to-live-photo/Sources/LivePhotoCore/AIEnhancer/RealESRGANProcessor.swift

//
//  RealESRGANProcessor.swift
//  LivePhotoCore
//
//  Core ML inference logic for Real-ESRGAN model.
//  This model requires fixed 512x512 input and outputs 2048x2048.
//

import Accelerate
import CoreML
import CoreVideo
import Foundation
import os

/// Real-ESRGAN Core ML model processor
/// Note: This model has fixed input size of 512x512
actor RealESRGANProcessor {
    private var model: MLModel?
    private let logger = Logger(subsystem: "LivePhotoCore", category: "RealESRGANProcessor")

    /// Fixed input size required by the model (512x512)
    static let inputSize: Int = 512

    /// Scale factor (4x for Real-ESRGAN x4plus)
    static let scaleFactor: Int = 4

    /// Output size (inputSize * scaleFactor = 2048)
    static let outputSize: Int = inputSize * scaleFactor  // 2048

    init() {}

    /// Load Core ML model from ODR or bundle
    func loadModel() async throws {
        guard model == nil else {
            logger.debug("Model already loaded")
            return
        }

        logger.info("Loading Real-ESRGAN Core ML model...")

        // 1. Try ODRManager first (supports both ODR download and bundle fallback)
        var modelURL = await ODRManager.shared.getModelURL()
        
        // 2. If ODRManager returns nil, try direct bundle lookup as fallback
        if modelURL == nil {
            let modelName = "RealESRGAN_x4plus"
            
            // Try main bundle
            if let url = Bundle.main.url(forResource: modelName, withExtension: "mlmodelc") {
                modelURL = url
            } else if let url = Bundle.main.url(forResource: modelName, withExtension: "mlpackage") {
                modelURL = url
            }
            
            // Try SPM bundle (development)
            #if SWIFT_PACKAGE
            if modelURL == nil {
                if let url = Bundle.module.url(forResource: modelName, withExtension: "mlmodelc") {
                    modelURL = url
                } else if let url = Bundle.module.url(forResource: modelName, withExtension: "mlpackage") {
                    modelURL = url
                }
            }
            #endif
        }

        guard let url = modelURL else {
            logger.error("Model not found. Please download the AI model first.")
            throw AIEnhanceError.modelNotFound
        }

        logger.info("Found model at: \(url.path)")

        // Configure model for optimal performance
        let config = MLModelConfiguration()
        config.computeUnits = .all  // Use Neural Engine when available

        do {
            model = try await MLModel.load(contentsOf: url, configuration: config)
            logger.info("Model loaded successfully")
        } catch {
            logger.error("Failed to load model: \(error.localizedDescription)")
            throw AIEnhanceError.modelLoadFailed(error.localizedDescription)
        }
    }

    /// Unload model from memory
    func unloadModel() {
        model = nil
        logger.info("Model unloaded from memory")
    }

    /// Process a 512x512 image through the model
    /// - Parameter pixelBuffer: Input image as CVPixelBuffer (must be 512x512, BGRA format)
    /// - Returns: Enhanced image as RGBA data array (2048x2048)
    func processImage(_ pixelBuffer: CVPixelBuffer) async throws -> [UInt8] {
        guard let model else {
            throw AIEnhanceError.modelNotFound
        }

        // Verify input size
        let width = CVPixelBufferGetWidth(pixelBuffer)
        let height = CVPixelBufferGetHeight(pixelBuffer)
        guard width == Self.inputSize, height == Self.inputSize else {
            throw AIEnhanceError.inferenceError(
                "Invalid input size \(width)x\(height), expected \(Self.inputSize)x\(Self.inputSize)"
            )
        }

        // Check for cancellation
        try Task.checkCancellation()

        logger.info("Running inference on \(width)x\(height) image...")

        // Capture actor-isolated state before entering non-isolated closure
        let localModel = model

        // Run inference on background queue (MLModel prediction is thread-safe)
        let output: [UInt8] = try await withCheckedThrowingContinuation { continuation in
            DispatchQueue.global(qos: .userInitiated).async {
                do {
                    // Create input feature from pixel buffer
                    let inputFeature = try MLFeatureValue(pixelBuffer: pixelBuffer)
                    let inputProvider = try MLDictionaryFeatureProvider(
                        dictionary: ["input": inputFeature]
                    )

                    // Run inference synchronously
                    let prediction = try localModel.prediction(from: inputProvider)

                    // Extract output from model
                    let rgbaData: [UInt8]

                    if let outputValue = prediction.featureValue(for: "activation_out") {
                        if let multiArray = outputValue.multiArrayValue {
                            rgbaData = try Self.multiArrayToRGBA(multiArray)
                        } else if let outputBuffer = outputValue.imageBufferValue {
                            rgbaData = try ImageFormatConverter.pixelBufferToRGBAData(outputBuffer)
                        } else {
                            continuation.resume(throwing: AIEnhanceError.inferenceError(
                                "Cannot extract data from model output"
                            ))
                            return
                        }
                    } else {
                        continuation.resume(throwing: AIEnhanceError.inferenceError(
                            "Model output 'activation_out' not found"
                        ))
                        return
                    }

                    continuation.resume(returning: rgbaData)
                } catch let error as AIEnhanceError {
                    continuation.resume(throwing: error)
                } catch {
                    continuation.resume(throwing: AIEnhanceError.inferenceError(error.localizedDescription))
                }
            }
        }

        logger.info("Inference completed, output size: \(output.count) bytes")
        return output
    }

    /// Convert MLMultiArray [C, H, W] to RGBA byte array
    /// - Parameter multiArray: Output from model with shape [3, H, W] (RGB channels)
    /// - Returns: RGBA byte array with shape [H * W * 4]
    private static func multiArrayToRGBA(_ multiArray: MLMultiArray) throws -> [UInt8] {
        let shape = multiArray.shape.map { $0.intValue }

        // Expect shape [3, H, W] for RGB
        guard shape.count == 3, shape[0] == 3 else {
            throw AIEnhanceError.inferenceError(
                "Unexpected output shape: \(shape), expected [3, H, W]"
            )
        }

        let height = shape[1]
        let width = shape[2]

        // Output array: RGBA format
        var rgbaData = [UInt8](repeating: 255, count: width * height * 4)

        // Get pointer to MultiArray data
        let dataPointer = multiArray.dataPointer.assumingMemoryBound(to: Float32.self)
        let channelStride = height * width

        // Convert CHW (channel-first) to RGBA (interleaved)
        // Model output is typically in range [0, 1] or [-1, 1], need to scale to [0, 255]
        for y in 0..<height {
            for x in 0..<width {
                let pixelIndex = y * width + x
                let rgbaIndex = pixelIndex * 4

                // Read RGB values from CHW layout
                let r = dataPointer[0 * channelStride + pixelIndex]
                let g = dataPointer[1 * channelStride + pixelIndex]
                let b = dataPointer[2 * channelStride + pixelIndex]

                // Clamp and convert to 0-255
                // Model typically outputs values in [0, 1] range
                rgbaData[rgbaIndex + 0] = UInt8(clamping: Int(max(0, min(1, r)) * 255))
                rgbaData[rgbaIndex + 1] = UInt8(clamping: Int(max(0, min(1, g)) * 255))
                rgbaData[rgbaIndex + 2] = UInt8(clamping: Int(max(0, min(1, b)) * 255))
                rgbaData[rgbaIndex + 3] = 255  // Alpha
            }
        }

        return rgbaData
    }
}