/**
* Apply the Scale2x effect on a group of rows. Used internally.
*/
static inline void stage_scale2x(void* dst0, void* dst1, const void* src0, const void* src1, const void* src2, unsigned pixel, unsigned pixel_per_row)
{
switch (pixel) {
#if defined(__GNUC__) && defined(__i386__)
case 1 :
scale2x_8_mmx(dst0, dst1, src0, src1, src2, pixel_per_row);
break;
case 2 :
scale2x_16_mmx(dst0, dst1, src0, src1, src2, pixel_per_row);
break;
case 4 :
scale2x_32_mmx(dst0, dst1, src0, src1, src2, pixel_per_row);
break;
#else
case 1 :
scale2x_8_def(dst0, dst1, src0, src1, src2, pixel_per_row);
break;
case 2 :
scale2x_16_def(dst0, dst1, src0, src1, src2, pixel_per_row);
break;
case 4 :
scale2x_32_def(dst0, dst1, src0, src1, src2, pixel_per_row);
break;
#endif
}
}
static bool Scale2x_16MMX( Texture *tex , sint32 sx, sint32 sy, sint32 sw, sint32 sh,
uint8* pixel, sint32 dw, sint32 dh, sint32 pitch, bool clamp_src)
{
// Must be at least 3 high
if (sh<3 && (clamp_src || tex->height<3)) return false;
// Source buffer pointers
uint16 *texel = reinterpret_cast<uint16*>(tex->buffer) + (sy * tex->width + sx);
int tpitch = tex->width;
// uint16 *tline_end = texel + sw;
uint16 *tex_end = texel + (sh-1)*tex->width;
bool clip_y = true;
if (sh+sy < tex->height && clamp_src == false)
{
clip_y = false;
tex_end = texel + sh*tex->width;
}
if (sy == 0) {
scale2x_16_mmx(reinterpret_cast<uint16*>(pixel),
reinterpret_cast<uint16*>(pixel+pitch),
texel, texel, texel+tpitch, sw);
pixel += pitch*2;
texel += tpitch;
}
// Src Loop Y
if (texel != tex_end) do {
scale2x_16_mmx(reinterpret_cast<uint16*>(pixel),
reinterpret_cast<uint16*>(pixel+pitch),
texel-tpitch, texel, texel+tpitch, sw);
pixel += pitch*2;
texel += tpitch;
} while (texel != tex_end);
if (clip_y) {
scale2x_16_mmx(reinterpret_cast<uint16*>(pixel),
reinterpret_cast<uint16*>(pixel+pitch),
texel-tpitch, texel, texel, sw);
}
scale2x_mmx_emms();
return true;
}
/**
* Apply the Scale2x effect on a group of rows. Used internally.
*/
static inline void stage_scale2x(void* dst0, void* dst1, const void* src0, const void* src1, const void* src2, unsigned pixel, unsigned pixel_per_row)
{
switch (pixel) {
#if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
case 1 : scale2x_8_mmx(SSDST(8,0), SSDST(8,1), SSSRC(8,0), SSSRC(8,1), SSSRC(8,2), pixel_per_row); break;
case 2 : scale2x_16_mmx(SSDST(16,0), SSDST(16,1), SSSRC(16,0), SSSRC(16,1), SSSRC(16,2), pixel_per_row); break;
case 4 : scale2x_32_mmx(SSDST(32,0), SSDST(32,1), SSSRC(32,0), SSSRC(32,1), SSSRC(32,2), pixel_per_row); break;
#else
case 1 : scale2x_8_def(SSDST(8,0), SSDST(8,1), SSSRC(8,0), SSSRC(8,1), SSSRC(8,2), pixel_per_row); break;
case 2 : scale2x_16_def(SSDST(16,0), SSDST(16,1), SSSRC(16,0), SSSRC(16,1), SSSRC(16,2), pixel_per_row); break;
case 4 : scale2x_32_def(SSDST(32,0), SSDST(32,1), SSSRC(32,0), SSSRC(32,1), SSSRC(32,2), pixel_per_row); break;
#endif
}
}