static inline void mmx_end(uint8_t *src3, uint8_t *src5,
uint8_t *dst, int X)
{
punpcklbw_m2r (mm_cpool[0], mm4);
punpckhbw_m2r (mm_cpool[0], mm5);
psubusw_r2r (mm2, mm0);
psubusw_r2r (mm3, mm1);
movq_m2r (src5[X], mm2);
movq_m2r (src5[X], mm3);
punpcklbw_m2r (mm_cpool[0], mm2);
punpckhbw_m2r (mm_cpool[0], mm3);
psubusw_r2r (mm2, mm0);
psubusw_r2r (mm3, mm1);
psrlw_i2r (3, mm0);
psrlw_i2r (3, mm1);
psubw_r2r (mm6, mm4);
psubw_r2r (mm7, mm5);
packuswb_r2r (mm1,mm0);
movq_r2r (mm4, mm6);
movq_r2r (mm5, mm7);
pcmpgtw_m2r (mm_lthr, mm4);
pcmpgtw_m2r (mm_lthr, mm5);
pcmpgtw_m2r (mm_hthr, mm6);
pcmpgtw_m2r (mm_hthr, mm7);
packsswb_r2r (mm5, mm4);
packsswb_r2r (mm7, mm6);
pxor_r2r (mm6, mm4);
movq_r2r (mm4, mm5);
pandn_r2r (mm0, mm4);
pand_m2r (src3[X], mm5);
por_r2r (mm4, mm5);
movq_r2m (mm5, dst[X]);
}
static void scale_uint16_x_4_x_generic_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i, j;
uint8_t * src, * dst, *src_start;
int32_t * factors;
// mmx_t tmp_mm;
/*
* mm0: Input
* mm1: factor_mask
* mm2: Factor
* mm3: Output
* mm4:
* mm5:
* mm6: 0
* mm7: scratch
*
*/
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
movq_m2r(factor_mask, mm1);
dst = dest_start;
for(i = 0; i < ctx->dst_size; i++)
{
src = src_start + 8*ctx->table_h.pixels[i].index;
factors = ctx->table_h.pixels[i].factor_i;
pxor_r2r(mm3, mm3);
for(j = 0; j < ctx->table_h.factors_per_pixel; j++)
{
/* Load pixels */
movq_m2r(*(src), mm0);
psrlw_i2r(1, mm0);
/* Load factors */
LOAD_FACTOR_1_4;
/* Multiply */
pmulhw_r2r(mm7, mm0);
paddw_r2r(mm0, mm3);
// DUMP_MM("mm3_2", mm3);
src += 8;
factors++;
}
pminsw_m2r(max_13, mm3);
pmaxsw_m2r(min_13, mm3);
psllw_i2r(3, mm3);
MOVQ_R2M(mm3, *dst);
dst+=8;
}
ctx->need_emms = 1;
}
static void
deinterlace_line_mmx (uint8_t * dst, uint8_t * lum_m4,
uint8_t * lum_m3, uint8_t * lum_m2,
uint8_t * lum_m1, uint8_t * lum, int size)
{
mmx_t rounder;
rounder.uw[0] = 4;
rounder.uw[1] = 4;
rounder.uw[2] = 4;
rounder.uw[3] = 4;
pxor_r2r (mm7, mm7);
movq_m2r (rounder, mm6);
for (; size > 3; size -= 4) {
movd_m2r (*lum_m4, mm0);
movd_m2r (*lum_m3, mm1);
movd_m2r (*lum_m2, mm2);
movd_m2r (*lum_m1, mm3);
movd_m2r (*lum, mm4);
punpcklbw_r2r (mm7, mm0);
punpcklbw_r2r (mm7, mm1);
punpcklbw_r2r (mm7, mm2);
punpcklbw_r2r (mm7, mm3);
punpcklbw_r2r (mm7, mm4);
paddw_r2r (mm3, mm1);
psllw_i2r (1, mm2);
paddw_r2r (mm4, mm0);
psllw_i2r (2, mm1); // 2
paddw_r2r (mm6, mm2);
paddw_r2r (mm2, mm1);
psubusw_r2r (mm0, mm1);
psrlw_i2r (3, mm1); // 3
packuswb_r2r (mm7, mm1);
movd_r2m (mm1, *dst);
lum_m4 += 4;
lum_m3 += 4;
lum_m2 += 4;
lum_m1 += 4;
lum += 4;
dst += 4;
}
emms ();
/* Handle odd widths */
if (size > 0)
deinterlace_line_c (dst, lum_m4, lum_m3, lum_m2, lum_m1, lum, size);
}
static void scale_uint16_x_1_x_bicubic_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i;
uint16_t * dst;
uint8_t * src, *src_start;
int32_t * factors;
mmx_t tmp_mm;
int32_t tmp;
// fprintf(stderr, "scale_uint8_x_1_x_bicubic_mmx\n");
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
dst = (uint16_t*)dest_start;
for(i = 0; i < ctx->dst_size; i++)
{
src = src_start + 2*ctx->table_h.pixels[i].index;
factors = ctx->table_h.pixels[i].factor_i;
/* Load pixels */
movq_m2r(*(src), mm0);
psrlw_i2r(1, mm0);
// DUMP_MM("mm0", mm0);
/* Load factors */
movq_m2r(*factors, mm2);
movq_m2r(*(factors+2), mm3);
packssdw_r2r(mm3, mm2);
/* Multiply */
pmaddwd_r2r(mm2, mm0);
MOVQ_R2M(mm0, tmp_mm);
tmp = tmp_mm.d[0] + tmp_mm.d[1];
tmp >>= 13;
RECLIP(tmp, ctx->plane);
*(dst++) = tmp;
}
ctx->need_emms = 1;
}
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;
}
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 interpolate_packed422_scanline_mmx( uint8_t *output, uint8_t *top,
uint8_t *bot, int width )
{
const mmx_t shiftmask = { 0xfefffefffefffeffULL }; /* To avoid shifting chroma to luma. */
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 );
pand_m2r( shiftmask, mm0 );
pand_m2r( shiftmask, mm1 );
pand_m2r( shiftmask, mm2 );
pand_m2r( shiftmask, mm3 );
pand_m2r( shiftmask, mm4 );
pand_m2r( shiftmask, mm5 );
pand_m2r( shiftmask, mm6 );
pand_m2r( shiftmask, mm7 );
psrlw_i2r( 1, mm0 );
psrlw_i2r( 1, mm1 );
psrlw_i2r( 1, mm2 );
psrlw_i2r( 1, mm3 );
psrlw_i2r( 1, mm4 );
psrlw_i2r( 1, mm5 );
psrlw_i2r( 1, mm6 );
psrlw_i2r( 1, mm7 );
paddb_r2r( mm1, mm0 );
paddb_r2r( mm3, mm2 );
paddb_r2r( mm5, mm4 );
paddb_r2r( mm7, mm6 );
movq_r2m( mm0, *output );
movq_r2m( mm2, *(output + 8) );
movq_r2m( mm4, *(output + 16) );
movq_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 );
pand_m2r( shiftmask, mm0 );
pand_m2r( shiftmask, mm1 );
psrlw_i2r( 1, mm0 );
psrlw_i2r( 1, mm1 );
paddb_r2r( mm1, mm0 );
movq_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;
}
emms();
}
static void scale_uint16_x_4_x_quadratic_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i;
uint8_t * src, * dst, *src_start;
int32_t * factors;
// mmx_t tmp_mm;
/*
* mm0: Input
* mm1: factor_mask
* mm2: Factor
* mm3: Output
* mm4:
* mm5:
* mm6: 0
* mm7: scratch
*
*/
// fprintf(stderr, "scale_uint8_x_1_x_bicubic_noclip_mmx\n");
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
movq_m2r(factor_mask, mm1);
dst = dest_start;
for(i = 0; i < ctx->dst_size; i++)
{
src = src_start + 8*ctx->table_h.pixels[i].index;
factors = ctx->table_h.pixels[i].factor_i;
/* Load pixels */
movq_m2r(*(src), mm0);
// punpcklbw_r2r(mm6, mm0);
psrlw_i2r(1, mm0);
/* Load factors */
LOAD_FACTOR_1_4_NOCLIP;
/* Multiply */
pmulhw_r2r(mm7, mm0);
movq_r2r(mm0, mm3);
// DUMP_MM("mm3_1", mm3);
src += 8;
factors++;
/* Load pixels */
movq_m2r(*(src), mm0);
// punpcklbw_r2r(mm6, mm0);
psrlw_i2r(1, mm0);
/* Load factors */
LOAD_FACTOR_1_4_NOCLIP;
/* Multiply */
pmulhw_r2r(mm7, mm0);
paddw_r2r(mm0, mm3);
// DUMP_MM("mm3_2", mm3);
src += 8;
factors++;
/* Load pixels */
movq_m2r(*(src), mm0);
// punpcklbw_r2r(mm6, mm0);
psrlw_i2r(1, mm0);
/* Load factors */
LOAD_FACTOR_1_4_NOCLIP;
/* Multiply */
pmulhw_r2r(mm7, mm0);
paddw_r2r(mm0, mm3);
// DUMP_MM("mm3_3", mm3);
src += 8;
psllw_i2r(3, mm3);
// packuswb_r2r(mm6, mm3);
MOVQ_R2M(mm3, *dst);
dst+=8;
}
ctx->need_emms = 1;
}
static void
deinterlace_greedy_scanline_mmx (GstDeinterlaceMethodGreedyL * self,
const guint8 * m0, const guint8 * t1,
const guint8 * b1, const guint8 * m2, guint8 * output, gint width)
{
mmx_t MaxComb;
mmx_t ShiftMask;
// 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;
ShiftMask.ub[0] = 0x7f;
ShiftMask.ub[1] = 0x7f;
ShiftMask.ub[2] = 0x7f;
ShiftMask.ub[3] = 0x7f;
ShiftMask.ub[4] = 0x7f;
ShiftMask.ub[5] = 0x7f;
ShiftMask.ub[6] = 0x7f;
ShiftMask.ub[7] = 0x7f;
// 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
movq_r2r (mm3, mm5); // L3
psrlw_i2r (1, mm4); // L1/2
pand_m2r (ShiftMask, mm4);
psrlw_i2r (1, mm5); // L3/2
pand_m2r (ShiftMask, mm5);
paddusb_r2r (mm5, 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
psubusb_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
psubusb_r2r (mm3, mm2); // - L3, with saturation
paddusb_r2r (mm3, mm2); // now = Max(L1,L3)
pcmpeqb_r2r (mm7, mm7); // all ffffffff
psubusb_r2r (mm1, mm7); // - L1
paddusb_r2r (mm7, mm3); // add, may sat at fff..
psubusb_r2r (mm7, 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
psubusb_r2r (mm3, mm4); // best - Min
paddusb_r2r (mm3, mm4); // now = Max(best,Min(L1,L3)
pcmpeqb_r2r (mm7, mm7); // all ffffffff
psubusb_r2r (mm4, mm7); // - Max(best,Min(best,L3)
paddusb_r2r (mm7, mm2); // add may sat at FFF..
psubusb_r2r (mm7, 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);
}
static void scale_uint8_x_1_x_bilinear_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i, imax, index;
uint8_t * src, * dst, *src_start;
mmx_t tmp_mm;
/*
* mm0: Input1 Input2
* mm1: Factor
* mm2:
* mm3:
* mm4:
* mm5:
* mm6: 0
* mm7: scratch
*
*/
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
dst = dest_start;
imax = ctx->dst_size / 4;
// imax = 0;
index = 0;
for(i = 0; i < imax; i++)
{
/* Load pixels */
src = src_start + ctx->table_h.pixels[index].index;
tmp_mm.uw[0] = *src;
tmp_mm.uw[1] = *(src+1);
src = src_start + ctx->table_h.pixels[index+1].index;
tmp_mm.uw[2] = *src;
tmp_mm.uw[3] = *(src+1);
movq_m2r(tmp_mm, mm0);
/* Load factors */
movq_m2r(ctx->table_h.pixels[index].factor_i[0], mm1);
movq_m2r(ctx->table_h.pixels[index+1].factor_i[0], mm7);
packssdw_r2r(mm7, mm1);
pmaddwd_r2r(mm0, mm1);
index += 2;
/* Load pixels */
src = src_start + ctx->table_h.pixels[index].index;
tmp_mm.uw[0] = *src;
tmp_mm.uw[1] = *(src+1);
src = src_start + ctx->table_h.pixels[index+1].index;
tmp_mm.uw[2] = *src;
tmp_mm.uw[3] = *(src+1);
movq_m2r(tmp_mm, mm0);
/* Load factors */
movq_m2r(ctx->table_h.pixels[index].factor_i[0], mm3);
movq_m2r(ctx->table_h.pixels[index+1].factor_i[0], mm7);
packssdw_r2r(mm7, mm3);
pmaddwd_r2r(mm0, mm3);
psrld_i2r(7, mm3);
psrld_i2r(7, mm1);
packssdw_r2r(mm3, mm1);
psrlw_i2r(7, mm1);
index += 2;
packuswb_r2r(mm6, mm1);
movd_r2m(mm1, *dst);
// *dst = tmp_mm.ub[0];
// *(dst+1) = tmp_mm.ub[4];
dst+=4;
}
ctx->need_emms = 1;
#if 1
imax = ctx->dst_size % 4;
for(i = 0; i < imax; i++)
{
src = (src_start + ctx->table_h.pixels[index].index);
*dst = (ctx->table_h.pixels[index].factor_i[0] * *src +
ctx->table_h.pixels[index].factor_i[1] * *(src+1)) >> 14;
dst++;
index++;
}
#endif
}
static void
deinterlace_scanline_linear_mmx (GstDeinterlaceMethod * self,
GstDeinterlace * parent, guint8 * out,
GstDeinterlaceScanlineData * scanlines, gint width)
{
const mmx_t shiftmask = { 0xfefffefffefffeffULL }; /* To avoid shifting chroma to luma. */
int 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);
pand_m2r (shiftmask, mm0);
pand_m2r (shiftmask, mm1);
pand_m2r (shiftmask, mm2);
pand_m2r (shiftmask, mm3);
pand_m2r (shiftmask, mm4);
pand_m2r (shiftmask, mm5);
pand_m2r (shiftmask, mm6);
pand_m2r (shiftmask, mm7);
psrlw_i2r (1, mm0);
psrlw_i2r (1, mm1);
psrlw_i2r (1, mm2);
psrlw_i2r (1, mm3);
psrlw_i2r (1, mm4);
psrlw_i2r (1, mm5);
psrlw_i2r (1, mm6);
psrlw_i2r (1, mm7);
paddb_r2r (mm1, mm0);
paddb_r2r (mm3, mm2);
paddb_r2r (mm5, mm4);
paddb_r2r (mm7, mm6);
movq_r2m (mm0, *out);
movq_r2m (mm2, *(out + 8));
movq_r2m (mm4, *(out + 16));
movq_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);
pand_m2r (shiftmask, mm0);
pand_m2r (shiftmask, mm1);
psrlw_i2r (1, mm0);
psrlw_i2r (1, mm1);
paddb_r2r (mm1, mm0);
movq_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_line( uint8_t *dst, uint8_t *lum_m4,
uint8_t *lum_m3, uint8_t *lum_m2,
uint8_t *lum_m1, uint8_t *lum, int size )
{
#if defined(__i386__) || defined(__x86_64__)
mmx_t rounder;
rounder.uw[0]=4;
rounder.uw[1]=4;
rounder.uw[2]=4;
rounder.uw[3]=4;
pxor_r2r(mm7,mm7);
movq_m2r(rounder,mm6);
for (;size > 3; size-=4) {
movd_m2r(lum_m4[0],mm0);
movd_m2r(lum_m3[0],mm1);
movd_m2r(lum_m2[0],mm2);
movd_m2r(lum_m1[0],mm3);
movd_m2r(lum[0],mm4);
punpcklbw_r2r(mm7,mm0);
punpcklbw_r2r(mm7,mm1);
punpcklbw_r2r(mm7,mm2);
punpcklbw_r2r(mm7,mm3);
punpcklbw_r2r(mm7,mm4);
paddw_r2r(mm3,mm1);
psllw_i2r(1,mm2);
paddw_r2r(mm4,mm0);
psllw_i2r(2,mm1);// 2
paddw_r2r(mm6,mm2);
paddw_r2r(mm2,mm1);
psubusw_r2r(mm0,mm1);
psrlw_i2r(3,mm1); // 3
packuswb_r2r(mm7,mm1);
movd_r2m(mm1,dst[0]);
lum_m4+=4;
lum_m3+=4;
lum_m2+=4;
lum_m1+=4;
lum+=4;
dst+=4;
}
emms();
#else
/**
* C implementation.
*/
int sum;
for(;size > 0;size--) {
sum = -lum_m4[0];
sum += lum_m3[0] << 2;
sum += lum_m2[0] << 1;
sum += lum_m1[0] << 2;
sum += -lum[0];
dst[0] = (sum + 4) >> 3; // This needs to be clipped at 0 and 255: cm[(sum + 4) >> 3];
lum_m4++;
lum_m3++;
lum_m2++;
lum_m1++;
lum++;
dst++;
}
#endif
}
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();
}
static void
deinterlace_scanline_linear_mmx (GstDeinterlaceSimpleMethod * self,
guint8 * out, const guint8 * bot, const guint8 * top, gint size)
{
const mmx_t shiftmask = { 0xfefffefffefffeffULL }; /* To avoid shifting chroma to luma. */
int 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);
pand_m2r (shiftmask, mm0);
pand_m2r (shiftmask, mm1);
pand_m2r (shiftmask, mm2);
pand_m2r (shiftmask, mm3);
pand_m2r (shiftmask, mm4);
pand_m2r (shiftmask, mm5);
pand_m2r (shiftmask, mm6);
pand_m2r (shiftmask, mm7);
psrlw_i2r (1, mm0);
psrlw_i2r (1, mm1);
psrlw_i2r (1, mm2);
psrlw_i2r (1, mm3);
psrlw_i2r (1, mm4);
psrlw_i2r (1, mm5);
psrlw_i2r (1, mm6);
psrlw_i2r (1, mm7);
paddb_r2r (mm1, mm0);
paddb_r2r (mm3, mm2);
paddb_r2r (mm5, mm4);
paddb_r2r (mm7, mm6);
movq_r2m (mm0, *out);
movq_r2m (mm2, *(out + 8));
movq_r2m (mm4, *(out + 16));
movq_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);
pand_m2r (shiftmask, mm0);
pand_m2r (shiftmask, mm1);
psrlw_i2r (1, mm0);
psrlw_i2r (1, mm1);
paddb_r2r (mm1, mm0);
movq_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 scale_uint16_x_1_x_generic_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i, j, jmax;
uint16_t * src, * dst;
uint8_t * src_start;
int32_t * factors;
mmx_t tmp_mm;
int tmp;
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
dst = (uint16_t*)dest_start;
for(i = 0; i < ctx->dst_size; i++)
{
src = (uint16_t*)(src_start + 2*ctx->table_h.pixels[i].index);
factors = ctx->table_h.pixels[i].factor_i;
jmax = ctx->table_h.factors_per_pixel / 4;
tmp = 0;
#if 1
pxor_r2r(mm4, mm4);
for(j = 0; j < jmax; j++)
{
/* Load pixels */
movq_m2r(*(src), mm0);
psrlw_i2r(1, mm0);
// DUMP_MM("mm0", mm0);
/* Load factors */
movq_m2r(*factors, mm2);
movq_m2r(*(factors+2), mm3);
packssdw_r2r(mm3, mm2);
/* Multiply */
pmaddwd_r2r(mm2, mm0);
paddd_r2r(mm0, mm4);
src += 4;
factors += 4;
}
MOVQ_R2M(mm4, tmp_mm);
tmp = tmp_mm.d[0] + tmp_mm.d[1];
jmax = ctx->table_h.factors_per_pixel % 4;
#else
jmax = ctx->table_h.factors_per_pixel;
#endif
for(j = 0; j < jmax; j++)
{
tmp += *factors * ((*src)>>1);
factors++;
src++;
}
// if(tmp > (255 << 14)) tmp = 255 << 14;
// if(tmp < 0) tmp = 0;
tmp >>= 13;
RECLIP(tmp, ctx->plane);
*(dst++) = tmp;
}
ctx->need_emms = 1;
}
static inline void mmx_yuv2rgb (uint8_t * py, uint8_t * pu, uint8_t * pv)
{
static mmx_t mmx_80w = {0x0080008000800080LL};
static mmx_t mmx_U_green = {0xf37df37df37df37dLL};
static mmx_t mmx_U_blue = {0x4093409340934093LL};
static mmx_t mmx_V_red = {0x3312331233123312LL};
static mmx_t mmx_V_green = {0xe5fce5fce5fce5fcLL};
static mmx_t mmx_10w = {0x1010101010101010LL};
static mmx_t mmx_00ffw = {0x00ff00ff00ff00ffLL};
static mmx_t mmx_Y_coeff = {0x253f253f253f253fLL};
movd_m2r (*pu, mm0); /* mm0 = 00 00 00 00 u3 u2 u1 u0 */
movd_m2r (*pv, mm1); /* mm1 = 00 00 00 00 v3 v2 v1 v0 */
movq_m2r (*py, mm6); /* mm6 = Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 */
pxor_r2r (mm4, mm4); /* mm4 = 0 */
/* XXX might do cache preload for image here */
/*
* Do the multiply part of the conversion for even and odd pixels
* register usage:
* mm0 -> Cblue, mm1 -> Cred, mm2 -> Cgreen even pixels
* mm3 -> Cblue, mm4 -> Cred, mm5 -> Cgreen odd pixels
* mm6 -> Y even, mm7 -> Y odd
*/
punpcklbw_r2r (mm4, mm0); /* mm0 = u3 u2 u1 u0 */
punpcklbw_r2r (mm4, mm1); /* mm1 = v3 v2 v1 v0 */
psubsw_m2r (mmx_80w, mm0); /* u -= 128 */
psubsw_m2r (mmx_80w, mm1); /* v -= 128 */
psllw_i2r (3, mm0); /* promote precision */
psllw_i2r (3, mm1); /* promote precision */
movq_r2r (mm0, mm2); /* mm2 = u3 u2 u1 u0 */
movq_r2r (mm1, mm3); /* mm3 = v3 v2 v1 v0 */
pmulhw_m2r (mmx_U_green, mm2); /* mm2 = u * u_green */
pmulhw_m2r (mmx_V_green, mm3); /* mm3 = v * v_green */
pmulhw_m2r (mmx_U_blue, mm0); /* mm0 = chroma_b */
pmulhw_m2r (mmx_V_red, mm1); /* mm1 = chroma_r */
paddsw_r2r (mm3, mm2); /* mm2 = chroma_g */
psubusb_m2r (mmx_10w, mm6); /* Y -= 16 */
movq_r2r (mm6, mm7); /* mm7 = Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 */
pand_m2r (mmx_00ffw, mm6); /* mm6 = Y6 Y4 Y2 Y0 */
psrlw_i2r (8, mm7); /* mm7 = Y7 Y5 Y3 Y1 */
psllw_i2r (3, mm6); /* promote precision */
psllw_i2r (3, mm7); /* promote precision */
pmulhw_m2r (mmx_Y_coeff, mm6); /* mm6 = luma_rgb even */
pmulhw_m2r (mmx_Y_coeff, mm7); /* mm7 = luma_rgb odd */
/*
* Do the addition part of the conversion for even and odd pixels
* register usage:
* mm0 -> Cblue, mm1 -> Cred, mm2 -> Cgreen even pixels
* mm3 -> Cblue, mm4 -> Cred, mm5 -> Cgreen odd pixels
* mm6 -> Y even, mm7 -> Y odd
*/
movq_r2r (mm0, mm3); /* mm3 = chroma_b */
movq_r2r (mm1, mm4); /* mm4 = chroma_r */
movq_r2r (mm2, mm5); /* mm5 = chroma_g */
paddsw_r2r (mm6, mm0); /* mm0 = B6 B4 B2 B0 */
paddsw_r2r (mm7, mm3); /* mm3 = B7 B5 B3 B1 */
paddsw_r2r (mm6, mm1); /* mm1 = R6 R4 R2 R0 */
paddsw_r2r (mm7, mm4); /* mm4 = R7 R5 R3 R1 */
paddsw_r2r (mm6, mm2); /* mm2 = G6 G4 G2 G0 */
paddsw_r2r (mm7, mm5); /* mm5 = G7 G5 G3 G1 */
packuswb_r2r (mm0, mm0); /* saturate to 0-255 */
packuswb_r2r (mm1, mm1); /* saturate to 0-255 */
packuswb_r2r (mm2, mm2); /* saturate to 0-255 */
packuswb_r2r (mm3, mm3); /* saturate to 0-255 */
packuswb_r2r (mm4, mm4); /* saturate to 0-255 */
packuswb_r2r (mm5, mm5); /* saturate to 0-255 */
punpcklbw_r2r (mm3, mm0); /* mm0 = B7 B6 B5 B4 B3 B2 B1 B0 */
punpcklbw_r2r (mm4, mm1); /* mm1 = R7 R6 R5 R4 R3 R2 R1 R0 */
punpcklbw_r2r (mm5, mm2); /* mm2 = G7 G6 G5 G4 G3 G2 G1 G0 */
}
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();
}