#include "nodes_audio_analysis.h"

namespace Rotor{
	bool Audio_thumbnailer::init(int _channels,int _bits,int _samples,int _rate) {
		//base_audio_processor::init(_channels,_bits,_samples);
		channels=_channels;
		bits=_bits;
		samples=_samples;
		samples_per_column=samples/width;
		offset=0x1<<(bits-1); //signed audio
		scale=1.0/offset;

		out_sample=0; //sample in whole track
		sample=0;
		samples=0;
		accum=0.0;
		return true;
	}
	int Audio_thumbnailer::process_frame(uint8_t *_data,int samples_in_frame){
		//begin by processing remaining samples
		//samples per column could be larger than a frame! (probably is)
		//but all we are doing is averaging
		int bytes=(bits>>3);
		int stride=channels*bytes;
		int in_sample=0;
		while (in_sample<samples_in_frame) {
			//continue the column
			while (sample<samples_per_column&&in_sample<samples_in_frame) {
				//accumulate samples for this column until we run out of samples
				for (int i=0;i<channels;i++) {
					unsigned int this_val=0;
					for (int j=0;j<bytes;j++) {
						this_val+=_data[(in_sample*stride)+(i*bytes)+j]<<(j*8);
					}
					//convert from integer data format - i.e s16p - to audio signal in -1..1 range
					//presume 16 bits for now...
					double val=((double)((int16_t)this_val))*scale;
					accum+=val*val;
					samples++;
				}
				in_sample++;
				sample++;
				out_sample++;
			}
			if (sample==samples_per_column) { //finished a column
				//get root-mean
				//why does valgrind complain here about uninitialised vars
				double mean=pow(accum/samples,0.5f);
				audiodata.push_back(mean);
				sample=0;
				samples=0;
				accum=0.0;
			}
		}
		return out_sample;
	}
	void Audio_thumbnailer::print_vector(xmlIO XML){
		string vdata;
		int i=0;
		for (auto sample: audiodata){
			if (i>0) vdata+=",";
			vdata+=toString(sample);
		}
		XML.addValue("data",vdata);
	}
	bool Vamp_node::init(int _channels,int _bits,int _samples, int _rate) {
		//need these to make sense of data
		channels=_channels;
		bits=_bits;
		samples=_samples;

		features.clear();
		return analyser.init(soname,id,_channels,_bits,_samples,_rate,outputNo,params);


		//attempt to load vamp plugin and prepare to receive frames of data
		//should the audio analysis contain a vamphost or should it inherit?
		//maybe neater to contain it in terms of headers etc

	}
	int Vamp_node::process_frame(uint8_t *data,int samples_in_frame) {
		analyser.process_frame(data,samples_in_frame);
		return 1;
	}
	void Vamp_node::cleanup() {
		analyser.cleanup();
		features=analyser.features;
	}
	string Vamp_node::get_features(){
		string data;
		for (auto i: features) {
			data=data+" ["+toString(i.second.number)+":"+toString(i.first);
			if (i.second.values.size()) {
				data+=" (";
				bool first=true;
				for (auto j: i.second.values) {
					if (first){
						first=false;
					}
					else data+=",";
					data=data+toString(j);
				}
				data+=") ";
			}
			data+="]";
		}
		return data;
	}
	bool sortsegments(std::pair<int,double> i,std::pair<int,double> j){
		return (i.second<j.second);
	}

	bool sortseggrps(std::pair<double,vector<pair<double,int> > > i,std::pair<double,vector<pair<double,int> > > j){
		return (i.first<j.first);
	}
	void Intensity_segmenter::cleanup(){
		//algorithm idea:
		//get average tempo and intensity for each segment and store them
		//scale by the range to get a value from 0.0 to 1.0
		//add tempo and intensity according to a weighting
		//score the results (ie 1st place, 2nd place) to end up with a set of integer numbers

		//how to group with similarity?
		//segments come with similarity groups
		// 1 - are the wanted git checksegments less than discovered?
		// N - do nothing
		// 2 - get intensity and tempo averages
		// 2 - count the groups
		// 3 - are the groups less than the discovered segments?
		// N - group by intensity as normal
		// 4 - are the groups less than the wanted levels?


		//for (auto a:analysers) a.second.cleanup(); //WHY NOT WORK - its as if the call is const
		analysers["segmenter"].cleanup();
		analysers["tempo"].cleanup();
		analysers["intensity"].cleanup();

		//combine with similarity numbers
		// 1. count similarity numbers

		map<int,vector<int> > similarities;

		//what do we want to know about these similarities?
		// how many are there? map.size()
		// how many members are in each one? map[item].size()
		// which members are they? auto m: map[item]

		uint32_t i=0;
		for (auto f:analysers["segmenter"].features) {
			if (f.second.values.size()) {
				int group=f.second.values[0];
				if (similarities.find(group)==similarities.end()){
					similarities[group]={};
				}
				similarities[group].push_back(i);
			}
			i++;
		}
		for (auto s:similarities) cerr<<"group "<<s.first<<" count: "<<s.second.size()<<endl;

		cerr<<analysers["segmenter"].features.size()<<" segments"<<endl;
		cerr<<analysers["tempo"].features.size()<<" tempo features"<<endl;
		cerr<<analysers["intensity"].features.size()<<" intensity features"<<endl;
		i=0;
		double min_tempo=9999999.0;
		double min_intensity=9999999.0;
		double max_tempo=0.0;
		double max_intensity=0.0;
		vector<double> tempos;
		vector<double> intensities;
		vector<double> times;
		auto g=++analysers["segmenter"].features.begin();
		for (auto f=analysers["segmenter"].features.begin();g!=analysers["segmenter"].features.end();f++,g++,i++){
			cerr<<"segment "<<i<<": "<<f->first<<" to "<<g->first<<endl;
			times.push_back(f->first);
			//integrate tempo and intensity algorithmically
			double tempo=0;
			if (analysers["tempo"].features.size()) {
				double pt=f->first;
				double pv=analysers["tempo"].get_value(f->first);
				for (auto u=analysers["tempo"].features.upper_bound(f->first);u!=analysers["tempo"].features.upper_bound(g->first);u++){
					tempo +=(u->first-pt)*(u->second.values[0]+pv)*0.5f; //area of the slice
					pt=u->first;
					pv=u->second.values[0];
				}
				tempo +=(g->first-pt)*(analysers["tempo"].get_value(g->first)+pv)*0.5f; //area of the last slice
				tempo /=g->first-f->first; //average value;
			}
			if (tempo>max_tempo) max_tempo=tempo;
			if (tempo<min_tempo) min_tempo=tempo;
			tempos.push_back(tempo);
			cerr<<"segment "<<i<<" average tempo: "<<tempo<<endl;

			double intensity=0;
			if (analysers["intensity"].features.size()) {
				double pt=f->first;
				double pv=analysers["intensity"].get_value(f->first);
				for (auto u=analysers["intensity"].features.upper_bound(f->first);u!=analysers["intensity"].features.upper_bound(g->first);u++){
					intensity +=(u->first-pt)*(u->second.values[0]+pv)*0.5f; //area of the slice
					pt=u->first;
					pv=u->second.values[0];
				}
				intensity +=(g->first-pt)*(analysers["intensity"].get_value(g->first)+pv)*0.5f; //area of the last slice
				intensity /=g->first-f->first; //average value;
			}
			if (intensity>max_intensity) max_intensity=intensity;
			if (intensity<min_intensity) min_intensity=intensity;
			intensities.push_back(intensity);
		}
		//make relative scale 0.0-1.0 and save weighted totals
		vector< pair<int,double>> totals;
		vector<double> totalsmap;
		for (i=0;i<tempos.size();i++){
			tempos[i]=(tempos[i]-min_tempo)/(max_tempo-min_tempo);
			intensities[i]=(intensities[i]-min_intensity)/(max_intensity-min_intensity);
			totals.push_back(make_pair(i,(tempos[i]*parameters["tempo_weight"]->value)+(intensities[i]*parameters["intensity_weight"]->value)));
			totalsmap.push_back((tempos[i]*parameters["tempo_weight"]->value)+(intensities[i]*parameters["intensity_weight"]->value));
		}

		
		//sort and convert to features
		std::sort(totals.begin(),totals.end(),sortsegments);
		for (i=0;i<totals.size();i++) {
			cerr<<"segment "<<totals[i].first<<" average intensity: "<<totals[i].second<<endl;
		}
		vector<double> bucketoffsets;
		for (auto t:totals) bucketoffsets.push_back(0.0);
		if (parameters["levels"]->value>0.0&&parameters["levels"]->value<totals.size()){
			//use bucketoffsets to redistribute into smaller number of buckets
			int numbertoredistribute=totals.size()-((int)parameters["levels"]->value);
			double numberperbin=((double)numbertoredistribute/totals.size());
			double toadd=0.5f;
			int added=0;
			for (int j=0;j<totals.size();j++){
				int numbertoadd=min(numbertoredistribute-added,(int)toadd);
				toadd=(toadd+numberperbin)-numbertoadd;
				added+=numbertoadd;
				bucketoffsets[j]=added;
			}
			if (numbertoredistribute>0) {
				cerr<<"reducing number of levels by "<<numbertoredistribute<<", offsets:"<<endl;
				for (auto o:bucketoffsets) {
					cerr<<o<<":";
				}
				cerr<<endl;
			}
		}

		for (i=0;i<totals.size();i++){
			vampHost::feature f;
			f.values.push_back((double)i-bucketoffsets[i]);
			features[times[totals[i].first]]=f;
		}

<<<<<<< HEAD
=======
		/*
sort intensity totals
find out how many segments will share levels apart from similarity levels
start with a structure:
map<inputnum,vector<pair<tempo,inputnum>>
start grouping by similarity
if there are more similarity groups than wantedgroups, start by grouping similarities
otherwise take biggest similarity groups and split them by intensity
if there are still too many groups, merge closest smallest groups
finally sort by intensity to map output

nned to retrieve total intensity by segment
		*/
		//  segment  group_intensity   seg_intense segment
		vector<pair<double,vector<pair<double,int> > > > seggrps;
		for (i=0;i<totalsmap.size();i++){
		    vector<pair<double,int> > data;
		    data.push_back(make_pair(totalsmap[i],i));
			seggrps.push_back(make_pair(totalsmap[i],data));
		}
		for (auto s:similarities){
			if (s.second.size()>1){
				for (int j=s.second.size()-1;j>0;j--){
				seggrps[s.second[0]].second.push_back(make_pair(totalsmap[s.second[j]],s.second[j]));
				//keep running average// should be by area?
				//seggrps[s.second[0]].first+=(totalsmap[s.second[j]]*(1.0/max(1,(int)seggrps[s.second[0]].second.size()-1)));
				//double div=seggrps[s.second[0]].second.size()==1?1.0:((double)seggrps[s.second[0]].second.size()-1/(double)seggrps[s.second[0]].second.size());
				//neat! this gives 1,1/2,2/3,3/4..
				//seggrps[s.second[0]].first*=div;

				//easier is to
				double avg=0.0f;
				for (auto p:seggrps[s.second[0]].second) avg+=p.first;
				avg/=seggrps[s.second[0]].second.size();
				seggrps[s.second[0]].first=avg;

				seggrps.erase(seggrps.begin()+s.second[j]);
				}
			}
		}
		cerr<<"similarities assigned, "<<(totalsmap.size()-seggrps.size())<<" segments merged"<<endl;
		//sort the contents by intensity
		std::sort(seggrps.begin(),seggrps.end(),sortseggrps);
		//possible mergers will be with groups with adjacent intensity
		if (((int)parameters["levels"]->value)>0) {
			while (seggrps.size()>(int)parameters["levels"]->value){
				//reduce similarity groups
				//decide the best 2 to merge
				vector<double> diffs;
				for (int j=0;j<seggrps.size()-1;j++) diffs.push_back(seggrps[j+1].first-seggrps[j].first);
				int smallest=0;
				for (int j=1;j<diffs.size();j++) if (diffs[j]<diffs[smallest]) smallest=j;
				for (int j=0;j<seggrps[smallest].second.size();j++) {
					seggrps[smallest+1].second.push_back(seggrps[smallest].second[j]);
				}
				//recalculate intensity average
				double avg=0.0f;
				for (auto p:seggrps[smallest+1].second) avg+=p.first;
				avg/=seggrps[smallest+1].second.size();
				seggrps[smallest+1].first=avg;

				seggrps.erase(seggrps.begin()+smallest);
			}
			cerr<<"intensities merged, "<<seggrps.size()<<" levels remain"<<endl;
			while (seggrps.size()<min((int)parameters["levels"]->value,(int)totalsmap.size())) {
				//split groups
				//calculate standard deviation of intensity variation
				vector<double> devs;
				for (int j=0;j<seggrps.size()-1;j++) {
					double avg=0.0;
					double dev=0.0;
					for (int k=0;k<seggrps[j].second.size();k++){
						avg+=k<seggrps[j].second[k].first;
					}
					avg/=seggrps[j].second.size();
					for (int k=0;k<seggrps[j].second.size();k++){
						dev+=pow(avg-k<seggrps[j].second[k].first,2.0);
					}
					dev/=seggrps[j].second.size();
					devs.push_back(pow(dev,0.5));
				}
				//find group with largest standard deviation
				int largest=0;
				for (int j=1;j<devs.size();j++) if (devs[j]>devs[largest]) largest=j;
				//sanity check: if there are any groups that can be split they will have larger SD than singleton groups
				//TODO actually split group
			}
		}

		map<int,int> outputvalues;
		for (int j=0;j<seggrps.size();j++){
			for (int k=0;k<seggrps[j].second.size();k++){
				outputvalues[seggrps[j].second[k].second]=j;
			}
		}


		for (i=0;i<totals.size();i++){
			vampHost::feature f;
			f.values.push_back(outputvalues[i]);
			features[times[totals[i].first]]=f;
		}
>>>>>>> intensitysimilarity
	}
}

/*
A data structure to represent segments and their mapping to output levels
how do we merge the intensity groups with the similarities?

we create a list




... we iterate through the list of segments and place the right output number


*/