% ATX_SUP_PARTIAL distributed (multiGPU) backprojector that allows to split the full volume into smaller pieces % this allows to solve datasets much larger than memory of used GPU or % spread calculations over several GPUs % % volData = Atx_sup_partial(projData, cfg, vectors, split, varargin) % % Inputs: % **projData - array Nlayers x width_sinogram x Nangles of back-projected data % **cfg - config structure generated by ASTRA_initialize % **vectors - orientation of projections generated by ASTRA_initialize % **split 3 or 4 elements vector, [split X, split Y, split Z, split angle ] % **deformation_fields: 3x1 cell contaning 3D arrays of local deformation of the object % **verbose - verbose <= 0 : quiet, verbose : standard info , verbose = 2: debug % **use_shared_memory - true - share data between processed by shared memory, false = use matlab parfor distribution % **max_memory_blocks - maximal size of used share memory memory % **varargin - for additional parameters see the code and als the astra.Axt_partial function % *returns* % ++volData - backprojected volume % % recompile commands % (Linux, GCC 4.8.5) mexcuda -outdir private +astra/ASTRA_GPU_wrapper/ASTRA_GPU_wrapper.cu +astra/ASTRA_GPU_wrapper/util3d.cu +astra/ASTRA_GPU_wrapper/par3d_fp.cu +astra/ASTRA_GPU_wrapper/par3d_bp.cu % (Windows) mexcuda -outdir private ASTRA_GPU_wrapper\ASTRA_GPU_wrapper.cu ASTRA_GPU_wrapper\util3d.cu ASTRA_GPU_wrapper\par3d_fp.cu ASTRA_GPU_wrapper\par3d_bp.cu %*-----------------------------------------------------------------------* %|                                                                       | %|  Except where otherwise noted, this work is licensed under a          | %|  Creative Commons Attribution-NonCommercial-ShareAlike 4.0            | %|  International (CC BY-NC-SA 4.0) license.                             | %|                                                                       | %|  Copyright (c) 2017 by Paul Scherrer Institute (http://www.psi.ch)    | %|                                                                       | %|      Author: CXS group, PSI  | %*-----------------------------------------------------------------------* % You may use this code with the following provisions: % % If the code is fully or partially redistributed, or rewritten in another % computing language this notice should be included in the redistribution. % % If this code, or subfunctions or parts of it, is used for research in a % publication or if it is fully or partially rewritten for another % computing language the authors and institution should be acknowledged % in written form in the publication: “Data processing was carried out % using the “cSAXS matlab package” developed by the CXS group, % Paul Scherrer Institut, Switzerland.” % Variations on the latter text can be incorporated upon discussion with % the CXS group if needed to more specifically reflect the use of the package % for the published work. % % A publication that focuses on describing features, or parameters, that % are already existing in the code should be first discussed with the % authors. % % This code and subroutines are part of a continuous development, they % are provided “as they are” without guarantees or liability on part % of PSI or the authors. It is the user responsibility to ensure its % proper use and the correctness of the results. function volData = Atx_sup_partial(projData, cfg, vectors, split, varargin) import utils.* import math.* par = inputParser; par.addOptional('deformation_fields', {}) % deformation_fields: 3x1 cell contaning 3D arrays of local deformation of the object par.addOptional('GPU', []) % list of GPUs to be used in reconstruction par.addOptional('split_sub', 1) % splitting of the sub block on smaller tasks in the Atx_partial method , 1 == no splitting , otherwise [split_x,split_y,split_z,split_angles] par.addOptional('verbose', 1) % verbose = 0 : quiet, verbose : standard info , verbose = 2: debug par.addOptional('use_shared_memory', []) % true - share data between processed by shared memory, false = use matlab parfor distribution par.addOptional('max_memory_blocks', min(50e9, utils.check_available_memory*1e6/4)) % maximal size of used share memory memory par.parse(varargin{:}) r = par.Results; if isscalar(split) split = split .* ones(1,3); end split_orig = split; if isempty(r.use_shared_memory) r.use_shared_memory = length(r.GPU) > 1; end if isscalar(r.split_sub) r.split_sub = r.split_sub .* ones(1,3); end gpu = gpuDevice; if isempty(r.GPU) r.GPU = gpu.Index; end N_GPU = length(r.GPU); if ~isempty(r.deformation_fields) % deformation field splitting not implemneted for split_sub; split(1:3) = split(1:3) .* r.split_sub(1:3); r.split_sub = 1; end %% if not splitting on this level is requirested, continue to tomo.Atx_partial if all(split(1:3) == 1) && N_GPU == 1 && gpu.AvailableMemory > 4*(numel(projData)/10+cfg.iVolX*cfg.iVolY*cfg.iVolZ) volData = astra.Atx_partial(projData, cfg, vectors, r.split_sub, ... 'GPU', r.GPU, 'deformation_fields', r.deformation_fields, 'verbose', r.verbose); return elseif all(split(1:2) == 1) && isempty(r.deformation_fields) && ... utils.check_available_memory*1e6 > 4*(cfg.iVolX*cfg.iVolY*cfg.iVolZ*N_GPU+numel(projData)) && ... cfg.iVolX*cfg.iVolY*cfg.iVolZ*4 < min(4*double(intmax('int32')),gpu.TotalMemory/2) %% if no splitting is needed and the volume is small enough then at least split the data on multiple GPUs by angles volData = Atx_angle_split(projData, cfg, vectors, r.split_sub, r.GPU, r.verbose); return end %% otherwise do proper checking of all inputs before splitting if ~(isa(projData, 'gpuArray') && strcmp(classUnderlying(projData), 'single')) && ... ~isa(projData, 'single') error('Only single precision input array supported') end projData = gather(projData); if any(cfg.pixel_scale ~= 1) error('Non integer pixel size is not implemented, try to use Atx_partial') end if cfg.skewness_angle ~= 0 error('Skewness is not working well with sup-split') end if length(cfg.pixel_scale) == 2 && cfg.pixel_scale(1) ~= cfg.pixel_scale(2) error('Variable pixel size for each axis is not implemented') end Nvol_orig = [cfg.iVolX,cfg.iVolY,cfg.iVolZ]; split(3) = max(split(3), ceil(split(3) * ( 4*(numel(projData)/5+cfg.iVolX*cfg.iVolY*cfg.iVolZ*(1+any(r.split_sub>1)) ) / gpu.AvailableMemory / prod(split) ))); split(3) = max(split(3), ceil(split(3) * ( ( cfg.iVolX*cfg.iVolY*cfg.iVolZ)/prod(split(1:3)) / double(intmax('int32')) ))); % make it equaly splitable among the GPUs split(3) = ceil(max(split(3), ceil(prod(split(1:3)) / N_GPU) * N_GPU) / prod(split(1:2))); split(1:3) = min(split(1:3), Nvol_orig(1:3)); % adjust splitting to make get equal blocks of the volume for i = 1:100 split(1:3) = ceil(Nvol_orig ./ floor(Nvol_orig ./ split(1:3))); end if prod(r.split_sub) > 1 % reduce the split_sub if possible r.split_sub(3) = ceil(r.split_sub(3) / 2^nextpow2(prod(split) / prod(split_orig)) ); end if length(split) < 4 split(4) = 1; end % initial parameters check Nvol_sub = Nvol_orig./split(1:3); Nproj_full = size(projData); assert(all(mod(Nvol_sub,1)==0), sprintf('Volume array cannot be divided to %i %i %i cubes', split(1:3))) assert(all(Nproj_full==[cfg.iProjV,cfg.iProjU,cfg.iProjAngles]), 'Wrong inputs size') assert(all(size(vectors)==[cfg.iProjAngles,12]), 'Wrong vectors size') % size of the subprojection of single subvolume Nproj_sub = [ (Nvol_sub(3)* sind(cfg.lamino_angle) + sqrt(sum(Nvol_sub(1:2).^2))*cosd(cfg.lamino_angle)), ... sqrt(sum(Nvol_sub(1:2).^2))]; % adjust sub projection size to account for inplane rotation of the geometry if cfg.tilt_angle ~= 0 Nproj_rot = [cosd(cfg.tilt_angle), -sind(cfg.tilt_angle); +sind(cfg.tilt_angle), cosd(cfg.tilt_angle)] * [0,0; Nproj_sub(1:2)]; % calculate projection window size after rotation Nproj_rot = max(Nproj_rot) - min(Nproj_rot); % add some extra padding Nproj_sub(1:2) = 2*(Nproj_rot - Nproj_sub(1:2)) + Nproj_sub(1:2); end % provide extra space for subpixel (linear) interpolation at the borders of the split volumes, needed only for noninteger CoR_offset Nproj_sub = Nproj_sub + 2*([split(3), max(split([1,2]))] - 1); Nproj_sub = ceil(Nproj_sub/16)*16; % make it easier splitable for ASTRA if cfg.lamino_angle == 90 Nproj_sub(1) = min(Nproj_sub(1), Nproj_full(1)); end % only if there is not split in the horizontal dimension if all(split(1:2) == 1) % do not take larger than size of the inputs Nproj_sub = min(Nproj_sub, Nproj_full(1:2)); end % avoid low RAM issues due to paralelization split(4) = max(split(4), ceil(4*prod(Nproj_sub)*cfg.iProjAngles/split(4)/(r.max_memory_blocks/N_GPU))); Nproj_sub(3) = ceil(cfg.iProjAngles/split(4)); cfg_small = cfg; cfg_small.iVolX = Nvol_sub(1); cfg_small.iVolY = Nvol_sub(2); cfg_small.iVolZ = Nvol_sub(3); % get new size of projections cfg_small.iProjU = Nproj_sub(2); cfg_small.iProjV = Nproj_sub(1); % calculate and store offset of the center of rotation, it will be used later offset = vectors(:,4:6) +(vectors(:,10:12).*cfg.iProjV/2+vectors(:,7:9).*cfg.iProjU/2 ); CoR_offset = -[dot(offset', vectors(:,10:12)') ./ dot(vectors(:,10:12)', vectors(:,10:12)'); dot(offset', vectors(:,7:9)') ./ dot(vectors(:,7:9)', vectors(:,7:9)')]' ; % remove centering offset && apply new offset shift_vec = vectors(:,10:12)*(cfg.iProjV/2-cfg_small.iProjV/2)+vectors(:,7:9)*(cfg.iProjU/2-cfg_small.iProjU/2); vectors(:,4:6) = vectors(:,4:6) + shift_vec; %% keep the volume in RAM , transfer only sub-blocks Nblocks = prod(split); %% write back to the preallocated shared array volData = zeros(Nvol_orig, 'single'); if N_GPU > 1 poolobj = gcp('nocreate'); if isempty(poolobj) || poolobj.NumWorkers < N_GPU delete(poolobj); poolobj = parpool(N_GPU); end poolobj.IdleTimeout = 600; % set idle timeout to 10 hours end % when the function is finished, make sure to execute following code global status status = true; if r.use_shared_memory out = onCleanup(@()myCleanupFun()); end % run blocks in series % run sub-blocks on each GPU in parallel t_total = tic(); clear output %% START OF OUTER GPU LOOP outputs_blocks = []; %% unitialize one solver per GPU for thread_id = 1:N_GPU % parse inputs and try to split them if possible [outputs_blocks, inputs_block{thread_id},cfg_all{thread_id}] = ... submit_block(thread_id, thread_id, outputs_blocks, projData, cfg, cfg_small, vectors,CoR_offset, split, Nproj_sub,Nvol_sub, r, varargin{:} ); end unprocessed_blocks = N_GPU+1:Nblocks; %% merge blocks back from GPUs and write to the shared array volData for ii = 1:Nblocks if r.verbose; utils.progressbar(ii, Nblocks); end % set values from the small blocks to the final output arrays [thread_id, timing, id] = gather_block( outputs_blocks,volData, cfg_all, split(4) > 1); if ~isempty(unprocessed_blocks) block_id = unprocessed_blocks(1); unprocessed_blocks(1) = []; if isa(outputs_blocks, 'parallel.FevalFuture') && sum([outputs_blocks.Read]) ~= 1 outputs_blocks keyboard end % submit a new job once the previous is finished [outputs_blocks, inputs_block{thread_id},cfg_all{block_id}] = ... submit_block(block_id, thread_id, outputs_blocks, projData, cfg, cfg_small, vectors, CoR_offset, split, Nproj_sub,Nvol_sub, r, varargin{:} ); if isa(outputs_blocks, 'parallel.FevalFuture') && any(cat(1,[outputs_blocks.Read])) outputs_blocks keyboard end end end if r.verbose > 1 fprintf('Timing system: GPU init %3.2gs shared_mem down %3.2gs upload on GPU %3.2gs tomo projection %3.2gs download from GPU %3.2gs shared_mem %3.2gs \n',sum(timing,2) ) if length(r.GPU) > 1 fprintf('Timing local: GPU init %3.2gs shared_mem down %3.2gs upload on GPU %3.2gs tomo projection %3.2gs download from GPU %3.2gs shared_mem %3.2gs \n ',sum(timing,2)/max(1,length(r.GPU)) ) fprintf('Total time %3.2fs, parfor overhead %3.2fs \n', t_total, t_total - sum(sum(timing,2)/max(1,length(r.GPU))) ) end end % finish GPU blocks % everything was fine -> no cleaning needed status = false; end function [outputs_blocks,inputs_block, cfg_out] = submit_block(block_id, thread_id, outputs_blocks, projData, cfg, cfg_small, vectors, CoR_offset,split, Nproj_sub,Nvol_sub, r, varargin ) % prepare blocks for asynchonous processing try inputs_block = prepare_block(block_id, projData, cfg, cfg_small, vectors,CoR_offset, split, Nproj_sub,Nvol_sub, r, varargin{:}); catch err disp(getReport(err)) keyboard end N_GPU = length(r.GPU); if isempty(outputs_blocks); clear outputs_blocks; end cfg_out = inputs_block{2}; try %% process preloaded data % no paralel toolbox if N_GPU <= 1 [outputs_blocks{thread_id}.volData_small,outputs_blocks{thread_id}.timing, outputs_blocks{thread_id}.id]=... run_partial_projector(inputs_block, block_id, 1,r.GPU,0); else % run it asynchronously if r.verbose > 3 ticBytes(gcp); end outputs_blocks(thread_id) = parfeval(@run_partial_projector, 3, inputs_block,block_id,thread_id, r.GPU,0); if r.verbose > 3 try; tocBytes(gcp); end end end catch err disp(getReport(err)) utils.check_available_memory keyboard end end function myCleanupFun() % destroy all shared memory that could have been left behind !ipcs -m | cut -d' ' -f2 | grep '^[0-9]' | while read x; do ipcrm -m $x; done end function [thread_id, timing, id] = gather_block(output_package,volData, cfg_all, add_values ) if isa(output_package, 'parallel.FevalFuture') % gather results from cluster , WAIT FOR CALCULATIONS TO BE FINISHED % [~, outputs_block, id] = fetchNext(output_package); %% my version of the fetchNext function, it seems faster id = []; assert(any(~[output_package.Read]), 'All blocks are already read') while true for thread_id =1:length(output_package) if strcmpi(output_package(thread_id).State, 'finished') && output_package(thread_id).Read == 0 try [volData_small,timing,id] = output_package(thread_id).fetchOutputs; catch err if strcmpi(err.identifier, 'parallel:fevalqueue:InvalidExecutionResult') warning('Unknown error, trying to restart parpool') delete(gcp('nocreate')); end if strcmpi(err.identifier, 'parallel:fevalqueue:InvalidExecutionResult') warning('Unknown error, trying to restart parpool') delete(gcp('nocreate')); end if ~isempty(output_package(thread_id).Diary) fprintf('============ THREAD %i FAILED, OUTPUT: ============= \n', thread_id) disp(output_package(thread_id).Diary) end fprintf('============ THREAD %i FAILED, ERROR: ============= \n', thread_id) disp(getReport(output_package(thread_id).Error)) keyboard output_package.cancel rethrow(err) end break end end if ~isempty(id); break; end pause(0.01) % wait for the data to be prepared end elseif iscell(output_package) thread_id = 1; id = output_package{thread_id}.id; volData_small = output_package{thread_id}.volData_small; timing = output_package{thread_id}.timing; else disp('FAILED ?? ') keyboard end if isempty(volData_small) warning('ASTRA projection probably failed') keyboard end if isa(volData_small, 'shm') % load data from shared memory [s,volData_small] = volData_small.attach; s.protected = false; % release shared memory elseif ~isnumeric(volData_small) keyboard end % write back to the full array stored in RAM positions = zeros(size(volData_small,3),2)+cfg_all{id}.volume_shift(1:2); indices = (1:size(volData_small,3)) + cfg_all{id}.volume_shift(3); % use a MEX code to speed it up utils.add_to_3D_projection(volData_small, volData,positions,indices, true, false); end function prepared_block = prepare_block(id, projData, cfg, cfg_small, vectors,CoR_offset, split, Nproj_sub, Nvol_sub ,r, varargin) [x,y,z,angle_block_id] = ind2sub(split,id); block_pos = [x,y,z]; shift = zeros(3,1); for n = 1:3 %% find optimal shift of the subvolume if mod(split(n),2)==1 %% odd shift(n) = (block_pos(n) - ceil(split(n)/2))*Nvol_sub(n); else shift(n) = (block_pos(n) - split(n)/2-1/2)*Nvol_sub(n); end end Nangle_per_blocks = ceil(cfg.iProjAngles / split(4)); %% for splitting to angular blocks using close angles, it makes ASTRA faster angle_ids = (1+(angle_block_id-1)*Nangle_per_blocks:min(cfg.iProjAngles, angle_block_id*Nangle_per_blocks)); cfg_small.iProjAngles = length(angle_ids); Nproj_sub(3) = cfg_small.iProjAngles; %% find optimal shift of the subvolume deform_fields_small = get_subdeform_fields(block_pos, r.deformation_fields, split); CoR_offset = [CoR_offset, zeros(cfg.iProjAngles,1)] ; % shift the sub-projections off center to create single % large projection after assembling, if shift == 0 => % projections will be rotationally centered vec = vectors; % apply optimal shift projection_shift = cfg.pixel_scale(1)^2.*[vec(:,10:12)*shift, vec(:,7:9)*shift, zeros(cfg.iProjAngles,1)]; projection_shift = bsxfun(@plus, projection_shift, [cfg.iProjV/2-cfg_small.iProjV/2,cfg.iProjU/2-cfg_small.iProjU/2,0] + CoR_offset); % calculate subpixel shifts projection_shift_subpix = projection_shift - round(projection_shift) - CoR_offset; projection_shift = round(projection_shift); % apply subpixel shifts vec(:,4:6) = vec(:,4:6) - ... ( bsxfun(@times,vectors(:,7:9),projection_shift_subpix(:,2))+ ... bsxfun(@times,vectors(:,10:12),projection_shift_subpix(:,1))); % just store shifts for later cfg_small.volume_shift = (block_pos-1).*Nvol_sub; cfg_small.projection_shift = projection_shift; % move data using custom made MEX routine % allocate a small sub array projData_small = zeros(Nproj_sub, 'single'); if r.use_shared_memory s = shm(); try s.allocate(projData_small) % attach the shared memory catch keyboard end [s, projsmall_shm] = s.attach(); % === write data ===== % use self-made MEX OMP function to move the data s utils.get_from_3D_projection(projsmall_shm, projData, projection_shift(angle_ids,1:2), angle_ids); % detach the shared memory projData_small = s; s.detach; else % use custom made MEX OMP function to move the data utils.get_from_3D_projection(projData_small, projData, projection_shift(angle_ids,1:2), angle_ids); end % return only vector for the use angles vec = vec(angle_ids,:); prepared_block = {projData_small, cfg_small, vec, r.split_sub, varargin{:}, ... 'deformation_fields', deform_fields_small, 'verbose', 0, 'GPU', []}; end function [vol, timing,block_id] = run_partial_projector(prepared_block, block_id, thread_id, GPU_list, verbose) t0 = tic; gpu_id = GPU_list(thread_id); % let matlab to choose which GPU use gpu = gpuDevice(); if ~isempty(GPU_list) && gpu.Index ~= gpu_id gpuDevice(gpu_id); % avoid unneeded initalization end t_init = toc(t0); timing = [t_init, 0,0,0,0,0]; t = tic; if isa(prepared_block{1}, 'shm') % data are attached to shared memory [s,projData_small] = prepared_block{1}.attach(); else % data are given directly to the worker projData_small = prepared_block{1}; end timing(2) = toc(t); t = tic; timing(3) = toc(t); % call the next level abstraction t = tic; is_remote = ~isempty(getCurrentTask()); % call the next level abstraction around ASTRA wrapper try for ii = 1:2 try vol = astra.Atx_partial(projData_small, prepared_block{2:end},'verbose', max(is_remote, verbose)); break catch err if ii == 2 rethrow(err) else warning('Atx_partial failed, trying again ... ') end end end catch err qkeyboard gpu = gpuDevice reset(gpu) fprintf('Error on GPU %i / %i', gpu.Index, gpuDeviceCount) disp( getReport(err, 'extended', 'hyperlinks', 'on')) % projData_small = single([]); rethrow(err) end %%%%%%%%%%% timing(4) = toc(t); t = tic; vol = gather(vol); % move to RAM timing(5) = toc(t); t = tic; if isscalar(prepared_block{1}) % data are distributed to shared memory s.detach(); s = shm(true); s.upload(vol); vol = s; end timing(6) = toc(t); end function deformation_fields_sub = get_subdeform_fields(block_pos, deform_fields, split) % crop the deformation field only int the region of interest if isempty(deform_fields) deformation_fields_sub = {}; return end if prod(split)==1 deformation_fields_sub = deform_fields; return end for ii = 1:3 N_full = size(deform_fields{1,ii}); N_small = ceil( N_full ./ reshape(split(1:3),[],1)'); for kk = 1:3 ind_def{kk} = (1+(block_pos(kk)-1)*N_small(kk)) : min(N_full(kk), (split(kk))*N_small(kk)); end for jj = 1:2 deformation_fields_sub{jj,ii} = deform_fields{jj,ii}(ind_def{:}); end end warning('Splitting of deformation field is not supported / recommended') end function volData = Atx_angle_split(projData, cfg, vectors, split, GPU, verbose) %% simplified version for multiGPU reconstruction of volumes that are small enough % simply split the projections by angles and apply each projection % block on a separated GPU Nangles = cfg.iProjAngles; N_GPU = length(GPU); Nblocks = ceil(Nangles / N_GPU); if length(split) == 4 split(4) = max(1, split(4) / N_GPU); end if verbose; utils.progressbar(1,4); end for id = 1:N_GPU ind{id} = 1+(id-1)*Nblocks:min(id*Nblocks,Nangles); projData_shm{id} = shm(); projData_shm{id} = tomo.get_from_array(projData, projData_shm{id}, ind{id}); volData_shm{id} = []; end if verbose; utils.progressbar(2,4); end parfor(id = 1:N_GPU, N_GPU) t = getCurrentTask(); gpu_id = GPU(1+mod(t.ID-1, N_GPU)); % let parfor to choose which GPU use gpu = gpuDevice(); if ~isempty(GPU) && gpu.Index ~= gpu_id gpuDevice(gpu_id); % avoid unneeded initalization end [s,projData_block] = projData_shm{id}.attach(); volData_block = astra.Atx_partial(projData_block, cfg, vectors(ind{id},:),split, 'keep_on_GPU', true, 'verbose',0); volData_block = gather(volData_block); s.upload(volData_block); s.protected = true; volData_shm{id} = s; end if verbose; utils.progressbar(3,4); end volData = single(0); for id = 1:N_GPU [s,volData_block] = volData_shm{id}.attach(); volData = volData + volData_block; s.free; end if verbose; utils.progressbar(4,4); end end