cuNDArray<float_complext> > reg_image(new cuNDArray<float_complext> (j->reg_host_.get()));
R_->compute(reg_image.get());
// Define preconditioning weights
boost::shared_ptr< cuNDArray<float> > _precon_weights = sum(abs_square(csm.get()).get(), 2);
boost::shared_ptr<cuNDArray<float> > R_diag = R_->get();
*R_diag *= float(kappa_);
*_precon_weights += *R_diag;
R_diag.reset();
reciprocal_sqrt_inplace(_precon_weights.get());
boost::shared_ptr< cuNDArray<float_complext> > precon_weights = real_to_complex<float_complext>( _precon_weights.get() );
_precon_weights.reset();
D_->set_weights( precon_weights );
*/
/*{
static int counter = 0;
char filename[256];
sprintf((char*)filename, "_traj_%d.real", counter);
write_nd_array<floatd2>( traj->to_host().get(), filename );
sprintf((char*)filename, "_dcw_%d.real", counter);
write_nd_array<float>( dcw->to_host().get(), filename );
sprintf((char*)filename, "_csm_%d.cplx", counter);
write_nd_array<float_complext>( csm->to_host().get(), filename );
sprintf((char*)filename, "_samples_%d.cplx", counter);
write_nd_array<float_complext>( device_samples->to_host().get(), filename );
sprintf((char*)filename, "_reg_%d.cplx", counter);
write_nd_array<float_complext>( reg_image->to_host().get(), filename );
counter++;
}*/
// Invoke solver
//
boost::shared_ptr< cuNDArray<float_complext> > cgresult;
{
boost::shared_ptr<GPUTimer> solve_timer;
if( output_timing_ )
solve_timer = boost::shared_ptr<GPUTimer>( new GPUTimer("gpuCgSpiritGadget::solve()") );
cgresult = cg_.solve(device_samples.get());
if( output_timing_ )
solve_timer.reset();
}
if (!cgresult.get()) {
GDEBUG("Iterative_spirit_compute failed\n");
return GADGET_FAIL;
}
/*
static int counter = 0;
char filename[256];
sprintf((char*)filename, "recon_%d.real", counter);
write_nd_array<float>( abs(cgresult.get())->to_host().get(), filename );
counter++;
*/
// If the recon matrix size exceeds the sequence matrix size then crop
if( matrix_size_seq_ != matrix_size_ )
cgresult = crop<float_complext,2>( (matrix_size_-matrix_size_seq_)>>1, matrix_size_seq_, cgresult.get() );
// Combine coil images
//
cgresult = real_to_complex<float_complext>(sqrt(sum(abs_square(cgresult.get()).get(), 3).get()).get()); // RSS
//cgresult = sum(cgresult.get(), 2);
// Pass on the reconstructed images
//
put_frames_on_que(frames,rotations,j,cgresult.get());
frame_counter_ += frames;
if( output_timing_ )
process_timer.reset();
m1->release();
return GADGET_OK;
}
int NFFT2DGadget::process(GadgetContainerMessage< ISMRMRD::AcquisitionHeader > *m1, // header
GadgetContainerMessage< hoNDArray< std::complex<float> > > *m2, // data
GadgetContainerMessage< hoNDArray<float> > *m3 ) // traj/dcw
{
// Throw away any noise samples if they have been allowed to pass this far down the chain...
//
bool is_noise = m1->getObjectPtr()->isFlagSet(ISMRMRD::ISMRMRD_ACQ_IS_NOISE_MEASUREMENT);
if (is_noise) {
m1->release();
return GADGET_OK;
}
// First pass initialization
//
if (frame_readout_queue_->message_count() == 0 ) {
samples_per_readout_ = m1->getObjectPtr()->number_of_samples;
num_coils_ = m1->getObjectPtr()->active_channels;
dimensions_.push_back(m1->getObjectPtr()->active_channels);
dimensions_.push_back(repetitions_);
num_trajectory_dims_ = m3->getObjectPtr()->get_size(0); // 2 for trajectories only, 3 for both trajectories + dcw
}
int samples = m1->getObjectPtr()->number_of_samples;
int readout = m1->getObjectPtr()->idx.kspace_encode_step_1;
int repetition = m1->getObjectPtr()->idx.kspace_encode_step_2;
// Enqueue incoming readouts and trajectories
//
frame_readout_queue_->enqueue_tail(duplicate_array(m2));
frame_traj_queue_->enqueue_tail(duplicate_array(m3));
// If the last readout for a slice has arrived then perform a reconstruction
//
bool is_last_scan_in_repetition =
m1->getObjectPtr()->isFlagSet(ISMRMRD::ISMRMRD_ACQ_LAST_IN_REPETITION);
if (is_last_scan_in_repetition) {
// Define the image header
//
GadgetContainerMessage<ISMRMRD::ImageHeader> *cm1 =
new GadgetContainerMessage<ISMRMRD::ImageHeader>();
GadgetContainerMessage< hoNDArray< std::complex<float> > > *cm2 =
new GadgetContainerMessage<hoNDArray< std::complex<float> > >();
cm1->getObjectPtr()->flags = 0;
cm1->cont(cm2);
cm1->getObjectPtr()->matrix_size[0] = dimensions_[0];
cm1->getObjectPtr()->matrix_size[1] = dimensions_[1];
cm1->getObjectPtr()->matrix_size[2] = 1;
cm1->getObjectPtr()->field_of_view[0] = field_of_view_[0];
cm1->getObjectPtr()->field_of_view[1] = field_of_view_[1];
cm1->getObjectPtr()->channels = num_coils_;
cm1->getObjectPtr()->repetition = m1->getObjectPtr()->idx.repetition;
memcpy(cm1->getObjectPtr()->position,
m1->getObjectPtr()->position,
sizeof(float)*3);
memcpy(cm1->getObjectPtr()->read_dir,
m1->getObjectPtr()->read_dir,
sizeof(float)*3);
memcpy(cm1->getObjectPtr()->phase_dir,
m1->getObjectPtr()->phase_dir,
sizeof(float)*3);
memcpy(cm1->getObjectPtr()->slice_dir,
m1->getObjectPtr()->slice_dir,
sizeof(float)*3);
memcpy(cm1->getObjectPtr()->patient_table_position,
m1->getObjectPtr()->patient_table_position, sizeof(float)*3);
cm1->getObjectPtr()->data_type = ISMRMRD::ISMRMRD_CXFLOAT;
cm1->getObjectPtr()->image_index = 0;
cm1->getObjectPtr()->image_series_index = 0;
//
// Perform reconstruction of repetition
//
// Get samples for frame
//
cuNDArray<float_complext> samples( extract_samples_from_queue( frame_readout_queue_.get()).get() );
// Get trajectories/dcw for frame
//
boost::shared_ptr< cuNDArray<floatd2> > traj(new cuNDArray<floatd2>);
boost::shared_ptr<cuNDArray<float> > dcw(new cuNDArray<float>);
extract_trajectory_and_dcw_from_queue( frame_traj_queue_.get(), traj.get(), dcw.get() );
//traj = compute_radial_trajectory_golden_ratio_2d<float>(samples_per_readout_,dimensions_[1],1,0,GR_ORIGINAL);
unsigned int num_profiles = samples.get_number_of_elements()/samples_per_readout_;
dcw = compute_radial_dcw_golden_ratio_2d<float>(samples_per_readout_,num_profiles,1.0,1.0f/samples_per_readout_/num_profiles,0,GR_ORIGINAL);
// Create output array
//
std::vector<size_t> img_dims(2);
img_dims[0] = dimensions_[0];
img_dims[1] = dimensions_[1];
cm2->getObjectPtr()->create(&img_dims);
cuNDArray<float_complext> image(&img_dims);
// Initialize plan
//
const float kernel_width = 5.5f;
cuNFFT_plan<float,2> plan( from_std_vector<size_t,2>(img_dims), from_std_vector<size_t,2>(img_dims)<<1, kernel_width );
plan.preprocess( traj.get(), cuNFFT_plan<float,2>::NFFT_PREP_NC2C );
/*
if( dcw->get_number_of_elements() == 0 ){
std::vector<size_t> dcw_dims; dcw_dims.push_back(samples_per_readout_);
hoNDArray<float> host_dcw( dcw_dims );
for( int i=0; i<(int)dcw_dims[0]; i++ )
host_dcw.get_data_ptr()[i]=abs(i-(int)dcw_dims[0]/2);
host_dcw.get_data_ptr()[dcw_dims[0]/2] = 0.25f; // ad hoc value (we do not want a DC component of 0)
dcw = expand(&host_dcw, traj->get_number_of_elements()/samples_per_readout_);
}
*/
// Gridder
//
plan.compute( &samples, &image,
(dcw->get_number_of_elements()>0) ? dcw.get() : 0x0,
cuNFFT_plan<float,2>::NFFT_BACKWARDS_NC2C );
// Download to host
//
image.to_host( (hoNDArray<float_complext>*)cm2->getObjectPtr() );
// Pass on data down the gadget chain
//
if (this->next()->putq(cm1) < 0) {
return GADGET_FAIL;
}
}
m1->release();
return GADGET_OK;
}
int gpuOsSenseGadget::process(GadgetContainerMessage<ISMRMRD::ImageHeader> *m1, GadgetContainerMessage<GenericReconJob> *m2)
{
// Is this data for this gadget's set/slice?
//
GDEBUG("Starting gpuOsSenseGadget\n");
if( m1->getObjectPtr()->set != set_number_ || m1->getObjectPtr()->slice != slice_number_ ) {
// No, pass it downstream...
return this->next()->putq(m1);
}
//GDEBUG("gpuOsSenseGadget::process\n");
//GPUTimer timer("gpuOsSenseGadget::process");
if (!is_configured_) {
GDEBUG("\nData received before configuration complete\n");
return GADGET_FAIL;
}
GenericReconJob* j = m2->getObjectPtr();
// Let's first check that this job has the required data...
if (!j->csm_host_.get() || !j->dat_host_.get() || !j->tra_host_.get() || !j->dcw_host_.get()) {
GDEBUG("Received an incomplete Sense job\n");
return GADGET_FAIL;
}
unsigned int samples = j->dat_host_->get_size(0);
unsigned int channels = j->dat_host_->get_size(1);
unsigned int rotations = samples / j->tra_host_->get_number_of_elements();
unsigned int frames = j->tra_host_->get_size(1)*rotations;
if( samples%j->tra_host_->get_number_of_elements() ) {
GDEBUG("Mismatch between number of samples (%d) and number of k-space coordinates (%d).\nThe first should be a multiplum of the latter.\n",
samples, j->tra_host_->get_number_of_elements());
return GADGET_FAIL;
}
boost::shared_ptr< cuNDArray<floatd2> > traj(new cuNDArray<floatd2> (j->tra_host_.get()));
boost::shared_ptr< cuNDArray<float> > dcw(new cuNDArray<float> (j->dcw_host_.get()));
sqrt_inplace(dcw.get());
boost::shared_ptr< cuNDArray<float_complext> > csm(new cuNDArray<float_complext> (j->csm_host_.get()));
boost::shared_ptr< cuNDArray<float_complext> > device_samples(new cuNDArray<float_complext> (j->dat_host_.get()));
// Take the reconstruction matrix size from the regulariaztion image.
// It could be oversampled from the sequence specified size...
matrix_size_ = uint64d2( j->reg_host_->get_size(0), j->reg_host_->get_size(1) );
cudaDeviceProp deviceProp;
if( cudaGetDeviceProperties( &deviceProp, device_number_ ) != cudaSuccess) {
GDEBUG( "\nError: unable to query device properties.\n" );
return GADGET_FAIL;
}
unsigned int warp_size = deviceProp.warpSize;
matrix_size_os_ =
uint64d2(((static_cast<unsigned int>(std::ceil(matrix_size_[0]*oversampling_factor_))+warp_size-1)/warp_size)*warp_size,
((static_cast<unsigned int>(std::ceil(matrix_size_[1]*oversampling_factor_))+warp_size-1)/warp_size)*warp_size);
GDEBUG("Matrix size : [%d,%d] \n", matrix_size_[0], matrix_size_[1]);
GDEBUG("Matrix size OS : [%d,%d] \n", matrix_size_os_[0], matrix_size_os_[1]);
std::vector<size_t> image_dims = to_std_vector(matrix_size_);
image_dims.push_back(frames);
E_->set_domain_dimensions(&image_dims);
E_->set_codomain_dimensions(device_samples->get_dimensions().get());
E_->set_csm(csm);
E_->setup( matrix_size_, matrix_size_os_, kernel_width_ );
E_->preprocess(traj.get());
{
auto precon = boost::make_shared<cuNDArray<float_complext>>(image_dims);
fill(precon.get(),float_complext(1.0f));
//solver_.set_preconditioning_image(precon);
}
reg_image_ = boost::shared_ptr< cuNDArray<float_complext> >(new cuNDArray<float_complext>(&image_dims));
// These operators need their domain/codomain set before being added to the solver
//
//E_->set_dcw(dcw);
GDEBUG("Prepared\n");
// Expand the average image to the number of frames
//
{
cuNDArray<float_complext> tmp(*j->reg_host_);
*reg_image_ = expand( tmp, frames );
}
PICS_->set_prior(reg_image_);
// Define preconditioning weights
//
//Apply weights
//*device_samples *= *dcw;
// Invoke solver
//
boost::shared_ptr< cuNDArray<float_complext> > result;
{
GDEBUG("Running NLCG solver\n");
GPUTimer timer("Running NLCG solver");
// Optionally, allow exclusive (per device) access to the solver
// This may not matter much in terms of speed, but it can in terms of memory consumption
//
if( exclusive_access_ )
_mutex[device_number_].lock();
result = solver_.solve(device_samples.get());
if( exclusive_access_ )
_mutex[device_number_].unlock();
}
// Provide some info about the scaling between the regularization and reconstruction.
// If it is not close to one, PICCS does not work optimally...
//
if( alpha_ > 0.0 ){
cuNDArray<float_complext> gpureg(j->reg_host_.get());
boost::shared_ptr< cuNDArray<float_complext> > gpurec = sum(result.get(),2);
*gpurec /= float(result->get_size(2));
float scale = abs(dot(gpurec.get(), gpurec.get())/dot(gpurec.get(),&gpureg));
GDEBUG("Scaling factor between regularization and reconstruction is %f.\n", scale);
}
if (!result.get()) {
GDEBUG("\nNon-linear conjugate gradient solver failed\n");
return GADGET_FAIL;
}
/*
static int counter = 0;
char filename[256];
sprintf((char*)filename, "recon_sb_%d.cplx", counter);
write_nd_array<float_complext>( sbresult->to_host().get(), filename );
counter++; */
// If the recon matrix size exceeds the sequence matrix size then crop
if( matrix_size_seq_ != matrix_size_ )
*result = crop<float_complext,2>( (matrix_size_-matrix_size_seq_)>>1, matrix_size_seq_, *result );
// Now pass on the reconstructed images
//
this->put_frames_on_que(frames,rotations,j,result.get(),channels);
frame_counter_ += frames;
m1->release();
return GADGET_OK;
}