static void scale_uint8_x_4_x_bilinear_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: Input1
* mm1: Factor mask
* mm2:
* mm3: Output
* mm4:
* mm5: Input2
* mm6: 0
* mm7: Factor
*
*/
// fprintf(stderr, "scale_uint8_x_4_x_bilinear_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 + 4*ctx->table_h.pixels[i].index;
factors = ctx->table_h.pixels[i].factor_i;
/* Load pixels */
movd_m2r(*(src), mm0);
punpcklbw_r2r(mm6, mm0);
psllw_i2r(6, mm0); /* 14 bit */
/* Load pixels */
movd_m2r(*(src+4), mm5);
punpcklbw_r2r(mm6, mm5);
psllw_i2r(6, mm5); /* 14 bit */
/* Load factors */
LOAD_FACTOR_1_4_NOCLIP; /* 14 bit */
/* Subtract */
psubsw_r2r(mm5, mm0); /* s1(mm0) - s2(mm5) -> mm0 (14 bit) */
pmulhw_r2r(mm7, mm0); /* factor * (s2 - s1) -> mm0 (12 bit) */
psllw_i2r(2, mm0); /* (14 bit) */
paddsw_r2r(mm5, mm0);/* (15 bit) */
psraw_i2r(6, mm0);/* (8 bit) */
packuswb_r2r(mm6, mm0);
movd_r2m(mm0, *dst);
dst+=4;
}
ctx->need_emms = 1;
}
static void scale_uint8_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 + 4*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 */
movd_m2r(*(src), mm0);
punpcklbw_r2r(mm6, mm0);
psllw_i2r(7, mm0);
/* Load factors */
LOAD_FACTOR_1_4;
/* Multiply */
pmulhw_r2r(mm7, mm0);
paddw_r2r(mm0, mm3);
// DUMP_MM("mm3_2", mm3);
src += 4;
factors++;
}
psraw_i2r(5, mm3);
packuswb_r2r(mm6, mm3);
movd_r2m(mm3, *dst);
dst+=4;
}
ctx->need_emms = 1;
}
static inline void XDeint8x8MergeMMXEXT( uint8_t *dst, int i_dst,
uint8_t *src1, int i_src1,
uint8_t *src2, int i_src2 )
{
static const uint64_t m_4 = INT64_C(0x0004000400040004);
int y, x;
/* Progressive */
pxor_r2r( mm7, mm7 );
for( y = 0; y < 8; y += 2 )
{
for( x = 0; x < 8; x +=4 )
{
movd_m2r( src1[x], mm0 );
movd_r2m( mm0, dst[x] );
movd_m2r( src2[x], mm1 );
movd_m2r( src1[i_src1+x], mm2 );
punpcklbw_r2r( mm7, mm0 );
punpcklbw_r2r( mm7, mm1 );
punpcklbw_r2r( mm7, mm2 );
paddw_r2r( mm1, mm1 );
movq_r2r( mm1, mm3 );
paddw_r2r( mm3, mm3 );
paddw_r2r( mm2, mm0 );
paddw_r2r( mm3, mm1 );
paddw_m2r( m_4, mm1 );
paddw_r2r( mm1, mm0 );
psraw_i2r( 3, mm0 );
packuswb_r2r( mm7, mm0 );
movd_r2m( mm0, dst[i_dst+x] );
}
dst += 2*i_dst;
src1 += i_src1;
src2 += i_src2;
}
}
static __inline__ int qblock_sad_mmxe(uint8_t *refblk,
uint32_t h,
uint32_t rowstride)
{
int res;
pxor_r2r (mm4,mm4);
movq_r2r (mm0,mm5); /* First row */
movd_m2r (*refblk, mm6);
pxor_r2r ( mm7, mm7);
refblk += rowstride;
punpcklbw_r2r ( mm7, mm5);
punpcklbw_r2r ( mm7, mm6);
psadbw_r2r ( mm5, mm6);
paddw_r2r ( mm6, mm4 );
movq_r2r (mm1,mm5); /* Second row */
movd_m2r (*refblk, mm6);
refblk += rowstride;
punpcklbw_r2r ( mm7, mm5);
punpcklbw_r2r ( mm7, mm6);
psadbw_r2r ( mm5, mm6);
paddw_r2r ( mm6, mm4 );
if( h == 4 )
{
movq_r2r (mm2,mm5); /* Third row */
movd_m2r (*refblk, mm6);
refblk += rowstride;
punpcklbw_r2r ( mm7, mm5);
punpcklbw_r2r ( mm7, mm6);
psadbw_r2r ( mm5, mm6);
paddw_r2r ( mm6, mm4 );
movq_r2r (mm3,mm5); /* Fourth row */
movd_m2r (*refblk, mm6);
punpcklbw_r2r ( mm7, mm5);
punpcklbw_r2r ( mm7, mm6);
psadbw_r2r ( mm5, mm6);
paddw_r2r ( mm6, mm4 );
}
movd_r2m ( mm4, res );
return res;
}
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 __inline__ void
mmx_sum_4_word_accs( mmx_t *accs, int32_t *res )
{
movq_m2r( *accs, mm1 );
movq_r2r( mm1, mm3 );
movq_r2r( mm1, mm2 );
/* Generate sign extensions for mm1 words! */
psraw_i2r( 15, mm3 );
punpcklwd_r2r( mm3, mm1 );
punpckhwd_r2r( mm3, mm2 );
paddd_r2r( mm1, mm2 );
movq_r2r( mm2, mm3);
psrlq_i2r( 32, mm2);
paddd_r2r( mm2, mm3);
movd_r2m( mm3, *res );
}
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;
}
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);
}
//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;
}
//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 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
}
/**
* 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 )
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
}
static inline int XDeint8x8DetectMMXEXT( uint8_t *src, int i_src )
{
int y, x;
int32_t ff, fr;
int fc;
/* Detect interlacing */
fc = 0;
pxor_r2r( mm7, mm7 );
for( y = 0; y < 9; y += 2 )
{
ff = fr = 0;
pxor_r2r( mm5, mm5 );
pxor_r2r( mm6, mm6 );
for( x = 0; x < 8; x+=4 )
{
movd_m2r( src[ x], mm0 );
movd_m2r( src[1*i_src+x], mm1 );
movd_m2r( src[2*i_src+x], mm2 );
movd_m2r( src[3*i_src+x], mm3 );
punpcklbw_r2r( mm7, mm0 );
punpcklbw_r2r( mm7, mm1 );
punpcklbw_r2r( mm7, mm2 );
punpcklbw_r2r( mm7, mm3 );
movq_r2r( mm0, mm4 );
psubw_r2r( mm1, mm0 );
psubw_r2r( mm2, mm4 );
psubw_r2r( mm1, mm2 );
psubw_r2r( mm1, mm3 );
pmaddwd_r2r( mm0, mm0 );
pmaddwd_r2r( mm4, mm4 );
pmaddwd_r2r( mm2, mm2 );
pmaddwd_r2r( mm3, mm3 );
paddd_r2r( mm0, mm2 );
paddd_r2r( mm4, mm3 );
paddd_r2r( mm2, mm5 );
paddd_r2r( mm3, mm6 );
}
movq_r2r( mm5, mm0 );
psrlq_i2r( 32, mm0 );
paddd_r2r( mm0, mm5 );
movd_r2m( mm5, fr );
movq_r2r( mm6, mm0 );
psrlq_i2r( 32, mm0 );
paddd_r2r( mm0, mm6 );
movd_r2m( mm6, ff );
if( ff < 6*fr/8 && fr > 32 )
fc++;
src += 2*i_src;
}
return fc;
}
