From 62ab6cd46c3f1c189328c8d41899db7444c7ac29 Mon Sep 17 00:00:00 2001 From: Daniil Kazantsev Date: Mon, 11 Sep 2017 09:36:13 +0100 Subject: 2 new GPU regularizers, FGP objective fixed, FISTA_REC updated --- demos/DemoRD2.m | 62 +++-- main_func/FISTA_REC.m | 138 ++++++++--- main_func/regularizers_CPU/FGP_TV.c | 47 +--- main_func/regularizers_CPU/FGP_TV_core.c | 49 ++++ main_func/regularizers_CPU/FGP_TV_core.h | 2 + .../Diffus_HO/Diff4thHajiaboli_GPU.cpp | 114 +++++++++ .../Diffus_HO/Diff4th_GPU_kernel.cu | 270 +++++++++++++++++++++ .../Diffus_HO/Diff4th_GPU_kernel.h | 6 + main_func/regularizers_GPU/NL_Regul/NLM_GPU.cpp | 171 +++++++++++++ .../regularizers_GPU/NL_Regul/NLM_GPU_kernel.cu | 239 ++++++++++++++++++ .../regularizers_GPU/NL_Regul/NLM_GPU_kernel.h | 6 + 11 files changed, 1016 insertions(+), 88 deletions(-) create mode 100644 main_func/regularizers_GPU/Diffus_HO/Diff4thHajiaboli_GPU.cpp create mode 100644 main_func/regularizers_GPU/Diffus_HO/Diff4th_GPU_kernel.cu create mode 100644 main_func/regularizers_GPU/Diffus_HO/Diff4th_GPU_kernel.h create mode 100644 main_func/regularizers_GPU/NL_Regul/NLM_GPU.cpp create mode 100644 main_func/regularizers_GPU/NL_Regul/NLM_GPU_kernel.cu create mode 100644 main_func/regularizers_GPU/NL_Regul/NLM_GPU_kernel.h diff --git a/demos/DemoRD2.m b/demos/DemoRD2.m index f177e26..717a55d 100644 --- a/demos/DemoRD2.m +++ b/demos/DemoRD2.m @@ -6,7 +6,7 @@ close all %% % % adding paths addpath('../data/'); -addpath('../main_func/'); addpath('../main_func/regularizers_CPU/'); +addpath('../main_func/'); addpath('../main_func/regularizers_CPU/'); addpath('../main_func/regularizers_GPU/NL_Regul/'); addpath('../main_func/regularizers_GPU/Diffus_HO/'); addpath('../supp/'); load('DendrRawData.mat') % load raw data of 3D dendritic set @@ -30,7 +30,7 @@ Weights3D = single(data_raw3D); % weights for PW model clear data_raw3D %% % set projection/reconstruction geometry here -Z_slices = 1; +Z_slices = 5; det_row_count = Z_slices; proj_geom = astra_create_proj_geom('parallel3d', 1, 1, det_row_count, size_det, angles_rad); vol_geom = astra_create_vol_geom(recon_size,recon_size,Z_slices); @@ -107,25 +107,46 @@ tic; [X_fista_GH_TVLLT, outputGH_TVLLT] = FISTA_REC(params); toc; figure; imshow(X_fista_GH_TVLLT(:,:,params.slice) , [0, 2.5]); title ('FISTA-OS-GH-TV-LLT reconstruction'); %% -% fprintf('%s\n', 'Reconstruction using FISTA-OS-PB...'); -% % very-slow on CPU -% clear params -% params.proj_geom = proj_geom; % pass geometry to the function -% params.vol_geom = vol_geom; -% params.sino = Sino3D(:,:,1); -% params.iterFISTA = 12; -% params.Regul_LambdaPatchBased = 1; % PB regularization parameter -% params.Regul_PB_h = 0.1; % threhsold parameter -% params.Ring_LambdaR_L1 = 0.002; % Soft-Thresh L1 ring variable parameter -% params.Ring_Alpha = 21; % to boost ring removal procedure -% params.weights = Weights3D(:,:,1); -% params.subsets = 16; % the number of ordered subsets -% params.show = 1; -% params.maxvalplot = 2.5; params.slice = 1; -% -% tic; [X_fista_GH_PB, outputPB] = FISTA_REC(params); toc; -% figure; imshow(X_fista_GH_PB(:,:,params.slice) , [0, 2.5]); title ('FISTA-OS-PB reconstruction'); +fprintf('%s\n', 'Reconstruction using FISTA-OS-GH-HigherOrderDiffusion...'); +% !GPU version! +clear params +params.proj_geom = proj_geom; % pass geometry to the function +params.vol_geom = vol_geom; +params.sino = Sino3D(:,:,1:5); +params.iterFISTA = 25; +params.Regul_LambdaDiffHO = 2; % DiffHO regularization parameter +params.Regul_DiffHO_EdgePar = 0.05; % threshold parameter +params.Regul_Iterations = 150; +params.Ring_LambdaR_L1 = 0.002; % Soft-Thresh L1 ring variable parameter +params.Ring_Alpha = 21; % to boost ring removal procedure +params.weights = Weights3D(:,:,1:5); +params.subsets = 16; % the number of ordered subsets +params.show = 1; +params.maxvalplot = 2.5; params.slice = 1; + +tic; [X_fista_GH_HO, outputHO] = FISTA_REC(params); toc; +figure; imshow(X_fista_GH_HO(:,:,params.slice) , [0, 2.5]); title ('FISTA-OS-HigherOrderDiffusion reconstruction'); +%% +fprintf('%s\n', 'Reconstruction using FISTA-PB...'); +% !GPU version! +clear params +params.proj_geom = proj_geom; % pass geometry to the function +params.vol_geom = vol_geom; +params.sino = Sino3D(:,:,1); +params.iterFISTA = 25; +params.Regul_LambdaPatchBased_GPU = 3; % PB regularization parameter +params.Regul_PB_h = 0.04; % threhsold parameter +params.Regul_PB_SearchW = 3; +params.Regul_PB_SimilW = 1; +params.Ring_LambdaR_L1 = 0.002; % Soft-Thresh L1 ring variable parameter +params.Ring_Alpha = 21; % to boost ring removal procedure +params.weights = Weights3D(:,:,1); +params.show = 1; +params.maxvalplot = 2.5; params.slice = 1; + +tic; [X_fista_GH_PB, outputPB] = FISTA_REC(params); toc; +figure; imshow(X_fista_GH_PB(:,:,params.slice) , [0, 2.5]); title ('FISTA-OS-PB reconstruction'); %% fprintf('%s\n', 'Reconstruction using FISTA-OS-GH-TGV...'); % still testing... @@ -146,6 +167,7 @@ params.maxvalplot = 2.5; params.slice = 1; tic; [X_fista_GH_TGV, outputTGV] = FISTA_REC(params); toc; figure; imshow(X_fista_GH_TGV(:,:,params.slice) , [0, 2.5]); title ('FISTA-OS-GH-TGV reconstruction'); + %% % fprintf('%s\n', 'Reconstruction using FISTA-Student-TV...'); % clear params diff --git a/main_func/FISTA_REC.m b/main_func/FISTA_REC.m index 00e59ab..edccbe1 100644 --- a/main_func/FISTA_REC.m +++ b/main_func/FISTA_REC.m @@ -1,6 +1,15 @@ function [X, output] = FISTA_REC(params) % <<<< FISTA-based reconstruction routine using ASTRA-toolbox >>>> +% This code solves regularised PWLS problem using FISTA approach. +% The code contains multiple regularisation penalties as well as it can be +% accelerated by using ordered-subset version. Various projection +% geometries supported. + +% DISCLAIMER +% It is recommended to use ASTRA version 1.8 or later in order to avoid +% crashing due to GPU memory overflow for big datasets + % ___Input___: % params.[] file: %----------------General Parameters------------------------ @@ -18,11 +27,17 @@ function [X, output] = FISTA_REC(params) % 2 .Regul_Lambda_SBTV (SplitBregman-TV regularization parameter) % 3 .Regul_LambdaLLT (Higher order LLT regularization parameter) % 3.1 .Regul_tauLLT (time step parameter for LLT (HO) term) -% 4 .Regul_LambdaPatchBased (Patch-based nonlocal regularization parameter) +% 4 .Regul_LambdaPatchBased_CPU (Patch-based nonlocal regularization parameter) % 4.1 .Regul_PB_SearchW (ratio of the searching window (e.g. 3 = (2*3+1) = 7 pixels window)) % 4.2 .Regul_PB_SimilW (ratio of the similarity window (e.g. 1 = (2*1+1) = 3 pixels window)) % 4.3 .Regul_PB_h (PB penalty function threshold) -% 5 .Regul_LambdaTGV (Total Generalized variation regularization parameter) +% 5 .Regul_LambdaPatchBased_GPU (Patch-based nonlocal regularization parameter) +% 5.1 .Regul_PB_SearchW (ratio of the searching window (e.g. 3 = (2*3+1) = 7 pixels window)) +% 5.2 .Regul_PB_SimilW (ratio of the similarity window (e.g. 1 = (2*1+1) = 3 pixels window)) +% 5.3 .Regul_PB_h (PB penalty function threshold) +% 6 .Regul_LambdaDiffHO (Higher-Order Diffusion regularization parameter) +% 6.1 .Regul_DiffHO_EdgePar (edge-preserving noise related parameter) +% 7 .Regul_LambdaTGV (Total Generalized variation regularization parameter) % - .Regul_tol (tolerance to terminate regul iterations, default 1.0e-04) % - .Regul_Iterations (iterations for the selected penalty, default 25) % - .Regul_Dimension ('2D' or '3D' way to apply regularization, '3D' is the default) @@ -173,11 +188,16 @@ if (isfield(params,'Regul_tauLLT')) else tauHO = 0.0001; end -if (isfield(params,'Regul_LambdaPatchBased')) - lambdaPB = params.Regul_LambdaPatchBased; +if (isfield(params,'Regul_LambdaPatchBased_CPU')) + lambdaPB = params.Regul_LambdaPatchBased_CPU; else lambdaPB = 0; end +if (isfield(params,'Regul_LambdaPatchBased_GPU')) + lambdaPB_GPU = params.Regul_LambdaPatchBased_GPU; +else + lambdaPB_GPU = 0; +end if (isfield(params,'Regul_PB_SearchW')) SearchW = params.Regul_PB_SearchW; else @@ -193,6 +213,16 @@ if (isfield(params,'Regul_PB_h')) else h_PB = 0.1; % default end +if (isfield(params,'Regul_LambdaDiffHO')) + LambdaDiff_HO = params.Regul_LambdaDiffHO; +else + LambdaDiff_HO = 0; +end +if (isfield(params,'Regul_DiffHO_EdgePar')) + LambdaDiff_HO_EdgePar = params.Regul_DiffHO_EdgePar; +else + LambdaDiff_HO_EdgePar = 0.01; +end if (isfield(params,'Regul_LambdaTGV')) LambdaTGV = params.Regul_LambdaTGV; else @@ -305,16 +335,16 @@ if (subsets == 0) t_old = t; r_old = r; - % if the geometry is parallel use slice-by-slice projection-backprojection routine + % if the geometry is parallel use slice-by-slice projection-backprojection routine if (strcmp(proj_geom.type,'parallel') || strcmp(proj_geom.type,'parallel3d')) - sino_updt = zeros(size(sino),'single'); - for kkk = 1:SlicesZ - [sino_id, sino_updt(:,:,kkk)] = astra_create_sino3d_cuda(X_t(:,:,kkk), proj_geomT, vol_geomT); - astra_mex_data3d('delete', sino_id); - end - else - % for divergent 3D geometry (for Matlab watch the GPU memory overflow) - [sino_id, sino_updt] = astra_create_sino3d_cuda(X_t, proj_geom, vol_geom); + sino_updt = zeros(size(sino),'single'); + for kkk = 1:SlicesZ + [sino_id, sino_updt(:,:,kkk)] = astra_create_sino3d_cuda(X_t(:,:,kkk), proj_geomT, vol_geomT); + astra_mex_data3d('delete', sino_id); + end + else + % for divergent 3D geometry (watch the GPU memory overflow in earlier ASTRA versions < 1.8) + [sino_id, sino_updt] = astra_create_sino3d_cuda(X_t, proj_geom, vol_geom); end if (lambdaR_L1 > 0) @@ -332,17 +362,17 @@ if (subsets == 0) residual = weights.*(sino_updt - sino); end - objective(i) = (0.5*norm(residual(:))^2)/(Detectors*anglesNumb*SlicesZ); % for the objective function output + objective(i) = (0.5*sum(residual(:).^2)); % for the objective function output - % if the geometry is parallel use slice-by-slice projection-backprojection routine + % if the geometry is parallel use slice-by-slice projection-backprojection routine if (strcmp(proj_geom.type,'parallel') || strcmp(proj_geom.type,'parallel3d')) - x_temp = zeros(size(X),'single'); - for kkk = 1:SlicesZ - [id, x_temp(:,:,kkk)] = astra_create_backprojection3d_cuda(squeeze(residual(:,:,kkk)), proj_geomT, vol_geomT); - astra_mex_data3d('delete', id); - end + x_temp = zeros(size(X),'single'); + for kkk = 1:SlicesZ + [id, x_temp(:,:,kkk)] = astra_create_backprojection3d_cuda(squeeze(residual(:,:,kkk)), proj_geomT, vol_geomT); + astra_mex_data3d('delete', id); + end else - [id, x_temp] = astra_create_backprojection3d_cuda(residual, proj_geom, vol_geom); + [id, x_temp] = astra_create_backprojection3d_cuda(residual, proj_geom, vol_geom); end X = X_t - (1/L_const).*x_temp; astra_mex_data3d('delete', sino_id); @@ -360,7 +390,7 @@ if (subsets == 0) % 3D regularization [X, f_val] = FGP_TV(single(X), lambdaFGP_TV, IterationsRegul, tol, 'iso'); end - objective(i) = objective(i) + f_val; + objective(i) = (objective(i) + f_val)./(Detectors*anglesNumb*SlicesZ); end if (lambdaSB_TV > 0) % Split Bregman regularization @@ -390,7 +420,7 @@ if (subsets == 0) end if (lambdaPB > 0) - % Patch-Based regularization (can be slow on CPU) + % Patch-Based regularization (can be very slow on CPU) if ((strcmp('2D', Dimension) == 1)) % 2D regularization for kkk = 1:SlicesZ @@ -400,13 +430,35 @@ if (subsets == 0) X = PatchBased_Regul(single(X), SearchW, SimilW, h_PB, lambdaPB); end end - if (LambdaTGV > 0) - % Total Generalized variation (currently only 2D) - lamTGV1 = 1.1; % smoothing trade-off parameters, see Pock's paper - lamTGV2 = 0.8; % second-order term + if (lambdaPB_GPU > 0) + % Patch-Based regularization (GPU CUDA implementation) + if ((strcmp('2D', Dimension) == 1)) + % 2D regularization for kkk = 1:SlicesZ - X(:,:,kkk) = TGV_PD(single(X(:,:,kkk)), LambdaTGV, lamTGV1, lamTGV2, IterationsRegul); + X(:,:,kkk) = NLM_GPU(single(X(:,:,kkk)), SearchW, SimilW, h_PB, lambdaPB_GPU); end + else + X = NLM_GPU(single(X), SearchW, SimilW, h_PB, lambdaPB_GPU); + end + end + if (LambdaDiff_HO > 0) + % Higher-order diffusion penalty (GPU CUDA implementation) + if ((strcmp('2D', Dimension) == 1)) + % 2D regularization + for kkk = 1:SlicesZ + X(:,:,kkk) = Diff4thHajiaboli_GPU(single(X(:,:,kkk)), LambdaDiff_HO_EdgePar, LambdaDiff_HO, IterationsRegul); + end + else + X = Diff4thHajiaboli_GPU(X, LambdaDiff_HO_EdgePar, LambdaDiff_HO, IterationsRegul); + end + end + if (LambdaTGV > 0) + % Total Generalized variation (currently only 2D) + lamTGV1 = 1.1; % smoothing trade-off parameters, see Pock's paper + lamTGV2 = 0.8; % second-order term + for kkk = 1:SlicesZ + X(:,:,kkk) = TGV_PD(single(X(:,:,kkk)), LambdaTGV, lamTGV1, lamTGV2, IterationsRegul); + end end if (lambdaR_L1 > 0) @@ -470,7 +522,7 @@ else % the ring removal part (Group-Huber fidelity) % first 2 iterations do additional work reconstructing whole dataset to ensure - % stablility + % the stablility if (i < 3) [sino_id2, sino_updt2] = astra_create_sino3d_cuda(X_t, proj_geom, vol_geom); astra_mex_data3d('delete', sino_id2); @@ -546,7 +598,7 @@ else % 3D regularization X2 = LLT_model(single(X), lambdaHO/subsets, tauHO/subsets, iterHO, 2.0e-05, 0); end - X = 0.5.*(X + X2); % averaged combination of two solutions + X = 0.5.*(X + X2); % the averaged combination of two solutions end if (lambdaPB > 0) % Patch-Based regularization (can be slow on CPU) @@ -559,14 +611,36 @@ else X = PatchBased_Regul(single(X), SearchW, SimilW, h_PB, lambdaPB/subsets); end end + if (lambdaPB_GPU > 0) + % Patch-Based regularization (GPU CUDA implementation) + if ((strcmp('2D', Dimension) == 1)) + % 2D regularization + for kkk = 1:SlicesZ + X(:,:,kkk) = NLM_GPU(single(X(:,:,kkk)), SearchW, SimilW, h_PB, lambdaPB_GPU); + end + else + X = NLM_GPU(single(X), SearchW, SimilW, h_PB, lambdaPB_GPU); + end + end + if (LambdaDiff_HO > 0) + % Higher-order diffusion penalty (GPU CUDA implementation) + if ((strcmp('2D', Dimension) == 1)) + % 2D regularization + for kkk = 1:SlicesZ + X(:,:,kkk) = Diff4thHajiaboli_GPU(single(X(:,:,kkk)), LambdaDiff_HO_EdgePar, LambdaDiff_HO, round(IterationsRegul/subsets)); + end + else + X = Diff4thHajiaboli_GPU(X, LambdaDiff_HO_EdgePar, LambdaDiff_HO, round(IterationsRegul/subsets)); + end + end if (LambdaTGV > 0) % Total Generalized variation (currently only 2D) lamTGV1 = 1.1; % smoothing trade-off parameters, see Pock's paper lamTGV2 = 0.5; % second-order term for kkk = 1:SlicesZ - X(:,:,kkk) = TGV_PD(single(X(:,:,kkk)), LambdaTGV/subsets, lamTGV1, lamTGV2, IterationsRegul); + X(:,:,kkk) = TGV_PD(single(X(:,:,kkk)), LambdaTGV/subsets, lamTGV1, lamTGV2, IterationsRegul); end - end + end if (lambdaR_L1 > 0) r = max(abs(r)-lambdaR_L1, 0).*sign(r); % soft-thresholding operator for ring vector diff --git a/main_func/regularizers_CPU/FGP_TV.c b/main_func/regularizers_CPU/FGP_TV.c index 5d8cfb9..cfe5b9e 100644 --- a/main_func/regularizers_CPU/FGP_TV.c +++ b/main_func/regularizers_CPU/FGP_TV.c @@ -54,7 +54,7 @@ void mexFunction( { int number_of_dims, iter, dimX, dimY, dimZ, ll, j, count, methTV; const int *dim_array; - float *A, *D=NULL, *D_old=NULL, *P1=NULL, *P2=NULL, *P3=NULL, *P1_old=NULL, *P2_old=NULL, *P3_old=NULL, *R1=NULL, *R2=NULL, *R3=NULL, lambda, tk, tkp1, re, re1, re_old, epsil, funcval; + float *A, *D=NULL, *D_old=NULL, *P1=NULL, *P2=NULL, *P3=NULL, *P1_old=NULL, *P2_old=NULL, *P3_old=NULL, *R1=NULL, *R2=NULL, *R3=NULL, lambda, tk, tkp1, re, re1, re_old, epsil; number_of_dims = mxGetNumberOfDimensions(prhs[0]); dim_array = mxGetDimensions(prhs[0]); @@ -78,7 +78,6 @@ void mexFunction( mxFree(penalty_type); } /*output function value (last iteration) */ - funcval = 0.0f; plhs[1] = mxCreateNumericMatrix(1, 1, mxSINGLE_CLASS, mxREAL); float *funcvalA = (float *) mxGetData(plhs[1]); @@ -117,7 +116,7 @@ void mexFunction( /*updating R and t*/ tkp1 = (1.0f + sqrt(1.0f + 4.0f*tk*tk))*0.5f; - Rupd_func2D(P1, P1_old, P2, P2_old, R1, R2, tkp1, tk, dimX, dimY); + Rupd_func2D(P1, P1_old, P2, P2_old, R1, R2, tkp1, tk, dimX, dimY); /* calculate norm */ re = 0.0f; re1 = 0.0f; @@ -129,23 +128,17 @@ void mexFunction( re = sqrt(re)/sqrt(re1); if (re < epsil) count++; if (count > 3) { - Obj_func2D(A, D, P1, P2, lambda, dimX, dimY); - funcval = 0.0f; - for(j=0; j 2) { if (re > re_old) { - Obj_func2D(A, D, P1, P2, lambda, dimX, dimY); - funcval = 0.0f; - for(j=0; j 3) { - Obj_func3D(A, D, P1, P2, P3,lambda, dimX, dimY, dimZ); - funcval = 0.0f; - for(j=0; j 2) { if (re > re_old) { - Obj_func3D(A, D, P1, P2, P3,lambda, dimX, dimY, dimZ); - funcval = 0.0f; - for(j=0; j +#include +#include +#include +#include +#include +#include "Diff4th_GPU_kernel.h" + +/* + * 2D and 3D CUDA implementation of the 4th order PDE denoising model by Hajiaboli + * + * Reference : + * "An anisotropic fourth-order diffusion filter for image noise removal" by M. Hajiaboli + * + * Example + * figure; + * Im = double(imread('lena_gray_256.tif'))/255; % loading image + * u0 = Im + .05*randn(size(Im)); % adding noise + * u = Diff4thHajiaboli_GPU(single(u0), 0.02, 150); + * subplot (1,2,1); imshow(u0,[ ]); title('Noisy Image') + * subplot (1,2,2); imshow(u,[ ]); title('Denoised Image') + * + * + * Linux/Matlab compilation: + * compile in terminal: nvcc -Xcompiler -fPIC -shared -o Diff4th_GPU_kernel.o Diff4th_GPU_kernel.cu + * then compile in Matlab: mex -I/usr/local/cuda-7.5/include -L/usr/local/cuda-7.5/lib64 -lcudart Diff4thHajiaboli_GPU.cpp Diff4th_GPU_kernel.o + */ + +void mexFunction( + int nlhs, mxArray *plhs[], + int nrhs, const mxArray *prhs[]) +{ + int numdims, dimZ, size; + float *A, *B, *A_L, *B_L; + const int *dims; + + numdims = mxGetNumberOfDimensions(prhs[0]); + dims = mxGetDimensions(prhs[0]); + + float sigma = (float)mxGetScalar(prhs[1]); /* edge-preserving parameter */ + float lambda = (float)mxGetScalar(prhs[2]); /* regularization parameter */ + int iter = (int)mxGetScalar(prhs[3]); /* iterations number */ + + if (numdims == 2) { + + int N, M, Z, i, j; + Z = 0; // for the 2D case + float tau = 0.01; // time step is sufficiently small for an explicit methods + + /*Input data*/ + A = (float*)mxGetData(prhs[0]); + N = dims[0] + 2; + M = dims[1] + 2; + A_L = (float*)mxGetData(mxCreateNumericMatrix(N, M, mxSINGLE_CLASS, mxREAL)); + B_L = (float*)mxGetData(mxCreateNumericMatrix(N, M, mxSINGLE_CLASS, mxREAL)); + + /*Output data*/ + B = (float*)mxGetData(plhs[0] = mxCreateNumericMatrix(dims[0], dims[1], mxSINGLE_CLASS, mxREAL)); + + // copy A to the bigger A_L with boundaries + #pragma omp parallel for shared(A_L, A) private(i,j) + for (i=0; i < N; i++) { + for (j=0; j < M; j++) { + if (((i > 0) && (i < N-1)) && ((j > 0) && (j < M-1))) A_L[i*M+j] = A[(i-1)*(dims[1])+(j-1)]; + }} + + // Running CUDA code here + Diff4th_GPU_kernel(A_L, B_L, N, M, Z, (float)sigma, iter, (float)tau, lambda); + + // copy the processed B_L to a smaller B + #pragma omp parallel for shared(B_L, B) private(i,j) + for (i=0; i < N; i++) { + for (j=0; j < M; j++) { + if (((i > 0) && (i < N-1)) && ((j > 0) && (j < M-1))) B[(i-1)*(dims[1])+(j-1)] = B_L[i*M+j]; + }} + } + if (numdims == 3) { + // 3D image denoising / regularization + int N, M, Z, i, j, k; + float tau = 0.0007; // Time Step is small for an explicit methods + A = (float*)mxGetData(prhs[0]); + N = dims[0] + 2; + M = dims[1] + 2; + Z = dims[2] + 2; + int N_dims[] = {N, M, Z}; + A_L = (float*)mxGetPr(mxCreateNumericArray(3, N_dims, mxSINGLE_CLASS, mxREAL)); + B_L = (float*)mxGetPr(mxCreateNumericArray(3, N_dims, mxSINGLE_CLASS, mxREAL)); + + /* output data */ + B = (float*)mxGetPr(plhs[0] = mxCreateNumericArray(3, dims, mxSINGLE_CLASS, mxREAL)); + + // copy A to the bigger A_L with boundaries + #pragma omp parallel for shared(A_L, A) private(i,j,k) + for (i=0; i < N; i++) { + for (j=0; j < M; j++) { + for (k=0; k < Z; k++) { + if (((i > 0) && (i < N-1)) && ((j > 0) && (j < M-1)) && ((k > 0) && (k < Z-1))) { + A_L[(N*M)*(k)+(i)*M+(j)] = A[(dims[0]*dims[1])*(k-1)+(i-1)*dims[1]+(j-1)]; + }}}} + + // Running CUDA kernel here for diffusivity + Diff4th_GPU_kernel(A_L, B_L, N, M, Z, (float)sigma, iter, (float)tau, lambda); + + // copy the processed B_L to a smaller B + #pragma omp parallel for shared(B_L, B) private(i,j,k) + for (i=0; i < N; i++) { + for (j=0; j < M; j++) { + for (k=0; k < Z; k++) { + if (((i > 0) && (i < N-1)) && ((j > 0) && (j < M-1)) && ((k > 0) && (k < Z-1))) { + B[(dims[0]*dims[1])*(k-1)+(i-1)*dims[1]+(j-1)] = B_L[(N*M)*(k)+(i)*M+(j)]; + }}}} + } +} \ No newline at end of file diff --git a/main_func/regularizers_GPU/Diffus_HO/Diff4th_GPU_kernel.cu b/main_func/regularizers_GPU/Diffus_HO/Diff4th_GPU_kernel.cu new file mode 100644 index 0000000..178af00 --- /dev/null +++ b/main_func/regularizers_GPU/Diffus_HO/Diff4th_GPU_kernel.cu @@ -0,0 +1,270 @@ +#include +#include +#include +#include "Diff4th_GPU_kernel.h" + +#define checkCudaErrors(err) __checkCudaErrors (err, __FILE__, __LINE__) + +inline void __checkCudaErrors(cudaError err, const char *file, const int line) +{ + if (cudaSuccess != err) + { + fprintf(stderr, "%s(%i) : CUDA Runtime API error %d: %s.\n", + file, line, (int)err, cudaGetErrorString(err)); + exit(EXIT_FAILURE); + } +} + +#define idivup(a, b) ( ((a)%(b) != 0) ? (a)/(b)+1 : (a)/(b) ) +#define sizeT (sizeX*sizeY*sizeZ) +#define epsilon 0.00000001 + +///////////////////////////////////////////////// +// 2D Image denosing - Second Step (The second derrivative) +__global__ void Diff4th2D_derriv(float* B, float* A, float *A0, int N, int M, float sigma, int iter, float tau, float lambda) +{ + float gradXXc = 0, gradYYc = 0; + int i = blockIdx.x*blockDim.x + threadIdx.x; + int j = blockIdx.y*blockDim.y + threadIdx.y; + + int index = j + i*N; + + if (((i < 1) || (i > N-2)) || ((j < 1) || (j > M-2))) { + return; } + + int indexN = (j)+(i-1)*(N); if (A[indexN] == 0) indexN = index; + int indexS = (j)+(i+1)*(N); if (A[indexS] == 0) indexS = index; + int indexW = (j-1)+(i)*(N); if (A[indexW] == 0) indexW = index; + int indexE = (j+1)+(i)*(N); if (A[indexE] == 0) indexE = index; + + gradXXc = B[indexN] + B[indexS] - 2*B[index] ; + gradYYc = B[indexW] + B[indexE] - 2*B[index] ; + A[index] = A[index] - tau*((A[index] - A0[index]) + lambda*(gradXXc + gradYYc)); +} + +// 2D Image denosing - The First Step +__global__ void Diff4th2D(float* A, float* B, int N, int M, float sigma, int iter, float tau) +{ + float gradX, gradX_sq, gradY, gradY_sq, gradXX, gradYY, gradXY, sq_sum, xy_2, V_norm, V_orth, c, c_sq; + + int i = blockIdx.x*blockDim.x + threadIdx.x; + int j = blockIdx.y*blockDim.y + threadIdx.y; + + int index = j + i*N; + + V_norm = 0.0f; V_orth = 0.0f; + + if (((i < 1) || (i > N-2)) || ((j < 1) || (j > M-2))) { + return; } + + int indexN = (j)+(i-1)*(N); if (A[indexN] == 0) indexN = index; + int indexS = (j)+(i+1)*(N); if (A[indexS] == 0) indexS = index; + int indexW = (j-1)+(i)*(N); if (A[indexW] == 0) indexW = index; + int indexE = (j+1)+(i)*(N); if (A[indexE] == 0) indexE = index; + int indexNW = (j-1)+(i-1)*(N); if (A[indexNW] == 0) indexNW = index; + int indexNE = (j+1)+(i-1)*(N); if (A[indexNE] == 0) indexNE = index; + int indexWS = (j-1)+(i+1)*(N); if (A[indexWS] == 0) indexWS = index; + int indexES = (j+1)+(i+1)*(N); if (A[indexES] == 0) indexES = index; + + gradX = 0.5f*(A[indexN]-A[indexS]); + gradX_sq = gradX*gradX; + gradXX = A[indexN] + A[indexS] - 2*A[index]; + + gradY = 0.5f*(A[indexW]-A[indexE]); + gradY_sq = gradY*gradY; + gradYY = A[indexW] + A[indexE] - 2*A[index]; + + gradXY = 0.25f*(A[indexNW] - A[indexNE] - A[indexWS] + A[indexES]); + xy_2 = 2.0f*gradX*gradY*gradXY; + sq_sum = gradX_sq + gradY_sq; + + if (sq_sum <= epsilon) { + V_norm = (gradXX*gradX_sq + xy_2 + gradYY*gradY_sq)/epsilon; + V_orth = (gradXX*gradY_sq - xy_2 + gradYY*gradX_sq)/epsilon; } + else { + V_norm = (gradXX*gradX_sq + xy_2 + gradYY*gradY_sq)/sq_sum; + V_orth = (gradXX*gradY_sq - xy_2 + gradYY*gradX_sq)/sq_sum; } + + c = 1.0f/(1.0f + sq_sum/sigma); + c_sq = c*c; + B[index] = c_sq*V_norm + c*V_orth; +} + +///////////////////////////////////////////////// +// 3D data parocerssing +__global__ void Diff4th3D_derriv(float *B, float *A, float *A0, int N, int M, int Z, float sigma, int iter, float tau, float lambda) +{ + float gradXXc = 0, gradYYc = 0, gradZZc = 0; + int xIndex = blockDim.x * blockIdx.x + threadIdx.x; + int yIndex = blockDim.y * blockIdx.y + threadIdx.y; + int zIndex = blockDim.z * blockIdx.z + threadIdx.z; + + int index = xIndex + M*yIndex + N*M*zIndex; + + if (((xIndex < 1) || (xIndex > N-2)) || ((yIndex < 1) || (yIndex > M-2)) || ((zIndex < 1) || (zIndex > Z-2))) { + return; } + + int indexN = (xIndex-1) + M*yIndex + N*M*zIndex; if (A[indexN] == 0) indexN = index; + int indexS = (xIndex+1) + M*yIndex + N*M*zIndex; if (A[indexS] == 0) indexS = index; + int indexW = xIndex + M*(yIndex-1) + N*M*zIndex; if (A[indexW] == 0) indexW = index; + int indexE = xIndex + M*(yIndex+1) + N*M*zIndex; if (A[indexE] == 0) indexE = index; + int indexU = xIndex + M*yIndex + N*M*(zIndex-1); if (A[indexU] == 0) indexU = index; + int indexD = xIndex + M*yIndex + N*M*(zIndex+1); if (A[indexD] == 0) indexD = index; + + gradXXc = B[indexN] + B[indexS] - 2*B[index] ; + gradYYc = B[indexW] + B[indexE] - 2*B[index] ; + gradZZc = B[indexU] + B[indexD] - 2*B[index] ; + + A[index] = A[index] - tau*((A[index] - A0[index]) + lambda*(gradXXc + gradYYc + gradZZc)); +} + +__global__ void Diff4th3D(float* A, float* B, int N, int M, int Z, float sigma, int iter, float tau) +{ + float gradX, gradX_sq, gradY, gradY_sq, gradZ, gradZ_sq, gradXX, gradYY, gradZZ, gradXY, gradXZ, gradYZ, sq_sum, xy_2, xyz_1, xyz_2, V_norm, V_orth, c, c_sq; + + int xIndex = blockDim.x * blockIdx.x + threadIdx.x; + int yIndex = blockDim.y * blockIdx.y + threadIdx.y; + int zIndex = blockDim.z * blockIdx.z + threadIdx.z; + + int index = xIndex + M*yIndex + N*M*zIndex; + V_norm = 0.0f; V_orth = 0.0f; + + if (((xIndex < 1) || (xIndex > N-2)) || ((yIndex < 1) || (yIndex > M-2)) || ((zIndex < 1) || (zIndex > Z-2))) { + return; } + + B[index] = 0; + + int indexN = (xIndex-1) + M*yIndex + N*M*zIndex; if (A[indexN] == 0) indexN = index; + int indexS = (xIndex+1) + M*yIndex + N*M*zIndex; if (A[indexS] == 0) indexS = index; + int indexW = xIndex + M*(yIndex-1) + N*M*zIndex; if (A[indexW] == 0) indexW = index; + int indexE = xIndex + M*(yIndex+1) + N*M*zIndex; if (A[indexE] == 0) indexE = index; + int indexU = xIndex + M*yIndex + N*M*(zIndex-1); if (A[indexU] == 0) indexU = index; + int indexD = xIndex + M*yIndex + N*M*(zIndex+1); if (A[indexD] == 0) indexD = index; + + int indexNW = (xIndex-1) + M*(yIndex-1) + N*M*zIndex; if (A[indexNW] == 0) indexNW = index; + int indexNE = (xIndex-1) + M*(yIndex+1) + N*M*zIndex; if (A[indexNE] == 0) indexNE = index; + int indexWS = (xIndex+1) + M*(yIndex-1) + N*M*zIndex; if (A[indexWS] == 0) indexWS = index; + int indexES = (xIndex+1) + M*(yIndex+1) + N*M*zIndex; if (A[indexES] == 0) indexES = index; + + int indexUW = (xIndex-1) + M*(yIndex) + N*M*(zIndex-1); if (A[indexUW] == 0) indexUW = index; + int indexUE = (xIndex+1) + M*(yIndex) + N*M*(zIndex-1); if (A[indexUE] == 0) indexUE = index; + int indexDW = (xIndex-1) + M*(yIndex) + N*M*(zIndex+1); if (A[indexDW] == 0) indexDW = index; + int indexDE = (xIndex+1) + M*(yIndex) + N*M*(zIndex+1); if (A[indexDE] == 0) indexDE = index; + + int indexUN = (xIndex) + M*(yIndex-1) + N*M*(zIndex-1); if (A[indexUN] == 0) indexUN = index; + int indexUS = (xIndex) + M*(yIndex+1) + N*M*(zIndex-1); if (A[indexUS] == 0) indexUS = index; + int indexDN = (xIndex) + M*(yIndex-1) + N*M*(zIndex+1); if (A[indexDN] == 0) indexDN = index; + int indexDS = (xIndex) + M*(yIndex+1) + N*M*(zIndex+1); if (A[indexDS] == 0) indexDS = index; + + gradX = 0.5f*(A[indexN]-A[indexS]); + gradX_sq = gradX*gradX; + gradXX = A[indexN] + A[indexS] - 2*A[index]; + + gradY = 0.5f*(A[indexW]-A[indexE]); + gradY_sq = gradY*gradY; + gradYY = A[indexW] + A[indexE] - 2*A[index]; + + gradZ = 0.5f*(A[indexU]-A[indexD]); + gradZ_sq = gradZ*gradZ; + gradZZ = A[indexU] + A[indexD] - 2*A[index]; + + gradXY = 0.25f*(A[indexNW] - A[indexNE] - A[indexWS] + A[indexES]); + gradXZ = 0.25f*(A[indexUW] - A[indexUE] - A[indexDW] + A[indexDE]); + gradYZ = 0.25f*(A[indexUN] - A[indexUS] - A[indexDN] + A[indexDS]); + + xy_2 = 2.0f*gradX*gradY*gradXY; + xyz_1 = 2.0f*gradX*gradZ*gradXZ; + xyz_2 = 2.0f*gradY*gradZ*gradYZ; + + sq_sum = gradX_sq + gradY_sq + gradZ_sq; + + if (sq_sum <= epsilon) { + V_norm = (gradXX*gradX_sq + gradYY*gradY_sq + gradZZ*gradZ_sq + xy_2 + xyz_1 + xyz_2)/epsilon; + V_orth = ((gradY_sq + gradZ_sq)*gradXX + (gradX_sq + gradZ_sq)*gradYY + (gradX_sq + gradY_sq)*gradZZ - xy_2 - xyz_1 - xyz_2)/epsilon; } + else { + V_norm = (gradXX*gradX_sq + gradYY*gradY_sq + gradZZ*gradZ_sq + xy_2 + xyz_1 + xyz_2)/sq_sum; + V_orth = ((gradY_sq + gradZ_sq)*gradXX + (gradX_sq + gradZ_sq)*gradYY + (gradX_sq + gradY_sq)*gradZZ - xy_2 - xyz_1 - xyz_2)/sq_sum; } + + c = 1; + if ((1.0f + sq_sum/sigma) != 0.0f) {c = 1.0f/(1.0f + sq_sum/sigma);} + + c_sq = c*c; + B[index] = c_sq*V_norm + c*V_orth; +} + +/******************************************************/ +/********* HOST FUNCTION*************/ +extern "C" void Diff4th_GPU_kernel(float* A, float* B, int N, int M, int Z, float sigma, int iter, float tau, float lambda) +{ + int deviceCount = -1; // number of devices + cudaGetDeviceCount(&deviceCount); + if (deviceCount == 0) { + fprintf(stderr, "No CUDA devices found\n"); + return; + } + + int BLKXSIZE, BLKYSIZE,BLKZSIZE; + float *Ad, *Bd, *Cd; + sigma = sigma*sigma; + + if (Z == 0){ + // 4th order diffusion for 2D case + BLKXSIZE = 8; + BLKYSIZE = 16; + + dim3 dimBlock(BLKXSIZE,BLKYSIZE); + dim3 dimGrid(idivup(N,BLKXSIZE), idivup(M,BLKYSIZE)); + + checkCudaErrors(cudaMalloc((void**)&Ad,N*M*sizeof(float))); + checkCudaErrors(cudaMalloc((void**)&Bd,N*M*sizeof(float))); + checkCudaErrors(cudaMalloc((void**)&Cd,N*M*sizeof(float))); + + checkCudaErrors(cudaMemcpy(Ad,A,N*M*sizeof(float),cudaMemcpyHostToDevice)); + checkCudaErrors(cudaMemcpy(Bd,A,N*M*sizeof(float),cudaMemcpyHostToDevice)); + checkCudaErrors(cudaMemcpy(Cd,A,N*M*sizeof(float),cudaMemcpyHostToDevice)); + + int n = 1; + while (n <= iter) { + Diff4th2D<<>>(Bd, Cd, N, M, sigma, iter, tau); + cudaDeviceSynchronize(); + checkCudaErrors( cudaPeekAtLastError() ); + Diff4th2D_derriv<<>>(Cd, Bd, Ad, N, M, sigma, iter, tau, lambda); + cudaDeviceSynchronize(); + checkCudaErrors( cudaPeekAtLastError() ); + n++; + } + checkCudaErrors(cudaMemcpy(B,Bd,N*M*sizeof(float),cudaMemcpyDeviceToHost)); + cudaFree(Ad); cudaFree(Bd); cudaFree(Cd); + } + + if (Z != 0){ + // 4th order diffusion for 3D case + BLKXSIZE = 8; + BLKYSIZE = 8; + BLKZSIZE = 8; + + dim3 dimBlock(BLKXSIZE,BLKYSIZE,BLKZSIZE); + dim3 dimGrid(idivup(N,BLKXSIZE), idivup(M,BLKYSIZE),idivup(Z,BLKXSIZE)); + + checkCudaErrors(cudaMalloc((void**)&Ad,N*M*Z*sizeof(float))); + checkCudaErrors(cudaMalloc((void**)&Bd,N*M*Z*sizeof(float))); + checkCudaErrors(cudaMalloc((void**)&Cd,N*M*Z*sizeof(float))); + + checkCudaErrors(cudaMemcpy(Ad,A,N*M*Z*sizeof(float),cudaMemcpyHostToDevice)); + checkCudaErrors(cudaMemcpy(Bd,A,N*M*Z*sizeof(float),cudaMemcpyHostToDevice)); + checkCudaErrors(cudaMemcpy(Cd,A,N*M*Z*sizeof(float),cudaMemcpyHostToDevice)); + + int n = 1; + while (n <= iter) { + Diff4th3D<<>>(Bd, Cd, N, M, Z, sigma, iter, tau); + cudaDeviceSynchronize(); + checkCudaErrors( cudaPeekAtLastError() ); + Diff4th3D_derriv<<>>(Cd, Bd, Ad, N, M, Z, sigma, iter, tau, lambda); + cudaDeviceSynchronize(); + checkCudaErrors( cudaPeekAtLastError() ); + n++; + } + checkCudaErrors(cudaMemcpy(B,Bd,N*M*Z*sizeof(float),cudaMemcpyDeviceToHost)); + cudaFree(Ad); cudaFree(Bd); cudaFree(Cd); + } +} \ No newline at end of file diff --git a/main_func/regularizers_GPU/Diffus_HO/Diff4th_GPU_kernel.h b/main_func/regularizers_GPU/Diffus_HO/Diff4th_GPU_kernel.h new file mode 100644 index 0000000..cfbb45a --- /dev/null +++ b/main_func/regularizers_GPU/Diffus_HO/Diff4th_GPU_kernel.h @@ -0,0 +1,6 @@ +#ifndef __DIFF_HO_H_ +#define __DIFF_HO_H_ + +extern "C" void Diff4th_GPU_kernel(float* A, float* B, int N, int M, int Z, float sigma, int iter, float tau, float lambda); + +#endif diff --git a/main_func/regularizers_GPU/NL_Regul/NLM_GPU.cpp b/main_func/regularizers_GPU/NL_Regul/NLM_GPU.cpp new file mode 100644 index 0000000..ff0cc90 --- /dev/null +++ b/main_func/regularizers_GPU/NL_Regul/NLM_GPU.cpp @@ -0,0 +1,171 @@ +#include "mex.h" +#include +#include +#include +#include +#include +#include +#include "NLM_GPU_kernel.h" + +/* CUDA implementation of the patch-based (PB) regularization for 2D and 3D images/volumes + * This method finds self-similar patches in data and performs one fixed point iteration to mimimize the PB penalty function + * + * References: 1. Yang Z. & Jacob M. "Nonlocal Regularization of Inverse Problems" + * 2. Kazantsev D. at. all "4D-CT reconstruction with unified spatial-temporal patch-based regularization" + * + * Input Parameters (mandatory): + * 1. Image/volume (2D/3D) + * 2. ratio of the searching window (e.g. 3 = (2*3+1) = 7 pixels window) + * 3. ratio of the similarity window (e.g. 1 = (2*1+1) = 3 pixels window) + * 4. h - parameter for the PB penalty function + * 5. lambda - regularization parameter + + * Output: + * 1. regularized (denoised) Image/volume (N x N x N) + * + * In matlab check what kind of GPU you have with "gpuDevice" command, + * then set your ComputeCapability, here I use -arch compute_35 + * + * Quick 2D denoising example in Matlab: + Im = double(imread('lena_gray_256.tif'))/255; % loading image + u0 = Im + .03*randn(size(Im)); u0(u0<0) = 0; % adding noise + ImDen = NLM_GPU(single(u0), 3, 2, 0.15, 1); + + * Linux/Matlab compilation: + * compile in terminal: nvcc -Xcompiler -fPIC -shared -o NLM_GPU_kernel.o NLM_GPU_kernel.cu + * then compile in Matlab: mex -I/usr/local/cuda-7.5/include -L/usr/local/cuda-7.5/lib64 -lcudart NLM_GPU.cpp NLM_GPU_kernel.o + * + * D. Kazantsev + * 2014-17 + * Harwell/Manchester UK + */ + +float pad_crop(float *A, float *Ap, int OldSizeX, int OldSizeY, int OldSizeZ, int NewSizeX, int NewSizeY, int NewSizeZ, int padXY, int switchpad_crop); + +void mexFunction( + int nlhs, mxArray *plhs[], + int nrhs, const mxArray *prhs[]) +{ + int N, M, Z, i_n, j_n, k_n, numdims, SearchW, SimilW, SearchW_real, padXY, newsizeX, newsizeY, newsizeZ, switchpad_crop, count, SearchW_full, SimilW_full; + const int *dims; + float *A, *B=NULL, *Ap=NULL, *Bp=NULL, *Eucl_Vec, h, h2, lambda, val, denh2; + + numdims = mxGetNumberOfDimensions(prhs[0]); + dims = mxGetDimensions(prhs[0]); + + N = dims[0]; + M = dims[1]; + Z = dims[2]; + + if ((numdims < 2) || (numdims > 3)) {mexErrMsgTxt("The input should be 2D image or 3D volume");} + if (mxGetClassID(prhs[0]) != mxSINGLE_CLASS) {mexErrMsgTxt("The input in single precision is required"); } + + if(nrhs != 5) mexErrMsgTxt("Five inputs reqired: Image(2D,3D), SearchW, SimilW, Threshold, Regularization parameter"); + + /*Handling inputs*/ + A = (float *) mxGetData(prhs[0]); /* the image to regularize/filter */ + SearchW_real = (int) mxGetScalar(prhs[1]); /* the searching window ratio */ + SimilW = (int) mxGetScalar(prhs[2]); /* the similarity window ratio */ + h = (float) mxGetScalar(prhs[3]); /* parameter for the PB filtering function */ + lambda = (float) mxGetScalar(prhs[4]); + + if (h <= 0) mexErrMsgTxt("Parmeter for the PB penalty function should be > 0"); + + SearchW = SearchW_real + 2*SimilW; + + SearchW_full = 2*SearchW + 1; /* the full searching window size */ + SimilW_full = 2*SimilW + 1; /* the full similarity window size */ + h2 = h*h; + + padXY = SearchW + 2*SimilW; /* padding sizes */ + newsizeX = N + 2*(padXY); /* the X size of the padded array */ + newsizeY = M + 2*(padXY); /* the Y size of the padded array */ + newsizeZ = Z + 2*(padXY); /* the Z size of the padded array */ + int N_dims[] = {newsizeX, newsizeY, newsizeZ}; + + /******************************2D case ****************************/ + if (numdims == 2) { + /*Handling output*/ + B = (float*)mxGetData(plhs[0] = mxCreateNumericMatrix(N, M, mxSINGLE_CLASS, mxREAL)); + /*allocating memory for the padded arrays */ + Ap = (float*)mxGetData(mxCreateNumericMatrix(newsizeX, newsizeY, mxSINGLE_CLASS, mxREAL)); + Bp = (float*)mxGetData(mxCreateNumericMatrix(newsizeX, newsizeY, mxSINGLE_CLASS, mxREAL)); + Eucl_Vec = (float*)mxGetData(mxCreateNumericMatrix(SimilW_full*SimilW_full, 1, mxSINGLE_CLASS, mxREAL)); + + /*Gaussian kernel */ + count = 0; + for(i_n=-SimilW; i_n<=SimilW; i_n++) { + for(j_n=-SimilW; j_n<=SimilW; j_n++) { + val = (float)(i_n*i_n + j_n*j_n)/(2*SimilW*SimilW); + Eucl_Vec[count] = exp(-val); + count = count + 1; + }} /*main neighb loop */ + + /**************************************************************************/ + /*Perform padding of image A to the size of [newsizeX * newsizeY] */ + switchpad_crop = 0; /*padding*/ + pad_crop(A, Ap, M, N, 0, newsizeY, newsizeX, 0, padXY, switchpad_crop); + + /* Do PB regularization with the padded array */ + NLM_GPU_kernel(Ap, Bp, Eucl_Vec, newsizeY, newsizeX, 0, numdims, SearchW, SimilW, SearchW_real, (float)h2, (float)lambda); + + switchpad_crop = 1; /*cropping*/ + pad_crop(Bp, B, M, N, 0, newsizeY, newsizeX, 0, padXY, switchpad_crop); + } + else + { + /******************************3D case ****************************/ + /*Handling output*/ + B = (float*)mxGetPr(plhs[0] = mxCreateNumericArray(3, dims, mxSINGLE_CLASS, mxREAL)); + /*allocating memory for the padded arrays */ + Ap = (float*)mxGetPr(mxCreateNumericArray(3, N_dims, mxSINGLE_CLASS, mxREAL)); + Bp = (float*)mxGetPr(mxCreateNumericArray(3, N_dims, mxSINGLE_CLASS, mxREAL)); + Eucl_Vec = (float*)mxGetData(mxCreateNumericMatrix(SimilW_full*SimilW_full*SimilW_full, 1, mxSINGLE_CLASS, mxREAL)); + + /*Gaussian kernel */ + count = 0; + for(i_n=-SimilW; i_n<=SimilW; i_n++) { + for(j_n=-SimilW; j_n<=SimilW; j_n++) { + for(k_n=-SimilW; k_n<=SimilW; k_n++) { + val = (float)(i_n*i_n + j_n*j_n + k_n*k_n)/(2*SimilW*SimilW*SimilW); + Eucl_Vec[count] = exp(-val); + count = count + 1; + }}} /*main neighb loop */ + /**************************************************************************/ + /*Perform padding of image A to the size of [newsizeX * newsizeY * newsizeZ] */ + switchpad_crop = 0; /*padding*/ + pad_crop(A, Ap, M, N, Z, newsizeY, newsizeX, newsizeZ, padXY, switchpad_crop); + + /* Do PB regularization with the padded array */ + NLM_GPU_kernel(Ap, Bp, Eucl_Vec, newsizeY, newsizeX, newsizeZ, numdims, SearchW, SimilW, SearchW_real, (float)h2, (float)lambda); + + switchpad_crop = 1; /*cropping*/ + pad_crop(Bp, B, M, N, Z, newsizeY, newsizeX, newsizeZ, padXY, switchpad_crop); + } /*end else ndims*/ +} + +float pad_crop(float *A, float *Ap, int OldSizeX, int OldSizeY, int OldSizeZ, int NewSizeX, int NewSizeY, int NewSizeZ, int padXY, int switchpad_crop) +{ + /* padding-cropping function */ + int i,j,k; + if (NewSizeZ > 1) { + for (i=0; i < NewSizeX; i++) { + for (j=0; j < NewSizeY; j++) { + for (k=0; k < NewSizeZ; k++) { + if (((i >= padXY) && (i < NewSizeX-padXY)) && ((j >= padXY) && (j < NewSizeY-padXY)) && ((k >= padXY) && (k < NewSizeZ-padXY))) { + if (switchpad_crop == 0) Ap[NewSizeX*NewSizeY*k + i*NewSizeY+j] = A[OldSizeX*OldSizeY*(k - padXY) + (i-padXY)*(OldSizeY)+(j-padXY)]; + else Ap[OldSizeX*OldSizeY*(k - padXY) + (i-padXY)*(OldSizeY)+(j-padXY)] = A[NewSizeX*NewSizeY*k + i*NewSizeY+j]; + } + }}} + } + else { + for (i=0; i < NewSizeX; i++) { + for (j=0; j < NewSizeY; j++) { + if (((i >= padXY) && (i < NewSizeX-padXY)) && ((j >= padXY) && (j < NewSizeY-padXY))) { + if (switchpad_crop == 0) Ap[i*NewSizeY+j] = A[(i-padXY)*(OldSizeY)+(j-padXY)]; + else Ap[(i-padXY)*(OldSizeY)+(j-padXY)] = A[i*NewSizeY+j]; + } + }} + } + return *Ap; +} \ No newline at end of file diff --git a/main_func/regularizers_GPU/NL_Regul/NLM_GPU_kernel.cu b/main_func/regularizers_GPU/NL_Regul/NLM_GPU_kernel.cu new file mode 100644 index 0000000..17da3a8 --- /dev/null +++ b/main_func/regularizers_GPU/NL_Regul/NLM_GPU_kernel.cu @@ -0,0 +1,239 @@ +#include +#include +#include +#include "NLM_GPU_kernel.h" + +#define checkCudaErrors(err) __checkCudaErrors (err, __FILE__, __LINE__) + +inline void __checkCudaErrors(cudaError err, const char *file, const int line) +{ + if (cudaSuccess != err) + { + fprintf(stderr, "%s(%i) : CUDA Runtime API error %d: %s.\n", + file, line, (int)err, cudaGetErrorString(err)); + exit(EXIT_FAILURE); + } +} + +extern __shared__ float sharedmem[]; + +// run PB den kernel here +__global__ void NLM_kernel(float *Ad, float* Bd, float *Eucl_Vec_d, int N, int M, int Z, int SearchW, int SimilW, int SearchW_real, int SearchW_full, int SimilW_full, int padXY, float h2, float lambda, dim3 imagedim, dim3 griddim, dim3 kerneldim, dim3 sharedmemdim, int nUpdatePerThread, float neighborsize) +{ + + int i1, j1, k1, i2, j2, k2, i3, j3, k3, i_l, j_l, k_l, count; + float value, Weight_norm, normsum, Weight; + + int bidx = blockIdx.x; + int bidy = blockIdx.y%griddim.y; + int bidz = (int)((blockIdx.y)/griddim.y); + + // global index for block endpoint + int beidx = __mul24(bidx,blockDim.x); + int beidy = __mul24(bidy,blockDim.y); + int beidz = __mul24(bidz,blockDim.z); + + int tid = __mul24(threadIdx.z,__mul24(blockDim.x,blockDim.y)) + + __mul24(threadIdx.y,blockDim.x) + threadIdx.x; + + #ifdef __DEVICE_EMULATION__ + printf("tid : %d", tid); + #endif + + // update shared memory + int nthreads = blockDim.x*blockDim.y*blockDim.z; + int sharedMemSize = sharedmemdim.x * sharedmemdim.y * sharedmemdim.z; + for(int i=0; i= padXY && idx < (imagedim.x - padXY) && + idy >= padXY && idy < (imagedim.y - padXY)) + { + int i_centr = threadIdx.x + (SearchW); /*indices of the centrilized (main) pixel */ + int j_centr = threadIdx.y + (SearchW); /*indices of the centrilized (main) pixel */ + + if ((i_centr > 0) && (i_centr < N) && (j_centr > 0) && (j_centr < M)) { + + Weight_norm = 0; value = 0.0; + /* Massive Search window loop */ + for(i1 = i_centr - SearchW_real ; i1 <= i_centr + SearchW_real; i1++) { + for(j1 = j_centr - SearchW_real ; j1<= j_centr + SearchW_real ; j1++) { + /* if inside the searching window */ + count = 0; normsum = 0.0; + for(i_l=-SimilW; i_l<=SimilW; i_l++) { + for(j_l=-SimilW; j_l<=SimilW; j_l++) { + i2 = i1+i_l; j2 = j1+j_l; + i3 = i_centr+i_l; j3 = j_centr+j_l; /*coordinates of the inner patch loop */ + if ((i2 > 0) && (i2 < N) && (j2 > 0) && (j2 < M)) { + if ((i3 > 0) && (i3 < N) && (j3 > 0) && (j3 < M)) { + normsum += Eucl_Vec_d[count]*pow((sharedmem[(j3)*sharedmemdim.x+(i3)] - sharedmem[j2*sharedmemdim.x+i2]), 2); + }} + count++; + }} + if (normsum != 0) Weight = (expf(-normsum/h2)); + else Weight = 0.0; + Weight_norm += Weight; + value += sharedmem[j1*sharedmemdim.x+i1]*Weight; + }} + + if (Weight_norm != 0) Bd[idz*imagedim.x*imagedim.y + idy*imagedim.x + idx] = value/Weight_norm; + else Bd[idz*imagedim.x*imagedim.y + idy*imagedim.x + idx] = Ad[idz*imagedim.x*imagedim.y + idy*imagedim.x + idx]; + } + } /*boundary conditions end*/ + } + else { + /*3D case*/ + /*checking boundaries to be within the image and avoid padded spaces */ + if( idx >= padXY && idx < (imagedim.x - padXY) && + idy >= padXY && idy < (imagedim.y - padXY) && + idz >= padXY && idz < (imagedim.z - padXY) ) + { + int i_centr = threadIdx.x + SearchW; /*indices of the centrilized (main) pixel */ + int j_centr = threadIdx.y + SearchW; /*indices of the centrilized (main) pixel */ + int k_centr = threadIdx.z + SearchW; /*indices of the centrilized (main) pixel */ + + if ((i_centr > 0) && (i_centr < N) && (j_centr > 0) && (j_centr < M) && (k_centr > 0) && (k_centr < Z)) { + + Weight_norm = 0; value = 0.0; + /* Massive Search window loop */ + for(i1 = i_centr - SearchW_real ; i1 <= i_centr + SearchW_real; i1++) { + for(j1 = j_centr - SearchW_real ; j1<= j_centr + SearchW_real ; j1++) { + for(k1 = k_centr - SearchW_real ; k1<= k_centr + SearchW_real ; k1++) { + /* if inside the searching window */ + count = 0; normsum = 0.0; + for(i_l=-SimilW; i_l<=SimilW; i_l++) { + for(j_l=-SimilW; j_l<=SimilW; j_l++) { + for(k_l=-SimilW; k_l<=SimilW; k_l++) { + i2 = i1+i_l; j2 = j1+j_l; k2 = k1+k_l; + i3 = i_centr+i_l; j3 = j_centr+j_l; k3 = k_centr+k_l; /*coordinates of the inner patch loop */ + if ((i2 > 0) && (i2 < N) && (j2 > 0) && (j2 < M) && (k2 > 0) && (k2 < Z)) { + if ((i3 > 0) && (i3 < N) && (j3 > 0) && (j3 < M) && (k3 > 0) && (k3 < Z)) { + normsum += Eucl_Vec_d[count]*pow((sharedmem[(k3)*sharedmemdim.x*sharedmemdim.y + (j3)*sharedmemdim.x+(i3)] - sharedmem[(k2)*sharedmemdim.x*sharedmemdim.y + j2*sharedmemdim.x+i2]), 2); + }} + count++; + }}} + if (normsum != 0) Weight = (expf(-normsum/h2)); + else Weight = 0.0; + Weight_norm += Weight; + value += sharedmem[k1*sharedmemdim.x*sharedmemdim.y + j1*sharedmemdim.x+i1]*Weight; + }}} /* BIG search window loop end*/ + + + if (Weight_norm != 0) Bd[idz*imagedim.x*imagedim.y + idy*imagedim.x + idx] = value/Weight_norm; + else Bd[idz*imagedim.x*imagedim.y + idy*imagedim.x + idx] = Ad[idz*imagedim.x*imagedim.y + idy*imagedim.x + idx]; + } + } /* boundary conditions end */ + } +} + +///////////////////////////////////////////////// +// HOST FUNCTION +extern "C" void NLM_GPU_kernel(float *A, float* B, float *Eucl_Vec, int N, int M, int Z, int dimension, int SearchW, int SimilW, int SearchW_real, float h2, float lambda) +{ + int deviceCount = -1; // number of devices + cudaGetDeviceCount(&deviceCount); + if (deviceCount == 0) { + fprintf(stderr, "No CUDA devices found\n"); + return; + } + +// cudaDeviceReset(); + + int padXY, SearchW_full, SimilW_full, blockWidth, blockHeight, blockDepth, nBlockX, nBlockY, nBlockZ, kernel_depth; + float *Ad, *Bd, *Eucl_Vec_d; + + if (dimension == 2) { + blockWidth = 16; + blockHeight = 16; + blockDepth = 1; + Z = 1; + kernel_depth = 0; + } + else { + blockWidth = 8; + blockHeight = 8; + blockDepth = 8; + kernel_depth = SearchW; + } + + // compute how many blocks are needed + nBlockX = ceil((float)N / (float)blockWidth); + nBlockY = ceil((float)M / (float)blockHeight); + nBlockZ = ceil((float)Z / (float)blockDepth); + + dim3 dimGrid(nBlockX,nBlockY*nBlockZ); + dim3 dimBlock(blockWidth, blockHeight, blockDepth); + dim3 imagedim(N,M,Z); + dim3 griddim(nBlockX,nBlockY,nBlockZ); + + dim3 kerneldim(SearchW,SearchW,kernel_depth); + dim3 sharedmemdim((SearchW*2)+blockWidth,(SearchW*2)+blockHeight,(kernel_depth*2)+blockDepth); + int sharedmemsize = sizeof(float)*sharedmemdim.x*sharedmemdim.y*sharedmemdim.z; + int updateperthread = ceil((float)(sharedmemdim.x*sharedmemdim.y*sharedmemdim.z)/(float)(blockWidth*blockHeight*blockDepth)); + float neighborsize = (2*SearchW+1)*(2*SearchW+1)*(2*kernel_depth+1); + + padXY = SearchW + 2*SimilW; /* padding sizes */ + + SearchW_full = 2*SearchW + 1; /* the full searching window size */ + SimilW_full = 2*SimilW + 1; /* the full similarity window size */ + + /*allocate space for images on device*/ + checkCudaErrors( cudaMalloc((void**)&Ad,N*M*Z*sizeof(float)) ); + checkCudaErrors( cudaMalloc((void**)&Bd,N*M*Z*sizeof(float)) ); + /*allocate space for vectors on device*/ + if (dimension == 2) { + checkCudaErrors( cudaMalloc((void**)&Eucl_Vec_d,SimilW_full*SimilW_full*sizeof(float)) ); + checkCudaErrors( cudaMemcpy(Eucl_Vec_d,Eucl_Vec,SimilW_full*SimilW_full*sizeof(float),cudaMemcpyHostToDevice) ); + } + else { + checkCudaErrors( cudaMalloc((void**)&Eucl_Vec_d,SimilW_full*SimilW_full*SimilW_full*sizeof(float)) ); + checkCudaErrors( cudaMemcpy(Eucl_Vec_d,Eucl_Vec,SimilW_full*SimilW_full*SimilW_full*sizeof(float),cudaMemcpyHostToDevice) ); + } + + /* copy data from the host to device */ + checkCudaErrors( cudaMemcpy(Ad,A,N*M*Z*sizeof(float),cudaMemcpyHostToDevice) ); + + // Run CUDA kernel here + NLM_kernel<<>>(Ad, Bd, Eucl_Vec_d, M, N, Z, SearchW, SimilW, SearchW_real, SearchW_full, SimilW_full, padXY, h2, lambda, imagedim, griddim, kerneldim, sharedmemdim, updateperthread, neighborsize); + + checkCudaErrors( cudaPeekAtLastError() ); +// gpuErrchk( cudaDeviceSynchronize() ); + + checkCudaErrors( cudaMemcpy(B,Bd,N*M*Z*sizeof(float),cudaMemcpyDeviceToHost) ); + cudaFree(Ad); cudaFree(Bd); cudaFree(Eucl_Vec_d); +} diff --git a/main_func/regularizers_GPU/NL_Regul/NLM_GPU_kernel.h b/main_func/regularizers_GPU/NL_Regul/NLM_GPU_kernel.h new file mode 100644 index 0000000..bc9d4a3 --- /dev/null +++ b/main_func/regularizers_GPU/NL_Regul/NLM_GPU_kernel.h @@ -0,0 +1,6 @@ +#ifndef __NLMREG_KERNELS_H_ +#define __NLMREG_KERNELS_H_ + +extern "C" void NLM_GPU_kernel(float *A, float* B, float *Eucl_Vec, int N, int M, int Z, int dimension, int SearchW, int SimilW, int SearchW_real, float denh2, float lambda); + +#endif -- cgit v1.2.3