Overhead of a switch statement in C

Asked 2009-05-29T18:18:06.670

I'm a fairly competent Java programmer who's very new to C. I am trying to optimize a routine that has four modes of operation.

I loop over all the pixels in an image and compute a new pixel value depending on the 'mode' passed.

My question regards the overhead of a switch statement within two nested for loops. I'd be interested in any links to documentation regarding the relative efficiency of basic C statements, math and logical operations.

The code would go as follows;

for (x = 0; x < width; x++) {
        for (y = 0; y < height; y++) {
             switch (mode)                  /* select the type of calculation */
             {
                case 0:
                weight = dCentre / maxDistanceEdge;
                case 1: 
                    weight = (float)x/width;             
                    break;
                case 2: 
                    weight = (float)y/height;
                    break;
                case 3: 
                    weight = dBottomLeft / maxDistanceCorner;
                    break;
                case 4: 
                    weight = dTopRight / maxDistanceCorner;
                    break;
                default: 
                weight = 1;
                break;
            }
             // Calculate the new pixel value given the weight
             ...
            }             

    }

Would you expect to see much overhead if this was over a 5000 x 5000 pixel image? I've tried to do some testing but my results are all over the place as the system (Mobile Device) has all sorts of stuff running in the background that may skew results.

The other option is to have a separate method for each mode, each with its own four loops. This would obviously introduce redundant code but efficiency is the name of the game here.

Thanks in advance!

Gav

c switch-statement overhead

Edit Report

3 Answers

In addition to Jim's advice, try swapping the order of the loops. Whether loop-swapping is ideal for case 1 would require testing, but I suspect it is. You almost always want your x coordinate inside your inner loop in order to improve paging performance, as this causes your function to have a better tendency to stay in the same general memory area each iteration. And a mobile device with limitted resources might have low enough ram that this difference will be emphasized.

answered 2009-05-29T19:24:59.903

Sorry to bump this thread, but it seems to me that the switch is FAR from the problem.

The real problem with the efficiency in that case is the divisions. It seems to me that all of denominators of the division operations are constants (width, height, max...) and these will not change throughout the course of the image. If my guess is right, then these are simple variables that can change based on the image loaded so that any size image can be used at runtime, now this allows for any image size to be loaded, but this also means the compiler cannot optimize them into the much simpler multiplication operation which it could do if they were declared "const". My suggestion would be to pre-calculate the inverses of these constants and multiply. As far as I can remember, the multiplication operation takes about 10 clock cycles, where as the division takes around 70. That's an increase of 60 cycles per pixel, and with the above mentioned 5000x5000, that's an estimated speed increase of 1.5 seconds on a 1 GHz CPU.

answered 2010-09-08T03:31:28.127

Depends on the chip and the compiler and the details of the code, and...but this will often be implemented as a jump table, which should be pretty fast.

BTW-- Understanding this kind of thing is a pretty good argument for spending a couple of weeks learning some assembly at some point in your career...

answered 2009-05-29T18:21:58.020

Overhead of a switch statement in C

3 Answers

Your Answer