void gpuNUFFT::GpuNUFFTOperator::initDeviceMemory(int n_coils)
{
if (gpuMemAllocated)
return;
gi_host = initAndCopyGpuNUFFTInfo();//
int data_count = (int)this->kSpaceTraj.count();
IndType imdata_count = this->imgDims.count();
int sector_count = (int)this->gridSectorDims.count();
if (DEBUG)
printf("allocate and copy data indices of size %d...\n",dataIndices.count());
allocateAndCopyToDeviceMem<IndType>(&data_indices_d,dataIndices.data,dataIndices.count());
if (DEBUG)
printf("allocate and copy data of size %d...\n",data_count);
allocateDeviceMem<DType2>(&data_sorted_d,data_count);
if (DEBUG)
printf("allocate and copy gdata of size %d...\n",gi_host->grid_width_dim);
allocateDeviceMem<CufftType>(&gdata_d,gi_host->grid_width_dim);
if (DEBUG)
printf("allocate and copy coords of size %d...\n",getImageDimensionCount()*data_count);
allocateAndCopyToDeviceMem<DType>(&crds_d,this->kSpaceTraj.data,getImageDimensionCount()*data_count);
if (DEBUG)
printf("allocate and copy kernel in const memory of size %d...\n",this->kernel.count());
initLookupTable();
//allocateAndCopyToDeviceMem<DType>(&kernel_d,kernel,kernel_count);
if (DEBUG)
printf("allocate and copy sectors of size %d...\n",sector_count+1);
allocateAndCopyToDeviceMem<IndType>(§ors_d,this->sectorDataCount.data,sector_count+1);
if (DEBUG)
printf("allocate and copy sector_centers of size %d...\n",getImageDimensionCount()*sector_count);
allocateAndCopyToDeviceMem<IndType>(§or_centers_d,(IndType*)this->getSectorCentersData(),getImageDimensionCount()*sector_count);
if (this->applyDensComp())
{
if (DEBUG)
printf("allocate and copy density compensation of size %d...\n",data_count);
allocateAndCopyToDeviceMem<DType>(&density_comp_d,this->dens.data,data_count);
}
if (this->applySensData())
{
if (DEBUG)
printf("allocate sens data of size %d...\n",imdata_count);
allocateDeviceMem<DType2>(&sens_d,imdata_count);
}
if (n_coils > 1)
{
if (DEBUG)
printf("allocate precompute deapofunction of size %d...\n",imdata_count);
allocateDeviceMem<DType>(&deapo_d,imdata_count);
precomputeDeapodization(deapo_d,gi_host);
}
if (DEBUG)
printf("sector pad width: %d\n",gi_host->sector_pad_width);
//Inverse fft plan and execution
if (DEBUG)
printf("creating cufft plan with %d,%d,%d dimensions\n",DEFAULT_VALUE(gi_host->gridDims.z),gi_host->gridDims.y,gi_host->gridDims.x);
cufftResult res = cufftPlan3d(&fft_plan, (int)DEFAULT_VALUE(gi_host->gridDims.z),(int)gi_host->gridDims.y,(int)gi_host->gridDims.x, CufftTransformType) ;
if (res != CUFFT_SUCCESS)
fprintf(stderr,"error on CUFFT Plan creation!!! %d\n",res);
gpuMemAllocated = true;
}
gpuNUFFT::GpuNUFFTInfo* gpuNUFFT::GpuNUFFTOperator::initGpuNUFFTInfo()
{
gpuNUFFT::GpuNUFFTInfo* gi_host = (gpuNUFFT::GpuNUFFTInfo*)malloc(sizeof(gpuNUFFT::GpuNUFFTInfo));
gi_host->data_count = (int)this->kSpaceTraj.count();
gi_host->sector_count = (int)this->gridSectorDims.count();
gi_host->sector_width = (int)sectorDims.width;
gi_host->kernel_width = (int)this->kernelWidth;
gi_host->kernel_widthSquared = (int)(this->kernelWidth * this->kernelWidth);
gi_host->kernel_count = (int)this->kernel.count();
gi_host->grid_width_dim = (int)this->getGridDims().count();
gi_host->grid_width_offset= (int)(floor(this->getGridDims().width / (DType)2.0));
gi_host->im_width_dim = (int)imgDims.count();
gi_host->im_width_offset.x = (int)(floor(imgDims.width / (DType)2.0));
gi_host->im_width_offset.y = (int)(floor(imgDims.height / (DType)2.0));
gi_host->im_width_offset.z = (int)(floor(imgDims.depth / (DType)2.0));
gi_host->imgDims.x = imgDims.width;
gi_host->imgDims.y = imgDims.height;
gi_host->imgDims.z = imgDims.depth;
gi_host->imgDims_count = imgDims.width*imgDims.height*DEFAULT_VALUE(imgDims.depth);//TODO check why not imgDims.count()
gi_host->gridDims.x = this->getGridDims().width;
gi_host->gridDims.y = this->getGridDims().height;
gi_host->gridDims.z = this->getGridDims().depth;
gi_host->gridDims_count = this->getGridDims().width*this->getGridDims().height*DEFAULT_VALUE(this->getGridDims().depth);//s.a.
//The largest value of the grid dimensions determines the kernel radius (resolution) in k-space units
int max_grid_dim = MAX(MAX(this->getGridDims().width,this->getGridDims().height),this->getGridDims().depth);
double kernel_radius = static_cast<double>(this->kernelWidth) / 2.0;
double radius = kernel_radius / static_cast<double>(max_grid_dim);
DType kernel_width_inv = (DType)1.0 / static_cast<DType>(this->kernelWidth);
double radiusSquared = radius * radius;
double kernelRadius_invSqr = 1.0 / radiusSquared;
DType dist_multiplier = (DType)((this->kernel.count() - 1) * kernelRadius_invSqr);
if (DEBUG)
printf("radius rel. to grid width %f\n",radius);
gpuNUFFT::Dimensions sectorPadDims = sectorDims + 2*(int)(floor(this->kernelWidth / (DType)2.0));
int sector_pad_width = (int)sectorPadDims.width;
int sector_dim = (int)sectorPadDims.count();
int sector_offset = (int)(floor(sector_pad_width / (DType)2.0));
gi_host->grid_width_inv.x = (DType)1.0 / static_cast<DType>(this->getGridDims().width);
gi_host->grid_width_inv.y = (DType)1.0 / static_cast<DType>(this->getGridDims().height);
gi_host->grid_width_inv.z = (DType)1.0 / static_cast<DType>(this->getGridDims().depth);
gi_host->kernel_widthInvSquared = kernel_width_inv * kernel_width_inv;
gi_host->osr = this->osf;
gi_host->kernel_radius = (DType)kernel_radius;
gi_host->sector_pad_width = sector_pad_width;
gi_host->sector_pad_max = sector_pad_width - 1;
gi_host->sector_dim = sector_dim;
gi_host->sector_offset = sector_offset;
gi_host->aniso_x_scale = ((DType)this->getGridDims().width/(DType)max_grid_dim);
gi_host->aniso_y_scale = ((DType)this->getGridDims().height/(DType)max_grid_dim);
gi_host->aniso_z_scale = ((DType)this->getGridDims().depth/(DType)max_grid_dim);
gi_host->radiusSquared = (DType)radiusSquared;
gi_host->radiusSquared_inv = (DType)kernelRadius_invSqr;
gi_host->dist_multiplier = dist_multiplier;
gi_host->is2Dprocessing = this->is2DProcessing();
return gi_host;
}
/** \brief Compute total count of array length
*
*/
IndType count()
{
return DEFAULT_VALUE(length) * DEFAULT_VALUE(width) *
DEFAULT_VALUE(height) * DEFAULT_VALUE(depth) *
DEFAULT_VALUE(channels) * DEFAULT_VALUE(frames);
}
