modules/video_filter/deinterlace/algo_phosphor.c

   1 /*****************************************************************************
   2  * algo_phosphor.c : Phosphor algorithm for the VLC deinterlacer
   3  *****************************************************************************
   4  * Copyright (C) 2011 VLC authors and VideoLAN
   5  * $Id$
   6  *
   7  * Author: Juha Jeronen <juha.jeronen@jyu.fi>
   8  *
   9  * This program is free software; you can redistribute it and/or modify it
  10  * under the terms of the GNU Lesser General Public License as published by
  11  * the Free Software Foundation; either version 2.1 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  17  * GNU Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public License
  20  * along with this program; if not, write to the Free Software Foundation,
  21  * Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
  22  *****************************************************************************/
  23
  24 #ifdef HAVE_CONFIG_H
  25 #   include "config.h"
  26 #endif
  27
  28 #ifdef CAN_COMPILE_MMXEXT
  29 #   include "mmx.h"
  30 #   include <stdalign.h>
  31 #endif
  32
  33 #include <stdint.h>
  34 #include <assert.h>
  35
  36 #include <vlc_common.h>
  37 #include <vlc_cpu.h>
  38 #include <vlc_picture.h>
  39 #include <vlc_filter.h>
  40
  41 #include "deinterlace.h" /* filter_sys_t */
  42 #include "helpers.h"     /* ComposeFrame() */
  43
  44 #include "algo_phosphor.h"
  45
  46 /*****************************************************************************
  47  * Internal functions
  48  *****************************************************************************/
  49
  50 /**
  51  * Internal helper function: dims (darkens) the given field
  52  * of the given picture.
  53  *
  54  * This is used for simulating CRT light output decay in RenderPhosphor().
  55  *
  56  * The strength "1" is recommended. It's a matter of taste,
  57  * so it's parametrized.
  58  *
  59  * Note on chroma formats:
  60  *   - If input is 4:2:2, all planes are processed.
  61  *   - If input is 4:2:0, only the luma plane is processed, because both fields
  62  *     have the same chroma. This will distort colours, especially for high
  63  *     filter strengths, especially for pixels whose U and/or V values are
  64  *     far away from the origin (which is at 128 in uint8 format).
  65  *
  66  * @param p_dst Input/output picture. Will be modified in-place.
  67  * @param i_field Darken which field? 0 = top, 1 = bottom.
  68  * @param i_strength Strength of effect: 1, 2 or 3 (division by 2, 4 or 8).
  69  * @see RenderPhosphor()
  70  * @see ComposeFrame()
  71  */
  72 static void DarkenField( picture_t *p_dst,
  73                          const int i_field, const int i_strength,
  74                          bool process_chroma )
  75 {
  76     assert( p_dst != NULL );
  77     assert( i_field == 0 || i_field == 1 );
  78     assert( i_strength >= 1 && i_strength <= 3 );
  79
  80     /* Bitwise ANDing with this clears the i_strength highest bits
  81        of each byte */
  82     const uint8_t  remove_high_u8 = 0xFF >> i_strength;
  83     const uint64_t remove_high_u64 = remove_high_u8 *
  84                                             INT64_C(0x0101010101010101);
  85
  86     /* Process luma.
  87
  88        For luma, the operation is just a shift + bitwise AND, so we vectorize
  89        even in the C version.
  90
  91        There is an MMX version too, because it performs about twice faster.
  92     */
  93     int i_plane = Y_PLANE;
  94     uint8_t *p_out, *p_out_end;
  95     int w = p_dst->p[i_plane].i_visible_pitch;
  96     p_out = p_dst->p[i_plane].p_pixels;
  97     p_out_end = p_out + p_dst->p[i_plane].i_pitch
  98                       * p_dst->p[i_plane].i_visible_lines;
  99
 100     /* skip first line for bottom field */
 101     if( i_field == 1 )
 102         p_out += p_dst->p[i_plane].i_pitch;
 103
 104     int wm8 = w % 8;   /* remainder */
 105     int w8  = w - wm8; /* part of width that is divisible by 8 */
 106     for( ; p_out < p_out_end ; p_out += 2*p_dst->p[i_plane].i_pitch )
 107     {
 108         uint64_t *po = (uint64_t *)p_out;
 109         int x = 0;
 110
 111         for( ; x < w8; x += 8, ++po )
 112             (*po) = ( ((*po) >> i_strength) & remove_high_u64 );
 113
 114         /* handle the width remainder */
 115         uint8_t *po_temp = (uint8_t *)po;
 116         for( ; x < w; ++x, ++po_temp )
 117             (*po_temp) = ( ((*po_temp) >> i_strength) & remove_high_u8 );
 118     }
 119
 120     /* Process chroma if the field chromas are independent.
 121
 122        The origin (black) is at YUV = (0, 128, 128) in the uint8 format.
 123        The chroma processing is a bit more complicated than luma,
 124        and needs MMX for vectorization.
 125     */
 126     if( process_chroma )
 127     {
 128         for( i_plane++ /* luma already handled*/;
 129              i_plane < p_dst->i_planes;
 130              i_plane++ )
 131         {
 132             int w = p_dst->p[i_plane].i_visible_pitch;
 133             p_out = p_dst->p[i_plane].p_pixels;
 134             p_out_end = p_out + p_dst->p[i_plane].i_pitch
 135                               * p_dst->p[i_plane].i_visible_lines;
 136
 137             /* skip first line for bottom field */
 138             if( i_field == 1 )
 139                 p_out += p_dst->p[i_plane].i_pitch;
 140
 141             for( ; p_out < p_out_end ; p_out += 2*p_dst->p[i_plane].i_pitch )
 142             {
 143                 /* Handle the width remainder */
 144                 uint8_t *po = p_out;
 145                 for( int x = 0; x < w; ++x, ++po )
 146                     (*po) = 128 + ( ((*po) - 128) / (1 << i_strength) );
 147             } /* for p_out... */
 148         } /* for i_plane... */
 149     } /* if process_chroma */
 150 }
 151
 152 #ifdef CAN_COMPILE_MMXEXT
 153 VLC_MMX
 154 static void DarkenFieldMMX( picture_t *p_dst,
 155                             const int i_field, const int i_strength,
 156                             bool process_chroma )
 157 {
 158     assert( p_dst != NULL );
 159     assert( i_field == 0 || i_field == 1 );
 160     assert( i_strength >= 1 && i_strength <= 3 );
 161
 162     uint64_t i_strength_u64 = i_strength; /* needs to know number of bits */
 163     const uint8_t  remove_high_u8 = 0xFF >> i_strength;
 164     const uint64_t remove_high_u64 = remove_high_u8 *
 165                                             INT64_C(0x0101010101010101);
 166
 167     int i_plane = Y_PLANE;
 168     uint8_t *p_out, *p_out_end;
 169     int w = p_dst->p[i_plane].i_visible_pitch;
 170     p_out = p_dst->p[i_plane].p_pixels;
 171     p_out_end = p_out + p_dst->p[i_plane].i_pitch
 172                       * p_dst->p[i_plane].i_visible_lines;
 173
 174     /* skip first line for bottom field */
 175     if( i_field == 1 )
 176         p_out += p_dst->p[i_plane].i_pitch;
 177
 178     int wm8 = w % 8;   /* remainder */
 179     int w8  = w - wm8; /* part of width that is divisible by 8 */
 180     for( ; p_out < p_out_end ; p_out += 2*p_dst->p[i_plane].i_pitch )
 181     {
 182         uint64_t *po = (uint64_t *)p_out;
 183         int x = 0;
 184
 185         movq_m2r( i_strength_u64,  mm1 );
 186         movq_m2r( remove_high_u64, mm2 );
 187         for( ; x < w8; x += 8 )
 188         {
 189             movq_m2r( (*po), mm0 );
 190
 191             psrlq_r2r( mm1, mm0 );
 192             pand_r2r(  mm2, mm0 );
 193
 194             movq_r2m( mm0, (*po++) );
 195         }
 196
 197         /* handle the width remainder */
 198         uint8_t *po_temp = (uint8_t *)po;
 199         for( ; x < w; ++x, ++po_temp )
 200             (*po_temp) = ( ((*po_temp) >> i_strength) & remove_high_u8 );
 201     }
 202
 203     /* Process chroma if the field chromas are independent.
 204
 205        The origin (black) is at YUV = (0, 128, 128) in the uint8 format.
 206        The chroma processing is a bit more complicated than luma,
 207        and needs MMX for vectorization.
 208     */
 209     if( process_chroma )
 210     {
 211         for( i_plane++ /* luma already handled */;
 212              i_plane < p_dst->i_planes;
 213              i_plane++ )
 214         {
 215             int w = p_dst->p[i_plane].i_visible_pitch;
 216             int wm8 = w % 8;   /* remainder */
 217             int w8  = w - wm8; /* part of width that is divisible by 8 */
 218
 219             p_out = p_dst->p[i_plane].p_pixels;
 220             p_out_end = p_out + p_dst->p[i_plane].i_pitch
 221                               * p_dst->p[i_plane].i_visible_lines;
 222
 223             /* skip first line for bottom field */
 224             if( i_field == 1 )
 225                 p_out += p_dst->p[i_plane].i_pitch;
 226
 227             for( ; p_out < p_out_end ; p_out += 2*p_dst->p[i_plane].i_pitch )
 228             {
 229                 int x = 0;
 230
 231                 /* See also easy-to-read C version below. */
 232                 static alignas (8) const mmx_t b128 = {
 233                     .uq = 0x8080808080808080ULL
 234                 };
 235
 236                 movq_m2r( b128, mm5 );
 237                 movq_m2r( i_strength_u64,  mm6 );
 238                 movq_m2r( remove_high_u64, mm7 );
 239
 240                 uint64_t *po8 = (uint64_t *)p_out;
 241                 for( ; x < w8; x += 8 )
 242                 {
 243                     movq_m2r( (*po8), mm0 );
 244
 245                     movq_r2r( mm5, mm2 ); /* 128 */
 246                     movq_r2r( mm0, mm1 ); /* copy of data */
 247                     psubusb_r2r( mm2, mm1 ); /* mm1 = max(data - 128, 0) */
 248                     psubusb_r2r( mm0, mm2 ); /* mm2 = max(128 - data, 0) */
 249
 250                     /* >> i_strength */
 251                     psrlq_r2r( mm6, mm1 );
 252                     psrlq_r2r( mm6, mm2 );
 253                     pand_r2r(  mm7, mm1 );
 254                     pand_r2r(  mm7, mm2 );
 255
 256                     /* collect results from pos./neg. parts */
 257                     psubb_r2r( mm2, mm1 );
 258                     paddb_r2r( mm5, mm1 );
 259
 260                     movq_r2m( mm1, (*po8++) );
 261                 }
 262
 263                 /* C version - handle the width remainder */
 264                 uint8_t *po = p_out;
 265                 for( ; x < w; ++x, ++po )
 266                     (*po) = 128 + ( ((*po) - 128) / (1 << i_strength) );
 267             } /* for p_out... */
 268         } /* for i_plane... */
 269     } /* if process_chroma */
 270
 271     emms();
 272 }
 273 #endif
 274
 275 /*****************************************************************************
 276  * Public functions
 277  *****************************************************************************/
 278
 279 /* See header for function doc. */
 280 int RenderPhosphor( filter_t *p_filter,
 281                     picture_t *p_dst, picture_t *p_pic,
 282                     int i_order, int i_field )
 283 {
 284     VLC_UNUSED(p_pic);
 285     assert( p_filter != NULL );
 286     assert( p_dst != NULL );
 287     assert( i_order >= 0 && i_order <= 2 ); /* 2 = soft field repeat */
 288     assert( i_field == 0 || i_field == 1 );
 289
 290     filter_sys_t *p_sys = p_filter->p_sys;
 291
 292     /* Last two input frames */
 293     picture_t *p_in  = p_sys->context.pp_history[HISTORY_SIZE-1];
 294     picture_t *p_old = p_sys->context.pp_history[HISTORY_SIZE-2];
 295
 296     /* Use the same input picture as "old" at the first frame after startup */
 297     if( !p_old )
 298         p_old = p_in;
 299
 300     /* If the history mechanism has failed, we can't do anything. */
 301     if( !p_in )
 302         return VLC_EGENERIC;
 303
 304     assert( p_old != NULL );
 305     assert( p_in != NULL );
 306
 307     /* Decide sources for top & bottom fields of output. */
 308     picture_t *p_in_top    = p_in;
 309     picture_t *p_in_bottom = p_in;
 310     /* For the first output field this frame,
 311        grab "old" field from previous frame. */
 312     if( i_order == 0 )
 313     {
 314         if( i_field == 0 ) /* rendering top field */
 315             p_in_bottom = p_old;
 316         else /* i_field == 1, rendering bottom field */
 317             p_in_top = p_old;
 318     }
 319
 320     compose_chroma_t cc = CC_ALTLINE;
 321     if( 2 * p_sys->chroma->p[1].h.num == p_sys->chroma->p[1].h.den &&
 322         2 * p_sys->chroma->p[2].h.num == p_sys->chroma->p[2].h.den )
 323     {
 324         /* Only 420 like chroma */
 325         switch( p_sys->phosphor.i_chroma_for_420 )
 326         {
 327         case PC_BLEND:
 328             cc = CC_MERGE;
 329             break;
 330         case PC_LATEST:
 331             if( i_field == 0 )
 332                 cc = CC_SOURCE_TOP;
 333             else /* i_field == 1 */
 334                 cc = CC_SOURCE_BOTTOM;
 335             break;
 336         case PC_ALTLINE:
 337             cc = CC_ALTLINE;
 338             break;
 339         case PC_UPCONVERT:
 340             cc = CC_UPCONVERT;
 341             break;
 342         default:
 343             /* The above are the only possibilities, if there are no bugs. */
 344             vlc_assert_unreachable();
 345             break;
 346         }
 347     }
 348     ComposeFrame( p_filter, p_dst, p_in_top, p_in_bottom, cc, p_filter->fmt_in.video.i_chroma == VLC_CODEC_YV12 );
 349
 350     /* Simulate phosphor light output decay for the old field.
 351
 352        The dimmer can also be switched off in the configuration, but that is
 353        more of a technical curiosity or an educational toy for advanced users
 354        than a useful deinterlacer mode (although it does make telecined
 355        material look slightly better than without any filtering).
 356
 357        In most use cases the dimmer is used.
 358     */
 359     if( p_sys->phosphor.i_dimmer_strength > 0 )
 360     {
 361 #ifdef CAN_COMPILE_MMXEXT
 362         if( vlc_CPU_MMXEXT() )
 363             DarkenFieldMMX( p_dst, !i_field, p_sys->phosphor.i_dimmer_strength,
 364                 p_sys->chroma->p[1].h.num == p_sys->chroma->p[1].h.den &&
 365                 p_sys->chroma->p[2].h.num == p_sys->chroma->p[2].h.den );
 366         else
 367 #endif
 368             DarkenField( p_dst, !i_field, p_sys->phosphor.i_dimmer_strength,
 369                 p_sys->chroma->p[1].h.num == p_sys->chroma->p[1].h.den &&
 370                 p_sys->chroma->p[2].h.num == p_sys->chroma->p[2].h.den );
 371     }
 372     return VLC_SUCCESS;
 373 }