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 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);
}
VLC_MMX
static int TestForMotionInBlockMMX( uint8_t *p_pix_p, uint8_t *p_pix_c,
int i_pitch_prev, int i_pitch_curr,
int* pi_top, int* pi_bot )
{
int32_t i_motion = 0;
int32_t i_top_motion = 0;
int32_t i_bot_motion = 0;
static alignas (8) const mmx_t bT = { .ub = { T, T, T, T, T, T, T, T } };
pxor_r2r( mm6, mm6 ); /* zero, used in psadbw */
movq_m2r( bT, mm5 );
pxor_r2r( mm3, mm3 ); /* score (top field) */
pxor_r2r( mm4, mm4 ); /* score (bottom field) */
for( int y = 0; y < 8; y+=2 )
{
/* top field */
movq_m2r( *((uint64_t*)p_pix_c), mm0 );
movq_m2r( *((uint64_t*)p_pix_p), mm1 );
movq_r2r( mm0, mm2 );
psubusb_r2r( mm1, mm2 );
psubusb_r2r( mm0, mm1 );
pcmpgtb_r2r( mm5, mm2 );
pcmpgtb_r2r( mm5, mm1 );
psadbw_r2r( mm6, mm2 );
psadbw_r2r( mm6, mm1 );
paddd_r2r( mm2, mm1 );
paddd_r2r( mm1, mm3 ); /* add to top field score */
p_pix_c += i_pitch_curr;
p_pix_p += i_pitch_prev;
/* bottom field - handling identical to top field, except... */
movq_m2r( *((uint64_t*)p_pix_c), mm0 );
movq_m2r( *((uint64_t*)p_pix_p), mm1 );
movq_r2r( mm0, mm2 );
psubusb_r2r( mm1, mm2 );
psubusb_r2r( mm0, mm1 );
pcmpgtb_r2r( mm5, mm2 );
pcmpgtb_r2r( mm5, mm1 );
psadbw_r2r( mm6, mm2 );
psadbw_r2r( mm6, mm1 );
paddd_r2r( mm2, mm1 );
paddd_r2r( mm1, mm4 ); /* ...here we add to bottom field score */
p_pix_c += i_pitch_curr;
p_pix_p += i_pitch_prev;
}
movq_r2r( mm3, mm7 ); /* score (total) */
paddd_r2r( mm4, mm7 );
movd_r2m( mm3, i_top_motion );
movd_r2m( mm4, i_bot_motion );
movd_r2m( mm7, i_motion );
/* The loop counts actual score * 255. */
i_top_motion /= 255;
i_bot_motion /= 255;
i_motion /= 255;
emms();
(*pi_top) = ( i_top_motion >= 8 );
(*pi_bot) = ( i_bot_motion >= 8 );
return (i_motion >= 8);
}
int bsad_mmx(uint8_t *pf, uint8_t *pb, uint8_t *p2, int lx, int hxf, int hyf, int hxb, int hyb, int h)
{
uint8_t *pfa,*pfb,*pfc,*pba,*pbb,*pbc;
int s, s1, s2;
pfa = pf + hxf;
pfb = pf + lx * hyf;
pfc = pfb + hxf;
pba = pb + hxb;
pbb = pb + lx * hyb;
pbc = pbb + hxb;
s = 0; /* the accumulator */
if (h > 0)
{
pxor_r2r(mm7, mm7);
pxor_r2r(mm6, mm6);
pcmpeqw_r2r(mm5, mm5);
psubw_r2r(mm5, mm6);
psllw_i2r(1, mm6);
do {
BSAD_LOAD(pf[0],mm0,mm1);
BSAD_LOAD_ACC(pfa[0],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pfb[0],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pfc[0],mm2,mm3,mm0,mm1);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm1);
BSAD_LOAD(pb[0],mm2,mm3);
BSAD_LOAD_ACC(pba[0],mm4,mm5,mm2,mm3);
BSAD_LOAD_ACC(pbb[0],mm4,mm5,mm2,mm3);
BSAD_LOAD_ACC(pbc[0],mm4,mm5,mm2,mm3);
paddw_r2r(mm6, mm2);
paddw_r2r(mm6, mm3);
psrlw_i2r(2, mm2);
psrlw_i2r(2, mm3);
paddw_r2r(mm2, mm0);
paddw_r2r(mm3, mm1);
psrlw_i2r(1, mm6);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psllw_i2r(1, mm6);
psrlw_i2r(1, mm0);
psrlw_i2r(1, mm1);
packuswb_r2r(mm1, mm0);
movq_m2r(p2[0], mm1);
movq_r2r(mm0, mm2);
psubusb_r2r(mm1, mm0);
psubusb_r2r(mm2, mm1);
por_r2r(mm1, mm0);
movq_r2r(mm0, mm1);
punpcklbw_r2r(mm7, mm0);
punpckhbw_r2r(mm7, mm1);
paddw_r2r(mm1, mm0);
movq_r2r(mm0, mm1);
punpcklwd_r2r(mm7, mm0);
punpckhwd_r2r(mm7, mm1);
paddd_r2r(mm1, mm0);
movd_r2g(mm0, s1);
psrlq_i2r(32, mm0);
movd_r2g(mm0, s2);
s += s1 + s2;
BSAD_LOAD(pf[8],mm0,mm1);
BSAD_LOAD_ACC(pfa[8],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pfb[8],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pfc[8],mm2,mm3,mm0,mm1);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm1);
BSAD_LOAD(pb[8],mm2,mm3);
BSAD_LOAD_ACC(pba[8],mm4,mm5,mm2,mm3);
BSAD_LOAD_ACC(pbb[8],mm4,mm5,mm2,mm3);
BSAD_LOAD_ACC(pbc[8],mm4,mm5,mm2,mm3);
paddw_r2r(mm6, mm2);
paddw_r2r(mm6, mm3);
psrlw_i2r(2, mm2);
psrlw_i2r(2, mm3);
paddw_r2r(mm2, mm0);
paddw_r2r(mm3, mm1);
psrlw_i2r(1, mm6);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psllw_i2r(1, mm6);
psrlw_i2r(1, mm0);
psrlw_i2r(1, mm1);
packuswb_r2r(mm1, mm0);
movq_m2r(p2[8], mm1);
movq_r2r(mm0, mm2);
psubusb_r2r(mm1, mm0);
psubusb_r2r(mm2, mm1);
por_r2r(mm1, mm0);
movq_r2r(mm0, mm1);
punpcklbw_r2r(mm7, mm0);
punpckhbw_r2r(mm7, mm1);
paddw_r2r(mm1, mm0);
movq_r2r(mm0, mm1);
punpcklwd_r2r(mm7, mm0);
punpckhwd_r2r(mm7, mm1);
paddd_r2r(mm1, mm0);
movd_r2g(mm0, s1);
psrlq_i2r(32, mm0);
movd_r2g(mm0, s2);
s += s1 + s2;
p2 += lx;
pf += lx;
pfa += lx;
pfb += lx;
pfc += lx;
pb += lx;
pba += lx;
pbb += lx;
pbc += lx;
h--;
} while (h > 0);
}
emms();
return s;
}
//VLC_MMX // sunqueen delete
static int TestForMotionInBlockMMX( uint8_t *p_pix_p, uint8_t *p_pix_c,
int i_pitch_prev, int i_pitch_curr,
int* pi_top, int* pi_bot )
{
int32_t i_motion = 0;
int32_t i_top_motion = 0;
int32_t i_bot_motion = 0;
uint64_t ui_pix_c, ui_pix_p; // sunqueen add
// static const mmx_t bT = { .ub = { T, T, T, T, T, T, T, T } };
__declspec(align(8)) static const mmx_t bT = { 0x0A0A0A0A0A0A0A0AULL }; // sunqueen modify
pxor_r2r( mm6, mm6 ); /* zero, used in psadbw */
movq_m2r( bT, mm5 );
pxor_r2r( mm3, mm3 ); /* score (top field) */
pxor_r2r( mm4, mm4 ); /* score (bottom field) */
for( int y = 0; y < 8; y+=2 )
{
/* top field */
// sunqueen add start
ui_pix_c = *((uint64_t*)p_pix_c);
movq_m2r( ui_pix_c, mm0 );
ui_pix_p = *((uint64_t*)p_pix_p);
movq_m2r( ui_pix_p, mm1 );
// sunqueen add end
#if 0 // sunqueen delete start
movq_m2r( *((uint64_t*)p_pix_c), mm0 );
movq_m2r( *((uint64_t*)p_pix_p), mm1 );
#endif // sunqueen delete end
movq_r2r( mm0, mm2 );
psubusb_r2r( mm1, mm2 );
psubusb_r2r( mm0, mm1 );
pcmpgtb_r2r( mm5, mm2 );
pcmpgtb_r2r( mm5, mm1 );
psadbw_r2r( mm6, mm2 );
psadbw_r2r( mm6, mm1 );
paddd_r2r( mm2, mm1 );
paddd_r2r( mm1, mm3 ); /* add to top field score */
p_pix_c += i_pitch_curr;
p_pix_p += i_pitch_prev;
/* bottom field - handling identical to top field, except... */
// sunqueen add start
ui_pix_c = *((uint64_t*)p_pix_c);
movq_m2r( ui_pix_c, mm0 );
ui_pix_p = *((uint64_t*)p_pix_p);
movq_m2r( ui_pix_p, mm1 );
// sunqueen add end
#if 0 // sunqueen delete start
movq_m2r( *((uint64_t*)p_pix_c), mm0 );
movq_m2r( *((uint64_t*)p_pix_p), mm1 );
#endif // sunqueen delete end
movq_r2r( mm0, mm2 );
psubusb_r2r( mm1, mm2 );
psubusb_r2r( mm0, mm1 );
pcmpgtb_r2r( mm5, mm2 );
pcmpgtb_r2r( mm5, mm1 );
psadbw_r2r( mm6, mm2 );
psadbw_r2r( mm6, mm1 );
paddd_r2r( mm2, mm1 );
paddd_r2r( mm1, mm4 ); /* ...here we add to bottom field score */
p_pix_c += i_pitch_curr;
p_pix_p += i_pitch_prev;
}
movq_r2r( mm3, mm7 ); /* score (total) */
paddd_r2r( mm4, mm7 );
movd_r2m( mm3, i_top_motion );
movd_r2m( mm4, i_bot_motion );
movd_r2m( mm7, i_motion );
/* The loop counts actual score * 255. */
i_top_motion /= 255;
i_bot_motion /= 255;
i_motion /= 255;
emms();
(*pi_top) = ( i_top_motion >= 8 );
(*pi_bot) = ( i_bot_motion >= 8 );
return (i_motion >= 8);
}
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);
}
/**
* Internal helper function for EstimateNumBlocksWithMotion():
* estimates whether there is motion in the given 8x8 block on one plane
* between two images. The block as a whole and its fields are evaluated
* separately, and use different motion thresholds.
*
* This is a low-level function only used by EstimateNumBlocksWithMotion().
* There is no need to call this function manually.
*
* For interpretation of pi_top and pi_bot, it is assumed that the block
* starts on an even-numbered line (belonging to the top field).
*
* The b_mmx parameter avoids the need to call vlc_CPU() separately
* for each block.
*
* @param[in] p_pix_p Base pointer to the block in previous picture
* @param[in] p_pix_c Base pointer to the same block in current picture
* @param i_pitch_prev i_pitch of previous picture
* @param i_pitch_curr i_pitch of current picture
* @param b_mmx (vlc_CPU() & CPU_CAPABILITY_MMXEXT) or false.
* @param[out] pi_top 1 if top field of the block had motion, 0 if no
* @param[out] pi_bot 1 if bottom field of the block had motion, 0 if no
* @return 1 if the block had motion, 0 if no
* @see EstimateNumBlocksWithMotion()
*/
static inline int TestForMotionInBlock( uint8_t *p_pix_p, uint8_t *p_pix_c,
int i_pitch_prev, int i_pitch_curr,
bool b_mmx,
int* pi_top, int* pi_bot )
{
/* Pixel luma/chroma difference threshold to detect motion. */
#define T 10
int32_t i_motion = 0;
int32_t i_top_motion = 0;
int32_t i_bot_motion = 0;
/* See below for the C version to see more quickly what this does. */
#ifdef CAN_COMPILE_MMXEXT
if( b_mmx )
{
static const mmx_t bT = { .ub = { T, T, T, T, T, T, T, T } };
pxor_r2r( mm6, mm6 ); /* zero, used in psadbw */
movq_m2r( bT, mm5 );
pxor_r2r( mm3, mm3 ); /* score (top field) */
pxor_r2r( mm4, mm4 ); /* score (bottom field) */
for( int y = 0; y < 8; y+=2 )
{
/* top field */
movq_m2r( *((uint64_t*)p_pix_c), mm0 );
movq_m2r( *((uint64_t*)p_pix_p), mm1 );
movq_r2r( mm0, mm2 );
psubusb_r2r( mm1, mm2 );
psubusb_r2r( mm0, mm1 );
pcmpgtb_r2r( mm5, mm2 );
pcmpgtb_r2r( mm5, mm1 );
psadbw_r2r( mm6, mm2 );
psadbw_r2r( mm6, mm1 );
paddd_r2r( mm2, mm1 );
paddd_r2r( mm1, mm3 ); /* add to top field score */
p_pix_c += i_pitch_curr;
p_pix_p += i_pitch_prev;
/* bottom field - handling identical to top field, except... */
movq_m2r( *((uint64_t*)p_pix_c), mm0 );
movq_m2r( *((uint64_t*)p_pix_p), mm1 );
movq_r2r( mm0, mm2 );
psubusb_r2r( mm1, mm2 );
psubusb_r2r( mm0, mm1 );
pcmpgtb_r2r( mm5, mm2 );
pcmpgtb_r2r( mm5, mm1 );
psadbw_r2r( mm6, mm2 );
psadbw_r2r( mm6, mm1 );
paddd_r2r( mm2, mm1 );
paddd_r2r( mm1, mm4 ); /* ...here we add to bottom field score */
p_pix_c += i_pitch_curr;
p_pix_p += i_pitch_prev;
}
movq_r2r( mm3, mm7 ); /* score (total) */
paddd_r2r( mm4, mm7 );
movd_r2m( mm3, i_top_motion );
movd_r2m( mm4, i_bot_motion );
movd_r2m( mm7, i_motion );
/* The loop counts actual score * 255. */
i_top_motion /= 255;
i_bot_motion /= 255;
i_motion /= 255;
emms();
}
else
#endif
{
for( int y = 0; y < 8; ++y )