From fba7cbdba2438ed73749cb09b5f28eb5f279220c Mon Sep 17 00:00:00 2001 From: "silas (home)" Date: Sun, 1 Feb 2026 13:53:20 +0100 Subject: [PATCH] Faster DP_fiber with GPU processing... Run test\gpu_cpu_comparison.m to see difference on your setup --- Classes/02_optical/DP_Fiber.m | 19 +- Classes/02_optical/Optical_Demultiplex.m | 45 +-- Classes/02_optical/Optical_Multiplex.m | 45 ++- Classes/02_optical/dp_fiber_lib/CNLSE_plain.m | 189 ++++++---- Classes/02_optical/dp_fiber_lib/canUseGPU.m | 20 + .../02_optical/dp_fiber_lib/getNLstepsize.m | 40 +- Classes/02_optical/dp_fiber_lib/lin_step.m | 64 ++-- Classes/DataBaseHandler/DBHandler.m | 13 +- .../Dissertation/mach_zehnder_modulator.m | 153 ++++++-- .../Auswertung_JLT/run_dsp_from_db.m | 5 +- projects/WDM/WDM_model_10km_queue.m | 2 +- test/bayesopt_ffe_tuning.m | 219 +++++++++++ test/gpu_cpu_comparison.m | 349 ++++++++++++++++++ test/gpu_processing_dpfiber.m | 265 +++++++++++++ test/gpu_processing_test.m | 70 ++++ 15 files changed, 1281 insertions(+), 217 deletions(-) create mode 100644 Classes/02_optical/dp_fiber_lib/canUseGPU.m create mode 100644 test/bayesopt_ffe_tuning.m create mode 100644 test/gpu_cpu_comparison.m create mode 100644 test/gpu_processing_dpfiber.m create mode 100644 test/gpu_processing_test.m diff --git a/Classes/02_optical/DP_Fiber.m b/Classes/02_optical/DP_Fiber.m index a5f3dce..c1f08ea 100644 --- a/Classes/02_optical/DP_Fiber.m +++ b/Classes/02_optical/DP_Fiber.m @@ -24,6 +24,8 @@ classdef DP_Fiber SS_dzmax % [m] max dz (adaptive SSFM) SS_dzmin % [m] min dz (adaptive SSFM) n_waveplates % number of PMD waveplates + useGPU % GPU acceleration: true, false, or 'auto' (default) + useSingle % Use single precision on GPU (default: false) % ---- Internal state (persistent between calls) ---- state % struct mirroring legacy 'state' @@ -56,6 +58,8 @@ classdef DP_Fiber options.SS_dzmax = 2e4 % m options.SS_dzmin = 100 % m options.n_waveplates = 100 + options.useGPU = 'auto' % 'auto', true, or false + options.useSingle = false % single precision GPU end % Copy provided options into properties @@ -208,7 +212,7 @@ classdef DP_Fiber % Frequency-dependent PMD phase term (legacy form) st.brf.db0 = (R.rand(st.wave_plates,1)*2*pi - pi) * brf_multiplier; st.brf.db1 = sqrt(3*pi/8)*(st.dgd/obj.fa)/st.wave_plates .* st.omega; - st.brf.simdgd = 0; + st.brf.simdgd = 0; % cumsum used in legacy only for debug; keep compatibility variable: ~cumsum(st.brf.db0); % no-op to mirror legacy path @@ -228,7 +232,18 @@ classdef DP_Fiber x_in = signal_in(:,1).'; y_in = signal_in(:,2).'; - [x_out, y_out, obj.state] = CNLSE_plain(x_in, y_in, obj.state); + % Determine GPU usage + if ischar(obj.useGPU) || isstring(obj.useGPU) + if strcmpi(obj.useGPU, 'auto') + gpuFlag = []; % Let CNLSE_plain auto-detect + else + error('DP_Fiber:InvalidGPU', 'useGPU must be true, false, or ''auto'''); + end + else + gpuFlag = logical(obj.useGPU); + end + + [x_out, y_out, obj.state] = CNLSE_plain(x_in, y_in, obj.state, gpuFlag, obj.useSingle); obj.state.propagated_length = obj.state.propagated_length + obj.state.L; diff --git a/Classes/02_optical/Optical_Demultiplex.m b/Classes/02_optical/Optical_Demultiplex.m index 7d7c39e..8e272d2 100644 --- a/Classes/02_optical/Optical_Demultiplex.m +++ b/Classes/02_optical/Optical_Demultiplex.m @@ -42,7 +42,7 @@ classdef Optical_Demultiplex < handle function signalclasses_out = process(obj, signalclass_in) - % ---- Infer wavelength: either given or from input total signal + % ---- Infer wavelength: either given or from input total signal if isempty(obj.wavelengthplan) obj.wavelengthplan = signalclass_in.lambda; %meter else @@ -81,7 +81,7 @@ classdef Optical_Demultiplex < handle obj signal_in end - + w = obj.fs_out ./ obj.fs_in ; blocklen_in = length(signal_in); blocklen_out = w*blocklen_in; @@ -119,30 +119,31 @@ classdef Optical_Demultiplex < handle N = size(lo,1); C = size(lo,2); - x_envelopes = zeros(N, C, 'like', signal_in); - y_envelopes = zeros(N, C, 'like', signal_in); + % ---- VECTORIZED: Process all channels in parallel ---- + % Batched FFT operates on each column simultaneously on GPU - s1 = signal_in(:,1); - s2 = signal_in(:,2); + % Extract polarization signals + s1 = signal_in(:,1); % X polarization [N×1] + s2 = signal_in(:,2); % Y polarization [N×1] - % Reusable work buffers (avoid reallocations) - wrk_time = zeros(N,1, 'like', signal_in); - wrk_freq = zeros(N,1, 'like', signal_in); + % Broadcast signal to all channels and multiply with LO + % s1, s2 are [N×1], lo is [N×C] → result is [N×C] + x_mixed = att .* s1 .* lo; % [N×C] + y_mixed = att .* s2 .* lo; % [N×C] + + % Batched FFT: each column computed in parallel + x_freq = fft(x_mixed); % [N×C] + y_freq = fft(y_mixed); % [N×C] + + % Apply filter (H is [N×1], broadcasts across columns) + x_filtered = x_freq .* H; % [N×C] + y_filtered = y_freq .* H; % [N×C] + + % Batched IFFT + x_envelopes = ifft(x_filtered); % [N×C] + y_envelopes = ifft(y_filtered); % [N×C] - for c = 1:C - % ---- X branch ---- - wrk_time(:) = att .* s1 .* lo(:,c); % N×1 - wrk_freq(:) = fft(wrk_time); % N×1 - wrk_freq(:) = wrk_freq .* H; % N×1 - x_envelopes(:,c) = ifft(wrk_freq); % N×1 - % ---- Y branch ---- - wrk_time(:) = att .* s2 .* lo(:,c); - wrk_freq(:) = fft(wrk_time); - wrk_freq(:) = wrk_freq .* H; - y_envelopes(:,c) = ifft(wrk_freq); - end - end end end diff --git a/Classes/02_optical/Optical_Multiplex.m b/Classes/02_optical/Optical_Multiplex.m index 1d722b0..7fc8a33 100644 --- a/Classes/02_optical/Optical_Multiplex.m +++ b/Classes/02_optical/Optical_Multiplex.m @@ -1,10 +1,10 @@ classdef Optical_Multiplex < handle % Takes a cell array of signals - % returns a total field signal - % WDM spacing is given in wavelength plan OR via delta_F - + % returns a total field signal + % WDM spacing is given in wavelength plan OR via delta_F + % The grid is stored in the output signal -> the demux will ideally - % look this up and use this as the demux frequencies... + % look this up and use this as the demux frequencies... % signal_cell = {Opt_sig_1, Opt_sig_2}; % Opt_sig_wdm = Optical_Multiplex("fs_in",Opt_sig.fs,"fs_out",4*Opt_sig.fs,... @@ -117,7 +117,7 @@ classdef Optical_Multiplex < handle % adapt frequency shifts to match the FFT grid! Find nearest grid point [glitch(o),pos] = min(abs( freqaxis-obj.df_T(o) )); obj.df_T(o) = freqaxis(pos); - + polrots = [polrots, data_in{o}.polrot]; end @@ -139,25 +139,36 @@ classdef Optical_Multiplex < handle x_envelopes = NaN([blocklen_out N]); y_envelopes = x_envelopes; - + + % ---- OPTIMIZED: Pre-compute all LO phases as [blocklen_out × N] matrix ---- + time_idx = (0:blocklen_out-1).'; % [blocklen_out × 1] + lo_phases = mod(2*pi * time_idx * obj.df_T / obj.fs_out, 2*pi); % [blocklen_out × N] + lo_all = cos(lo_phases) + 1i*sin(lo_phases); % [blocklen_out × N] + + % Collect all resampled signals first (still requires loop due to cell array) + x_signals = zeros(blocklen_out, N); + y_signals = zeros(blocklen_out, N); + for o = 1:N - - pha = mod(2*pi*(0:blocklen_out-1)*obj.df_T(o)/obj.fs_out,2*pi).'; - lo = cos(pha)+1i*sin(pha); - data_in_resampled = data_in{o}.resample("fs_out",obj.fs_out); - - res_env = ifft(fft(data_in_resampled.signal(:,1)).*H); - x_envelopes(:,o) = att.*res_env.*lo; - - res_env = ifft(fft(data_in_resampled.signal(:,2)).*H); - y_envelopes(:,o) = att.*res_env.*lo; - + data_in_resampled = data_in{o}.resample("fs_out", obj.fs_out); + x_signals(:, o) = data_in_resampled.signal(:, 1); + y_signals(:, o) = data_in_resampled.signal(:, 2); end + % ---- VECTORIZED: Batched FFT/IFFT for all channels ---- + % Apply filter to all channels at once + x_filtered = ifft(fft(x_signals) .* H); % [blocklen_out × N] + y_filtered = ifft(fft(y_signals) .* H); % [blocklen_out × N] + + % Apply attenuation and LO shift to all channels + x_envelopes = att .* x_filtered .* lo_all; % [blocklen_out × N] + y_envelopes = att .* y_filtered .* lo_all; % [blocklen_out × N] + data_out = data_in_resampled; data_out.signal = [sum(x_envelopes,2), sum(y_envelopes,2)]; data_out.lambda = obj.lambda_T; data_out.polrot = polrots; + end end diff --git a/Classes/02_optical/dp_fiber_lib/CNLSE_plain.m b/Classes/02_optical/dp_fiber_lib/CNLSE_plain.m index cfa7294..463adca 100644 --- a/Classes/02_optical/dp_fiber_lib/CNLSE_plain.m +++ b/Classes/02_optical/dp_fiber_lib/CNLSE_plain.m @@ -1,41 +1,72 @@ -function [opt_out_x,opt_out_y,state] = CNLSE_plain(opt_in_x,opt_in_y,state) - +function [opt_out_x,opt_out_y,state] = CNLSE_plain(opt_in_x,opt_in_y,state,useGPU,useSingle) +% GPU auto-detection if not specified +if nargin < 4 || isempty(useGPU) + useGPU = canUseGPU(); +end +if nargin < 5 || isempty(useSingle) + useSingle = false; +end - %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% - % pre calculations - %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +% pre calculations +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% - state.common_beta=struct('X',0,'Y',0); +state.common_beta=struct('X',0,'Y',0); - for n=1:2 - % get current polarization name and contrary one - curPol = state.polNames{n}; +for n=1:2 + % get current polarization name and contrary one + curPol = state.polNames{n}; - % extend linear transfer function depending on beta values for the current polarization - % Was ist der Sinn dieser komischen beta notation? zB. state.beta.X = [0.3142 0 -9.1105e-28 5.1068e-41] - for n_beta = 1:length(state.beta.(curPol)) - state.common_beta.(curPol) = state.common_beta.(curPol) + state.beta.(curPol)(n_beta) * (state.omega).^(n_beta-1) / factorial(n_beta-1); - end - - %opt_out_struct.(curPol)=opt_in_struct.(curPol).envelope; + % extend linear transfer function depending on beta values for the current polarization + % Was ist der Sinn dieser komischen beta notation? zB. state.beta.X = [0.3142 0 -9.1105e-28 5.1068e-41] + for n_beta = 1:length(state.beta.(curPol)) + state.common_beta.(curPol) = state.common_beta.(curPol) + state.beta.(curPol)(n_beta) * (state.omega).^(n_beta-1) / factorial(n_beta-1); end - beta_const = state.beta.('X')(1); - beta_1 = state.beta.('X')(2); - beta_2 = state.beta.('X')(3); - beta_3 = state.beta.('X')(4); - deltaomega = state.omega; - beta_x = beta_const + beta_1 * deltaomega + 1/2 * beta_2 * deltaomega.^2 + 1/6 *beta_3 * deltaomega.^3; - -% opt_in_x = gpuArray(opt_in_x); -% opt_in_y = gpuArray(opt_in_y); + %opt_out_struct.(curPol)=opt_in_struct.(curPol).envelope; +end - - %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% - % Split Step Method - %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +beta_const = state.beta.('X')(1); +beta_1 = state.beta.('X')(2); +beta_2 = state.beta.('X')(3); +beta_3 = state.beta.('X')(4); +deltaomega = state.omega; +beta_x = beta_const + beta_1 * deltaomega + 1/2 * beta_2 * deltaomega.^2 + 1/6 *beta_3 * deltaomega.^3; + +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +% GPU Transfer (if enabled) +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +if useGPU + % Convert to single precision if requested (faster on most GPUs) + if useSingle + opt_in_x = gpuArray(single(opt_in_x)); + opt_in_y = gpuArray(single(opt_in_y)); + state.common_beta.X = gpuArray(single(state.common_beta.X)); + state.common_beta.Y = gpuArray(single(state.common_beta.Y)); + state.brf.db1 = gpuArray(single(state.brf.db1)); + state.brf.db0 = gpuArray(single(state.brf.db0)); + for k = 1:numel(state.brf.matR) + state.brf.matR{k} = gpuArray(single(state.brf.matR{k})); + end + else + opt_in_x = gpuArray(opt_in_x); + opt_in_y = gpuArray(opt_in_y); + state.common_beta.X = gpuArray(state.common_beta.X); + state.common_beta.Y = gpuArray(state.common_beta.Y); + state.brf.db1 = gpuArray(state.brf.db1); + state.brf.db0 = gpuArray(state.brf.db0); + for k = 1:numel(state.brf.matR) + state.brf.matR{k} = gpuArray(state.brf.matR{k}); + end + end +end + + +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +% Split Step Method +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % [opt_out_x,opt_out_y] = split_step_loop(state.L,opt_in_x,opt_in_y,state.gamma,state.SS_dzmin,state.SS_dzmax,state.SS_dphimax,state.alpha_lin,... % state.lin_z_test,state.corr_length,state.n_plates_done,state.missing_dz,state.brf,state.common_beta,.... @@ -44,35 +75,33 @@ function [opt_out_x,opt_out_y,state] = CNLSE_plain(opt_in_x,opt_in_y,state) % [opt_out_x,opt_out_y] = split_step_loop_mex(state.L,opt_in_x,opt_in_y,state.gamma,state.SS_dzmin,state.SS_dzmax,state.SS_dphimax,state.alpha_lin,... % state.lin_z_test,state.corr_length,state.n_plates_done,state.missing_dz,state.brf,state.common_beta,.... % state.chi,state.manakov,state.beat_len); - - % get nonlinear step size - [state.dz] = getNLstepsize(opt_in_x,opt_in_y,state.gamma,state.SS_dzmin,state.SS_dzmax,state.SS_dphimax,state.alpha_lin); - %[state.dz] = getNLstepsize_original(state,opt_out_struct); - - state.n_step = 0; - state.z_prop = 0; - state.test_dz = []; - state.powers = []; - tic +% get nonlinear step size +[state.dz] = getNLstepsize(opt_in_x,opt_in_y,state.gamma,state.SS_dzmin,state.SS_dzmax,state.SS_dphimax,state.alpha_lin); +%[state.dz] = getNLstepsize_original(state,opt_out_struct); - while state.z_prop < state.L +state.n_step = 0; +state.z_prop = 0; +state.test_dz = []; +state.powers = []; - % reduce step length (dz) if we are to overshoot the fiber length - % (L) in the next step - if state.z_prop + state.dz > state.L - state.dz = state.L - state.z_prop; - end - - % update step number (n) - state.n_step=state.n_step+1; +while state.z_prop < state.L - % append current step length to logbook (dzs) - state.dzs(state.n_step)=state.dz; + % reduce step length (dz) if we are to overshoot the fiber length + % (L) in the next step + if state.z_prop + state.dz > state.L + state.dz = state.L - state.z_prop; + end - - % half linear step - [opt_in_x,opt_in_y,state.z_prop,state.lin_z_test,... + % update step number (n) + state.n_step=state.n_step+1; + + % append current step length to logbook (dzs) + state.dzs(state.n_step)=state.dz; + + + % half linear step + [opt_in_x,opt_in_y,state.z_prop,state.lin_z_test,... state.corr_length,state.n_plates_done,state.missing_dz,state.n_step,... state.test_plates,state.test_plate_numbers,state.brf,state.common_beta.X,... state.common_beta.Y,state.alpha_lin.X,state.alpha_lin.X]... @@ -82,12 +111,12 @@ function [opt_out_x,opt_out_y,state] = CNLSE_plain(opt_in_x,opt_in_y,state) state.test_plates,state.test_plate_numbers,state.brf,state.common_beta.X,... state.common_beta.Y,state.alpha_lin.X,state.alpha_lin.X); - % complete nonlinear step - - [opt_in_x,opt_in_y] = nl_step(opt_in_x,opt_in_y, state.dz, state.gamma, state.chi, state.manakov, state.beat_len ,state.alpha_lin.X, state.alpha_lin.Y); - - % half linear step - [opt_in_x,opt_in_y,state.z_prop,state.lin_z_test,... + % complete nonlinear step + + [opt_in_x,opt_in_y] = nl_step(opt_in_x,opt_in_y, state.dz, state.gamma, state.chi, state.manakov, state.beat_len ,state.alpha_lin.X, state.alpha_lin.Y); + + % half linear step + [opt_in_x,opt_in_y,state.z_prop,state.lin_z_test,... state.corr_length,state.n_plates_done,state.missing_dz,state.n_step,... state.test_plates,state.test_plate_numbers,state.brf,state.common_beta.X,... state.common_beta.Y,state.alpha_lin.X,state.alpha_lin.X]... @@ -97,24 +126,34 @@ function [opt_out_x,opt_out_y,state] = CNLSE_plain(opt_in_x,opt_in_y,state) state.test_plates,state.test_plate_numbers,state.brf,state.common_beta.X,... state.common_beta.Y,state.alpha_lin.X,state.alpha_lin.X); - % get nonlinear step size - [state.dz] = getNLstepsize(opt_in_x,opt_in_y,state.gamma,state.SS_dzmin,state.SS_dzmax,state.SS_dphimax,state.alpha_lin); - %[state.dz] = getNLstepsize_original(state,opt_out_struct); + % get nonlinear step size + [state.dz] = getNLstepsize(opt_in_x,opt_in_y,state.gamma,state.SS_dzmin,state.SS_dzmax,state.SS_dphimax,state.alpha_lin); + %[state.dz] = getNLstepsize_original(state,opt_out_struct); - + +end + + + +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +% GPU Gather (if enabled) +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +if useGPU + opt_out_x = gather(opt_in_x); + opt_out_y = gather(opt_in_y); + + % Gather state arrays back to CPU + state.common_beta.X = gather(state.common_beta.X); + state.common_beta.Y = gather(state.common_beta.Y); + state.brf.db1 = gather(state.brf.db1); + state.brf.db0 = gather(state.brf.db0); + for k = 1:numel(state.brf.matR) + state.brf.matR{k} = gather(state.brf.matR{k}); end - - toc +else + opt_out_x = opt_in_x; + opt_out_y = opt_in_y; +end - - - - - opt_out_x = (opt_in_x); - opt_out_y = (opt_in_y); - -% opt_out_x = gather(opt_in_x); -% opt_out_y = gather(opt_in_y); - end \ No newline at end of file diff --git a/Classes/02_optical/dp_fiber_lib/canUseGPU.m b/Classes/02_optical/dp_fiber_lib/canUseGPU.m new file mode 100644 index 0000000..84666bb --- /dev/null +++ b/Classes/02_optical/dp_fiber_lib/canUseGPU.m @@ -0,0 +1,20 @@ +function canUse = canUseGPU() +%CANUSEGPU Check if a compatible GPU is available for parallel computing +% Returns true if MATLAB Parallel Computing Toolbox is available and +% a CUDA-capable GPU is detected. + +canUse = false; + +% Check if Parallel Computing Toolbox is installed +if ~license('test', 'Distrib_Computing_Toolbox') + return; +end + +% Check for GPU device +try + gpu = gpuDevice(); + canUse = gpu.DeviceSupported; +catch + canUse = false; +end +end diff --git a/Classes/02_optical/dp_fiber_lib/getNLstepsize.m b/Classes/02_optical/dp_fiber_lib/getNLstepsize.m index a70edb9..fd42eba 100644 --- a/Classes/02_optical/dp_fiber_lib/getNLstepsize.m +++ b/Classes/02_optical/dp_fiber_lib/getNLstepsize.m @@ -1,25 +1,25 @@ function [rDZ] = getNLstepsize(ux,uy,gamma,dzmin,dzmax,dphimax,alpha_lin) - - maxPow = max(gamma.*max(real(ux).^2+imag(ux).^2+real(uy).^2+imag(uy).^2)); - - Leff = dphimax/maxPow; - alpha_lin = max([alpha_lin.X alpha_lin.Y]); - nl_att_len_ratio = alpha_lin*Leff; - - if nl_att_len_ratio >= 1 - rDZ = dzmax; +% gather() handles gpuArray inputs - ensures scalar is on CPU for comparisons +maxPow = gather(max(gamma.*max(real(ux).^2+imag(ux).^2+real(uy).^2+imag(uy).^2))); + +Leff = dphimax/maxPow; +alpha_lin = max([alpha_lin.X alpha_lin.Y]); +nl_att_len_ratio = alpha_lin*Leff; + +if nl_att_len_ratio >= 1 + rDZ = dzmax; +else + if alpha_lin == 0 + step = Leff; else - if alpha_lin == 0 - step = Leff; - else - %effective length? - step = -1/alpha_lin*log(1-nl_att_len_ratio); - end - - rDZ = min([step dzmax]); - rDZ = max([rDZ dzmin]); - end - + %effective length? + step = -1/alpha_lin*log(1-nl_att_len_ratio); + end + + rDZ = min([step dzmax]); + rDZ = max([rDZ dzmin]); +end + end \ No newline at end of file diff --git a/Classes/02_optical/dp_fiber_lib/lin_step.m b/Classes/02_optical/dp_fiber_lib/lin_step.m index eea8a1e..cdb3562 100644 --- a/Classes/02_optical/dp_fiber_lib/lin_step.m +++ b/Classes/02_optical/dp_fiber_lib/lin_step.m @@ -22,37 +22,37 @@ n_plates_left = n_plates - n_plates_done; % compute last plate size ( if it fits, it should be 0) if missing_dz > aStepSize - + last_plate = aStepSize; missing_dz = missing_dz-aStepSize; plate_sizes = last_plate; plate_numbers = n_plates; - + else - + last_plate = aStepSize - missing_dz - (n_plates_left-1)*corr_length; - + if missing_dz == 0 missing_dz = []; end - + %build vector of plate lengths with missing plate part from prev. %iterartion , then some normal plates and finally a fraction of a plate %to fit into the step length plate_sizes = [missing_dz corr_length*ones(1,n_plates_left-1) last_plate]; - + if n_plates_done == 0 plate_numbers =[(n_plates_done+1):(n_plates-1) n_plates]; else plate_numbers = [n_plates_done (n_plates_done+1):(n_plates-1) n_plates]; % not wrking yet end - - %remember for next step + + %remember for next step missing_dz = corr_length - last_plate; - + end -plate_steps = repmat(n_step,1,length(plate_sizes)); +plate_steps = repmat(n_step,1,length(plate_sizes)); %#ok % %figure;stem(plate_sizes); @@ -66,50 +66,39 @@ test_plate_numbers = [test_plate_numbers, plate_numbers]; opt_x=fft(opt_x); opt_y=fft(opt_y); -% db1 = gpuArray(brf.db1); -% db0 = gpuArray(brf.db0); -% common_beta_x = gpuArray(common_beta_x); -% common_beta_y = gpuArray(common_beta_y); - -db1 = (brf.db1); -db0 = (brf.db0); -common_beta_x = (common_beta_x); -common_beta_y = (common_beta_y); +% Note: db1, db0, common_beta_x, common_beta_y are already on GPU when +% useGPU=true (transferred in CNLSE_plain) +db1 = brf.db1; +db0 = brf.db0; % process every waveplate with given sizes in plate_sizes for n=1:length(plate_sizes) dz = plate_sizes(n); - - % figure(87);subplot(2,1,1);plot(real(x(900:1150)));subplot(2,1,2);plot(real(y(900:1150))); - % MOV1=[MOV1 getframe(87)]; - + % extract rotation matrix from pre calculated matrices matR = brf.matR{plate_numbers(n)}; - + % transform to eigenvalue of of fiber segment tOpt.X = conj(matR(1,1))*opt_x + conj(matR(2,1))*opt_y; tOpt.Y = conj(matR(1,2))*opt_x + conj(matR(2,2))*opt_y; - + % calculate statistical delta beta for pmd delta_beta = 0.5*(db1+db0(n))/corr_length; - % build transfer function with delta beta - %common.beta = beta1+beta2*omega^2 - + %accumulate delta beta for log... brf.simdgd = brf.simdgd + (db1(length(db1)/2+1)+db0(n))/corr_length; - + h.X = exp(-1j*(common_beta_x-delta_beta)*dz); h.Y = exp(-1j*(common_beta_y+delta_beta)*dz); - % delta_beta has to be added to the transfer function - + % process with transfer function tOpt.X = h.X.*tOpt.X ; tOpt.Y = h.Y.*tOpt.Y ; - + % rotate back opt_x = matR(1,1)*tOpt.X + matR(1,2)*tOpt.Y; opt_y = matR(2,1)*tOpt.X + matR(2,2)*tOpt.Y; - + end lin_z_test = lin_z_test + sum(plate_sizes,2); @@ -117,9 +106,8 @@ lin_z_test = lin_z_test + sum(plate_sizes,2); %update the number of processed plates so far n_plates_done = n_plates_done + n_plates_left; -% attanuate the signal each linear state with alpha -% ( 0.2dB = 4.6052e-05 ) -rOpt_x=ifft(exp(-alpha_lin_x*aStepSize/2).*opt_x); % /2 not sure why (have to find it in formulas) -rOpt_y=ifft(exp(-alpha_lin_y*aStepSize/2).*opt_y); % but not relevant for now +% attenuate the signal each linear step with alpha +rOpt_x=ifft(exp(-alpha_lin_x*aStepSize/2).*opt_x); +rOpt_y=ifft(exp(-alpha_lin_y*aStepSize/2).*opt_y); -end \ No newline at end of file +end diff --git a/Classes/DataBaseHandler/DBHandler.m b/Classes/DataBaseHandler/DBHandler.m index 5569746..b7ba026 100644 --- a/Classes/DataBaseHandler/DBHandler.m +++ b/Classes/DataBaseHandler/DBHandler.m @@ -23,6 +23,12 @@ classdef DBHandler < handle arguments options.dataBase = ""; % Default value for pathToDB if not provided options.type = "mysql"; + options.server = ""; + options.port = 3306; + options.user = ""; + options.password = ""; + + end % Assign values to class properties based on input arguments @@ -46,12 +52,13 @@ classdef DBHandler < handle obj.conn = database( ... string(obj.dataBase), ... % Database name - "silas", ... % Username - "silas", ... % Password (or getSecret) + options.user, ... % Username + options.password, ... % Password (or getSecret) "Vendor", "MySQL", ... - "Server", "134.245.243.254", ... + "Server", options.server, ... "PortNumber", 3306, ... "JDBCDriverLocation", "C:\Users\Silas\Documents\mysql-connector-j-9.3.0\mysql-connector-j-9.3.0.jar"); + end catch e diff --git a/Functions/Theory/Dissertation/mach_zehnder_modulator.m b/Functions/Theory/Dissertation/mach_zehnder_modulator.m index 7b379c2..f9eca58 100644 --- a/Functions/Theory/Dissertation/mach_zehnder_modulator.m +++ b/Functions/Theory/Dissertation/mach_zehnder_modulator.m @@ -1,16 +1,4 @@ -% Minimal MZM transfer-function demo (sinusoidal drive) — aligned with your notation -% -% Implements exactly: -% E_out(t) = E0 * exp(j*w0*t) * exp(-j*w0*L*n_eff/c0) * 1/2 * [ exp(-j*phi1(t)) + rho*exp(-j*phi2(t)) ] -% with phi_{1,2}(t) = pi * v_{1,2}(t)/Vpi -% -% Push-pull: -% v1(t) = +v_drive(t)/2 , v2(t) = -v_drive(t)/2 => phi1 = +pi/2 * v_drive/Vpi, phi2 = -pi/2 * v_drive/Vpi -% -% And the ideal TF (rho=1): -% E_out/E_in = exp(-j*w0*L*n_eff/c0) * cos( (pi/2) * v_drive/Vpi ) -% -% Note: E_in(t) = E0 * exp(j*w0*t) in this script. + % Parameters c0 = physconst('lightspeed'); % [m/s] @@ -25,27 +13,60 @@ Vpi = 3.2; % [V] half-wave voltage (your V_pi) % Drive f0 = 1e9; % [Hz] -fs = 100e9; % [Hz] -Nper = 1; % number of periods -Vpp = 0.5*Vpi; % [V] peak-to-peak of v_drive(t) +fs = 200e9; % [Hz] +Nper = 2; % number of periods +Vpp = 0.6*Vpi; % [V] peak-to-peak of v_drive(t) -biasV = 2; % [V] differential bias added to v_drive - -% Analytic -v_ = linspace(-1,2, 2001); -% Field transfer function (amplitude) -Field_mzm_analytic = cos((pi/2)*v_); -% Power transfer function (intensity) -P_mzm_analytic = Field_mzm_analytic.^2; - -% Imbalance factor in YOUR notation: -rho = 1; % rho=1 -> ideal balanced MZM (collapses to ideal TF) +biasV = 1.1; % [V] differential bias added to v_drive % Time axis + differential drive voltage v_drive(t) T = Nper/f0; t = (0:1/fs:T-1/fs).'; -v_drive = biasV + (Vpp/2)*sin(2*pi*f0*t); % v_drive(t) (peak = Vpp/2) + +if 1 + % SINE + v_drive = biasV + (Vpp/2)*sin(2*pi*f0*t); % v_drive(t) (peak = Vpp/2) + +else + + % --- Generate PAM-4 Sequence --- + symbols = linspace(-0.5, 0.5, 4); + num_symbols = 12; % Increased slightly for better visual + rng(44); + random_data = symbols(randi(4, 1, num_symbols)); + + % Create time axis (Note: T is your period from the sine code) + sps = round(T * fs); + t = (0:1/fs:(num_symbols*T)-1/fs).'; + + % Upsample to rectangular waveform + v_pam = repelem(random_data, sps).'; + + % Apply swing and bias: Resulting range is [biasV-Vpp/2, biasV+Vpp/2] + v_drive_rect = biasV + (v_pam * Vpp); + + % --- Round the edges --- + filter_span = round(sps/1.5); % Increased span for smoother "rounding" + window = gausswin(filter_span); + window = window / sum(window); + + % Apply filter (using 'same' to keep vector length, but be aware of edge transients) + v_drive = conv(v_drive_rect, window, 'same'); + +end + + +% Analytic +v_ = linspace(-1,2, 2001); +% Field transfer function (amplitude) +Field_mzm_analytic = cos((pi/2)*v_); + +% Power transfer function (intensity) +P_mzm_analytic = Field_mzm_analytic.^2; + +% Imbalance factor in YOUR notation: +rho = 1; % Push-pull branch voltages (consistent with v_drive = v1 - v2) v1 = +0.5*v_drive; % arm 1 @@ -69,8 +90,6 @@ H_num = E_out ./ E_in; % Power (normalized) Pnorm_num = abs(H_num).^2; % since |E_out/E_in|^2 - - % Ideal TF (analytic) for comparison (rho=1, push-pull) H_ideal = common_phase .* cos( (pi/2) * (v_drive./Vpi) ); @@ -78,9 +97,6 @@ Pnorm_ideal = abs(H_ideal).^2; Pnorm_math = cos( (pi/2) * (v_drive./Vpi) ).^2; - - - set(groot, 'defaultLegendInterpreter', 'tex'); set(groot, 'defaultAxesTickLabelInterpreter', 'tex'); set(groot, 'defaultTextInterpreter', 'tex'); @@ -100,8 +116,11 @@ plot(v_norm,t*1e9, 'LineWidth', 1.0,'Color',colvdrive); grid on; ylabel('t [ns]'); xlabel('v_{drive}(t)/V_\pi'); title('Drive voltage (normalized)'); xlim([min(v_) max(v_)]); +% mat2tikz_improved('C:\Users\Silas\Documents\6971e0b65b380ca6d71c837f\02_IMDD_System\tikz\linear_casee\mzm_input_signal.tex'); + %% IN/OUT (static transfer) — normalized x-axis + analytic curve +if 0 figure(2); clf plot(v_, Field_mzm_analytic, 'LineWidth', 1.2,'LineStyle','--','Color',colfield); hold on;% analytic power TF plot(v_, P_mzm_analytic, 'LineWidth', 1.2, 'Color',colpow); hold on;% analytic power TF @@ -123,18 +142,78 @@ xlim([min(v_) max(v_)+1]); ylim([-1 1]); % mat2tikz_improved('C:\Users\Silas\Documents\6971e0b65b380ca6d71c837f\02_IMDD_System\tikz\mzm.tex'); - - +end %% -% % FIELD TF (only field here; do not mix power into this figure) + figure(3); clf +plot(v_, Field_mzm_analytic, 'LineWidth', 1.2,'LineStyle','--','Color',colfield); hold on;% analytic power TF +plot(v_, P_mzm_analytic, 'LineWidth', 1.2, 'Color',colpow); hold on;% analytic power TF + +scatter(v_norm, Pnorm_num, 12, '.', 'LineWidth', 1,'MarkerEdgeColor',colvdrive); +scatter(biasV./Vpi,(cos((pi/2)*biasV./Vpi)^2),10,'Marker','o'); +line([min(v_drive), min(v_drive)]./Vpi,[(cos((pi/2)*min(v_drive)./Vpi)^2), -2],'linewidth',0.5,'color','black','linestyle','--'); +line([max(v_drive) max(v_drive)]./Vpi,[(cos((pi/2)*max(v_drive)./Vpi)^2), -2],'linewidth',0.5,'color','black','linestyle','--'); +xline([min(v_norm) max(v_norm)]) + +grid on; +xlabel('v_{drive}(t)/V_\pi'); ylabel('|E_{out}/E_{in}|^2'); +% legend +xlim([min(v_) max(v_)]); +ylim([-1 1]); + +% mat2tikz_improved('C:\Users\Silas\Documents\6971e0b65b380ca6d71c837f\02_IMDD_System\tikz\mzm_tramsfer_function_matlab.tex'); + + +%% + +figure(4); clf +% plot(v_, Field_mzm_analytic, 'LineWidth', 1.2,'LineStyle','--','Color',colfield); hold on;% analytic power TF +plot(v_, P_mzm_analytic, 'LineWidth', 1.2, 'Color','black'); hold on;% analytic power TF +input_dots = linspace(min(v_drive),max(v_drive),4)./Vpi; +% input_dots = unique(v_drive_rect)./Vpi; +output_dots = (cos((pi/2)*input_dots).^2); +scatter(input_dots,output_dots,'Marker','x','LineWidth',1,'MarkerEdgeColor','black'); +scatter(input_dots,zeros(size(input_dots)),'Marker','^','LineWidth',2,'MarkerEdgeColor','black'); +% scatter(ones(size(input_dots)),output_dots,'Marker','<','LineWidth',2,'MarkerEdgeColor','black'); + +for i = 1:numel(input_dots) + % Draw the dashed projection lines + line([input_dots(i), input_dots(i)], [output_dots(i), 0], 'linewidth', 0.5, 'color', 'black', 'linestyle', '--', 'handlevisibility', 'off'); + line([input_dots(i), 1], [output_dots(i), output_dots(i)], 'linewidth', 0.5, 'color', 'black', 'linestyle', '--', 'handlevisibility', 'off'); + + % Add the level annotation boxes near the output (y-axis) + % Adjust the '1.05' to move the box further right or 'output_dots(i)' for height + j = 3-(i-1)*2; + text(1, output_dots(i), sprintf('Level %d', j), ... + 'FontSize', 8, ... + 'EdgeColor', 'black', ... + 'BackgroundColor', 'white', ... + 'Margin', 2); +end + +xlim([0,1.5]); +ylim([0,1]) + +% line([min(v_drive), min(v_drive)]./Vpi,[(cos((pi/2)*min(v_drive)./Vpi)^2), 0],'linewidth',0.5,'color','black','linestyle','--'); +% line([max(v_drive), max(v_drive)]./Vpi,[(cos((pi/2)*max(v_drive)./Vpi)^2), 0],'linewidth',0.5,'color','black','linestyle','--'); + +% mat2tikz_improved('C:\Users\Silas\Documents\6971e0b65b380ca6d71c837f\02_IMDD_System\tikz\linear_casee\mzm_tf.tex'); +% xticks(sort(input_dots)); +% yticks(sort(output_dots)); +grid off + + +%% +% % FIELD TF (only field here; do not mix power into this figure) +figure(5); clf % plot(t*1e9, real(H_num), 'LineWidth', 1.0); hold on; % plot(t*1e9, real(H_ideal), '--', 'LineWidth', 1.0,'DisplayName','Field','Color',colfield); hold on; plot(t*1e9, Pnorm_num, 'LineWidth', 1.0,'DisplayName','Intensity', 'Color',colpow); hold on; grid on; xlabel('t [ns]'); ylabel('Re\{E_{out}/E_{in}\}'); legend -mat2tikz_improved('C:\Users\Silas\Documents\6971e0b65b380ca6d71c837f\02_IMDD_System\tikz\mzm_out.tex'); +yticks(sort(output_dots)); +% mat2tikz_improved('C:\Users\Silas\Documents\6971e0b65b380ca6d71c837f\02_IMDD_System\tikz\linear_casee\mzm_output_signal.tex'); diff --git a/projects/HighSpeedExperiment_2024/Auswertung_JLT/run_dsp_from_db.m b/projects/HighSpeedExperiment_2024/Auswertung_JLT/run_dsp_from_db.m index ab31451..67a2ff8 100644 --- a/projects/HighSpeedExperiment_2024/Auswertung_JLT/run_dsp_from_db.m +++ b/projects/HighSpeedExperiment_2024/Auswertung_JLT/run_dsp_from_db.m @@ -10,10 +10,11 @@ if dsp_options.mode == "load_run_id" if experiment == "highspeed_2024" - dsp_options.database_type = 'mysql'; + dsp_options.database_type = "mysql"; dsp_options.dataBase = 'labor_highspeed';%'C:\Users\Silas\Documents\MATLAB\Datensätze\sioe_labor\silas_labor_newdsp_newstructure.db'; dsp_options.storage_path = 'Z:\2024\sioe_labor\'; - db = DBHandler("dataBase", [dsp_options.dataBase], "type", dsp_options.database_type); + db = DBHandler("dataBase", [dsp_options.dataBase],... + "type", dsp_options.database_type,"server","192.168.178.192","user","silas","password","silas"); elseif experiment == "mpi_ecoc_2025" diff --git a/projects/WDM/WDM_model_10km_queue.m b/projects/WDM/WDM_model_10km_queue.m index a41cb02..8aae2ea 100644 --- a/projects/WDM/WDM_model_10km_queue.m +++ b/projects/WDM/WDM_model_10km_queue.m @@ -312,7 +312,7 @@ for realiz = 1:s.num_realiz Opt_sig_wdm_fib = DP_Fiber("L",segment_length,"D",Dvec(seg),"Dpmd",s.pmd,"Ds",0.07, ... "beat_len",10,"corr_len",100,"dz",1,"manakov",0, ... "gamma",s.gamma,"lambda",zdw,"n_waveplates",10,"SS_dphimax",0.01, ... - "SS_dzmax",50,"SS_dzmin",10,"X_alpha",0.3,"X_beta",0,"rng",1).process(Opt_sig_wdm_fib); + "SS_dzmax",50,"SS_dzmin",10,"X_alpha",0.3,"X_beta",0,"rng",1,"useGPU",1,"useSingle",1).process(Opt_sig_wdm_fib); % -------- Evaluate at intermediate distance (enqueue jobs) -------- if eval_ptr <= nEval && seg == eval_seg(eval_ptr) diff --git a/test/bayesopt_ffe_tuning.m b/test/bayesopt_ffe_tuning.m new file mode 100644 index 0000000..adedb3d --- /dev/null +++ b/test/bayesopt_ffe_tuning.m @@ -0,0 +1,219 @@ +%% Bayesian Optimization for FFE Parameter Tuning +% This script uses bayesopt to find optimal mu_dd and mu_tr values +% that minimize BER for the FFE equalizer. + +clear; clc; + +%% Setup - Same as gpu_processing_dpfiber.m +s.wavelengthplan = calcWavelengthPlan(4, 400e9, 1310); +link_length = 10; +s.pmd = 0.1; +s.gamma = 0.0023; + +s.M = 4; +fsym = 112e9; +fdac = 2*fsym; +fadc = 120000000000; +s.random_key = 1; + +% Laser / Modulator +vbias_rel = 0.5; +u_pi = 4.6; +vbias = -vbias_rel*u_pi; +laser_linewidth = 0e6; + +duob_mode = db_mode.no_db; +rcalpha = 0.05; +Pform = Pulseformer("fsym",fsym,"fdac",4*fsym,"pulse","rc","pulselength",16,"alpha",rcalpha); + +s.chirpalpha = 0; +s.p_launch = 3; +s.p = "co"; + +N = numel(s.wavelengthplan); + +switch s.p + case "co" + pol_rot = 100.*ones(1,N); + d_local = 0; +end + +f_plan = physconst('lightspeed')./(s.wavelengthplan.*1e-9); +margin = 25e12; +f_span = (max(f_plan)+margin)-(min(f_plan)-margin); +f_nyq = f_span/2; + +kover = 4; +upsample_required = f_nyq./(fdac*kover/2); +upsample_pow = 2^nextpow2(upsample_required); + +s.f_opt = fdac*kover*upsample_pow; +s.f_opt_nyq = s.f_opt/2; + +s.rop = -8; % Fixed ROP for optimization + +%% Generate TX signals (run once) +fprintf('Generating TX signals...\n'); +for l = 1:N + [Digi_sig,Symbols{l},Tx_bits{l}] = PAMsource( ... + "fsym",fsym,"M",s.M,"order",15,"useprbs",0, ... + "fs_out",fdac, ... + "applyclipping",0,"clipfactor",1.5, ... + "applypulseform",1,"pulseformer",Pform, ... + "randkey",s.random_key+l, ... + "mrds_code",0,"mrds_blocklength",512,"duobinary_mode",duob_mode ... + ).process(); + + Lp_awg = Filter('filtdegree',3,"f_cutoff",56e9,"fs",fdac*kover, ... + "filterType",filtertypes.gaussian,"active",true); + + El_sig = AWG("fdac",fdac,"f_cutoff",fsym,"lpf_active",1,"kover",kover, ... + "bit_resolution",6,"upsampling_method","samplehold","precomp_sinc_rolloff",0, ... + "H_lpf",Lp_awg,"dac_max",0.6,"dac_min",-0.6).process(Digi_sig); + + El_sig = El_sig.normalize("mode","oneone"); + scaling = 0.6*(u_pi/2-abs(vbias-u_pi/2)); + El_sig = El_sig .* scaling; + + Eml_out = EML("mode",eml_mode.im_cosinus,"power",3,"fsimu",El_sig.fs, ... + "lambda",s.wavelengthplan(l),"bias",vbias,"u_pi",u_pi, ... + "linewidth",laser_linewidth,"randomkey",s.random_key+l,"alpha",s.chirpalpha).process(El_sig); + + signal_cell{l} = Polarization_Controller("mode","rot_power","desired_power",pol_rot(l)).process(Eml_out); +end + +%% WDM mux + launch +Opt_sig_wdm = Optical_Multiplex("fs_in",fdac*kover,"fs_out",upsample_pow*fdac*kover, ... + "lambda_center",1310,"random_key",0,"filtype",1,"B",120e9).process(signal_cell); + +Opt_sig_wdm = Amplifier("amp_mode","ideal_no_noise","gain_mode","output_power", ... + "amplification_db",s.p_launch+10*log10(N)).process(Opt_sig_wdm); + +%% Fiber propagation +segment_length = 1; +nSegments = link_length/segment_length; +nSegments = round(nSegments); + +zdw = 1310; +randomize_D = true; +Dvec = getDispersionVector(nSegments, d_local, zdw, randomize_D, s.random_key); + +Opt_sig_wdm_fib = Opt_sig_wdm; +fprintf('Running fiber propagation...\n'); +for seg = 1:nSegments + fprintf('Segment %d/%d\n', seg, nSegments); + Opt_sig_wdm_fib = DP_Fiber("L",segment_length,"D",Dvec(seg),"Dpmd",s.pmd,"Ds",0.07, ... + "beat_len",10,"corr_len",100,"dz",1,"manakov",0, ... + "gamma",s.gamma,"lambda",zdw,"n_waveplates",10,"SS_dphimax",0.01, ... + "SS_dzmax",50,"SS_dzmin",10,"X_alpha",0.3,"X_beta",0,"rng",1,"useGPU",true,"useSingle",true).process(Opt_sig_wdm_fib); +end + +%% Pre-process to get Rx_sig (do demux once) +fprintf('Pre-processing receiver chain...\n'); +l = 1; % Use channel 1 for optimization + +Opt_sig_demux = Optical_Demultiplex("attenuation",0,"B",200e9,"filtype",1, ... + "fs_out",fdac*kover,"fs_in",fdac*kover*upsample_pow,"lambda_center",1310).process(Opt_sig_wdm_fib); + +Opt_sig_rx = Amplifier("amp_mode","ideal_no_noise","gain_mode","output_power", ... + "amplification_db",s.rop).process(Opt_sig_demux{l}); + +PD_sig = Photodiode("fsimu",fdac*kover,"dark_current",2e-08,"responsivity",1,"temperature",20, ... + "nep",1.8e-11,"randomkey",s.random_key+l).process(Opt_sig_rx); + +rx_bwl = 100e9; +PD_sig = Filter('filtdegree',4,"f_cutoff",rx_bwl,"fs",fdac*kover, ... + "filterType",filtertypes.butterworth,"active",true).process(PD_sig); + +Lp_scpe = Filter('filtdegree',4,"f_cutoff",80e9,"fs",fadc,"filterType",filtertypes.butterworth,"active",true); +Scpe_sig = Scope("fsimu",fdac*kover,"fadc",fadc, ... + "delay",0,"fixed_delay",0,"filtertype",filtertypes.butterworth, ... + "samplingdelay",0,"rand_samplingdelay",0,"freq_offset",0,"samp_jitter",0, ... + "adcresolution",8,"quantbuffer",0.1,'block_dc',1,'lpf_active',0,'H_lpf',Lp_scpe).process(PD_sig); + +Scpe_sig_2sps = Scpe_sig.resample("fs_out",2*fsym); + +[~, Scpe_cell, ~, ~] = Scpe_sig_2sps.tsynch("reference", Symbols{l}, "fs_ref", fsym, "debug_plots", 0); +Rx_sig = Scpe_cell{1}; +Rx_sig = Rx_sig.normalize("mode","rms"); + +fprintf('Receiver pre-processing complete. Ready for optimization.\n\n'); + +%% Define the objective function for bayesopt +function ber = ffe_objective(params, Rx_sig, Symbols_l, Tx_bits_l, M, duob_mode) +mu_dd = params.mu_dd; +mu_tr = params.mu_tr; + +try + eq_ffe = FFE("epochs_tr", 5, "epochs_dd", 2, "len_tr", 2^13, ... + "mu_dd", mu_dd, "mu_tr", mu_tr, ... + "order", 50, "sps", 2, "decide", 0, ... + "adaption", adaption_method.nlms, "dd_mode", 1); + + ffe_results = ffe(eq_ffe, M, Rx_sig, Symbols_l, Tx_bits_l, ... + "precode_mode", duob_mode, ... + 'showAnalysis', 0, ... + "postFFE", [], ... + "eth_style_symbol_mapping", 0); + + ber = ffe_results.metrics.BER; + + if ber == 0 + ber = 1e-10; + end + + if ~isfinite(ber) + ber = 0.5; + end + + fprintf(' mu_dd=%.4e, mu_tr=%.4e -> BER=%.4e\n', mu_dd, mu_tr, ber); + +catch ME + fprintf(' mu_dd=%.4e, mu_tr=%.4e -> FAILED (%s)\n', mu_dd, mu_tr, ME.message); + ber = 0.5; +end +end + +%% Define optimizable variables +mu_dd_var = optimizableVariable('mu_dd', [1e-5, 0.1], 'Transform', 'log'); +mu_tr_var = optimizableVariable('mu_tr', [1e-5, 0.1], 'Transform', 'log'); + +%% Run Bayesian Optimization +fprintf('========== Starting Bayesian Optimization ==========\n'); +fprintf('Optimizing mu_dd and mu_tr to minimize BER\n'); +fprintf('Search range: mu_dd=[1e-5, 0.1], mu_tr=[1e-5, 0.1]\n\n'); + +objective_fn = @(params) ffe_objective(params, Rx_sig, Symbols{l}, Tx_bits{l}, s.M, duob_mode); + +results = bayesopt(objective_fn, [mu_dd_var, mu_tr_var], ... + 'MaxObjectiveEvaluations', 30, ... + 'AcquisitionFunctionName', 'expected-improvement-plus', ... + 'IsObjectiveDeterministic', false, ... + 'ExplorationRatio', 0.5, ... + 'Verbose', 1, ... + 'PlotFcn', []); + +%% Display Results +fprintf('\n========== FFE Optimization Complete ==========\n'); +fprintf('Best FFE parameters found:\n'); +fprintf(' mu_dd = %.6e\n', results.XAtMinObjective.mu_dd); +fprintf(' mu_tr = %.6e\n', results.XAtMinObjective.mu_tr); +fprintf(' BER = %.6e\n', results.MinObjective); + +%% Verify with optimal parameters +fprintf('\nVerifying optimal FFE parameters...\n'); +best_mu_dd = results.XAtMinObjective.mu_dd; +best_mu_tr = results.XAtMinObjective.mu_tr; + +eq_ffe_best = FFE("epochs_tr", 5, "epochs_dd", 2, "len_tr", 2^13, ... + "mu_dd", best_mu_dd, "mu_tr", best_mu_tr, ... + "order", 50, "sps", 2, "decide", 0, ... + "adaption", adaption_method.nlms, "dd_mode", 1); + +ffe_results_best = ffe(eq_ffe_best, s.M, Rx_sig, Symbols{l}, Tx_bits{l}, ... + "precode_mode", duob_mode, ... + 'showAnalysis', 1, ... + "postFFE", [], ... + "eth_style_symbol_mapping", 0); + +fprintf('\nFinal FFE BER with optimal parameters: %.6e\n', ffe_results_best.metrics.BER); diff --git a/test/gpu_cpu_comparison.m b/test/gpu_cpu_comparison.m new file mode 100644 index 0000000..02d40ba --- /dev/null +++ b/test/gpu_cpu_comparison.m @@ -0,0 +1,349 @@ +%% GPU vs CPU Comparison Test for DP_Fiber +% This script runs the fiber simulation with and without GPU acceleration +% and compares the numerical results. + +% clear; clc; + +%% Setup (same as gpu_processing_dpfiber.m but simplified) +s.wavelengthplan = calcWavelengthPlan(4, 400e9, 1310); +link_length = 10; +s.pmd = 0.1; +s.gamma = 0.0023; + +s.M = 4; +fsym = 112e9; +fdac = 2*fsym; +fadc = 120000000000; +s.random_key = 1; + +% Laser / Modulator +vbias_rel = 0.5; +u_pi = 4.6; +vbias = -vbias_rel*u_pi; +laser_linewidth = 0e6; + +% DB Stuff +duob_mode = db_mode.no_db; + +rcalpha = 0.05; +Pform = Pulseformer("fsym",fsym,"fdac",4*fsym,"pulse","rc","pulselength",16,"alpha",rcalpha); + +s.chirpalpha = 0; +s.p_launch = 3; +s.p = "co"; + +N = numel(s.wavelengthplan); + +switch s.p + case "co" + pol_rot = 100.*ones(1,N); + d_local = 0; + case "pair" + pol_rot = repmat([100,100,0,0],1,N/4); + d_local = 0; + case "alt" + pol_rot = repmat([100,0,100,0],1,N/4); + d_local = 0; + case "seg" + pol_rot = 100.*ones(1,N); + d_local = 3; + otherwise + error('Unknown fwm_mitigation_technique: %s', string(s.p)); +end + +f_plan = physconst('lightspeed')./(s.wavelengthplan.*1e-9); +margin = 5e12; +f_span = (max(f_plan)+margin)-(min(f_plan)-margin); +f_nyq = f_span/2; + +kover = 4; +upsample_required = f_nyq./(fdac*kover/2); +upsample_pow = 2^nextpow2(upsample_required); + +s.f_opt = fdac*kover*upsample_pow; +s.f_opt_nyq = s.f_opt/2; + +%% TX per channel +for l = 1:N + [Digi_sig,Symbols{l},Tx_bits{l}] = PAMsource( ... + "fsym",fsym,"M",s.M,"order",15,"useprbs",0, ... + "fs_out",fdac, ... + "applyclipping",0,"clipfactor",1.5, ... + "applypulseform",1,"pulseformer",Pform, ... + "randkey",s.random_key+l, ... + "mrds_code",0,"mrds_blocklength",512,"duobinary_mode",duob_mode ... + ).process(); + + Lp_awg = Filter('filtdegree',3,"f_cutoff",56e9,"fs",fdac*kover, ... + "filterType",filtertypes.gaussian,"active",true); + + El_sig = AWG("fdac",fdac,"f_cutoff",fsym,"lpf_active",1,"kover",kover, ... + "bit_resolution",6,"upsampling_method","samplehold","precomp_sinc_rolloff",0, ... + "H_lpf",Lp_awg,"dac_max",0.6,"dac_min",-0.6).process(Digi_sig); + + clear Digi_sig + + El_sig = El_sig.normalize("mode","oneone"); + scaling = 0.6*(u_pi/2-abs(vbias-u_pi/2)); + El_sig = El_sig .* scaling; + + Eml_out = EML("mode",eml_mode.im_cosinus,"power",3,"fsimu",El_sig.fs, ... + "lambda",s.wavelengthplan(l),"bias",vbias,"u_pi",u_pi, ... + "linewidth",laser_linewidth,"randomkey",s.random_key+l,"alpha",s.chirpalpha).process(El_sig); + + clear El_sig + + signal_cell{l} = Polarization_Controller("mode","rot_power","desired_power",pol_rot(l)).process(Eml_out); + + clear Eml_out Lp_awg +end + +disp('Signal generated for all channels.'); + +%% WDM mux + launch +Opt_sig_wdm = Optical_Multiplex("fs_in",fdac*kover,"fs_out",upsample_pow*fdac*kover, ... + "lambda_center",1310,"random_key",0,"filtype",1,"B",120e9).process(signal_cell); + +Opt_sig_wdm = Amplifier("amp_mode","ideal_no_noise","gain_mode","output_power", ... + "amplification_db",s.p_launch+10*log10(N)).process(Opt_sig_wdm); + +% Save input for both runs +Opt_sig_input = Opt_sig_wdm; + +segment_length = 1; +nSegments = link_length/segment_length; +if abs(nSegments - round(nSegments)) > 1e-12 + error('fiber_length_km=%g must be an integer multiple of segment_length=%g km.', link_length, segment_length); +end +nSegments = round(nSegments); + +zdw = 1310; +randomize_D = true; + +if nSegments > 0 + Dvec = getDispersionVector(nSegments, d_local, zdw, randomize_D, s.random_key); +else + Dvec = []; +end + +%% Run WITHOUT GPU +fprintf('\n========== Running WITHOUT GPU (CPU) ==========\n'); +Opt_sig_cpu = Opt_sig_input; + +tic; +for seg = 1:nSegments + fprintf('CPU Segment %d/%d \n',seg, nSegments); + + Opt_sig_cpu = DP_Fiber("L",segment_length,"D",Dvec(seg),"Dpmd",s.pmd,"Ds",0.07, ... + "beat_len",10,"corr_len",100,"dz",1,"manakov",0, ... + "gamma",s.gamma,"lambda",zdw,"n_waveplates",10,"SS_dphimax",0.01, ... + "SS_dzmax",50,"SS_dzmin",10,"X_alpha",0.3,"X_beta",0,"rng",1,"useGPU",false).process(Opt_sig_cpu); +end +time_cpu = toc; +fprintf('CPU Time: %.3f seconds\n', time_cpu); + +%% Run WITH GPU (Double Precision) +fprintf('\n========== Running WITH GPU (Double Precision) ==========\n'); +Opt_sig_gpu_double = Opt_sig_input; + +tic; +for seg = 1:nSegments + fprintf('GPU-Double Segment %d/%d \n',seg, nSegments); + + Opt_sig_gpu_double = DP_Fiber("L",segment_length,"D",Dvec(seg),"Dpmd",s.pmd,"Ds",0.07, ... + "beat_len",10,"corr_len",100,"dz",1,"manakov",0, ... + "gamma",s.gamma,"lambda",zdw,"n_waveplates",10,"SS_dphimax",0.01, ... + "SS_dzmax",50,"SS_dzmin",10,"X_alpha",0.3,"X_beta",0,"rng",1,"useGPU",true,"useSingle",false).process(Opt_sig_gpu_double); +end +time_gpu_double = toc; +fprintf('GPU Double Time: %.3f seconds\n', time_gpu_double); + +%% Run WITH GPU (Single Precision) +fprintf('\n========== Running WITH GPU (Single Precision) ==========\n'); +Opt_sig_gpu_single = Opt_sig_input; + +tic; +for seg = 1:nSegments + fprintf('GPU-Single Segment %d/%d \n',seg, nSegments); + + Opt_sig_gpu_single = DP_Fiber("L",segment_length,"D",Dvec(seg),"Dpmd",s.pmd,"Ds",0.07, ... + "beat_len",10,"corr_len",100,"dz",1,"manakov",0, ... + "gamma",s.gamma,"lambda",zdw,"n_waveplates",10,"SS_dphimax",0.01, ... + "SS_dzmax",50,"SS_dzmin",10,"X_alpha",0.3,"X_beta",0,"rng",1,"useGPU",true,"useSingle",true).process(Opt_sig_gpu_single); +end +time_gpu_single = toc; +fprintf('GPU Single Time: %.3f seconds\n', time_gpu_single); + +%% Compare Results +fprintf('\n========== Numerical Comparison ==========\n'); + +sig_cpu = Opt_sig_cpu.signal; +sig_gpu_double = Opt_sig_gpu_double.signal; +sig_gpu_single = Opt_sig_gpu_single.signal; + +% Check dimensions +fprintf('CPU signal size: [%d x %d]\n', size(sig_cpu,1), size(sig_cpu,2)); +fprintf('GPU Double signal size: [%d x %d]\n', size(sig_gpu_double,1), size(sig_gpu_double,2)); +fprintf('GPU Single signal size: [%d x %d]\n', size(sig_gpu_single,1), size(sig_gpu_single,2)); + +% CPU vs GPU Double +fprintf('\n--- CPU vs GPU Double ---\n'); +diff_cpu_double = abs(sig_cpu - sig_gpu_double); +max_diff_cpu_double = max(diff_cpu_double(:)); +mean_diff_cpu_double = mean(diff_cpu_double(:)); +rel_diff_cpu_double = max_diff_cpu_double / max(abs(sig_cpu(:))); +fprintf('Max absolute difference: %.6e\n', max_diff_cpu_double); +fprintf('Mean absolute difference: %.6e\n', mean_diff_cpu_double); +fprintf('Max relative difference: %.6e\n', rel_diff_cpu_double); + +% CPU vs GPU Single +fprintf('\n--- CPU vs GPU Single ---\n'); +diff_cpu_single = abs(sig_cpu - sig_gpu_single); +max_diff_cpu_single = max(diff_cpu_single(:)); +mean_diff_cpu_single = mean(diff_cpu_single(:)); +rel_diff_cpu_single = max_diff_cpu_single / max(abs(sig_cpu(:))); +fprintf('Max absolute difference: %.6e\n', max_diff_cpu_single); +fprintf('Mean absolute difference: %.6e\n', mean_diff_cpu_single); +fprintf('Max relative difference: %.6e\n', rel_diff_cpu_single); + +% GPU Double vs GPU Single +fprintf('\n--- GPU Double vs GPU Single ---\n'); +diff_double_single = abs(sig_gpu_double - sig_gpu_single); +max_diff_double_single = max(diff_double_single(:)); +mean_diff_double_single = mean(diff_double_single(:)); +rel_diff_double_single = max_diff_double_single / max(abs(sig_gpu_double(:))); +fprintf('Max absolute difference: %.6e\n', max_diff_double_single); +fprintf('Mean absolute difference: %.6e\n', mean_diff_double_single); +fprintf('Max relative difference: %.6e\n', rel_diff_double_single); + +% Check tolerances +fprintf('\n--- Tolerance Check ---\n'); +tol_double = 1e-10; +tol_single = 1e-5; % Single precision has ~7 significant digits + +if max_diff_cpu_double < tol_double + fprintf('✓ CPU vs GPU Double: EQUIVALENT (diff < %.0e)\n', tol_double); +else + fprintf('✗ CPU vs GPU Double: DIFFER beyond tolerance (%.0e)\n', tol_double); +end + +if max_diff_cpu_single < tol_single + fprintf('✓ CPU vs GPU Single: ACCEPTABLE (diff < %.0e)\n', tol_single); +else + fprintf('⚠ CPU vs GPU Single: Precision loss detected (diff = %.2e, tol = %.0e)\n', max_diff_cpu_single, tol_single); +end + +% Performance comparison +fprintf('\n========== Performance Summary ==========\n'); +fprintf('CPU Time: %.3f s\n', time_cpu); +fprintf('GPU Double Time: %.3f s\n', time_gpu_double); +fprintf('GPU Single Time: %.3f s\n', time_gpu_single); +fprintf('\n'); +fprintf('Speedup (GPU Double vs CPU): %.2fx\n', time_cpu/time_gpu_double); +fprintf('Speedup (GPU Single vs CPU): %.2fx\n', time_cpu/time_gpu_single); +fprintf('Speedup (GPU Single vs GPU Double): %.2fx\n', time_gpu_double/time_gpu_single); + +%% ========== BER Comparison ========== +% Process each fiber output through simplified receiver to check if +% single-precision affects actual BER performance + +fprintf('\n========== BER Comparison ==========\n'); +fprintf('Processing signals through receiver chain...\n'); + +% Receiver parameters +rop = -7; % Received optical power [dBm] +len_tr = 4096; % Training length +mu_dc = 0.005; +mu_ffe = [0.0001 0.0008 0.001]; +mu_dfe = 0.0004; + +% Helper function to process through receiver and get BER +function ber = process_receiver(Opt_sig_fib, l, Symbols, Tx_bits, ... + fdac, kover, upsample_pow, fsym, fadc, rop, s, len_tr, mu_dc, mu_ffe, mu_dfe, duob_mode) + +% Demux single channel +Opt_sig_demux = Optical_Demultiplex("attenuation",0,"B",200e9,"filtype",1, ... + "fs_out",fdac*kover,"fs_in",fdac*kover*upsample_pow,"lambda_center",1310).process(Opt_sig_fib); + +% ROP amplifier +Opt_sig_rx = Amplifier("amp_mode","ideal_no_noise","gain_mode","output_power", ... + "amplification_db",rop).process(Opt_sig_demux{l}); + +% Photodiode +PD_sig = Photodiode("fsimu",fdac*kover,"dark_current",2e-08,"responsivity",1,"temperature",20, ... + "nep",1.8e-11,"randomkey",s.random_key+l).process(Opt_sig_rx); + +% Low-pass filter +rx_bwl = 100e9; +PD_sig = Filter('filtdegree',4,"f_cutoff",rx_bwl,"fs",fdac*kover, ... + "filterType",filtertypes.butterworth,"active",true).process(PD_sig); + +% Scope +Lp_scpe = Filter('filtdegree',4,"f_cutoff",80e9,"fs",fadc,"filterType",filtertypes.butterworth,"active",true); +Scpe_sig = Scope("fsimu",fdac*kover,"fadc",fadc, ... + "delay",0,"fixed_delay",0,"filtertype",filtertypes.butterworth, ... + "samplingdelay",0,"rand_samplingdelay",0,"freq_offset",0,"samp_jitter",0, ... + "adcresolution",8,"quantbuffer",0.1,'block_dc',1,'lpf_active',0,'H_lpf',Lp_scpe).process(PD_sig); + +% Resample to 2 sps +Scpe_sig_2sps = Scpe_sig.resample("fs_out",2*fsym); + +% Time sync +[~, Scpe_cell, ~, ~] = Scpe_sig_2sps.tsynch("reference", Symbols{l}, "fs_ref", fsym, "debug_plots", 0); +Rx_sig = Scpe_cell{1}; +Rx_sig = Rx_sig.normalize("mode","rms"); + +% FFE Equalizer +ffe_order = [50, 0, 0]; +eq_ffe = EQ("Ne",ffe_order,"Nb",[0,0,0],"training_length",len_tr,"training_loops",5,"dd_loops",5, ... + "K",2,"DCmu",mu_dc,"DDmu",[mu_ffe mu_dfe],"DFEmu",0.005,"FFEmu",0,"plotfinal",0,"ideal_dfe",0); + +ffe_results = ffe(eq_ffe,s.M,Rx_sig,Symbols{l},Tx_bits{l}, ... + "precode_mode",duob_mode, ... + 'showAnalysis',0, ... + "postFFE",[], ... + "eth_style_symbol_mapping",0); + +ber = ffe_results.metrics.BER; +end + +% Process each mode for channel 1 +l = 4; % Use first channel for comparison + +fprintf('Processing CPU result...\n'); +ber_cpu = process_receiver(Opt_sig_cpu, l, Symbols, Tx_bits, ... + fdac, kover, upsample_pow, fsym, fadc, rop, s, len_tr, mu_dc, mu_ffe, mu_dfe, duob_mode); + +fprintf('Processing GPU Double result...\n'); +ber_gpu_double = process_receiver(Opt_sig_gpu_double, l, Symbols, Tx_bits, ... + fdac, kover, upsample_pow, fsym, fadc, rop, s, len_tr, mu_dc, mu_ffe, mu_dfe, duob_mode); + +fprintf('Processing GPU Single result...\n'); +ber_gpu_single = process_receiver(Opt_sig_gpu_single, l, Symbols, Tx_bits, ... + fdac, kover, upsample_pow, fsym, fadc, rop, s, len_tr, mu_dc, mu_ffe, mu_dfe, duob_mode); + +% Display BER results +fprintf('\n========== BER Results (Channel %d, ROP = %d dBm) ==========\n', l, rop); +fprintf('CPU: BER = %.4e\n', ber_cpu); +fprintf('GPU Double: BER = %.4e\n', ber_gpu_double); +fprintf('GPU Single: BER = %.4e\n', ber_gpu_single); + +fprintf('\n--- BER Comparison ---\n'); +if ber_cpu == 0 && ber_gpu_double == 0 && ber_gpu_single == 0 + fprintf('✓ All BERs are zero (no errors detected)\n'); +else + ber_diff_double = abs(ber_cpu - ber_gpu_double); + ber_diff_single = abs(ber_cpu - ber_gpu_single); + fprintf('|BER_cpu - BER_gpu_double| = %.4e\n', ber_diff_double); + fprintf('|BER_cpu - BER_gpu_single| = %.4e\n', ber_diff_single); + + if ber_diff_double < 1e-6 && ber_diff_single < 1e-6 + fprintf('✓ BER differences are negligible\n'); + elseif ber_diff_single > ber_diff_double * 10 + fprintf('⚠ Single precision shows measurable BER impact\n'); + else + fprintf('✓ BER differences within acceptable range\n'); + end +end + +fprintf('\n========== Test Complete ==========\n'); diff --git a/test/gpu_processing_dpfiber.m b/test/gpu_processing_dpfiber.m new file mode 100644 index 0000000..317833e --- /dev/null +++ b/test/gpu_processing_dpfiber.m @@ -0,0 +1,265 @@ + + +s.wavelengthplan = calcWavelengthPlan(16, 400e9, 1310); +N = numel(s.wavelengthplan); +link_length = 10; +s.pmd = 0.1;%0.1; +s.gamma = 0.0023; + +s.M = 4; +fsym = 112e9; +fdac = 2*fsym; +fadc = 120000000000; +s.random_key = 1; + +% Laser / s.Modulator +vbias_rel = 0.5; +u_pi = 4.6; +vbias = -vbias_rel*u_pi; +laser_linewidth = 0e6; + +% DB Stuff +duob_mode = db_mode.no_db; + +rcalpha = 0.05; +Pform = Pulseformer("fsym",fsym,"fdac",4*fsym,"pulse","rc","pulselength",16,"alpha",rcalpha); + +s.chirpalpha = 0; + +s.p_launch = 3; +s.p = "co"; + +switch s.p + case "co" + pol_rot = 100.*ones(1,N); + d_local = 0; + case "pair" + pol_rot = repmat([100,100,0,0],1,N/4); + d_local = 0; + case "alt" + pol_rot = repmat([100,0,100,0],1,N/4); + d_local = 0; + case "seg" + pol_rot = 100.*ones(1,N); + d_local = 3; + otherwise + error('Unknown fwm_mitigation_technique: %s', string(s.p)); +end + +f_plan = physconst('lightspeed')./(s.wavelengthplan.*1e-9); +margin = 25e12; % some THz left and right +f_span = (max(f_plan)+margin)-(min(f_plan)-margin); +f_nyq = f_span/2; + +kover = 4; +upsample_required = f_nyq./(fdac*kover/2); +upsample_pow = 2^nextpow2(upsample_required); + +s.f_opt = fdac*kover*upsample_pow; +s.f_opt_nyq = s.f_opt/2; + +s.rop = -10:1:0; + +profile on + +%% ---------- TX per channel ---------- +for l = 1:N + + [Digi_sig,Symbols{l},Tx_bits{l}] = PAMsource( ... + "fsym",fsym,"M",s.M,"order",17,"useprbs",0, ... + "fs_out",fdac, ... + "applyclipping",0,"clipfactor",1.5, ... + "applypulseform",1,"pulseformer",Pform, ... + "randkey",s.random_key+l, ... + "mrds_code",0,"mrds_blocklength",512,"duobinary_mode",duob_mode ... + ).process(); + + Lp_awg = Filter('filtdegree',3,"f_cutoff",56e9,"fs",fdac*kover, ... + "filterType",filtertypes.gaussian,"active",true); + + El_sig = AWG("fdac",fdac,"f_cutoff",fsym,"lpf_active",1,"kover",kover, ... + "bit_resolution",6,"upsampling_method","samplehold","precomp_sinc_rolloff",0, ... + "H_lpf",Lp_awg,"dac_max",0.6,"dac_min",-0.6).process(Digi_sig); + + % Digi_sig not needed after AWG + clear Digi_sig + + % Electrical Driver Amplifier + El_sig = El_sig.normalize("mode","oneone"); + scaling = 0.6*(u_pi/2-abs(vbias-u_pi/2)); + El_sig = El_sig .* scaling; + + % E/O Conversion + Eml_out = EML("mode",eml_mode.im_cosinus,"power",3,"fsimu",El_sig.fs, ... + "lambda",s.wavelengthplan(l),"bias",vbias,"u_pi",u_pi, ... + "linewidth",laser_linewidth,"randomkey",s.random_key+l,"alpha",s.chirpalpha).process(El_sig); + + % El_sig not needed after EML + clear El_sig + + signal_cell{l} = Polarization_Controller("mode","rot_power","desired_power",pol_rot(l)).process(Eml_out); + + % Eml_out not needed after pol controller + clear Eml_out Lp_awg +end + +disp('Signal generated for all channels.'); + +%% ---------- WDM mux + launch ---------- +Opt_sig_wdm = Optical_Multiplex("fs_in",fdac*kover,"fs_out",upsample_pow*fdac*kover, ... + "lambda_center",1310,"random_key",0,"filtype",1,"B",120e9).process(signal_cell); + +Opt_sig_wdm.spectrum(); + +Opt_sig_wdm = Amplifier("amp_mode","ideal_no_noise","gain_mode","output_power", ... + "amplification_db",s.p_launch+10*log10(N)).process(Opt_sig_wdm); + +Opt_sig_wdm_fib = Opt_sig_wdm; + +segment_length = 1; +nSegments = link_length/segment_length; +if abs(nSegments - round(nSegments)) > 1e-12 + error('fiber_length_km=%g must be an integer multiple of segment_length=%g km.', link_length, segment_length); +end +nSegments = round(nSegments); + +zdw = 1310; +randomize_D = true; + +% Guard for 0 km: avoid calling getDispersionVector(0,...) if it doesn't support it +if nSegments > 0 + Dvec = getDispersionVector(nSegments, d_local, zdw, randomize_D, s.random_key); +else + Dvec = []; +end + +for seg = 1:nSegments + + fprintf('Segment %d/%d \n',seg, nSegments); + + Opt_sig_wdm_fib = DP_Fiber("L",segment_length,"D",Dvec(seg),"Dpmd",s.pmd,"Ds",0.07, ... + "beat_len",10,"corr_len",100,"dz",1,"manakov",0, ... + "gamma",s.gamma,"lambda",zdw,"n_waveplates",10,"SS_dphimax",0.01, ... + "SS_dzmax",50,"SS_dzmin",10,"X_alpha",0.3,"X_beta",0,"rng",1,"useGPU",true,"useSingle",1).process(Opt_sig_wdm_fib); + +end + +profile off +profile viewer + +Opt_sig_wdm_fib.spectrum(); + +%% ========== BER Evaluation ========== +fprintf('\n========== BER Evaluation ==========\n'); +fprintf('Processing signals through receiver chain...\n'); + +% Receiver parameters +len_tr = 4096; % Training length +mu_dc = 0.005; +mu_ffe = [0.0001 0.0008 0.001]; +mu_dfe = 0.0004; + +% Helper function to process through receiver and get BER +function ber = process_receiver(Opt_sig_fib, l, Symbols, Tx_bits, ... + fdac, kover, upsample_pow, fsym, fadc, rop, s, len_tr, mu_dc, mu_ffe, mu_dfe, duob_mode) + +% Demux single channel +Opt_sig_demux = Optical_Demultiplex("attenuation",0,"B",200e9,"filtype",1, ... + "fs_out",fdac*kover,"fs_in",fdac*kover*upsample_pow,"lambda_center",1310).process(Opt_sig_fib); + +% ROP amplifier +Opt_sig_rx = Amplifier("amp_mode","ideal_no_noise","gain_mode","output_power", ... + "amplification_db",rop).process(Opt_sig_demux{l}); + +% Photodiode +PD_sig = Photodiode("fsimu",fdac*kover,"dark_current",2e-08,"responsivity",1,"temperature",20, ... + "nep",1.8e-11,"randomkey",s.random_key+l).process(Opt_sig_rx); + +% Low-pass filter +rx_bwl = 100e9; +PD_sig = Filter('filtdegree',4,"f_cutoff",rx_bwl,"fs",fdac*kover, ... + "filterType",filtertypes.butterworth,"active",true).process(PD_sig); + +% Scope +Lp_scpe = Filter('filtdegree',4,"f_cutoff",80e9,"fs",fadc,"filterType",filtertypes.butterworth,"active",true); +Scpe_sig = Scope("fsimu",fdac*kover,"fadc",fadc, ... + "delay",0,"fixed_delay",0,"filtertype",filtertypes.butterworth, ... + "samplingdelay",0,"rand_samplingdelay",0,"freq_offset",0,"samp_jitter",0, ... + "adcresolution",8,"quantbuffer",0.1,'block_dc',1,'lpf_active',0,'H_lpf',Lp_scpe).process(PD_sig); + +% Resample to 2 sps +Scpe_sig_2sps = Scpe_sig.resample("fs_out",2*fsym); + +% Time sync +[~, Scpe_cell, ~, ~] = Scpe_sig_2sps.tsynch("reference", Symbols{l}, "fs_ref", fsym, "debug_plots", 0); +Rx_sig = Scpe_cell{1}; +Rx_sig = Rx_sig.normalize("mode","rms"); + +% FFE Equalizer +ffe_order = [50, 0, 0]; +eq_ffe = EQ("Ne",ffe_order,"Nb",[0,0,0],"training_length",len_tr,"training_loops",5,"dd_loops",5, ... + "K",2,"DCmu",mu_dc,"DDmu",[mu_ffe mu_dfe],"DFEmu",0.005,"FFEmu",0,"plotfinal",0,"ideal_dfe",0); +eq_ffe = FFE("epochs_tr",5,"epochs_dd",2,"len_tr",2^13,"mu_dd",6.624e-05,"mu_tr",0.058136,"order",50,"sps",2,"decide",0, "adaption",adaption_method.nlms,"dd_mode",1); + +ffe_results = ffe(eq_ffe,s.M,Rx_sig,Symbols{l},Tx_bits{l}, ... + "precode_mode",duob_mode, ... + 'showAnalysis',0, ... + "postFFE",[], ... + "eth_style_symbol_mapping",0); + +ber = ffe_results.metrics.BER; +end + +% Process each ROP value for selected channels using parfor +ber_results = zeros(length(s.rop), N); + +% Flatten loop for parfor: iterate over all (ROP, channel) combinations +num_rop = length(s.rop); +rop_vals = s.rop; +ber_flat = zeros(num_rop * N, 1); + +parfor idx = 1:(num_rop * N) + % Convert linear index to (ri, l) subscripts + ri = ceil(idx / N); + l = mod(idx - 1, N) + 1; + + fprintf('ROP %d dBm, Channel %d/%d\n', rop_vals(ri), l, N); + ber_flat(idx) = process_receiver(Opt_sig_wdm_fib, l, Symbols, Tx_bits, ... + fdac, kover, upsample_pow, fsym, fadc, rop_vals(ri), s, len_tr, mu_dc, mu_ffe, mu_dfe, duob_mode); +end + +% Reshape back to [num_rop × N] matrix +ber_results = reshape(ber_flat, [N, num_rop]).'; + + +%% Display BER Results +fprintf('\n========== BER Results ==========\n'); +fprintf('ROP [dBm] | '); +for l = 1:N + fprintf('Ch%d | ', l); +end +fprintf('\n'); + +for ri = 1:length(s.rop) + fprintf('%8d | ', s.rop(ri)); + for l = 1:N + fprintf('%.2e | ', ber_results(ri, l)); + end + fprintf('\n'); +end + +% Plot BER vs ROP +figure; +semilogy(s.rop, mean(ber_results, 2), '-o', 'LineWidth', 2); +hold on; +for l = 1:N + semilogy(s.rop, ber_results(:, l), '--', 'LineWidth', 1); +end +hold off; +xlabel('ROP [dBm]'); +ylabel('BER'); +title('BER vs Received Optical Power (GPU Single Precision)'); +legend(['Mean', arrayfun(@(x) sprintf('Ch%d', x), 1:N, 'UniformOutput', false)]); +grid on; + +fprintf('\n========== Test Complete ==========\n'); diff --git a/test/gpu_processing_test.m b/test/gpu_processing_test.m new file mode 100644 index 0000000..b08c03f --- /dev/null +++ b/test/gpu_processing_test.m @@ -0,0 +1,70 @@ +% 1. Setup Data OUTSIDE the timer +d = gpuDevice; +N = 10000; + +fprintf('Preparing data...\n'); +A_cpu_double = rand(N, N); % Create double on CPU +A_cpu_single = single(A_cpu_double); % Create single on CPU + +% Warmup run (wakes up the GPU from idle state) +A_warm = gpuArray.rand(1000, 1000, 'single'); +B_warm = A_warm * A_warm; +wait(d); + +fprintf('------------------------------------------------\n'); + +% TEST 1: Double Precision (The "Slow" way) +% We move data to GPU first so we only measure calculation time +A_gpu_double = gpuArray(A_cpu_double); +wait(d); % Ensure transfer is done before starting timer + +fprintf('Running DOUBLE precision test... '); +tic; +B_gpu = A_gpu_double * A_gpu_double; +wait(d); % FORCE MATLAB TO WAIT FOR GPU +time_double = toc; +fprintf('Done.\n'); +fprintf('Double Precision Time: %.4f seconds\n', time_double); + +% TEST 2: Single Precision (The "Fast" way) +A_gpu_single = gpuArray(A_cpu_single); +wait(d); % Ensure transfer is done + +fprintf('Running SINGLE precision test... '); +tic; +B_gpu = A_gpu_single * A_gpu_single; +wait(d); % FORCE MATLAB TO WAIT FOR GPU +time_single = toc; +fprintf('Done.\n'); +fprintf('Single Precision Time: %.4f seconds\n', time_single); + +% Calculate Speedup +fprintf('------------------------------------------------\n'); +fprintf('Speedup Factor using single precision: %.2fx\n', time_double / time_single); + +% +% Create 10,000 small matrices (10x10) stacked in a 3D array +A_stack = gpuArray.rand(10, 10, 10000, 'single'); +B_stack = gpuArray.rand(10, 10, 10000, 'single'); + +% BAD: Looping (GPU overhead kills you) +tic; +for i=1:10000 + C(:,:,i) = A_stack(:,:,i) * B_stack(:,:,i); +end +wait(d); +loop_time = toc; + +% GOOD: Pagefun (Executes all 10,000 mults simultaneously) +tic; +C_stack = pagefun(@mtimes, A_stack, B_stack); +wait(d); +pagefun_time = toc; + +fprintf('------------------------------------------------\n'); +fprintf('Speedup Factor using pagefun: %.2fx\n', loop_time / pagefun_time); + + +fprintf('Total VRAM: %.2f GB\n', d.TotalMemory / 1e9); +fprintf('Available VRAM: %.2f GB\n', d.AvailableMemory / 1e9); +fprintf('Usage: %.1f%%\n', 100 * (1 - d.AvailableMemory / d.TotalMemory)); \ No newline at end of file