static inline void mmx_unpack_32rgb (uint8_t * image, const int cpu)
{
/*
* convert RGB plane to RGB packed format,
* mm0 -> B, mm1 -> R, mm2 -> G, mm3 -> 0,
* mm4 -> GB, mm5 -> AR pixel 4-7,
* mm6 -> GB, mm7 -> AR pixel 0-3
*/
pxor_r2r (mm3, mm3);
movq_r2r (mm0, mm6);
movq_r2r (mm1, mm7);
movq_r2r (mm0, mm4);
movq_r2r (mm1, mm5);
punpcklbw_r2r (mm2, mm6);
punpcklbw_r2r (mm3, mm7);
punpcklwd_r2r (mm7, mm6);
movntq (mm6, *image);
movq_r2r (mm0, mm6);
punpcklbw_r2r (mm2, mm6);
punpckhwd_r2r (mm7, mm6);
movntq (mm6, *(image+8));
punpckhbw_r2r (mm2, mm4);
punpckhbw_r2r (mm3, mm5);
punpcklwd_r2r (mm5, mm4);
movntq (mm4, *(image+16));
movq_r2r (mm0, mm4);
punpckhbw_r2r (mm2, mm4);
punpckhwd_r2r (mm5, mm4);
movntq (mm4, *(image+24));
}
static inline void mmx_average_2_U8 (uint8_t * dest,
uint8_t * src1, uint8_t * src2)
{
/* *dest = (*src1 + *src2 + 1)/ 2; */
movq_m2r (*src1, mm1); // load 8 src1 bytes
movq_r2r (mm1, mm2); // copy 8 src1 bytes
movq_m2r (*src2, mm3); // load 8 src2 bytes
movq_r2r (mm3, mm4); // copy 8 src2 bytes
punpcklbw_r2r (mm0, mm1); // unpack low src1 bytes
punpckhbw_r2r (mm0, mm2); // unpack high src1 bytes
punpcklbw_r2r (mm0, mm3); // unpack low src2 bytes
punpckhbw_r2r (mm0, mm4); // unpack high src2 bytes
paddw_r2r (mm3, mm1); // add lows to mm1
paddw_m2r (round1, mm1);
psraw_i2r (1, mm1); // /2
paddw_r2r (mm4, mm2); // add highs to mm2
paddw_m2r (round1, mm2);
psraw_i2r (1, mm2); // /2
packuswb_r2r (mm2, mm1); // pack (w/ saturation)
movq_r2m (mm1, *dest); // store result in dest
}
static inline void mmx_unpack_16rgb (uint8_t * image, const int cpu)
{
static mmx_t mmx_bluemask = {0xf8f8f8f8f8f8f8f8LL};
static mmx_t mmx_greenmask = {0xfcfcfcfcfcfcfcfcLL};
static mmx_t mmx_redmask = {0xf8f8f8f8f8f8f8f8LL};
/*
* convert RGB plane to RGB 16 bits
* mm0 -> B, mm1 -> R, mm2 -> G
* mm4 -> GB, mm5 -> AR pixel 4-7
* mm6 -> GB, mm7 -> AR pixel 0-3
*/
pand_m2r (mmx_bluemask, mm0); /* mm0 = b7b6b5b4b3______ */
pand_m2r (mmx_greenmask, mm2); /* mm2 = g7g6g5g4g3g2____ */
pand_m2r (mmx_redmask, mm1); /* mm1 = r7r6r5r4r3______ */
psrlq_i2r (3, mm0); /* mm0 = ______b7b6b5b4b3 */
pxor_r2r (mm4, mm4); /* mm4 = 0 */
movq_r2r (mm0, mm5); /* mm5 = ______b7b6b5b4b3 */
movq_r2r (mm2, mm7); /* mm7 = g7g6g5g4g3g2____ */
punpcklbw_r2r (mm4, mm2);
punpcklbw_r2r (mm1, mm0);
psllq_i2r (3, mm2);
por_r2r (mm2, mm0);
movntq (mm0, *image);
punpckhbw_r2r (mm4, mm7);
punpckhbw_r2r (mm1, mm5);
psllq_i2r (3, mm7);
por_r2r (mm7, mm5);
movntq (mm5, *(image+8));
}
static inline void mmx_average_4_U8 (uint8_t * dest, const uint8_t * src1,
const uint8_t * src2,
const uint8_t * src3,
const uint8_t * src4)
{
/* *dest = (*src1 + *src2 + *src3 + *src4 + 2)/ 4; */
movq_m2r (*src1, mm1); /* load 8 src1 bytes */
movq_r2r (mm1, mm2); /* copy 8 src1 bytes */
punpcklbw_r2r (mm0, mm1); /* unpack low src1 bytes */
punpckhbw_r2r (mm0, mm2); /* unpack high src1 bytes */
movq_m2r (*src2, mm3); /* load 8 src2 bytes */
movq_r2r (mm3, mm4); /* copy 8 src2 bytes */
punpcklbw_r2r (mm0, mm3); /* unpack low src2 bytes */
punpckhbw_r2r (mm0, mm4); /* unpack high src2 bytes */
paddw_r2r (mm3, mm1); /* add lows */
paddw_r2r (mm4, mm2); /* add highs */
/* now have partials in mm1 and mm2 */
movq_m2r (*src3, mm3); /* load 8 src3 bytes */
movq_r2r (mm3, mm4); /* copy 8 src3 bytes */
punpcklbw_r2r (mm0, mm3); /* unpack low src3 bytes */
punpckhbw_r2r (mm0, mm4); /* unpack high src3 bytes */
paddw_r2r (mm3, mm1); /* add lows */
paddw_r2r (mm4, mm2); /* add highs */
movq_m2r (*src4, mm5); /* load 8 src4 bytes */
movq_r2r (mm5, mm6); /* copy 8 src4 bytes */
punpcklbw_r2r (mm0, mm5); /* unpack low src4 bytes */
punpckhbw_r2r (mm0, mm6); /* unpack high src4 bytes */
paddw_r2r (mm5, mm1); /* add lows */
paddw_r2r (mm6, mm2); /* add highs */
/* now have subtotal in mm1 and mm2 */
paddw_m2r (round4, mm1);
psraw_i2r (2, mm1); /* /4 */
paddw_m2r (round4, mm2);
psraw_i2r (2, mm2); /* /4 */
packuswb_r2r (mm2, mm1); /* pack (w/ saturation) */
movq_r2m (mm1, *dest); /* store result in dest */
}
static inline void mmx_average_4_U8 (uint8_t * dest,
uint8_t * src1, uint8_t * src2,
uint8_t * src3, uint8_t * src4)
{
/* *dest = (*src1 + *src2 + *src3 + *src4 + 2)/ 4; */
movq_m2r (*src1, mm1); // load 8 src1 bytes
movq_r2r (mm1, mm2); // copy 8 src1 bytes
punpcklbw_r2r (mm0, mm1); // unpack low src1 bytes
punpckhbw_r2r (mm0, mm2); // unpack high src1 bytes
movq_m2r (*src2, mm3); // load 8 src2 bytes
movq_r2r (mm3, mm4); // copy 8 src2 bytes
punpcklbw_r2r (mm0, mm3); // unpack low src2 bytes
punpckhbw_r2r (mm0, mm4); // unpack high src2 bytes
paddw_r2r (mm3, mm1); // add lows
paddw_r2r (mm4, mm2); // add highs
/* now have partials in mm1 and mm2 */
movq_m2r (*src3, mm3); // load 8 src3 bytes
movq_r2r (mm3, mm4); // copy 8 src3 bytes
punpcklbw_r2r (mm0, mm3); // unpack low src3 bytes
punpckhbw_r2r (mm0, mm4); // unpack high src3 bytes
paddw_r2r (mm3, mm1); // add lows
paddw_r2r (mm4, mm2); // add highs
movq_m2r (*src4, mm5); // load 8 src4 bytes
movq_r2r (mm5, mm6); // copy 8 src4 bytes
punpcklbw_r2r (mm0, mm5); // unpack low src4 bytes
punpckhbw_r2r (mm0, mm6); // unpack high src4 bytes
paddw_r2r (mm5, mm1); // add lows
paddw_r2r (mm6, mm2); // add highs
/* now have subtotal in mm1 and mm2 */
paddw_m2r (round4, mm1);
psraw_i2r (2, mm1); // /4
paddw_m2r (round4, mm2);
psraw_i2r (2, mm2); // /4
packuswb_r2r (mm2, mm1); // pack (w/ saturation)
movq_r2m (mm1, *dest); // store result in dest
}
static void frame_i2f_sse(u_char *src,float *dst,int l)
{
int i;
pxor_r2r(mm7,mm7);
for( i=0; i<l; i+=8 ) {
movq_m2r(*src,mm0);
movq_r2r(mm0, mm2);
punpcklbw_r2r(mm7, mm0);
punpckhbw_r2r(mm7, mm2);
movq_r2r(mm0, mm1);
movq_r2r(mm2, mm3);
punpcklwd_r2r(mm7, mm0);
punpckhwd_r2r(mm7, mm1);
punpcklwd_r2r(mm7, mm2);
punpckhwd_r2r(mm7, mm3);
cvtpi2ps_r2r(mm0,xmm0);
cvtpi2ps_r2r(mm1,xmm1);
cvtpi2ps_r2r(mm2,xmm2);
cvtpi2ps_r2r(mm3,xmm3);
movlps_r2m(xmm0,dst[0]);
movlps_r2m(xmm1,dst[2]);
movlps_r2m(xmm2,dst[4]);
movlps_r2m(xmm3,dst[6]);
src+=8;
dst+=8;
}
emms();
}
int main(int ac, char **av)
{
int i, j, k, n;
unsigned char dat0[8] = { 0x01, 0xf2, 0x03, 0x04, 0x05, 0x06, 0xf7, 0x08 };
long long *datp = (long long *)&dat0;
int16_t dat1[8] = { 0x10, 0x20, -0x130, -0x140, 0x50, -0x160, -0x170, 0x80 };
volatile uint8_t *rfp = dat0;
volatile int16_t *bp = dat1;
unsigned char ans1[8], ans2[8];
n = 0;
for( i=-32768; i<32768; ++i ) {
j = 0;
while( j < 256 ) {
for( k=0; k<8; ++k ) {
dat0[k] = i;
dat1[k] = j++;
}
movq_m2r(m_(&rfp[0]),mm1); /* rfp[0..7] */
pxor_r2r(mm3,mm3);
pxor_r2r(mm4,mm4);
movq_m2r(m_(&bp[0]),mm5); /* bp[0..3] */
movq_r2r(mm1,mm2);
movq_m2r(m_(&bp[4]),mm6); /* bp[4..7] */
punpcklbw_r2r(mm3,mm1); /* rfp[0,2,4,6] */
punpckhbw_r2r(mm3,mm2); /* rfp[1,3,5,7] */
paddsw_r2r(mm5,mm1); /* bp[0..3] */
paddsw_r2r(mm6,mm2); /* bp[4..7] */
pcmpgtw_r2r(mm1,mm3);
pcmpgtw_r2r(mm2,mm4);
pandn_r2r(mm1,mm3);
pandn_r2r(mm2,mm4);
packuswb_r2r(mm4,mm3);
movq_r2m(mm3,m_(&ans1[0]));
emms();
ans2[0] = clip(bp[0] + rfp[0]);
ans2[1] = clip(bp[1] + rfp[1]);
ans2[2] = clip(bp[2] + rfp[2]);
ans2[3] = clip(bp[3] + rfp[3]);
ans2[4] = clip(bp[4] + rfp[4]);
ans2[5] = clip(bp[5] + rfp[5]);
ans2[6] = clip(bp[6] + rfp[6]);
ans2[7] = clip(bp[7] + rfp[7]);
if( *(uint64_t *)&ans1[0] != *(uint64_t *)&ans2[0] )
{
printf(" i=%5d %02x %02x %02x %02x %02x %02x %02x %02x\n", i,
ans1[0], ans1[1], ans1[2], ans1[3], ans1[4], ans1[5], ans1[6], ans1[7]);
printf(" j=%5d %02x %02x %02x %02x %02x %02x %02x %02x\n", j,
ans2[0], ans2[1], ans2[2], ans2[3], ans2[4], ans2[5], ans2[6], ans2[7]);
// exit(0);
}
n += 8;
}
}
printf("n=%d\n",n);
return 0;
}
VLC_MMX
static int CalculateInterlaceScoreMMX( const picture_t* p_pic_top,
const picture_t* p_pic_bot )
{
assert( p_pic_top->i_planes == p_pic_bot->i_planes );
/* Amount of bits must be known for MMX, thus int32_t.
Doesn't hurt the C implementation. */
int32_t i_score_mmx = 0; /* this must be divided by 255 when finished */
int32_t i_score_c = 0; /* this counts as-is (used for non-MMX parts) */
pxor_r2r( mm7, mm7 ); /* we will keep score in mm7 */
for( int i_plane = 0 ; i_plane < p_pic_top->i_planes ; ++i_plane )
{
/* Sanity check */
if( p_pic_top->p[i_plane].i_visible_lines !=
p_pic_bot->p[i_plane].i_visible_lines )
return -1;
const int i_lasty = p_pic_top->p[i_plane].i_visible_lines-1;
const int w = FFMIN( p_pic_top->p[i_plane].i_visible_pitch,
p_pic_bot->p[i_plane].i_visible_pitch );
const int wm8 = w % 8; /* remainder */
const int w8 = w - wm8; /* part of width that is divisible by 8 */
/* Current line / neighbouring lines picture pointers */
const picture_t *cur = p_pic_bot;
const picture_t *ngh = p_pic_top;
int wc = cur->p[i_plane].i_pitch;
int wn = ngh->p[i_plane].i_pitch;
/* Transcode 1.1.5 only checks every other line. Checking every line
works better for anime, which may contain horizontal,
one pixel thick cartoon outlines.
*/
for( int y = 1; y < i_lasty; ++y )
{
uint8_t *p_c = &cur->p[i_plane].p_pixels[y*wc]; /* this line */
uint8_t *p_p = &ngh->p[i_plane].p_pixels[(y-1)*wn]; /* prev line */
uint8_t *p_n = &ngh->p[i_plane].p_pixels[(y+1)*wn]; /* next line */
int x = 0;
/* Easy-to-read C version further below.
Assumptions: 0 < T < 127
# of pixels < (2^32)/255
Note: calculates score * 255
*/
static alignas (8) const mmx_t b0 = {
.uq = 0x0000000000000000ULL };
static alignas (8) const mmx_t b128 = {
.uq = 0x8080808080808080ULL };
static alignas (8) const mmx_t bT = {
.ub = { T, T, T, T, T, T, T, T } };
for( ; x < w8; x += 8 )
{
movq_m2r( *((int64_t*)p_c), mm0 );
movq_m2r( *((int64_t*)p_p), mm1 );
movq_m2r( *((int64_t*)p_n), mm2 );
psubb_m2r( b128, mm0 );
psubb_m2r( b128, mm1 );
psubb_m2r( b128, mm2 );
psubsb_r2r( mm0, mm1 );
psubsb_r2r( mm0, mm2 );
pxor_r2r( mm3, mm3 );
pxor_r2r( mm4, mm4 );
pxor_r2r( mm5, mm5 );
pxor_r2r( mm6, mm6 );
punpcklbw_r2r( mm1, mm3 );
punpcklbw_r2r( mm2, mm4 );
punpckhbw_r2r( mm1, mm5 );
punpckhbw_r2r( mm2, mm6 );
pmulhw_r2r( mm3, mm4 );
pmulhw_r2r( mm5, mm6 );
packsswb_r2r(mm4, mm6);
pcmpgtb_m2r( bT, mm6 );
psadbw_m2r( b0, mm6 );
paddd_r2r( mm6, mm7 );
p_c += 8;
p_p += 8;
p_n += 8;
}
for( ; x < w; ++x )
{
/* Worst case: need 17 bits for "comb". */
int_fast32_t C = *p_c;
int_fast32_t P = *p_p;
int_fast32_t N = *p_n;
/* Comments in Transcode's filter_ivtc.c attribute this
combing metric to Gunnar Thalin.
The idea is that if the picture is interlaced, both
expressions will have the same sign, and this comes
up positive. The value T = 100 has been chosen such
that a pixel difference of 10 (on average) will
trigger the detector.
*/
int_fast32_t comb = (P - C) * (N - C);
if( comb > T )
++i_score_c;
++p_c;
++p_p;
++p_n;
}
/* Now the other field - swap current and neighbour pictures */
const picture_t *tmp = cur;
cur = ngh;
ngh = tmp;
int tmp_pitch = wc;
wc = wn;
wn = tmp_pitch;
}
}
movd_r2m( mm7, i_score_mmx );
emms();
return i_score_mmx/255 + i_score_c;
}
#endif
/* See header for function doc. */
int CalculateInterlaceScore( const picture_t* p_pic_top,
const picture_t* p_pic_bot )
{
/*
We use the comb metric from the IVTC filter of Transcode 1.1.5.
This was found to work better for the particular purpose of IVTC
than RenderX()'s comb metric.
Note that we *must not* subsample at all in order to catch interlacing
in telecined frames with localized motion (e.g. anime with characters
talking, where only mouths move and everything else stays still.)
*/
assert( p_pic_top != NULL );
assert( p_pic_bot != NULL );
if( p_pic_top->i_planes != p_pic_bot->i_planes )
return -1;
#ifdef CAN_COMPILE_MMXEXT
if (vlc_CPU_MMXEXT())
return CalculateInterlaceScoreMMX( p_pic_top, p_pic_bot );
#endif
int32_t i_score = 0;
for( int i_plane = 0 ; i_plane < p_pic_top->i_planes ; ++i_plane )
{
/* Sanity check */
if( p_pic_top->p[i_plane].i_visible_lines !=
p_pic_bot->p[i_plane].i_visible_lines )
return -1;
const int i_lasty = p_pic_top->p[i_plane].i_visible_lines-1;
const int w = FFMIN( p_pic_top->p[i_plane].i_visible_pitch,
p_pic_bot->p[i_plane].i_visible_pitch );
/* Current line / neighbouring lines picture pointers */
const picture_t *cur = p_pic_bot;
const picture_t *ngh = p_pic_top;
int wc = cur->p[i_plane].i_pitch;
int wn = ngh->p[i_plane].i_pitch;
/* Transcode 1.1.5 only checks every other line. Checking every line
works better for anime, which may contain horizontal,
one pixel thick cartoon outlines.
*/
for( int y = 1; y < i_lasty; ++y )
{
uint8_t *p_c = &cur->p[i_plane].p_pixels[y*wc]; /* this line */
uint8_t *p_p = &ngh->p[i_plane].p_pixels[(y-1)*wn]; /* prev line */
uint8_t *p_n = &ngh->p[i_plane].p_pixels[(y+1)*wn]; /* next line */
for( int x = 0; x < w; ++x )
{
/* Worst case: need 17 bits for "comb". */
int_fast32_t C = *p_c;
int_fast32_t P = *p_p;
int_fast32_t N = *p_n;
/* Comments in Transcode's filter_ivtc.c attribute this
combing metric to Gunnar Thalin.
The idea is that if the picture is interlaced, both
expressions will have the same sign, and this comes
up positive. The value T = 100 has been chosen such
that a pixel difference of 10 (on average) will
trigger the detector.
*/
int_fast32_t comb = (P - C) * (N - C);
if( comb > T )
++i_score;
++p_c;
++p_p;
++p_n;
}
/* Now the other field - swap current and neighbour pictures */
const picture_t *tmp = cur;
cur = ngh;
ngh = tmp;
int tmp_pitch = wc;
wc = wn;
wn = tmp_pitch;
}
}
return i_score;
}
static inline void mmx_interp_average_4_U8 (uint8_t * dest,
const uint8_t * src1,
const uint8_t * src2,
const uint8_t * src3,
const uint8_t * src4)
{
/* *dest = (*dest + (*src1 + *src2 + *src3 + *src4 + 2)/ 4 + 1)/ 2; */
movq_m2r (*src1, mm1); /* load 8 src1 bytes */
movq_r2r (mm1, mm2); /* copy 8 src1 bytes */
punpcklbw_r2r (mm0, mm1); /* unpack low src1 bytes */
punpckhbw_r2r (mm0, mm2); /* unpack high src1 bytes */
movq_m2r (*src2, mm3); /* load 8 src2 bytes */
movq_r2r (mm3, mm4); /* copy 8 src2 bytes */
punpcklbw_r2r (mm0, mm3); /* unpack low src2 bytes */
punpckhbw_r2r (mm0, mm4); /* unpack high src2 bytes */
paddw_r2r (mm3, mm1); /* add lows */
paddw_r2r (mm4, mm2); /* add highs */
/* now have partials in mm1 and mm2 */
movq_m2r (*src3, mm3); /* load 8 src3 bytes */
movq_r2r (mm3, mm4); /* copy 8 src3 bytes */
punpcklbw_r2r (mm0, mm3); /* unpack low src3 bytes */
punpckhbw_r2r (mm0, mm4); /* unpack high src3 bytes */
paddw_r2r (mm3, mm1); /* add lows */
paddw_r2r (mm4, mm2); /* add highs */
movq_m2r (*src4, mm5); /* load 8 src4 bytes */
movq_r2r (mm5, mm6); /* copy 8 src4 bytes */
punpcklbw_r2r (mm0, mm5); /* unpack low src4 bytes */
punpckhbw_r2r (mm0, mm6); /* unpack high src4 bytes */
paddw_r2r (mm5, mm1); /* add lows */
paddw_r2r (mm6, mm2); /* add highs */
paddw_m2r (round4, mm1);
psraw_i2r (2, mm1); /* /4 */
paddw_m2r (round4, mm2);
psraw_i2r (2, mm2); /* /4 */
/* now have subtotal/4 in mm1 and mm2 */
movq_m2r (*dest, mm3); /* load 8 dest bytes */
movq_r2r (mm3, mm4); /* copy 8 dest bytes */
packuswb_r2r (mm2, mm1); /* pack (w/ saturation) */
movq_r2r (mm1,mm2); /* copy subresult */
pxor_r2r (mm1, mm3); /* xor srcavg and dest */
pand_m2r (mask1, mm3); /* mask lower bits */
psrlq_i2r (1, mm3); /* /2 */
por_r2r (mm2, mm4); /* or srcavg and dest */
psubb_r2r (mm3, mm4); /* subtract subresults */
movq_r2m (mm4, *dest); /* store result in dest */
}
static __inline__ void
sum_sumsq_8bytes( uint8_t *cur_lum_mb,
uint8_t *pred_lum_mb,
mmx_t *sumtop_accs,
mmx_t *sumbot_accs,
mmx_t *sumsqtop_accs,
mmx_t *sumsqbot_accs,
mmx_t *sumxprod_accs
)
{
pxor_r2r(mm0,mm0);
/* Load pixels from top field into mm1.w,mm2.w
*/
movq_m2r( *((mmx_t*)cur_lum_mb), mm1 );
movq_m2r( *((mmx_t*)pred_lum_mb), mm2 );
/* mm3 := mm1 mm4 := mm2
mm1.w[0..3] := mm1.b[0..3]-mm2.b[0..3]
*/
movq_r2r( mm1, mm3 );
punpcklbw_r2r( mm0, mm1 );
movq_r2r( mm2, mm4 );
punpcklbw_r2r( mm0, mm2 );
psubw_r2r( mm2, mm1 );
/* mm3.w[0..3] := mm3.b[4..7]-mm4.b[4..7]
*/
punpckhbw_r2r( mm0, mm3 );
punpckhbw_r2r( mm0, mm4 );
psubw_r2r( mm4, mm3 );
/* sumtop_accs->w[0..3] += mm1.w[0..3];
sumtop_accs->w[0..3] += mm3.w[0..3];
mm6 = mm1; mm7 = mm3;
*/
movq_m2r( *sumtop_accs, mm5 );
paddw_r2r( mm1, mm5 );
paddw_r2r( mm3, mm5 );
movq_r2r( mm1, mm6 );
movq_r2r( mm3, mm7 );
movq_r2m( mm5, *sumtop_accs );
/*
*sumsq_top_acc += mm1.w[0..3] * mm1.w[0..3];
*sumsq_top_acc += mm3.w[0..3] * mm3.w[0..3];
*/
pmaddwd_r2r( mm1, mm1 );
movq_m2r( *sumsqtop_accs, mm5 );
pmaddwd_r2r( mm3, mm3 );
paddd_r2r( mm1, mm5 );
paddd_r2r( mm3, mm5 );
movq_r2m( mm5, *sumsqtop_accs );
/* Load pixels from bot field into mm1.w,mm2.w
*/
movq_m2r( *((mmx_t*)(cur_lum_mb+opt->phy_width)), mm1 );
movq_m2r( *((mmx_t*)(pred_lum_mb+opt->phy_width)), mm2 );
/* mm2 := mm1 mm4 := mm2
mm1.w[0..3] := mm1.b[0..3]-mm2.b[0..3]
*/
movq_r2r( mm1, mm3 );
punpcklbw_r2r( mm0, mm1 );
movq_r2r( mm2, mm4 );
punpcklbw_r2r( mm0, mm2 );
psubw_r2r( mm2, mm1 );
/* mm3.w[0..3] := mm3.b[4..7]-mm4.b[4..7]
*/
punpckhbw_r2r( mm0, mm3 );
punpckhbw_r2r( mm0, mm4 );
psubw_r2r( mm4, mm3 );
/*
sumbot_accs->w[0..3] += mm1.w[0..3];
sumbot_accs->w[0..3] += mm3.w[0..3];
mm2 := mm1; mm4 := mm3;
*/
movq_m2r( *sumbot_accs, mm5 );
paddw_r2r( mm1, mm5 );
movq_r2r( mm1, mm2 );
paddw_r2r( mm3, mm5 );
movq_r2r( mm3, mm4 );
movq_r2m( mm5, *sumbot_accs );
/*
*sumsqbot_acc += mm1.w[0..3] * mm1.w[0..3];
*sumsqbot_acc += mm3.w[0..3] * mm3.w[0..3];
*/
pmaddwd_r2r( mm1, mm1 );
movq_m2r( *sumsqbot_accs, mm5 );
pmaddwd_r2r( mm3, mm3 );
paddd_r2r( mm1, mm5 );
paddd_r2r( mm3, mm5 );
movq_r2m( mm5, *sumsqbot_accs );
/* Accumulate cross-product
*sum_xprod_acc += mm1.w[0..3] * mm6[0..3];
*sum_xprod_acc += mm3.w[0..3] * mm7[0..3];
*/
movq_m2r( *sumxprod_accs, mm5 );
pmaddwd_r2r( mm6, mm2);
pmaddwd_r2r( mm7, mm4);
paddd_r2r( mm2, mm5 );
paddd_r2r( mm4, mm5 );
movq_r2m( mm5, *sumxprod_accs );
emms();
}
static void
_evas_yv12torgb_sse(unsigned char **yuv, unsigned char *rgb, int w, int h)
{
#ifdef BUILD_MMX
int xx, yy;
register unsigned char *yp1, *up, *vp;
unsigned char *dp1;
/* destination pointers */
dp1 = rgb;
for (yy = 0; yy < h; yy++)
{
/* plane pointers */
yp1 = yuv[yy];
up = yuv[h + (yy / 2)];
vp = yuv[h + (h / 2) + (yy / 2)];
for (xx = 0; xx < (w - 7); xx += 8)
{
movd_m2r(*up, mm3);
movd_m2r(*vp, mm2);
movq_m2r(*yp1, mm0);
pxor_r2r(mm7, mm7);
punpcklbw_r2r(mm7, mm2);
punpcklbw_r2r(mm7, mm3);
movq_r2r(mm0, mm1);
psrlw_i2r(8, mm0);
psllw_i2r(8, mm1);
psrlw_i2r(8, mm1);
movq_m2r(CONST_16, mm4);
psubsw_r2r(mm4, mm0);
psubsw_r2r(mm4, mm1);
movq_m2r(CONST_128, mm5);
psubsw_r2r(mm5, mm2);
psubsw_r2r(mm5, mm3);
movq_m2r(CONST_YMUL, mm4);
pmullw_r2r(mm4, mm0);
pmullw_r2r(mm4, mm1);
movq_m2r(CONST_CRVCRV, mm7);
pmullw_r2r(mm3, mm7);
movq_m2r(CONST_CBUCBU, mm6);
pmullw_r2r(mm2, mm6);
movq_m2r(CONST_CGUCGU, mm5);
pmullw_r2r(mm2, mm5);
movq_m2r(CONST_CGVCGV, mm4);
pmullw_r2r(mm3, mm4);
movq_r2r(mm0, mm2);
paddsw_r2r(mm7, mm2);
paddsw_r2r(mm1, mm7);
psraw_i2r(RES, mm2);
psraw_i2r(RES, mm7);
packuswb_r2r(mm7, mm2);
pxor_r2r(mm7, mm7);
movq_r2r(mm2, mm3);
punpckhbw_r2r(mm7, mm2);
punpcklbw_r2r(mm3, mm7);
por_r2r(mm7, mm2);
movq_r2r(mm0, mm3);
psubsw_r2r(mm5, mm3);
psubsw_r2r(mm4, mm3);
paddsw_m2r(CONST_32, mm3);
movq_r2r(mm1, mm7);
psubsw_r2r(mm5, mm7);
psubsw_r2r(mm4, mm7);
paddsw_m2r(CONST_32, mm7);
psraw_i2r(RES, mm3);
psraw_i2r(RES, mm7);
packuswb_r2r(mm7, mm3);
pxor_r2r(mm7, mm7);
movq_r2r(mm3, mm4);
punpckhbw_r2r(mm7, mm3);
punpcklbw_r2r(mm4, mm7);
por_r2r(mm7, mm3);
movq_m2r(CONST_32, mm4);
paddsw_r2r(mm6, mm0);
paddsw_r2r(mm6, mm1);
paddsw_r2r(mm4, mm0);
paddsw_r2r(mm4, mm1);
psraw_i2r(RES, mm0);
psraw_i2r(RES, mm1);
packuswb_r2r(mm1, mm0);
pxor_r2r(mm7, mm7);
movq_r2r(mm0, mm5);
punpckhbw_r2r(mm7, mm0);
punpcklbw_r2r(mm5, mm7);
por_r2r(mm7, mm0);
movq_m2r(CONST_FF, mm1);
movq_r2r(mm0, mm5);
movq_r2r(mm3, mm6);
movq_r2r(mm2, mm7);
punpckhbw_r2r(mm3, mm2);
punpcklbw_r2r(mm6, mm7);
punpckhbw_r2r(mm1, mm0);
punpcklbw_r2r(mm1, mm5);
movq_r2r(mm7, mm1);
punpckhwd_r2r(mm5, mm7);
punpcklwd_r2r(mm5, mm1);
movq_r2r(mm2, mm4);
punpckhwd_r2r(mm0, mm2);
punpcklwd_r2r(mm0, mm4);
movntq_r2m(mm1, *(dp1));
movntq_r2m(mm7, *(dp1 + 8));
movntq_r2m(mm4, *(dp1 + 16));
movntq_r2m(mm2, *(dp1 + 24));
yp1 += 8;
up += 4;
vp += 4;
dp1 += 8 * 4;
}
/* cleanup pixles that arent a multiple of 8 pixels wide */
if (xx < w)
{
int y, u, v, r, g, b;
for (; xx < w; xx += 2)
{
u = (*up++) - 128;
v = (*vp++) - 128;
y = RZ(YMUL) * ((*yp1++) - 16);
r = LUT_CLIP((y + (_crv * v)) >> RES);
g = LUT_CLIP((y - (_cgu * u) - (_cgv * v) + RZ(OFF)) >> RES);
b = LUT_CLIP((y + (_cbu * u) + RZ(OFF)) >> RES);
*((DATA32 *) dp1) = 0xff000000 + RGB_JOIN(r,g,b);
dp1 += 4;
y = RZ(YMUL) * ((*yp1++) - 16);
r = LUT_CLIP((y + (_crv * v)) >> RES);
g = LUT_CLIP((y - (_cgu * u) - (_cgv * v) + RZ(OFF)) >> RES);
b = LUT_CLIP((y + (_cbu * u) + RZ(OFF)) >> RES);
*((DATA32 *) dp1) = 0xff000000 + RGB_JOIN(r,g,b);
dp1 += 4;
}
}
}
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 CalculateInterlaceScoreMMX( const picture_t* p_pic_top,
const picture_t* p_pic_bot )
{
assert( p_pic_top->i_planes == p_pic_bot->i_planes );
/* Amount of bits must be known for MMX, thus int32_t.
Doesn't hurt the C implementation. */
int32_t i_score_mmx = 0; /* this must be divided by 255 when finished */
int32_t i_score_c = 0; /* this counts as-is (used for non-MMX parts) */
pxor_r2r( mm7, mm7 ); /* we will keep score in mm7 */
for( int i_plane = 0 ; i_plane < p_pic_top->i_planes ; ++i_plane )
{
/* Sanity check */
if( p_pic_top->p[i_plane].i_visible_lines !=
p_pic_bot->p[i_plane].i_visible_lines )
return -1;
const int i_lasty = p_pic_top->p[i_plane].i_visible_lines-1;
const int w = FFMIN( p_pic_top->p[i_plane].i_visible_pitch,
p_pic_bot->p[i_plane].i_visible_pitch );
const int wm8 = w % 8; /* remainder */
const int w8 = w - wm8; /* part of width that is divisible by 8 */
/* Current line / neighbouring lines picture pointers */
const picture_t *cur = p_pic_bot;
const picture_t *ngh = p_pic_top;
int wc = cur->p[i_plane].i_pitch;
int wn = ngh->p[i_plane].i_pitch;
/* Transcode 1.1.5 only checks every other line. Checking every line
works better for anime, which may contain horizontal,
one pixel thick cartoon outlines.
*/
for( int y = 1; y < i_lasty; ++y )
{
uint8_t *p_c = &cur->p[i_plane].p_pixels[y*wc]; /* this line */
uint8_t *p_p = &ngh->p[i_plane].p_pixels[(y-1)*wn]; /* prev line */
uint8_t *p_n = &ngh->p[i_plane].p_pixels[(y+1)*wn]; /* next line */
int64_t i_p_c, i_p_p, i_p_n; // sunqueen add
int x = 0;
/* Easy-to-read C version further below.
Assumptions: 0 < T < 127
# of pixels < (2^32)/255
Note: calculates score * 255
*/
// static const mmx_t b0 = { .uq = 0x0000000000000000ULL };
__declspec(align(8)) static const mmx_t b0 = { /*.uq =*/ 0x0000000000000000ULL }; // sunqueen modify
// static const mmx_t b128 = { .uq = 0x8080808080808080ULL };
__declspec(align(8)) static const mmx_t b128 = { /*.uq =*/ 0x8080808080808080ULL }; // sunqueen modify
// static const mmx_t bT = { .ub = { T, T, T, T, T, T, T, T } };
__declspec(align(8)) static const mmx_t bT = { 0x6464646464646464ULL }; // sunqueen modify
for( ; x < w8; x += 8 )
{
// sunqueen add start
i_p_c = *((int64_t*)p_c);
movq_m2r( i_p_c, mm0 );
i_p_p = *((int64_t*)p_p);
movq_m2r( i_p_p, mm1 );
i_p_n = *((int64_t*)p_n);
movq_m2r( i_p_n, mm2 );
// sunqueen add end
#if 0 // sunqueen delete start
movq_m2r( *((int64_t*)p_c), mm0 );
movq_m2r( *((int64_t*)p_p), mm1 );
movq_m2r( *((int64_t*)p_n), mm2 );
#endif // sunqueen delete end
psubb_m2r( b128, mm0 );
psubb_m2r( b128, mm1 );
psubb_m2r( b128, mm2 );
psubsb_r2r( mm0, mm1 );
psubsb_r2r( mm0, mm2 );
pxor_r2r( mm3, mm3 );
pxor_r2r( mm4, mm4 );
pxor_r2r( mm5, mm5 );
pxor_r2r( mm6, mm6 );
punpcklbw_r2r( mm1, mm3 );
punpcklbw_r2r( mm2, mm4 );
punpckhbw_r2r( mm1, mm5 );
punpckhbw_r2r( mm2, mm6 );
pmulhw_r2r( mm3, mm4 );
pmulhw_r2r( mm5, mm6 );
packsswb_r2r(mm4, mm6);
pcmpgtb_m2r( bT, mm6 );
psadbw_m2r( b0, mm6 );
paddd_r2r( mm6, mm7 );
p_c += 8;
p_p += 8;
p_n += 8;
}
for( ; x < w; ++x )
{
/* Worst case: need 17 bits for "comb". */
int_fast32_t C = *p_c;
int_fast32_t P = *p_p;
int_fast32_t N = *p_n;
/* Comments in Transcode's filter_ivtc.c attribute this
combing metric to Gunnar Thalin.
The idea is that if the picture is interlaced, both
expressions will have the same sign, and this comes
up positive. The value T = 100 has been chosen such
that a pixel difference of 10 (on average) will
trigger the detector.
*/
int_fast32_t comb = (P - C) * (N - C);
if( comb > T )
++i_score_c;
++p_c;
++p_p;
++p_n;
}
/* Now the other field - swap current and neighbour pictures */
const picture_t *tmp = cur;
cur = ngh;
ngh = tmp;
int tmp_pitch = wc;
wc = wn;
wn = tmp_pitch;
}
}
movd_r2m( mm7, i_score_mmx );
emms();
return i_score_mmx/255 + i_score_c;
}
void
yuv411planar_to_rgb_mmx (const unsigned char *yuv, unsigned char *rgb,
unsigned int w, unsigned int h)
{
unsigned int xx, yy;
register const unsigned char *yp1, *up, *vp;
unsigned char *dp1;
/* plane pointers */
yp1 = yuv;
up = yuv + (w * h);
vp = up + (w * (h / 4));
/* destination pointers */
dp1 = rgb;
yp1 = yuv;
up = yuv + (w * h);
vp = up + ((w / 2) * (h / 2));
dp1 = rgb;
for (yy = 0; yy < h; yy++)
{
for (xx = 0; xx < w; xx += 8)
{
movq_m2r(*yp1, mm0);
movq_r2r(mm0, mm1);
psrlw_i2r(8, mm0);
psllw_i2r(8, mm1);
psrlw_i2r(8, mm1);
pxor_r2r(mm7, mm7);
movd_m2r(*up, mm3);
movd_m2r(*vp, mm2);
punpcklbw_r2r(mm7, mm2);
punpcklbw_r2r(mm7, mm3);
movq_m2r(CONST_16, mm4);
psubsw_r2r(mm4, mm0);
psubsw_r2r(mm4, mm1);
movq_m2r(CONST_128, mm5);
psubsw_r2r(mm5, mm2);
psubsw_r2r(mm5, mm3);
movq_m2r(CONST_YMUL, mm4);
pmullw_r2r(mm4, mm0);
pmullw_r2r(mm4, mm1);
movq_m2r(CONST_CRVCRV, mm7);
pmullw_r2r(mm3, mm7);
movq_m2r(CONST_CBUCBU, mm6);
pmullw_r2r(mm2, mm6);
movq_m2r(CONST_CGUCGU, mm5);
pmullw_r2r(mm2, mm5);
movq_m2r(CONST_CGVCGV, mm4);
pmullw_r2r(mm3, mm4);
movq_r2r(mm0, mm2);
paddsw_r2r(mm7, mm2);
paddsw_r2r(mm1, mm7);
psraw_i2r(RES, mm2);
psraw_i2r(RES, mm7);
packuswb_r2r(mm7, mm2);
pxor_r2r(mm7, mm7);
movq_r2r(mm2, mm3);
punpckhbw_r2r(mm7, mm2);
punpcklbw_r2r(mm3, mm7);
por_r2r(mm7, mm2);
movq_r2r(mm0, mm3);
psubsw_r2r(mm5, mm3);
psubsw_r2r(mm4, mm3);
paddsw_m2r(CONST_32, mm3);
movq_r2r(mm1, mm7);
psubsw_r2r(mm5, mm7);
psubsw_r2r(mm4, mm7);
paddsw_m2r(CONST_32, mm7);
psraw_i2r(RES, mm3);
psraw_i2r(RES, mm7);
packuswb_r2r(mm7, mm3);
pxor_r2r(mm7, mm7);
movq_r2r(mm3, mm4);
punpckhbw_r2r(mm7, mm3);
punpcklbw_r2r(mm4, mm7);
por_r2r(mm7, mm3);
movq_m2r(CONST_32, mm4);
paddsw_r2r(mm6, mm0);
paddsw_r2r(mm6, mm1);
paddsw_r2r(mm4, mm0);
paddsw_r2r(mm4, mm1);
psraw_i2r(RES, mm0);
psraw_i2r(RES, mm1);
packuswb_r2r(mm1, mm0);
pxor_r2r(mm7, mm7);
movq_r2r(mm0, mm5);
punpckhbw_r2r(mm7, mm0);
punpcklbw_r2r(mm5, mm7);
por_r2r(mm7, mm0);
pxor_r2r(mm1, mm1);
movq_r2r(mm0, mm5);
movq_r2r(mm3, mm6);
movq_r2r(mm2, mm7);
punpckhbw_r2r(mm3, mm2);
punpcklbw_r2r(mm6, mm7);
punpckhbw_r2r(mm1, mm0);
punpcklbw_r2r(mm1, mm5);
movq_r2r(mm7, mm1);
punpckhwd_r2r(mm5, mm7);
punpcklwd_r2r(mm5, mm1);
movq_r2r(mm2, mm4);
punpckhwd_r2r(mm0, mm2);
punpcklwd_r2r(mm0, mm4);
movntq_r2m(mm1, *(dp1));
movntq_r2m(mm7, *(dp1 + 8));
movntq_r2m(mm4, *(dp1 + 16));
movntq_r2m(mm2, *(dp1 + 24));
yp1 += 8;
up += 4;
vp += 4;
dp1 += 8 * 4;
}
if (yy & 0x1)
{
up -= w / 2;
vp -= w / 2;
}
}
emms();
}
