Functionally useless implementation of overlapping FFT's. The overlaps should be calculated right but seem to not be rendered. (Rendered result is same as without any overlapping.)

Originally committed to SVN as r1302.

aegisub/audio_spectrum.cpp

@@ -44,7 +44,6 @@
 
 // Audio spectrum FFT data cache
 
-AudioSpectrumCache::LineAge AudioSpectrumCache::cache_line_age_limit = 0x10000; // TODO: make this less arbitrary
 AudioSpectrumCache::CacheLine AudioSpectrumCache::null_line;
 unsigned long AudioSpectrumCache::line_length;
 
@@ -61,44 +60,35 @@ class FinalSpectrumCache : public AudioSpectrumCache {
 private:
 	std::vector<CacheLine> data;
 	unsigned long start, length; // start and end of range
-	unsigned long last_access; // aging parameter
+	unsigned int overlaps;
 
 public:
-	CacheLine& GetLine(unsigned long i, LineAge aging_time)
+	CacheLine& GetLine(unsigned long i, unsigned int overlap)
 	{
-		last_access = aging_time;
-
 		// This check ought to be redundant
 		if (i >= start && i-start < length)
-			return data[i - start];
+			return data[i - start + overlap*length];
 		else
 			return null_line;
 	}
 
-	bool Age(LineAge aging_time)
+	FinalSpectrumCache(AudioProvider *provider, unsigned long _start, unsigned long _length, unsigned int _overlaps)
-	{
-		assert(aging_time >= last_access);
-
-		return aging_time - last_access <= cache_line_age_limit;
-	}
-
-	FinalSpectrumCache(AudioProvider *provider, unsigned long _start, unsigned long _length)
 	{
 		start = _start;
 		length = _length;
-		last_access = 0;
+		overlaps = _overlaps;
+
+		if (overlaps <  1) overlaps = 1;
+		// Add an upper limit to number of overlaps or trust user to do sane things?
+		// Any limit should probably be a function of length
 
 		assert(length > 2);
 
 		// First fill the data vector with blanks
 		// Both start and end are included in the range stored, so we have end-start+1 elements
-		data.resize(length, null_line);
+		data.resize(length*overlaps, null_line);
 
-		// Start sample number of the next line calculated
+		unsigned int overlap_offset = length / overlaps / 2;
-		// line_length is half of the number of samples used to calculate a line, since half of the output from
-		// a Fourier transform of real data is redundant, and not interesting for the purpose of creating
-		// a frequenmcy/power spectrum.
-		__int64 sample = start * line_length*2;
 
 		// Raw sample data
 		short *raw_sample_data = new short[line_length*2];
@@ -107,23 +97,31 @@ public:
 		float *out_r = new float[line_length*2];
 		float *out_i = new float[line_length*2];
 
-		FFT fft; // TODO: use FFTW instead? A wavelet?
+		FFT fft; // Use FFTW instead? A wavelet?
 
-		for (unsigned long i = 0; i < length; ++i) {
+		for (unsigned int overlap = 0; overlap < overlaps; ++overlap) {
-			provider->GetAudio(raw_sample_data, sample, line_length*2);
+			// Start sample number of the next line calculated
-			for (size_t j = 0; j < line_length; ++j) {
+			// line_length is half of the number of samples used to calculate a line, since half of the output from
-				sample_data[j*2] = (float)raw_sample_data[j*2];
+			// a Fourier transform of real data is redundant, and not interesting for the purpose of creating
-				sample_data[j*2+1] = (float)raw_sample_data[j*2+1];
+			// a frequenmcy/power spectrum.
+			__int64 sample = start * line_length*2 + overlap*overlap_offset;
+
+			for (unsigned long i = 0; i < length; ++i) {
+				provider->GetAudio(raw_sample_data, sample, line_length*2);
+				for (size_t j = 0; j < line_length; ++j) {
+					sample_data[j*2] = (float)raw_sample_data[j*2];
+					sample_data[j*2+1] = (float)raw_sample_data[j*2+1];
+				}
+
+				fft.Transform(line_length*2, sample_data, out_r, out_i);
+
+				CacheLine &line = data[i + length*overlap];
+				for (size_t j = 0; j < line_length; ++j) {
+					line[j] = sqrt(out_r[j]*out_r[j] + out_i[j]*out_i[j]);
+				}
+
+				sample += line_length*2;
 			}
-
-			fft.Transform(line_length*2, sample_data, out_r, out_i);
-
-			CacheLine &line = data[i];
-			for (size_t j = 0; j < line_length; ++j) {
-				line[j] = sqrt(out_r[j]*out_r[j] + out_i[j]*out_i[j]);
-			}
-
-			sample += line_length*2;
 		}
 
 		delete[] raw_sample_data;
@@ -145,12 +143,13 @@ class IntermediateSpectrumCache : public AudioSpectrumCache {
 private:
 	std::vector<AudioSpectrumCache*> sub_caches;
 	unsigned long start, length, subcache_length;
+	unsigned int overlaps;
 	bool subcaches_are_final;
 	int depth;
 	AudioProvider *provider;
 
 public:
-	CacheLine &GetLine(unsigned long i, LineAge aging_time)
+	CacheLine &GetLine(unsigned long i, unsigned int overlap)
 	{
 		if (i >= start && i-start <= length) {
 			// Determine which sub-cache this line resides in
@@ -159,44 +158,24 @@ public:
 
 			if (!sub_caches[subcache]) {
 				if (subcaches_are_final) {
-					sub_caches[subcache] = new FinalSpectrumCache(provider, start+subcache*subcache_length, subcache_length);
+					sub_caches[subcache] = new FinalSpectrumCache(provider, start+subcache*subcache_length, subcache_length, overlaps);
 				} else {
-					sub_caches[subcache] = new IntermediateSpectrumCache(provider, start+subcache*subcache_length, subcache_length, depth+1);
+					sub_caches[subcache] = new IntermediateSpectrumCache(provider, start+subcache*subcache_length, subcache_length, overlaps, depth+1);
 				}
 			}
 
-			return sub_caches[subcache]->GetLine(i, aging_time);
+			return sub_caches[subcache]->GetLine(i, overlap);
 		} else {
 			return null_line;
 		}
 	}
 
-	bool Age(LineAge aging_time)
+	IntermediateSpectrumCache(AudioProvider *_provider, unsigned long _start, unsigned long _length, unsigned int _overlaps, int _depth)
-	{
-		bool living_caches = false;
-
-		// Age every active sub-cache
-		for (size_t i = 0; i < sub_caches.size(); ++i) {
-			if (sub_caches[i]) {
-				if (sub_caches[i]->Age(aging_time)) {
-					// The sub-cache is still fresh, so we are too
-					living_caches = true;
-				} else {
-					// Sub-cache is too old, remove it
-					delete sub_caches[i];
-					sub_caches[i] = 0;
-				}
-			}
-		}
-
-		return living_caches;
-	}
-
-	IntermediateSpectrumCache(AudioProvider *_provider, unsigned long _start, unsigned long _length, int _depth)
 	{
 		provider = _provider;
 		start = _start;
 		length = _length;
+		overlaps = _overlaps;
 		depth = _depth;
 
 		// FIXME: this calculation probably needs tweaking
@@ -234,16 +213,16 @@ AudioSpectrum::AudioSpectrum(AudioProvider *_provider, unsigned long _line_lengt
 	//assert (_num_lines < (1<<31)); // hope it fits into 32 bits...
 	num_lines = (unsigned long)_num_lines;
 
+	fft_overlaps = Options.AsInt(_T("Audio Spectrum Overlaps"));
+	fft_overlaps = MAX(1, fft_overlaps);
 	AudioSpectrumCache::SetLineLength(line_length);
-	cache = new IntermediateSpectrumCache(provider, 0, num_lines, 0);
+	cache = new IntermediateSpectrumCache(provider, 0, num_lines, fft_overlaps, 0);
-	cache_age = 0;
 
 	power_scale = 1;
 	minband = Options.AsInt(_T("Audio Spectrum Cutoff"));
 	maxband = line_length - minband * 2/3; // TODO: make this customisable?
 
 	// Generate colour maps
-	// TODO? allow selecting between several colour maps
 	unsigned char *palptr = colours_normal;
 	for (int i = 0; i < 256; i++) {
 		//hsl_to_rgb(170 + i * 2/3, 128 + i/2, i, palptr+0, palptr+1, palptr+2);	// Previous
@@ -267,8 +246,8 @@ AudioSpectrum::~AudioSpectrum()
 
 void AudioSpectrum::RenderRange(__int64 range_start, __int64 range_end, bool selected, unsigned char *img, int imgleft, int imgwidth, int imgpitch, int imgheight)
 {
-	unsigned long first_line = (unsigned long)(range_start / line_length / 2);
+	unsigned long first_line = (unsigned long)(fft_overlaps * range_start / line_length / 2);
-	unsigned long last_line = (unsigned long)(range_end / line_length / 2);
+	unsigned long last_line = (unsigned long)(fft_overlaps * range_end / line_length / 2);
 
 	float *power = new float[line_length];
 
@@ -287,17 +266,23 @@ void AudioSpectrum::RenderRange(__int64 range_start, __int64 range_end, bool sel
 	const double onethirdmaxpower = maxpower / 3, twothirdmaxpower = maxpower * 2/3;
 	const double logoverscale = log(maxpower*upscale - twothirdmaxpower);
 
+	// Note that here "lines" are actually bands of power data
+	unsigned long baseline = first_line / fft_overlaps;
+	unsigned int overlap = first_line % fft_overlaps;
 	for (unsigned long i = first_line; i <= last_line; ++i) {
 		// Handle horizontal compression and don't unneededly re-render columns
 		int imgcol = imgleft + imgwidth * (i - first_line) / (last_line - first_line + 1);
 		if (imgcol <= last_imgcol_rendered)
 			continue;
 
-		AudioSpectrumCache::CacheLine &line = cache->GetLine(i, cache_age);
+		AudioSpectrumCache::CacheLine &line = cache->GetLine(baseline, overlap);
-		cache_age++;
+		++overlap;
+		if (overlap >= fft_overlaps) {
+			overlap = 0;
+			++baseline;
+		}
 
-		// Calculate the signal power over frequency
+		// Apply a "compressed" scaling to the signal power
-		// "Compressed" scale
 		for (unsigned int j = 0; j < line_length; j++) {
 			// First do a simple linear scale power calculation -- 8 gives a reasonable default scaling
 			power[j] = line[j] * upscale;
@@ -361,8 +346,6 @@ void AudioSpectrum::RenderRange(__int64 range_start, __int64 range_end, bool sel
 	}
 
 	delete[] power;
-
-	cache->Age(cache_age);
 }
 
 

aegisub/audio_spectrum.h

@@ -45,24 +45,21 @@
 // Spectrum cache basically caches the raw result of FFT
 class AudioSpectrumCache {
 public:
+	// Type of a single FFT result line
 	typedef std::vector<float> CacheLine;
-	typedef unsigned long LineAge;
 
-	// Get the i'th cache line from the tree and propagate down the current "age time"
+	// Get the overlap'th overlapping FFT in FFT group i, generating it if needed
-	virtual CacheLine& GetLine(unsigned long i, LineAge aging_time) = 0;
+	virtual CacheLine& GetLine(unsigned long i, unsigned int overlap) = 0;
 
-	// Set the desired length of cache lines
+	// Set the FFT size used
 	static void SetLineLength(unsigned long new_length);
 
-	// Perform cache aging process
-	// Return true if the object Age was called on is still fresh, false if it's old and should be purged
-	virtual bool Age(LineAge aging_time) = 0;
-
 	virtual ~AudioSpectrumCache() {};
 
 protected:
-	static LineAge cache_line_age_limit;
+	// A cache line containing only zero-values
 	static CacheLine null_line;
+	// The FFT size
 	static unsigned long line_length;
 };
 
@@ -71,7 +68,6 @@ class AudioSpectrum {
 private:
 	// Data provider
 	AudioSpectrumCache *cache;
-	AudioSpectrumCache::LineAge cache_age;
 
 	// Colour pallettes
 	unsigned char colours_normal[256*3];
@@ -81,6 +77,7 @@ private:
 
 	unsigned long line_length; // number of frequency components per line (half of number of samples)
 	unsigned long num_lines; // number of lines needed for the audio
+	unsigned int fft_overlaps; // number of overlaps used in FFT
 	float power_scale; // amplification of displayed power
 	int minband; // smallest frequency band displayed
 	int maxband; // largest frequency band displayed

aegisub/dialog_options.cpp

@@ -529,6 +529,7 @@ DialogOptions::DialogOptions(wxWindow *parent)
 		AddTextControl(audioAdvPage,audioAdvSizer1,_("HD cache name"),_T("Audio HD CAche Name"));
 		AddTextControl(audioAdvPage,audioAdvSizer1,_("Spectrum cutoff"),_T("Audio spectrum cutoff"),TEXT_TYPE_NUMBER);
 		AddTextControl(audioAdvPage,audioAdvSizer1,_("Spectrum FFT window exponent"),_T("Audio spectrum window"),TEXT_TYPE_NUMBER);
+		AddTextControl(audioAdvPage,audioAdvSizer1,_("Spectrum FFT window overlaps"),_T("Audio spectrum overlaps"),TEXT_TYPE_NUMBER);
 		audioAdvSizer1->AddGrowableCol(0,1);
 
 		// Main sizer

aegisub/options.cpp

@@ -189,8 +189,10 @@ void OptionsManager::LoadDefaults(bool onlyDefaults) {
 	SetText(_T("Audio Alsa Device"), _T("plughw:0,0"));
 	SetText(_T("Audio HD Cache Location"),_T("default"));
 	SetText(_T("Audio HD Cache Name"),_T("audio%02i.tmp"));
+	// Technically these can do with just the spectrum class being re-created
 	SetInt(_T("Audio Spectrum Cutoff"),0);
 	SetInt(_T("Audio Spectrum Window"),8);
+	SetInt(_T("Audio Spectrum Overlaps"),1);
 
 	// Automation
 	// The path changes only take effect when a script is (re)loaded but Automatic should be good enough, it certainly doesn't warrart a restart