c@69: /*
c@69:     Constant-Q library
c@69:     Copyright (c) 2013-2014 Queen Mary, University of London
c@69: 
c@69:     Permission is hereby granted, free of charge, to any person
c@69:     obtaining a copy of this software and associated documentation
c@69:     files (the "Software"), to deal in the Software without
c@69:     restriction, including without limitation the rights to use, copy,
c@69:     modify, merge, publish, distribute, sublicense, and/or sell copies
c@69:     of the Software, and to permit persons to whom the Software is
c@69:     furnished to do so, subject to the following conditions:
c@69: 
c@69:     The above copyright notice and this permission notice shall be
c@69:     included in all copies or substantial portions of the Software.
c@69: 
c@69:     THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
c@69:     EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
c@69:     MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
c@69:     NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY
c@69:     CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
c@69:     CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
c@69:     WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
c@69: 
c@69:     Except as contained in this notice, the names of the Centre for
c@69:     Digital Music; Queen Mary, University of London; and Chris Cannam
c@69:     shall not be used in advertising or otherwise to promote the sale,
c@69:     use or other dealings in this Software without prior written
c@69:     authorization.
c@69: */
c@10: 
c@37: module cqt;
c@10: 
c@10: cqtkernel = load cqtkernel;
c@10: resample = load may.stream.resample;
c@10: manipulate = load may.stream.manipulate;
c@72: mat = load may.matrix;
c@10: cm = load may.matrix.complex;
c@10: framer = load may.stream.framer;
c@10: cplx = load may.complex;
c@10: fft = load may.transform.fft;
c@10: vec = load may.vector;
c@42: ch = load may.stream.channels;
c@10: 
c@10: { pow, round, floor, ceil, log2, nextPowerOfTwo } = load may.mathmisc;
c@10: 
c@37: cqt { maxFreq, minFreq, binsPerOctave } str =
c@10:    (sampleRate = str.sampleRate;
c@10:     octaves = ceil (log2 (maxFreq / minFreq));
c@65: //    actualMinFreq = (maxFreq / (pow 2 octaves)) * (pow 2 (1/binsPerOctave));
c@10: 
c@41:     kdata = cqtkernel.makeKernel { sampleRate, maxFreq, binsPerOctave };
c@10: 
c@63: //    eprintln "sampleRate = \(sampleRate), maxFreq = \(maxFreq), minFreq = \(minFreq), actualMinFreq = \(actualMinFreq), octaves = \(octaves), binsPerOctave = \(binsPerOctave), fftSize = \(kdata.fftSize), hop = \(kdata.fftHop)";
c@10: 
c@63: //    eprintln "atomsPerFrame = \(kdata.atomsPerFrame)";
c@11: 
c@41:     padding = (kdata.fftSize * (pow 2 (octaves-1)));
c@40: 
c@63: //    eprintln "padding = \(padding)";
c@40: 
c@40:     str = manipulate.paddedBy padding str;
c@40: 
c@10:     streams = manipulate.duplicated octaves str;
c@10: 
c@10:     //!!! can't be right!
c@10:     kernel = cm.transposed (cm.conjugateTransposed kdata.kernel);
c@10: 
c@63: //    eprintln "have kernel";
c@10: 
c@10:     fftFunc = fft.forward kdata.fftSize;
c@10: 
c@10:     cqblocks =
c@10:         map do octave:
c@42:             frames = map ch.mixedDown //!!! mono for now
c@42:                (framer.frames kdata.fftSize [ Hop kdata.fftHop, Padded false ]
c@42:                    (resample.decimated (pow 2 octave) streams[octave]));
c@10:             map do frame:
c@10:                 freq = fftFunc (cplx.complexArray frame (vec.zeros kdata.fftSize));
c@43: // eprintln "octave = \(octave), frame = \(vec.list frame)";
c@43: // eprintln "octave = \(octave), freq = \(freq)";
c@10:                 cm.product kernel (cm.newComplexColumnVector freq);
c@10:             done frames;
c@10:         done [0..octaves-1];
c@10: 
c@13:     // The cqblocks list is a list<list<matrix>>. Each top-level list
c@11:     // corresponds to an octave, from highest to lowest, each having
c@11:     // twice as many elements in its list as the next octave. The
c@11:     // sub-lists are sampled in time with an effective spacing of
c@11:     // fftSize * 2^(octave-1) audio frames, and the matrices are row
c@11:     // vectors with atomsPerFrame * binsPerOctave complex elements.
c@13:     //
c@13:     // ***
c@13:     // 
c@13:     // In a typical constant-Q structure, each (2^(octaves-1) *
c@13:     // fftHop) input frames gives us an output structure conceptually
c@13:     // like this:
c@10:     //
c@10:     // [][][][][][][][]   <- fftHop frames per highest-octave output value
c@10:     // [][][][][][][][]      layered as many times as binsPerOctave (here 2)
c@10:     // [--][--][--][--]   <- fftHop*2 frames for the next lower octave
c@10:     // [--][--][--][--]      etc
c@10:     // [------][------]
c@10:     // [------][------]
c@10:     // [--------------]
c@10:     // [--------------]
c@10:     //
c@13:     // ***
c@13:     //
c@13:     // But the kernel we're using here has more than one temporally
c@13:     // spaced atom; each individual cell is a row vector with
c@13:     // atomsPerFrame * binsPerOctave elements, but that actually
c@13:     // represents a rectangular matrix of result cells with width
c@13:     // atomsPerFrame and height binsPerOctave. The columns of this
c@13:     // matrix (the atoms) then need to be spaced by 2^(octave-1)
c@13:     // relative to those from the highest octave.
c@10: 
c@15:     // Reshape each row vector into the appropriate rectangular matrix
c@21:     // and split into single-atom columns
c@19: 
c@44:     emptyHops = kdata.firstCentre / kdata.atomSpacing; //!!! int? round?
c@65: //    maxDrop = emptyHops * (pow 2 (octaves-1)) - emptyHops;
c@63: //    eprintln "maxDrop = \(maxDrop)";
c@21: 
c@47:     cqblocks =
c@47:         map do octlist:
c@47:             concatMap do rv:
c@21:                 cm.asColumns
c@21:                    (cm.generate do row col:
c@21:                         cm.at rv ((row * kdata.atomsPerFrame) + col) 0
c@21:                     done {
c@21:                         rows = kdata.binsPerOctave,
c@21:                         columns = kdata.atomsPerFrame
c@21:                     })
c@47:             done octlist
c@47:         done cqblocks;
c@21: 
c@21:     cqblocks = array (map2 do octlist octave:
c@21:         d = emptyHops * (pow 2 (octaves-octave)) - emptyHops;
c@63: //        eprintln "dropping \(d)";
c@21:         drop d octlist;
c@21:     done cqblocks [1..octaves]);
c@14: 
c@17:     assembleBlock bits =
c@59:        (//eprintln "assembleBlock: structure of bits is:";
c@59:         //eprintln (map length bits);
c@19: 
c@19:         rows = octaves * kdata.binsPerOctave;
c@19:         columns = (pow 2 (octaves - 1)) * kdata.atomsPerFrame;
c@19: 
c@18:         cm.generate do row col:
c@19: 
c@19:             // bits structure: [1,2,4,8,...]
c@19: 
c@19:             // each elt of bits is a list of the chunks that should
c@19:             // make up this block in that octave (lowest octave first)
c@19: 
c@19:             // each chunk has atomsPerFrame * binsPerOctave elts in it
c@19: 
c@19:             // row is disposed with 0 at the top, highest octave (in
c@19:             // both pitch and index into bits structure)
c@19: 
c@18:             oct = int (row / binsPerOctave);
c@19:             binNo = row % kdata.binsPerOctave;
c@21: 
c@19:             chunks = pow 2 oct;
c@21:             colsPerAtom = int (columns / (chunks * kdata.atomsPerFrame));
c@21:             atomNo = int (col / colsPerAtom);
c@21:             atomOffset = col % colsPerAtom;
c@18: 
c@40:             if atomOffset == 0 and atomNo < length bits[oct] then
c@21:                 bits[oct][atomNo][binNo];
c@20:             else
c@20:                 cplx.zero
c@20:             fi;
c@19: 
c@19:         done { rows, columns };
c@19:         );
c@15: 
c@72:     assembleBlockSpectrogram bits =
c@72:        (// As assembleBlock, but producing a dense magnitude
c@72:         // spectrogram (rather than a complex output with zeros
c@72:         // between the cell values in lower octaves). (todo: smoothing)
c@72: 
c@72:         //eprintln "assembleBlockSpectrogram: structure of bits is:";
c@72:         //eprintln (map length bits);
c@72: 
c@72:         rows = octaves * kdata.binsPerOctave;
c@72:         columns = (pow 2 (octaves - 1)) * kdata.atomsPerFrame;
c@72: 
c@72:         mat.generate do row col:
c@72: 
c@72:             oct = int (row / binsPerOctave);
c@72:             binNo = row % kdata.binsPerOctave;
c@72: 
c@72:             chunks = pow 2 oct;
c@72:             colsPerAtom = int (columns / (chunks * kdata.atomsPerFrame));
c@72:             atomNo = int (col / colsPerAtom);
c@72: 
c@72:             if atomNo < length bits[oct] then
c@72:                 cplx.magnitude bits[oct][atomNo][binNo];
c@72:             else 
c@72:                 0
c@72:             fi;
c@72: 
c@72:         done { rows, columns };
c@72:         );
c@72: 
c@72:     processOctaveLists assembler octs =
c@17:         case octs[0] of
c@17:         block::rest:
c@19:            (toAssemble = array 
c@19:                (map do oct:
c@21:                     n = kdata.atomsPerFrame * pow 2 oct;
c@17:                     if not empty? octs[oct] then
c@19:                         forBlock = array (take n octs[oct]);
c@17:                         octs[oct] := drop n octs[oct];
c@17:                         forBlock
c@17:                     else
c@19:                         array []
c@17:                     fi
c@19:                 done (keys octs));
c@72:             assembler toAssemble :. \(processOctaveLists assembler octs));
c@17:          _: []
c@15:         esac;
c@15: 
c@63: //eprintln "cqblocks has \(length cqblocks) entries";
c@15: 
c@17:     octaveLists = [:];
c@19: 
c@19:     cqblocks = array cqblocks;
c@17:     for [1..octaves] do oct:
c@17:         octaveLists[octaves - oct] := cqblocks[oct-1];
c@17:     done;
c@17: /*
c@17:     \() (map2 do octlist octave:
c@17: println "oct \(octaves) - \(octave) = \(octaves - octave)";
c@17:              octaveLists[octaves - octave] := octlist 
c@17:          done cqblocks [1..octaves]);
c@17: */
c@63: //eprintln "octaveLists keys are: \(keys octaveLists)";
c@15: 
c@40:     {
c@40:         kernel = kdata with {
c@47:             binFrequencies = array
c@47:                (concatMap do octave:
c@40:                     map do freq:
c@40:                         freq / (pow 2 octave);
c@40:                     done (reverse (list kdata.binFrequencies))
c@47:                 done [0..octaves-1])
c@40:         },
c@72:         octaves,
c@72:         output type =
c@72:             case type of
c@72:             ComplexCQ ():
c@72:                 Complex (processOctaveLists assembleBlock octaveLists);
c@72:             Spectrogram ():
c@72:                 Real (processOctaveLists assembleBlockSpectrogram octaveLists);
c@72:             esac
c@40:     }
c@10:     );
c@10: 
c@37: { cqt }
c@10: