I found an interesting webpage regarding rounding precision with machine-level floating point numbers in gcc. In DOSBox, sdlmain.cpp casts a double to integer, where 240 can end up as 239 because internally the number is something like 239.9999999

As mentioned in the article, compiling with the "-mfpmath=sse -msse2" options forces the use of the SSE floating point unit instead of the x87 FPU. The "-march=pentium4" compile option has the same effect. Of course you need a Pentium 4 or later Intel processor for it to work. I use -march=pentium4 in my own builds, and don't get the rounding problem (i.e. the window size is 320x240, or 640x480 with a 2x scaler)

The article also has an example of how to switch the x87 FPU to using double precision rounding, but it seems kind of awkward to use. Another option is to pass floating point types through a rounding function such as rint() before casting to int, but then you have to modify the code in quite a few places.

maybe add some +0.5 or +0,25 to them by default ? should get rounded correctly then

That would work. Adding .5 would be like straight-rounding, but adding a very small amount would be more like compensating for the rounding error. You'd still have to find all the places to apply it. Perhaps a define in a common header would look cleaner:

1#define ROUND_ERR 0.0000001
2
3      sdl.clip.x=0;sdl.clip.y=0;
4      sdl.clip.w=(Bit16u)(sdl.draw.width*sdl.draw.scalex+ROUND_ERR);
5      sdl.clip.h=(Bit16u)(sdl.draw.height*sdl.draw.scaley+ROUND_ERR);
6      sdl.surface=SDL_SetVideoMode(sdl.clip.w,sdl.clip.h,bpp,sdl_flags);
7      return sdl.surface;

Instead of:

1#include <math.h>
2
3      sdl.clip.x=0;sdl.clip.y=0;
4      sdl.clip.w=(Bit16u)rint(sdl.draw.width*sdl.draw.scalex);
5      sdl.clip.h=(Bit16u)rint(sdl.draw.height*sdl.draw.scaley);
6      sdl.surface=SDL_SetVideoMode(sdl.clip.w,sdl.clip.h,bpp,sdl_flags);
7      return sdl.surface;

Either one has the desired result.

rint() (and similar) is usually faster than casting on x87. Not that it would matter that much here 😀

Fixed point math!

should be fixed in the cvs.