/* For a 16*h block, this computes
(((((*pf + *pf2 + 1)>>1) + ((*pb + *pb2 + 1)>>1) + 1)>>1) + *p2 + 1)>>1
*/
static int bsad_0quad_mmxe(uint8_t *pf,uint8_t *pf2,uint8_t *pb,uint8_t *pb2,uint8_t *p2,int lx,int h)
{
int32_t s=0;
pxor_r2r(mm7, mm7);
do {
movq_m2r(pf2[0],mm0);
movq_m2r(pf2[8],mm2);
movq_m2r(pb2[0],mm1);
movq_m2r(pb2[8],mm3);
pavgb_m2r(pf[0],mm0);
pavgb_m2r(pf[8],mm2);
pavgb_m2r(pb[0],mm1);
pavgb_m2r(pb[8],mm3);
pavgb_r2r(mm1,mm0);
pavgb_r2r(mm3,mm2);
psadbw_m2r(p2[0],mm0);
psadbw_m2r(p2[8],mm2);
paddd_r2r(mm0,mm7);
paddd_r2r(mm2,mm7);
pf+=lx;
pf2+=lx;
pb+=lx;
pb2+=lx;
p2+=lx;
h--;
} while (h);
movd_r2g(mm7,s);
emms();
return s;
}
static void
deinterlace_scanline_linear_mmxext (GstDeinterlaceMethod * self,
GstDeinterlace * parent, guint8 * out,
GstDeinterlaceScanlineData * scanlines, gint width)
{
gint i;
guint8 *bot = scanlines->b0, *top = scanlines->t0;
for (i = width / 16; i; --i) {
movq_m2r (*bot, mm0);
movq_m2r (*top, mm1);
movq_m2r (*(bot + 8), mm2);
movq_m2r (*(top + 8), mm3);
movq_m2r (*(bot + 16), mm4);
movq_m2r (*(top + 16), mm5);
movq_m2r (*(bot + 24), mm6);
movq_m2r (*(top + 24), mm7);
pavgb_r2r (mm1, mm0);
pavgb_r2r (mm3, mm2);
pavgb_r2r (mm5, mm4);
pavgb_r2r (mm7, mm6);
movntq_r2m (mm0, *out);
movntq_r2m (mm2, *(out + 8));
movntq_r2m (mm4, *(out + 16));
movntq_r2m (mm6, *(out + 24));
out += 32;
top += 32;
bot += 32;
}
width = (width & 0xf);
for (i = width / 4; i; --i) {
movq_m2r (*bot, mm0);
movq_m2r (*top, mm1);
pavgb_r2r (mm1, mm0);
movntq_r2m (mm0, *out);
out += 8;
top += 8;
bot += 8;
}
width = width & 0x7;
/* Handle last few pixels. */
for (i = width * 2; i; --i) {
*out++ = ((*top++) + (*bot++)) >> 1;
}
emms ();
}
static void
deinterlace_scanline_linear_mmxext (GstDeinterlaceSimpleMethod * self,
guint8 * out, const guint8 * bot, const guint8 * top, gint size)
{
gint i;
for (i = size / 32; i; --i) {
movq_m2r (*bot, mm0);
movq_m2r (*top, mm1);
movq_m2r (*(bot + 8), mm2);
movq_m2r (*(top + 8), mm3);
movq_m2r (*(bot + 16), mm4);
movq_m2r (*(top + 16), mm5);
movq_m2r (*(bot + 24), mm6);
movq_m2r (*(top + 24), mm7);
pavgb_r2r (mm1, mm0);
pavgb_r2r (mm3, mm2);
pavgb_r2r (mm5, mm4);
pavgb_r2r (mm7, mm6);
movntq_r2m (mm0, *out);
movntq_r2m (mm2, *(out + 8));
movntq_r2m (mm4, *(out + 16));
movntq_r2m (mm6, *(out + 24));
out += 32;
top += 32;
bot += 32;
}
size = (size & 0x1f);
for (i = size / 8; i; --i) {
movq_m2r (*bot, mm0);
movq_m2r (*top, mm1);
pavgb_r2r (mm1, mm0);
movntq_r2m (mm0, *out);
out += 8;
top += 8;
bot += 8;
}
emms ();
size = size & 0xf;
/* Handle last few pixels. */
for (i = size; i; --i) {
*out++ = ((*top++) + (*bot++)) >> 1;
}
}
static void interpolate_packed422_scanline_mmxext( uint8_t *output, uint8_t *top,
uint8_t *bot, int width )
{
int i;
for( i = width/16; i; --i ) {
movq_m2r( *bot, mm0 );
movq_m2r( *top, mm1 );
movq_m2r( *(bot + 8), mm2 );
movq_m2r( *(top + 8), mm3 );
movq_m2r( *(bot + 16), mm4 );
movq_m2r( *(top + 16), mm5 );
movq_m2r( *(bot + 24), mm6 );
movq_m2r( *(top + 24), mm7 );
pavgb_r2r( mm1, mm0 );
pavgb_r2r( mm3, mm2 );
pavgb_r2r( mm5, mm4 );
pavgb_r2r( mm7, mm6 );
movntq_r2m( mm0, *output );
movntq_r2m( mm2, *(output + 8) );
movntq_r2m( mm4, *(output + 16) );
movntq_r2m( mm6, *(output + 24) );
output += 32;
top += 32;
bot += 32;
}
width = (width & 0xf);
for( i = width/4; i; --i ) {
movq_m2r( *bot, mm0 );
movq_m2r( *top, mm1 );
pavgb_r2r( mm1, mm0 );
movntq_r2m( mm0, *output );
output += 8;
top += 8;
bot += 8;
}
width = width & 0x7;
/* Handle last few pixels. */
for( i = width * 2; i; --i ) {
*output++ = ((*top++) + (*bot++)) >> 1;
}
sfence();
emms();
}
static inline void XDeint8x8FieldEMMXEXT( uint8_t *dst, int i_dst,
uint8_t *src, int i_src )
{
int y;
/* Interlaced */
for( y = 0; y < 8; y += 2 )
{
movq_m2r( src[0], mm0 );
movq_r2m( mm0, dst[0] );
dst += i_dst;
movq_m2r( src[2*i_src], mm1 );
pavgb_r2r( mm1, mm0 );
movq_r2m( mm0, dst[0] );
dst += 1*i_dst;
src += 2*i_src;
}
}
static int bsad_1quad_mmxe(uint8_t *pf, uint8_t *pb, uint8_t *pb2, uint8_t *p2, int lx, int h)
{
int s;
s = 0; /* the accumulator */
if (h > 0)
{
pcmpeqw_r2r(mm6, mm6);
psrlw_i2r(15, mm6);
paddw_r2r(mm6, mm6);
pxor_r2r(mm7, mm7);
pxor_r2r(mm5, mm5);
do {
BSAD_LOAD(pf[0],mm0,mm1);
BSAD_LOAD_ACC(pf[1],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pf[lx],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pf[lx+1],mm2,mm3,mm0,mm1);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm1);
packuswb_r2r(mm1, mm0);
movq_m2r(pb2[0],mm1);
pavgb_m2r(pb[0],mm1);
pavgb_r2r(mm1, mm0);
psadbw_m2r(p2[0],mm0);
paddd_r2r(mm0,mm5);
BSAD_LOAD(pf[8],mm0,mm1);
BSAD_LOAD_ACC(pf[9],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pf[lx+8],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pf[lx+9],mm2,mm3,mm0,mm1);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm1);
packuswb_r2r(mm1, mm0);
movq_m2r(pb2[8],mm1);
pavgb_m2r(pb[8],mm1);
pavgb_r2r(mm1, mm0);
psadbw_m2r(p2[8],mm0);
paddd_r2r(mm0,mm5);
p2 += lx;
pf += lx;
pb += lx;
pb2 += lx;
h--;
} while (h > 0);
}
movd_r2g(mm5,s);
emms();
return s;
}
static void
deinterlace_greedy_scanline_mmxext (GstDeinterlaceMethodGreedyL *
self, const guint8 * m0, const guint8 * t1, const guint8 * b1,
const guint8 * m2, guint8 * output, gint width)
{
mmx_t MaxComb;
// How badly do we let it weave? 0-255
MaxComb.ub[0] = self->max_comb;
MaxComb.ub[1] = self->max_comb;
MaxComb.ub[2] = self->max_comb;
MaxComb.ub[3] = self->max_comb;
MaxComb.ub[4] = self->max_comb;
MaxComb.ub[5] = self->max_comb;
MaxComb.ub[6] = self->max_comb;
MaxComb.ub[7] = self->max_comb;
// L2 == m0
// L1 == t1
// L3 == b1
// LP2 == m2
movq_m2r (MaxComb, mm6);
for (; width > 7; width -= 8) {
movq_m2r (*t1, mm1); // L1
movq_m2r (*m0, mm2); // L2
movq_m2r (*b1, mm3); // L3
movq_m2r (*m2, mm0); // LP2
// average L1 and L3 leave result in mm4
movq_r2r (mm1, mm4); // L1
pavgb_r2r (mm3, mm4); // (L1 + L3)/2
// get abs value of possible L2 comb
movq_r2r (mm2, mm7); // L2
psubusb_r2r (mm4, mm7); // L2 - avg
movq_r2r (mm4, mm5); // avg
psubusb_r2r (mm2, mm5); // avg - L2
por_r2r (mm7, mm5); // abs(avg-L2)
// get abs value of possible LP2 comb
movq_r2r (mm0, mm7); // LP2
psubusb_r2r (mm4, mm7); // LP2 - avg
psubusb_r2r (mm0, mm4); // avg - LP2
por_r2r (mm7, mm4); // abs(avg-LP2)
// use L2 or LP2 depending upon which makes smaller comb
psubusb_r2r (mm5, mm4); // see if it goes to zero
pxor_r2r (mm5, mm5); // 0
pcmpeqb_r2r (mm5, mm4); // if (mm4=0) then FF else 0
pcmpeqb_r2r (mm4, mm5); // opposite of mm4
// if Comb(LP2) <= Comb(L2) then mm4=ff, mm5=0 else mm4=0, mm5 = 55
pand_r2r (mm2, mm5); // use L2 if mm5 == ff, else 0
pand_r2r (mm0, mm4); // use LP2 if mm4 = ff, else 0
por_r2r (mm5, mm4); // may the best win
// Now lets clip our chosen value to be not outside of the range
// of the high/low range L1-L3 by more than abs(L1-L3)
// This allows some comb but limits the damages and also allows more
// detail than a boring oversmoothed clip.
movq_r2r (mm1, mm2); // copy L1
pmaxub_r2r (mm3, mm2); // now = Max(L1,L3)
pminub_r2r (mm1, mm3); // now = Min(L1,L3)
// allow the value to be above the high or below the low by amt of MaxComb
paddusb_r2r (mm6, mm2); // increase max by diff
psubusb_r2r (mm6, mm3); // lower min by diff
pmaxub_r2r (mm3, mm4); // now = Max(best,Min(L1,L3)
pminub_r2r (mm4, mm2); // now = Min( Max(best, Min(L1,L3)), L2 )=L2 clipped
movq_r2m (mm2, *output); // move in our clipped best
// Advance to the next set of pixels.
output += 8;
m0 += 8;
t1 += 8;
b1 += 8;
m2 += 8;
}
emms ();
if (width > 0)
deinterlace_greedy_scanline_c (self, m0, t1, b1, m2, output, width);
}
