NavidromeApp/iOS/Views/Visualizer/OfflineAudioAnalyzer.swift

import Foundation
import AVFoundation
import Accelerate

/// Processes an entire audio file faster than real-time, producing per-frame FFT data
/// that can be cached and played back in sync with the audio.
actor OfflineAudioAnalyzer {
    static let shared = OfflineAudioAnalyzer()
    
    /// Progress callback (0.0 to 1.0)
    typealias ProgressCallback = @Sendable (Float) -> Void
    
    /// Analyze an audio file and return an array of FFT frames.
    /// Each frame is an array of `pointsCount` floats (0.0-1.0) representing frequency band amplitudes.
    func analyze(
        url: URL,
        pointsCount: Int = 20,
        fps: Double = 30.0,
        cutoff: Int = 90,
        eqBoostFactor: Float = 3.5,
        progress: ProgressCallback? = nil
    ) throws -> [[Float]] {
        let file = try AVAudioFile(forReading: url)
        let format = file.processingFormat
        let sampleRate = format.sampleRate
        let totalFrames = file.length
        
        // How many audio frames per visualizer frame
        let audioFramesPerVisFrame = AVAudioFrameCount(sampleRate / fps)
        
        // Use power-of-2 buffer for FFT
        let fftSize = 1024
        let bufferSize = max(AVAudioFrameCount(fftSize), audioFramesPerVisFrame)
        
        guard let buffer = AVAudioPCMBuffer(pcmFormat: format, frameCapacity: bufferSize) else {
            throw NSError(domain: "OfflineAnalyzer", code: 1, userInfo: [NSLocalizedDescriptionKey: "Failed to create buffer"])
        }
        
        let log2n = vDSP_Length(log2(Double(fftSize)))
        guard let fftSetup = vDSP_create_fftsetup(log2n, Int32(kFFTRadix2)) else {
            throw NSError(domain: "OfflineAnalyzer", code: 2, userInfo: [NSLocalizedDescriptionKey: "Failed to create FFT setup"])
        }
        defer { vDSP_destroy_fftsetup(fftSetup) }
        
        let halfSize = fftSize / 2
        var visualizerData: [[Float]] = []
        
        // Estimate total frames for progress
        let estimatedVisFrames = Int(Double(totalFrames) / Double(audioFramesPerVisFrame))
        visualizerData.reserveCapacity(estimatedVisFrames)
        
        // Reusable buffers
        var window = [Float](repeating: 0, count: fftSize)
        vDSP_hann_window(&window, vDSP_Length(fftSize), Int32(vDSP_HANN_NORM))
        
        var frameIndex = 0
        
        while file.framePosition < totalFrames {
            // Read a chunk
            let framesToRead = min(bufferSize, AVAudioFrameCount(totalFrames - file.framePosition))
            buffer.frameLength = 0
            try file.read(into: buffer, frameCount: framesToRead)
            
            guard let channelData = buffer.floatChannelData?[0] else { continue }
            let actualFrames = Int(buffer.frameLength)
            guard actualFrames >= fftSize else {
                // Pad with zeros for the last partial buffer
                if actualFrames > 0 {
                    let frame = processFFTFrame(
                        channelData: channelData,
                        frameCount: actualFrames,
                        fftSize: fftSize,
                        halfSize: halfSize,
                        window: window,
                        fftSetup: fftSetup,
                        pointsCount: pointsCount,
                        cutoff: cutoff,
                        eqBoostFactor: eqBoostFactor
                    )
                    visualizerData.append(frame)
                }
                break
            }
            
            // Process one or more vis frames from this buffer
            var sampleOffset = 0
            while sampleOffset + fftSize <= actualFrames {
                let frame = processFFTFrame(
                    channelData: channelData.advanced(by: sampleOffset),
                    frameCount: fftSize,
                    fftSize: fftSize,
                    halfSize: halfSize,
                    window: window,
                    fftSetup: fftSetup,
                    pointsCount: pointsCount,
                    cutoff: cutoff,
                    eqBoostFactor: eqBoostFactor
                )
                visualizerData.append(frame)
                sampleOffset += Int(audioFramesPerVisFrame)
                frameIndex += 1
                
                // Report progress every 50 frames
                if frameIndex % 50 == 0, let progress = progress {
                    let pct = Float(file.framePosition) / Float(totalFrames)
                    progress(pct)
                }
            }
        }
        
        progress?(1.0)
        return visualizerData
    }
    
    /// Process a single FFT frame from raw audio samples
    private func processFFTFrame(
        channelData: UnsafePointer<Float>,
        frameCount: Int,
        fftSize: Int,
        halfSize: Int,
        window: [Float],
        fftSetup: FFTSetup,
        pointsCount: Int,
        cutoff: Int,
        eqBoostFactor: Float
    ) -> [Float] {
        let n = min(frameCount, fftSize)
        
        // 1. Apply Hann window
        var windowed = [Float](repeating: 0, count: fftSize)
        if n < fftSize {
            // Zero-pad if short
            for i in 0..<n { windowed[i] = channelData[i] * window[i] }
        } else {
            vDSP_vmul(channelData, 1, window, 1, &windowed, 1, vDSP_Length(fftSize))
        }
        
        // 2. FFT
        var realp = [Float](repeating: 0, count: halfSize)
        var imagp = [Float](repeating: 0, count: halfSize)
        var magnitudes = [Float](repeating: 0, count: halfSize)
        
        realp.withUnsafeMutableBufferPointer { realpBuf in
            imagp.withUnsafeMutableBufferPointer { imagpBuf in
                var splitComplex = DSPSplitComplex(
                    realp: realpBuf.baseAddress!,
                    imagp: imagpBuf.baseAddress!
                )
                
                windowed.withUnsafeBytes { raw in
                    let ptr = raw.bindMemory(to: DSPComplex.self).baseAddress!
                    vDSP_ctoz(ptr, 2, &splitComplex, 1, vDSP_Length(halfSize))
                }
                
                vDSP_fft_zrip(fftSetup, &splitComplex, 1, vDSP_Length(log2(Double(fftSize))), FFTDirection(FFT_FORWARD))
                vDSP_zvmags(&splitComplex, 1, &magnitudes, 1, vDSP_Length(halfSize))
            }
        }
        
        // 3. Normalize
        let fftSizeF = Float(fftSize)
        var scale: Float = 1.0 / (fftSizeF * fftSizeF)
        vDSP_vsmul(magnitudes, 1, &scale, &magnitudes, 1, vDSP_Length(halfSize))
        
        // sqrt for perceptual amplitude
        for i in 0..<halfSize {
            magnitudes[i] = sqrt(magnitudes[i])
        }
        
        // 4. Logarithmic binning with EQ boost
        var framePoints = [Float](repeating: 0, count: pointsCount)
        let maxUsefulBin = min(halfSize - 1, cutoff)
        
        for i in 0..<pointsCount {
            let normalizedIndex = Float(i + 1) / Float(pointsCount)
            let logIndex = log10(normalizedIndex * 9.0 + 1.0)
            let centerBin = logIndex * Float(maxUsefulBin)
            let binWidth = max(1.0, Float(maxUsefulBin) / Float(pointsCount) * logIndex)
            
            let startBin = max(1, Int(centerBin - binWidth / 2))
            let endBin = min(maxUsefulBin, Int(centerBin + binWidth / 2))
            
            var sum: Float = 0
            var countInBand = 0
            for j in startBin...endBin where j < magnitudes.count {
                sum += magnitudes[j]
                countInBand += 1
            }
            
            let average = countInBand > 0 ? (sum / Float(countInBand)) : 0
            let eqBoost: Float = 1.0 + (Float(i) / Float(pointsCount)) * eqBoostFactor
            framePoints[i] = average * eqBoost
        }
        
        return framePoints
    }
}