7

I'm currently developing a project with SDL. It basically draws and moves images (surfaces) on the screen.

To move an image without leaving a trail, you must first clear the screen surface, pretty much like glClear(), and I'm currently doing it with a simple for loop iterating over the surface's pixels (also drawing a black box on the surface or memset).

While the previous solutions work fine for small surfaces, they get increasingly slower as the surface grows bigger so i was looking for the fastest way I could clear (zero) a memory block.

Also, a friend pointed out that using SIMD instructions could do the work really fast but the last time I've done ASM was on a 8085, any insight on this could also be useful.

NeonMan
  • 623
  • 10
  • 24

2 Answers2

15

The fastest way is to use memset.

memset(ptr, 0, length);

This automatically uses SIMD on architectures that support it*. You are not going to beat it. It is already memory bound, so it's writing zeroes as fast as the processor can spit them out. I don't know who told you that memset is slower for larger blocks, but you should stop listening to that person.

*There are some toolchains that don't give you a fast memset. It is unlikely that you are using one.

Dietrich Epp
  • 205,541
  • 37
  • 345
  • 415
  • 1
    Actually, it only applies to SDL functions, didn't thoroughly test memset. <*There are some toolchains that don't give you a fast memset> Some PIC MCU toolchains, for example (not the case obviously) – NeonMan Dec 21 '11 at 01:50
  • Well, it's a little strange to run SDL on a PIC microcontroller, after all, and PIC doesn't have SIMD. – Dietrich Epp Dec 21 '11 at 02:11
  • YMMV. My experience is that memset is commonly pretty slow. The ones I've looked at source for just do byte sets. I've had great success writing my own that (for larger areas) use larger writes for the middle sections (just using small writes at the ends as necesary). – Brian Knoblauch Dec 21 '11 at 13:39
  • @BrianKnoblauch: Which implementations have you looked at? I've looked at the libraries both for Linux (glibc) and Mac OS X (libc) and both of them have a handful of implementations: one in C, then several in assembly language. Both of these libraries even have three different x86 implementations: one for x86, one for x86+SSE2, and one for x86-64, all written entirely in assembly. If you can show me a benchmark on your platform, that would be great. – Dietrich Epp Dec 21 '11 at 17:50
3

You should try memset, the implementation should be highly optimized to take advantage of whatever instructions are available on your system.

TJD
  • 11,800
  • 1
  • 26
  • 34