Swift，数组，图像处理。使用 array.map() 是处理数组中所有数据的最快方法吗？

Question

我有一个包含数百万个元素（7201 x 7201 个数据点）的数组，我正在将数据转换为灰度图像。

var imagePixels = heights.map { terrainToColorScale([=12=].height, gradient: .Linear) }
let _ = writeImageToFile(imagePixels, height: latStrides, width: lonStrides, imgColorSpace: .BW, to: imageURL, as: .png)

这段代码大约需要 11 秒才能完成（CPU=2.3Ghz 8 核 i9），但我希望尽可能获得更好的性能。代码目前运行在单线程中。

简单地将我的 heights 数组分成块（比如 100 个块）和运行每个块都有一个任务的 TaskGroup 是否会得到不错的改进？还是我正在考虑进入 Metal 和着色器（我对 Metal 了解为零！！）以获得更好的结果？

出于兴趣，生成的典型图像是... （图像已缩减采样，因为太大而无法在此处上传。）

更新：添加与 terrainToColorScale 关联的代码基本上，对于线性转换，它将采用地形高度（通常为 0...9000）并将其缩放到 return 0...255 之间的值我还有非线性实现（下面未显示），它将显示主要是 low/high 地形高程的数据集的更多细节。

let terrainLowerCutOff: Double = 0.0     // Mean Sea Level
let terrainUpperCutOff: Double = 9000.0  // Value in meters, just higher that Everest

func terrainToColorScale(_ elev: Double, lowerCutOff: Double = terrainLowerCutOff, upperCutOff: Double = terrainUpperCutOff, gradient: ImageColorGradient = .Linear) -> UInt8 {
  switch gradient {
  case .Linear:
    return linearColorScale(elev, lowerCutOff: lowerCutOff, upperCutOff: upperCutOff)
  case .LinearInverse:
    return linearInverseColorScale(elev, lowerCutOff: lowerCutOff, upperCutOff: upperCutOff)
  case .CurveEmphasiseLows:
    return reciprocalPowerColorScale(elev, lowerCutOff: lowerCutOff, upperCutOff: upperCutOff)
  case .CurveEmphasiseLowsInverse:
    return reciprocalInversePowerColorScale(elev, lowerCutOff: lowerCutOff, upperCutOff: upperCutOff)
  case .CurveEmphasiseHighs:
    return powerColorScale(elev, lowerCutOff: lowerCutOff, upperCutOff: upperCutOff)
  case .CurveEmphasiseHighsInverse:
    return powerInverseColorScale(elev, lowerCutOff: lowerCutOff, upperCutOff: upperCutOff)
  }
}

fileprivate func linearColorScale(_ value: Double, lowerCutOff: Double, upperCutOff: Double) -> UInt8 {
    return UInt8( 255 * normaliseColorScale(value, lowerCutOff: lowerCutOff, upperCutOff: upperCutOff) )
}

fileprivate func normaliseColorScale(_ value: Double, lowerCutOff: Double, upperCutOff: Double) -> Double {
  switch value {
  case _ where value <= lowerCutOff :
    return 0.0
  case _ where value >= upperCutOff :
    return 1.0
  default :
    return (value - lowerCutOff) / (upperCutOff - lowerCutOff)
  }
}

Answer 1

这不是您问题的完整答案，但我认为它应该能让您了解下一步该怎么做。 vDSP 是 Accelerate 的一部分，它旨在加速数组的数学运算。此代码使用多个步骤，因此可能会更加优化，并且它没有考虑除线性之外的任何其他过滤器，但我没有足够的知识来使这些步骤更有效。然而，在我的机器上，vDSP 在以下处理中比 map 快 4 倍：

import Foundation
import Accelerate

let count = 7200 * 7200
var date = Date()
print("Generating")
let test: [CGPoint] = (0..<count).map {
    CGPoint(x: [=10=], y: Int.random(in: -2000...10000))
}
print("Generating took \(Date().timeIntervalSince(date))")
date = Date()
print("Mapping")
let heights: [Float] = test.map { Float([=10=].y) }
print("Mapping took \(Date().timeIntervalSince(date))")
date = Date()
print("Converting via vDSP")
let clipped = vDSP.clip(heights, to: 0...9000)
let scaled = vDSP.divide(clipped, 9000)
let multiplied = vDSP.multiply(255, scaled)
let integers = vDSP.floatingPointToInteger(multiplied, integerType: UInt8.self, rounding: .towardNearestInteger)
print("Converting via DSP took \(Date().timeIntervalSince(date))")
date = Date()

print("Converting via map")
let mappedIntegers = heights.map { height -> UInt8 in
    let clipped: Float
    if height < 0 {
        clipped = 0
    } else if height > 9000 {
        clipped = 9000
    } else {
        clipped = height
    }
    let scaled = clipped / 9000
    let multiplied = scaled * 255
    return UInt8(multiplied.rounded())
}
print("Converting via map took \(Date().timeIntervalSince(date))")

Swift，数组，图像处理。使用 array.map() 是处理数组中所有数据的最快方法吗？

Swift, array, image processing. Is using array.map() the fastest way to process all data in an array?

arrays

optimization

array-map

swift

accelerate