Speed of horizontal flip

[John Colvin via Digitalmars-d-learn]

On Thursday, 2 April 2015 at 09:55:15 UTC, Rikki Cattermole wrote:
> On 2/04/2015 10:47 p.m., Rikki Cattermole wrote:
>> On 2/04/2015 2:52 a.m., tchaloupka wrote:
>>> Hi,
>>> I have a bunch of square r16 and png images which I need to 
>>> flip
>>> horizontally.
>>>
>>> My flip method looks like this:
>>> void hFlip(T)(T[] data, int w)
>>> {
>>>    import std.datetime : StopWatch;
>>>
>>>    StopWatch sw;
>>>    sw.start();
>>>
>>>    foreach(int i; 0..w)
>>>    {
>>>      auto row = data[i*w..(i+1)*w];
>>>      row.reverse();
>>>    }
>>>
>>>    sw.stop();
>>>    writeln("Img flipped in: ", sw.peek().msecs, "[ms]");
>>> }
>>>
>>> With simple r16 file format its pretty fast, but with RGB PNG
>>> files (2048x2048) I noticed its somewhat slow so I tried to
>>> compare it with C# and was pretty surprised by the results.
>>>
>>> C#:
>>> PNG load - 90ms
>>> PNG flip - 10ms
>>> PNG save - 380ms
>>>
>>> D using dlib (http://code.dlang.org/packages/dlib):
>>> PNG load - 500ms
>>> PNG flip - 30ms
>>> PNG save - 950ms
>>>
>>> D using imageformats
>>> (http://code.dlang.org/packages/imageformats):
>>> PNG load - 230ms
>>> PNG flip - 30ms
>>> PNG save - 1100ms
>>>
>>> I used dmd-2.0.67 with -release -inline -O
>>> C# was just with debug and VisualStudio attached to process 
>>> for
>>> debugging and even with that it is much faster.
>>>
>>> I know that System.Drawing is using Windows GDI+, that can be
>>> used with D too, but not on linux.
>>> If we ignore the PNG loading and saving (didn't tried libpng
>>> yet), even flip method itself is 3 times slower - I don't 
>>> know D
>>> enough to be sure if there isn't some more effecient way to 
>>> make
>>> the flip. I like how the slices can be used here.
>>>
>>> For a C# user who is expecting things to just work as fast as
>>> possible from a system level programming language this can be
>>> somewhat disappointing to see that pure D version is about 3
>>> times slower.
>>>
>>> Am I doing something utterly wrong?
>>> Note that this example is not critical for me, it's just a 
>>> simple
>>> hobby script I use to move and flip some images - I can wait. 
>>> But
>>> I post it to see if this can be taken somewhat closer to what 
>>> can
>>> be expected from a system level programming language.
>>>
>>> dlib:
>>> auto im = loadPNG(name);
>>> hFlip(cast(ubyte[3][])im.data, cast(int)im.width);
>>> savePNG(im, newName);
>>>
>>> imageformats:
>>> auto im = read_image(name);
>>> hFlip(cast(ubyte[3][])im.pixels, cast(int)im.w);
>>> write_image(newName, im.w, im.h, im.pixels);
>>>
>>> C# code:
>>> static void Main(string[] args)
>>>          {
>>>              var files = Directory.GetFiles(args[0]);
>>>
>>>              foreach (var f in files)
>>>              {
>>>                  var sw = Stopwatch.StartNew();
>>>                  var img = Image.FromFile(f);
>>>
>>>                  Debug.WriteLine("Img loaded in {0}[ms]",
>>> (int)sw.Elapsed.TotalMilliseconds);
>>>                  sw.Restart();
>>>
>>>                  
>>> img.RotateFlip(RotateFlipType.RotateNoneFlipX);
>>>                  Debug.WriteLine("Img flipped in {0}[ms]",
>>> (int)sw.Elapsed.TotalMilliseconds);
>>>                  sw.Restart();
>>>
>>>                  img.Save(Path.Combine(args[0], "test_" +
>>> Path.GetFileName(f)));
>>>                  Debug.WriteLine("Img saved in {0}[ms]",
>>> (int)sw.Elapsed.TotalMilliseconds);
>>>                  sw.Stop();
>>>              }
>>>          }
>>
>>
>> Assuming I've done it correctly, Devisualization.Image takes 
>> around 8ms
>> in debug mode to flip horizontally using dmd. But 3ms for 
>> release.
>>
>> module test;
>>
>> void main() {
>>     import devisualization.image;
>>     import devisualization.image.mutable;
>>     import devisualization.util.core.linegraph;
>>
>>     import std.stdio;
>>
>>     writeln("===============\nREAD\n===============");
>>     Image img = imageFromFile("test/large.png");
>>     img = new MutableImage(img);
>>
>>     import std.datetime : StopWatch;
>>
>>     StopWatch sw;
>>     sw.start();
>>
>>     foreach(i; 0 .. 1000) {
>>         img.flipHorizontal;
>>     }
>>
>>     sw.stop();
>>
>>     writeln("Img flipped in: ", sw.peek().msecs / 1000, 
>> "[ms]");
>> }
>>
>> I was planning on doing this earlier. But I discovered a PR I 
>> pulled
>> which fixed for 2.067 broke chunk types reading.
>
> My bad, forgot I decreased test image resolution to 256x256. 
> I'm totally out of the running. I have some serious work to do 
> by the looks.

Have you considered just being able to grab an object with 
changed iteration order instead of actually doing the flip? The 
same goes for transposes and 90º rotations. Sure, sometimes you 
do need actually rearrange the memory and in a subset of those 
cases you need it to be done fast, but a lot of the time you're 
better off* just using a different iteration scheme (which, for 
ranges, should probably be part of the type to avoid checking the 
scheme every iteration).

*for speed and memory reasons. Need to keep the original and the 
transpose? No need to for any duplicates

Note that this is what numpy does with transposes. The .T and 
.transpose methods of ndarray don't actually modify the data, 
they just set the memory order** whereas the transpose function 
actually moves memory around.

**using a runtime flag, which is ok for them because internal 
iteration lets you only branch once on it.