thirdparty/scummvm
2
0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Rest of patch #774556 (small optimization for aspect ratio correction)

Browse source 
svn-id: r9103
This commit is contained in:
Max Horn 2003-07-21 00:07:27 +00:00
parent 88810dc9bd
commit a7391d0a9a
1 changed files with 30 additions and 24 deletions
Download patch file Download diff file Expand all files Collapse all files

54
common/scaler.cpp
View file

@ -685,30 +685,19 @@ static inline void interpolate5Line(uint16 *dst, const uint16 *srcA, const uint1
#if ASPECT_MODE == kFastAndNiceAspectMode #if ASPECT_MODE == kFastAndNiceAspectMode
template<int scale> template<int scale>
static inline void interpolate5Line(uint16 *dst, const uint16 *srcA, const uint16 *srcB, int width) { static inline void interpolate5Line(uint16 *dst, const uint16 *srcA, const uint16 *srcB, int width) {
// For efficiency reasons we blit two pixels at a time, so it is
// important that makeRectStretchable() guarantees that the width is
// even and that the rect starts on a well-aligned address. (Even
// where unaligned memory access is allowed there may be a speed
// penalty for it.)
// TODO: This code may not work correctly on architectures that require alignment. // These asserts are disabled for maximal speed; but I leave them in here in case
// And even on those which accept it, reading/writing 32 bits words from odd memory // other people want to test if the memory alignment (to an address divisibl by 4)
// locations usually has a speed penalty. Hence, it might be wise to first check // are really effective.
// if the dst address is odd, in which case we blit one pixel first; then we //assert(((int)dst & 3) == 0);
// blit pixel pairs, till we get to the end, at which point we may have to blit //assert(((int)srcA & 3) == 0);
// again a single seperate pixel. This would of course cause additional overhead //assert(((int)srcB & 3) == 0);
// for each blitted line. Some of that overhead can be avoid by moving //assert((width & 1) == 0);
// this logic into stretch200To240, since whether dst is at an odd position, and
// whether the last pixel has to be blitted seperately or not, is identical for
// each blitted line.
//
if (width & 1) {
// For efficency reasons we normally blit two pixels at a time; but if the
// width is odd, we first blit a single pixel.
width--;
if (scale == 1) {
uint32 B = *srcB++;
*dst++ = (uint16)Q_INTERPOLATE(*srcA++, B, B, B);
} else {
*dst++ = (uint16)INTERPOLATE(*srcA++, *srcB++);
}
}
width /= 2; width /= 2;
const uint32 *sA = (const uint32 *)srcA; const uint32 *sA = (const uint32 *)srcA;
@ -728,15 +717,32 @@ static inline void interpolate5Line(uint16 *dst, const uint16 *srcA, const uint1
#endif #endif
void makeRectStretchable(int &x, int &y, int &w, int &h) { void makeRectStretchable(int &x, int &y, int &w, int &h) {
#if ASPECT_MODE != kVeryFastAndUglyAspectMode
int m = real2Aspect(y) % 6; int m = real2Aspect(y) % 6;
// Ensure that the rect will start on a line that won't have its // Ensure that the rect will start on a line that won't have its
// colours changed by the stretching function. // colours changed by the stretching function.
if (m != 0 && m != 5) { if (m != 0 && m != 5) {
y -= m; y -= m;
h += m; h += m;
} }
#if ASPECT_MODE == kFastAndNiceAspectMode
// Force x to be even, to ensure aligned memory access (this assumes
// that each line starts at an even memory location, but that should
// be the case on every target anyway).
if (x & 1) {
x--;
w++;
}
// Finally force the width to be even, since we blit 2 pixels at a time.
// While this means we may sometimes blit one column more than necessary,
// this should actually be faster than having the check for the
if (w & 1)
w++;
#endif
#endif
} }
/** /**