CGBitmapContext 2x as slow compared to CALayer's draw() - swift

Question

I have some code I can't change that expects to be able to draw at any time. It's the main() function in BackgroundThread below - pretend it can't be modified in any way. Running this will use 70-80% CPU.
If instead of running the thread I replicate what it is doing in View::draw() (i.e. draw 5000 white rectangles at random positions), this will use about 30% CPU.
Where's the difference coming from? Looking at Instruments, although the call stack is the same starting from CGContextFillRect, the View::draw() version only spends 16% of the time doing memset() whereas the threaded version spends 80% of the time.
The code below is the FAST version. Comment out the FAST lines and uncomment the SLOW lines to switch to the SLOW (threaded) version. Compile with swiftc test.swift -otest && ./test. I'm on macOS 10.13, integrated graphics, if that matters.
Is there anything I can do to make the threaded version as fast as the View::draw() version?
import Cocoa
let NSApp = NSApplication.shared,
vwaitSem = DispatchSemaphore(value: 0)
var
mainWindow: NSWindow?,
screen: CGContext?,
link: CVDisplayLink?
class View: NSView, CALayerDelegate {
var lastTime: CFTimeInterval = 0
override var acceptsFirstResponder: Bool {return true}
required init(coder aDecoder: NSCoder) {fatalError("This class does not support NSCoding")}
override func makeBackingLayer() -> CALayer {return CALayer()}
override init(frame: CGRect) {
super.init(frame: frame)
self.wantsLayer = true
self.layer?.contentsScale = 2.0
self.layer?.backgroundColor = CGColor(red:0, green:0, blue:0, alpha: 1)
self.layerContentsRedrawPolicy = NSView.LayerContentsRedrawPolicy.onSetNeedsDisplay // FAST
}
func draw(_ layer: CALayer, in ctx: CGContext) {
let now = CACurrentMediaTime(), timePassed = ((now-lastTime)*1000).rounded()
// NSLog("\(timePassed)")
lastTime = now
ctx.setFillColor(CGColor.white)
ctx.setStrokeColor(CGColor.white)
for _ in 0...5000 {
let rect = CGRect(x: CGFloat(arc4random_uniform(640)+1), y: CGFloat(arc4random_uniform(480)+1), width:6, height:6)
ctx.setFillColor(CGColor.white)
ctx.fill(rect)
}
}
}
func displayLinkOutputCallback(_ displayLink: CVDisplayLink, _ nowPtr: UnsafePointer<CVTimeStamp>,
_ outputTimePtr: UnsafePointer<CVTimeStamp>, _ flagsIn: CVOptionFlags, _ flagsOut: UnsafeMutablePointer<CVOptionFlags>,
_ displayLinkContext: UnsafeMutableRawPointer?) -> CVReturn {
DispatchQueue.main.async {
// mainWindow!.contentView!.layer!.contents = screen!.makeImage() // SLOW
mainWindow!.contentView!.display() // FAST
vwaitSem.signal()
}
return kCVReturnSuccess
}
class BackgroundThread: Thread {
var lastTime: CFTimeInterval = 0
override func main() {
while true {
let now = CACurrentMediaTime(), timePassed = ((now-lastTime)*1000).rounded()
// NSLog("\(timePassed)")
lastTime = now
screen?.clear(CGRect(x:0, y:0, width:640*2, height:480*2))
for _ in 0...5000 {
screen?.setFillColor(CGColor.white)
screen?.setStrokeColor(CGColor.white)
screen?.fill(CGRect(x: CGFloat(arc4random_uniform(640*2)+1), y: CGFloat(arc4random_uniform(480*2)+1), width: 6*2, height: 6*2))
}
vwaitSem.wait()
}
}
}
let width = 640, height = 480,
appMenuItem = NSMenuItem(),
quitMenuItem = NSMenuItem(title:"Quit",
action:#selector(NSApplication.terminate), keyEquivalent:"q"),
window = NSWindow(contentRect:NSMakeRect(0,0, CGFloat(width), CGFloat(height)),
styleMask:[.closable,.titled], backing:.buffered, defer:false),
colorProfile = ColorSyncProfileCreateWithDisplayID(0),
colorSpace = CGColorSpace(platformColorSpaceRef: colorProfile!.toOpaque()),
screen_ = CGContext(data: nil, width: Int(width)*2, height:Int(height)*2, bitsPerComponent:8, bytesPerRow: 0,
space: colorSpace!, bitmapInfo: CGImageAlphaInfo.premultipliedLast.rawValue),
backgroundThread = BackgroundThread()
NSApp.setActivationPolicy(NSApplication.ActivationPolicy.regular)
NSApp.mainMenu = NSMenu()
NSApp.mainMenu?.addItem(appMenuItem)
appMenuItem.submenu = NSMenu()
appMenuItem.submenu?.addItem(quitMenuItem)
window.cascadeTopLeft(from:NSMakePoint(20,20))
window.makeKeyAndOrderFront(nil)
window.contentView = View()
window.makeFirstResponder(window.contentView)
NSApp.activate(ignoringOtherApps:true)
mainWindow = window
screen = screen_
CVDisplayLinkCreateWithCGDisplay(CGMainDisplayID(), &link)
CVDisplayLinkSetOutputCallback(link!, displayLinkOutputCallback, UnsafeMutableRawPointer(Unmanaged.passUnretained(window).toOpaque()))
CVDisplayLinkStart(link!)
// backgroundThread.start() // SLOW
NSApp.run()

Answer 1

I misread the note in the documentation for makeImage() and thought it would not copy the data unless it really had to. Well, Instruments shows it does copy the data. Every single frame.
So I switched to Metal and now I can draw from the background thread with the same performance/CPU usage as with CGContext alone, with no copies as far as I can tell.
Here's some working code:
import Cocoa
import MetalKit
class View: MTKView {
var screen: CGContext?
var commandQueue: MTLCommandQueue?
var buffer: MTLBuffer?
var texture: MTLTexture?
var vwaitSem = DispatchSemaphore(value: 0)
var backgroundThread: Thread?
var allocationSize = 0
func alignUp(size: Int, align: Int) -> Int {return (size+(align-1)) & ~(align-1)}
override var acceptsFirstResponder: Bool {return true}
required init(coder aDecoder: NSCoder) {fatalError("This class does not support NSCoding")}
init() {super.init(frame: CGRect(x:0, y:0, width:0, height: 0), device: MTLCreateSystemDefaultDevice())}
override func viewDidMoveToWindow() {
layer?.contentsScale = NSScreen.main!.backingScaleFactor
let metalLayer = layer as! CAMetalLayer
let pixelRowAlignment = metalLayer.device!.minimumLinearTextureAlignment(for: metalLayer.pixelFormat)
let bytesPerRow = alignUp(size: Int(layer!.frame.width)*Int(layer!.contentsScale)*4, align: pixelRowAlignment)
let pagesize = Int(getpagesize())
var data: UnsafeMutableRawPointer? = nil
allocationSize = alignUp(size: bytesPerRow*Int(layer!.frame.height)*Int(layer!.contentsScale), align: pagesize)
posix_memalign(&data, pagesize, allocationSize)
let colorProfile = ColorSyncProfileCreateWithDisplayID(0),
colorSpace = CGColorSpace(platformColorSpaceRef: colorProfile!.toOpaque()),
screen_ = CGContext(data: data,
width: Int(layer!.frame.width)*Int(layer!.contentsScale),
height: Int(layer!.frame.height)*Int(layer!.contentsScale),
bitsPerComponent:8, bytesPerRow: bytesPerRow,
space: colorSpace!, bitmapInfo: CGImageAlphaInfo.premultipliedLast.rawValue)!,
buffer_ = metalLayer.device!.makeBuffer(bytesNoCopy: data!, length: allocationSize, options: .storageModeManaged,
deallocator: { pointer, length in free(self.screen!.data!) })!,
textureDescriptor = MTLTextureDescriptor()
textureDescriptor.pixelFormat = metalLayer.pixelFormat
textureDescriptor.width = screen_.width
textureDescriptor.height = screen_.height
textureDescriptor.storageMode = buffer_.storageMode
textureDescriptor.usage = MTLTextureUsage(rawValue: MTLTextureUsage.shaderRead.rawValue)
texture = buffer_.makeTexture(descriptor: textureDescriptor, offset: 0, bytesPerRow: screen_.bytesPerRow)
commandQueue = device?.makeCommandQueue()
screen = screen_
buffer = buffer_
backgroundThread = BackgroundThread(screen: screen!, vwaitSem: vwaitSem)
backgroundThread!.start()
}
override func draw(_ dirtyRect: NSRect) {
if let drawable = currentDrawable {
buffer!.didModifyRange(0..<allocationSize)
texture!.replace(region: MTLRegionMake2D(0,0, screen!.width, screen!.height),
mipmapLevel:0, slice:0, withBytes: screen!.data!, bytesPerRow: screen!.bytesPerRow, bytesPerImage: 0)
let commandBuffer = commandQueue!.makeCommandBuffer()!
let blitPass = commandBuffer.makeBlitCommandEncoder()!
blitPass.copy(from: texture!, sourceSlice:0, sourceLevel:0, sourceOrigin: MTLOrigin(x:0,y:0,z:0),
sourceSize: MTLSize(width:screen!.width, height:screen!.height, depth: 1),
to: drawable.texture, destinationSlice:0, destinationLevel:0, destinationOrigin: MTLOrigin(x:0,y:0,z:0))
blitPass.endEncoding()
if let renderPass = currentRenderPassDescriptor {
renderPass.colorAttachments[0].texture = drawable.texture
renderPass.colorAttachments[0].loadAction = .load
commandBuffer.makeRenderCommandEncoder(descriptor: renderPass)!.endEncoding()
commandBuffer.addCompletedHandler {cb in self.vwaitSem.signal()}
commandBuffer.present(drawable)
commandBuffer.commit()
}
}
}
}
class BackgroundThread: Thread {
var screen: CGContext
var vwaitSem: DispatchSemaphore
var x = 0
init(screen:CGContext, vwaitSem:DispatchSemaphore) {
self.screen = screen
self.vwaitSem = vwaitSem
}
override func main() {
while true {
// screen.clear(CGRect(x:0,y:0, width:screen.width, height:screen.height))
// screen.setFillColor(CGColor.white)
// screen.fill(CGRect(x:x, y:0, width:100, height:100))
// x += 1
screen.clear(CGRect(x:0,y:0, width:screen.width, height:screen.height))
screen.setFillColor(CGColor.white)
let screenWidth = UInt32(screen.width), screenHeight = UInt32(screen.height)
for _ in 0...5000 {
let rect = CGRect(x: CGFloat(arc4random_uniform(screenWidth+1)),
y: CGFloat(arc4random_uniform(screenHeight+1)), width:6, height:6)
screen.fill(rect)
}
vwaitSem.wait()
}
}
}
let width = 640, height = 480,
appMenuItem = NSMenuItem(),
quitMenuItem = NSMenuItem(title:"Quit",
action:#selector(NSApplication.terminate), keyEquivalent:"q"),
window = NSWindow(contentRect:NSMakeRect(0,0, CGFloat(width), CGFloat(height)),
styleMask:[.closable,.titled], backing:.buffered, defer:false)
NSApp.setActivationPolicy(NSApplication.ActivationPolicy.regular)
NSApp.mainMenu = NSMenu()
NSApp.mainMenu?.addItem(appMenuItem)
appMenuItem.submenu = NSMenu()
appMenuItem.submenu?.addItem(quitMenuItem)
window.cascadeTopLeft(from:NSMakePoint(20,20))
window.makeKeyAndOrderFront(nil)
window.contentView = View()
window.makeFirstResponder(window.contentView)
NSApp.activate(ignoringOtherApps:true)
NSApp.run()