Can anyone write CUDA code for the following MATLAB code?

Question

0 个投票

function dR = FDK(x,y,z,u_off,v_off,du,dv,Beta,P,A,SDD,SAD,varargin);
%where
 % x = 1x512 double
 % y = 1x512 double
 % z = 1x512 double
 % u_off = 1x1 double
 % v_off = 1x1 double
 % du = 1x1 double
 % dv = 1x1 double
 % beta = 1x1 double
 % P = 512x512 double
 % A = 3x4 double
 % SDD = 1x1 double
 % SAD = 1x1 double
 % varargin = 1x1 cell
 % varargin{1} = 1x512 double

Outputs:

After fixing the projection angle, the routine returns the

value calculated by FDK equation for each voxel.

Description:

Reconstruct single Cone-Beam projection using 3D Feldkamp

 %#codegen
% Dimensions of reconstruction grid
nx = length(x);
ny = length(y);
nz = length(z);
%Remark:There may be some savings in fft computation if P is 
%transposed
[nv,nu] = size(P);
u = (-u_off + [0:nu-1]')*du;
v = (-v_off + [0:nv-1]')*dv;
%Remark: if P is transposed,'meshgrid'should be changed to 'ndgrid'
[uu,vv] = meshgrid(u,v);
weight = SDD./sqrt(SDD^2 + vv.^2 + uu.^2);
% Remark: P is over-written to conserve memory in following   
% The interpretation at each stage should be clear from context
P = P .* weight;
% Filtering:
Phihat = varargin{1};
%if ~isempty(Phihat), % Skip to backprojection is filter is empty
%Remark:This call to fft can likely be optimised usingthe transpose
%of P instead to capitalise on Matlab's column-oriented storage.
   P = fft( P, length(Phihat), 2 ); % Rows zero-padded automatically
      % Remark: This loop can likely be vectorised using repmat
      for j=1:nv
          P(j,:) = P(j,:) .* Phihat ;
      end;
%Remark:This call to fft can likely be optimised usingthe transpose
% of P instead to capitalise on Matlab's column-oriented storage.
      P = real( ifft( P, [], 2 ) );
      P = P(:,1:nu); % Trim zero-padding on rows
%end % of if-block
%Increment to add to reconstruction in backprojection stage
dR = zeros(ny,nx,nz);
% Vectorised computation of backprojection
[yy, xx, zz] = meshgrid( y, x, z);
% Use projection matrix to project reconstruction (x,y,z) grid
% into detector (u,v) grid
UV = A * [ xx(:)'; yy(:)'; zz(:)'; ones(1,nx*ny*nz) ];
%Nearest neighbour interpolation to find detector coordinates (u,v)
% that are closest to projections of voxels (x,y,z)
U = round( UV(1,:)./UV(3,:) ) + 1 ;
V = round( UV(2,:)./UV(3,:) ) + 1 ;
%Identify indices of voxels with projections strictly within 
%detector grid
ind = find( (U>=1) & (V>=1) & (U<=nu) & (V<=nv) );
% Remark: This will have to be changed if P is transposed
P_ind = ( U(ind) - 1 )*nv + V(ind) ;
co = cos(Beta*pi/180);
si = sin(Beta*pi/180);
%Grid-dependent weighting factors (only computed for voxels needed)
W = ( SAD ./ ( SAD - co*xx(ind) - si*yy(ind) ) ).^2;
dR( ind ) = W .* P( P_ind );
return