When should I use PixelFormats.Pbgra32 in WPF application?

Question

Consider following setup: I have some images on disk. I cannot modify those images since they belong to another project. I just read them.

I need them processed to look optimal in my app. I mean pixel shader processed. When I'm done with processing, the output image will be used as a thumbnail in some sort of ListView or other control.

I chose PixelFormats.Bgra32 as the output format, because it's pretty straightforward to build pixel shader for it. I read however PixelFormats.Pbgra32 is natively used by Windows and thus is faster.

Is it still faster, when fade effects are used on the image? Does it make sense to pre-multiply channels in managed code instead leaving this to the framework?

Answer 1

So I tested it how I could (compile with /unsafe):

using System;
using System.Collections.Generic;
using System.Linq;
using System.Threading.Tasks;
using System.Windows;
using System.Windows.Controls;
using System.Windows.Media;
using System.Windows.Media.Imaging;
using System.Windows.Threading;

namespace BgraVsPbgra {

    public partial class MainWindow : Window {

        readonly WrapPanel Panel;
        readonly TextBlock Output;
        readonly PixelFormat FN = PixelFormats.Bgra32;
        readonly PixelFormat FP = PixelFormats.Pbgra32;
        readonly List<long> TN = new List<long>();
        readonly List<long> TP = new List<long>();
        DateTime T0;
        long DT => (DateTime.Now - T0).Ticks;
        DateTime Now => DateTime.Now;

        public MainWindow() {
            InitializeComponent();
            SizeToContent = SizeToContent.WidthAndHeight;
            Title = "Bgra32 vs Pbgra Benchmark";
            Panel = new WrapPanel {
                Width = 512,
                Height = 512
            };
            Output = new TextBlock {
                HorizontalAlignment = HorizontalAlignment.Center,
                VerticalAlignment = VerticalAlignment.Center,
                Foreground = new SolidColorBrush(Colors.White)
            };
            (Content as Grid).Children.Add(Panel);
            (Content as Grid).Children.Add(Output);
            Dispatcher.BeginInvoke(new Action(() => Start()), DispatcherPriority.SystemIdle);
        }

        private async void Start() {
            Output.Text += "Bgra32 vs Pbgra32 benchmark started.\r\n\r\n";
            var t0 = DateTime.Now;
            var n = 3;
            var i = 16384;
            for (var p = 1; p <= n; p++) {
                Output.Text += $"\r\nPass {p}/{n}.\r\n\r\n";
                await Benchmark(i / 4, 0.75, true);
                await Benchmark(i / 4, 0.75, false);
                await Benchmark(i / 4, 1, true);
                await Benchmark(i / 4, 1, false);
            }
            var t = (DateTime.Now - t0).TotalSeconds;
            Output.Text += $"\r\nDone in {t:0.0}s.\r\n";
        }

        private async Task Benchmark(int n = 256, double opacity = 1, bool cheat = false) {
            Output.Text += $"Benchmarking with opacity = {opacity}...   ";
            if (cheat) Output.Text += "CHEATING!...   ";
            //Output.Text += "\r\n";
            for (int i = 0; i < n; i++) await Dispatcher.BeginInvoke(new Action(async () => await DrawTestAsync(opacity, cheat)), DispatcherPriority.Send);
            Output.Text += ($"Pbgra32 advantage: {(TN.Average() / TP.Average() * 100d - 100d):0.0}%\r\n");
        }

        async Task DrawTestAsync(double opacity = 1, bool cheat = false) {
            await Dispatcher.BeginInvoke(new Action(() => {
                Panel.Children.Clear();
                for (int i = 0; i < 8; i++) {
                    T0 = Now; Panel.Children.Add(new Image { Source = CreateBitmap(FP, cheat), Width = 128, Height = 128, Opacity = opacity }); TP.Add(DT);
                    T0 = Now; Panel.Children.Add(new Image { Source = CreateBitmap(FN, cheat), Width = 128, Height = 128, Opacity = opacity }); TN.Add(DT);
                }
            }), DispatcherPriority.Send);

        }



        BitmapSource CreateBitmap(PixelFormat pixelFormat, bool cheat = false) {
            var bitmap = new WriteableBitmap(256, 256, 96, 96, pixelFormat, null);
            bitmap.Lock();
            unsafe
            {
                var pixels = (uint*)bitmap.BackBuffer;
                if (pixelFormat == FN || cheat)
                    for (uint y = 0; y < 256; y++) for (uint x = 0; x < 256; x++) pixels[x + y * 256] = 0x80000000u | (x << 16) | y;
                else
                    for (uint y = 0; y < 256; y++) for (uint x = 0; x < 256; x++) pixels[x + y * 256] = 0x80000000u | (x / 2 << 16) | y / 2;
            }
            bitmap.Unlock();
            bitmap.Freeze();
            return bitmap;
        }

    }

}

I get quite random results, but most of the times using Pbgra32 pixel format seems to be a little (like 3%) faster than using Bgra32 pixel format.

When using Pbgra32 format some additional operations must be made for each channel, but 2 additional divisions seem not to make any measurable difference.

So if you want to make your bitmaps circa 3% faster, use Pbgra32 format and premultiply channels like pr = r * a / 255 which might look as ugly as:

uint pixel = a << 24 | r * a / 255 << 16 | g * a / 255 << 8 | b * a / 255;

And do not make the pixel function, or you kill the performance. Speaking of performance, unsafe pixel manipulation seems to be 2 up to 3 times faster than using safe CopyPixels() / WritePixels() methods.

Nah, I wouldn't use Pbgra32 for creating bitmaps or pixel shaders for bitmaps. My opinion is: not worth it. Correct me if I'm wrong.