static inline void MC_avg4_16 (int height, uint8_t * dest, const uint8_t * ref,
const int stride, const int cpu)
{
do {
movq_m2r (*ref, mm0);
movq_m2r (*(ref+stride+1), mm1);
movq_r2r (mm0, mm7);
movq_m2r (*(ref+1), mm2);
pxor_r2r (mm1, mm7);
movq_m2r (*(ref+stride), mm3);
movq_r2r (mm2, mm6);
pxor_r2r (mm3, mm6);
pavg_r2r (mm1, mm0);
pavg_r2r (mm3, mm2);
por_r2r (mm6, mm7);
movq_r2r (mm0, mm6);
pxor_r2r (mm2, mm6);
pand_r2r (mm6, mm7);
pand_m2r (mask_one, mm7);
pavg_r2r (mm2, mm0);
psubusb_r2r (mm7, mm0);
movq_m2r (*dest, mm1);
pavg_r2r (mm1, mm0);
movq_r2m (mm0, *dest);
movq_m2r (*(ref+8), mm0);
movq_m2r (*(ref+stride+9), mm1);
movq_r2r (mm0, mm7);
movq_m2r (*(ref+9), mm2);
pxor_r2r (mm1, mm7);
movq_m2r (*(ref+stride+8), mm3);
movq_r2r (mm2, mm6);
pxor_r2r (mm3, mm6);
pavg_r2r (mm1, mm0);
pavg_r2r (mm3, mm2);
por_r2r (mm6, mm7);
movq_r2r (mm0, mm6);
pxor_r2r (mm2, mm6);
pand_r2r (mm6, mm7);
pand_m2r (mask_one, mm7);
pavg_r2r (mm2, mm0);
psubusb_r2r (mm7, mm0);
movq_m2r (*(dest+8), mm1);
pavg_r2r (mm1, mm0);
ref += stride;
movq_r2m (mm0, *(dest+8));
dest += stride;
} while (--height);
}
static inline void mean8(unsigned char *refpix,unsigned char *pixel,int radius_count,int row_stride,int threshold,int8_t *diff,unsigned char *count)
{
int a,b;
pxor_r2r(mm6,mm6); // mm6 (aka count) = 0
pxor_r2r(mm7,mm7); // mm7 (aka diff) = 0
movq_m2r(*refpix,mm3); // mm3 = refpix[0]
movd_g2r(0x80808080,mm4); // mm4 = 128
punpcklbw_r2r(mm4,mm4);
pxor_r2r(mm4,mm3); // mm3 = refpix[0]-128
movd_g2r(threshold,mm5); // mm5 = threshold
punpcklbw_r2r(mm5,mm5);
punpcklbw_r2r(mm5,mm5);
punpcklbw_r2r(mm5,mm5);
for( b=0; b<radius_count; b++ ) {
for( a=0; a<radius_count; a++ ) {
movq_m2r(*pixel,mm0); // mm0 = pixel[0]
pxor_r2r(mm4,mm0); // mm0 = pixel[0]-128
movq_r2r(mm3,mm2); // mm2 = refpix[0]-128
psubsb_r2r(mm0,mm2); // mm2 = refpix[0]-pixel[0]
psubsb_r2r(mm3,mm0); // mm0 = pixel[0]-refpix[0]
pminub_r2r(mm0,mm2); // mm2 = abs(pixel[0]-refpix[0])
movq_r2r(mm5,mm1); // mm1 = threshold
pcmpgtb_r2r(mm2,mm1); // mm1 = (threshold > abs(pixel[0]-refpix[0])) ? -1 : 0
psubb_r2r(mm1,mm6); // mm6 += (threshold > abs(pixel[0]-refpix[0]))
pand_r2r(mm1,mm0); // mm0 = (threshold > abs(pixel[0]-refpix[0])) ? pixel[0]-refpix[0] : 0
paddb_r2r(mm0,mm7); // mm7 += (threshold > abs(pixel[0]-refpix[0])) ? pixel[0]-refpix[0] : 0
++pixel;
}
pixel += row_stride - radius_count;
}
movq_r2m(mm6,*count);
movq_r2m(mm7,*diff);
emms();
}
static inline void MC_put4_8 (int height, uint8_t * dest, const uint8_t * ref,
const int stride, const int cpu)
{
movq_m2r (*ref, mm0);
movq_m2r (*(ref+1), mm1);
movq_r2r (mm0, mm7);
pxor_r2r (mm1, mm7);
pavg_r2r (mm1, mm0);
ref += stride;
do {
movq_m2r (*ref, mm2);
movq_r2r (mm0, mm5);
movq_m2r (*(ref+1), mm3);
movq_r2r (mm2, mm6);
pxor_r2r (mm3, mm6);
pavg_r2r (mm3, mm2);
por_r2r (mm6, mm7);
pxor_r2r (mm2, mm5);
pand_r2r (mm5, mm7);
pavg_r2r (mm2, mm0);
pand_m2r (mask_one, mm7);
psubusb_r2r (mm7, mm0);
ref += stride;
movq_r2m (mm0, *dest);
dest += stride;
movq_r2r (mm6, mm7); /* unroll ! */
movq_r2r (mm2, mm0); /* unroll ! */
} while (--height);
}
static void vfilter_chroma_332_packed422_scanline_mmx( uint8_t *output, int width,
uint8_t *m, uint8_t *t, uint8_t *b )
{
int i;
const mmx_t ymask = { 0x00ff00ff00ff00ffULL };
const mmx_t cmask = { 0xff00ff00ff00ff00ULL };
// Get width in bytes.
width *= 2;
i = width / 8;
width -= i * 8;
movq_m2r( ymask, mm7 );
movq_m2r( cmask, mm6 );
while ( i-- )
{
movq_m2r( *t, mm0 );
movq_m2r( *b, mm1 );
movq_m2r( *m, mm2 );
movq_r2r ( mm2, mm3 );
pand_r2r ( mm7, mm3 );
pand_r2r ( mm6, mm0 );
pand_r2r ( mm6, mm1 );
pand_r2r ( mm6, mm2 );
psrlq_i2r( 8, mm0 );
psrlq_i2r( 7, mm1 );
psrlq_i2r( 8, mm2 );
movq_r2r ( mm0, mm4 );
psllw_i2r( 1, mm4 );
paddw_r2r( mm4, mm0 );
movq_r2r ( mm2, mm4 );
psllw_i2r( 1, mm4 );
paddw_r2r( mm4, mm2 );
paddw_r2r( mm0, mm2 );
paddw_r2r( mm1, mm2 );
psllw_i2r( 5, mm2 );
pand_r2r( mm6, mm2 );
por_r2r ( mm3, mm2 );
movq_r2m( mm2, *output );
output += 8;
t += 8;
b += 8;
m += 8;
}
output++; t++; b++; m++;
while ( width-- )
{
*output = (3 * *t + 3 * *m + 2 * *b) >> 3;
output +=2; t+=2; b+=2; m+=2;
}
emms();
}
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);
}
void deinterlace_bob_yuv_mmx(uint8_t *pdst, uint8_t *psrc,
int width, int height )
{
int Line;
long long* YVal1;
long long* YVal2;
long long* YVal3;
long long* Dest;
uint8_t* pEvenLines = psrc;
uint8_t* pOddLines = psrc+width;
int LineLength = width;
int Pitch = width * 2;
int IsOdd = 1;
long EdgeDetect = 625;
long JaggieThreshold = 73;
int n;
unsigned long long qwEdgeDetect;
unsigned long long qwThreshold;
const unsigned long long Mask = 0xfefefefefefefefeULL;
const unsigned long long YMask = 0x00ff00ff00ff00ffULL;
qwEdgeDetect = EdgeDetect;
qwEdgeDetect += (qwEdgeDetect << 48) + (qwEdgeDetect << 32) + (qwEdgeDetect << 16);
qwThreshold = JaggieThreshold;
qwThreshold += (qwThreshold << 48) + (qwThreshold << 32) + (qwThreshold << 16);
// copy first even line no matter what, and the first odd line if we're
// processing an odd field.
ac_memcpy(pdst, pEvenLines, LineLength);
if (IsOdd)
ac_memcpy(pdst + LineLength, pOddLines, LineLength);
height = height / 2;
for (Line = 0; Line < height - 1; ++Line)
{
if (IsOdd)
{
YVal1 = (long long *)(pOddLines + Line * Pitch);
YVal2 = (long long *)(pEvenLines + (Line + 1) * Pitch);
YVal3 = (long long *)(pOddLines + (Line + 1) * Pitch);
Dest = (long long *)(pdst + (Line * 2 + 2) * LineLength);
}
else
{
YVal1 = (long long *)(pEvenLines + Line * Pitch);
YVal2 = (long long *)(pOddLines + Line * Pitch);
YVal3 = (long long *)(pEvenLines + (Line + 1) * Pitch);
Dest = (long long *)(pdst + (Line * 2 + 1) * LineLength);
}
// For ease of reading, the comments below assume that we're operating on an odd
// field (i.e., that bIsOdd is true). The exact same processing is done when we
// operate on an even field, but the roles of the odd and even fields are reversed.
// It's just too cumbersome to explain the algorithm in terms of "the next odd
// line if we're doing an odd field, or the next even line if we're doing an
// even field" etc. So wherever you see "odd" or "even" below, keep in mind that
// half the time this function is called, those words' meanings will invert.
// Copy the odd line to the overlay verbatim.
ac_memcpy((char *)Dest + LineLength, YVal3, LineLength);
n = LineLength >> 3;
while( n-- )
{
movq_m2r (*YVal1++, mm0);
movq_m2r (*YVal2++, mm1);
movq_m2r (*YVal3++, mm2);
// get intensities in mm3 - 4
movq_r2r ( mm0, mm3 );
movq_r2r ( mm1, mm4 );
movq_r2r ( mm2, mm5 );
pand_m2r ( *&YMask, mm3 );
pand_m2r ( *&YMask, mm4 );
pand_m2r ( *&YMask, mm5 );
// get average in mm0
pand_m2r ( *&Mask, mm0 );
pand_m2r ( *&Mask, mm2 );
psrlw_i2r ( 01, mm0 );
psrlw_i2r ( 01, mm2 );
paddw_r2r ( mm2, mm0 );
// work out (O1 - E) * (O2 - E) / 2 - EdgeDetect * (O1 - O2) ^ 2 >> 12
// result will be in mm6
psrlw_i2r ( 01, mm3 );
psrlw_i2r ( 01, mm4 );
psrlw_i2r ( 01, mm5 );
movq_r2r ( mm3, mm6 );
psubw_r2r ( mm4, mm6 ); //mm6 = O1 - E
movq_r2r ( mm5, mm7 );
psubw_r2r ( mm4, mm7 ); //mm7 = O2 - E
pmullw_r2r ( mm7, mm6 ); // mm6 = (O1 - E) * (O2 - E)
movq_r2r ( mm3, mm7 );
psubw_r2r ( mm5, mm7 ); // mm7 = (O1 - O2)
pmullw_r2r ( mm7, mm7 ); // mm7 = (O1 - O2) ^ 2
psrlw_i2r ( 12, mm7 ); // mm7 = (O1 - O2) ^ 2 >> 12
pmullw_m2r ( *&qwEdgeDetect, mm7 );// mm7 = EdgeDetect * (O1 - O2) ^ 2 >> 12
psubw_r2r ( mm7, mm6 ); // mm6 is what we want
pcmpgtw_m2r ( *&qwThreshold, mm6 );
movq_r2r ( mm6, mm7 );
pand_r2r ( mm6, mm0 );
pandn_r2r ( mm1, mm7 );
por_r2r ( mm0, mm7 );
movq_r2m ( mm7, *Dest++ );
}
}
// Copy last odd line if we're processing an even field.
if (! IsOdd)
{
ac_memcpy(pdst + (height * 2 - 1) * LineLength,
pOddLines + (height - 1) * Pitch,
LineLength);
}
// clear out the MMX registers ready for doing floating point
// again
emms();
}
