Faster DP_fiber with GPU processing... Run test\gpu_cpu_comparison.m to see difference on your setup
This commit is contained in:
silas (home)
219
test/bayesopt_ffe_tuning.m
Normal file
219
test/bayesopt_ffe_tuning.m
Normal file
@@ -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);
|
||||
349
test/gpu_cpu_comparison.m
Normal file
349
test/gpu_cpu_comparison.m
Normal file
@@ -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');
|
||||
265
test/gpu_processing_dpfiber.m
Normal file
265
test/gpu_processing_dpfiber.m
Normal file
@@ -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');
|
||||
70
test/gpu_processing_test.m
Normal file
70
test/gpu_processing_test.m
Normal file
@@ -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));
|
||||
Reference in New Issue
Block a user