git.sesse.net Git - vlc/blob - plugins/yuv/video_yuvmmx.c

   1 /*****************************************************************************
   2  * video_yuvmmx.c: MMX YUV transformation functions
   3  * Provides functions to perform the YUV conversion.
   4  *****************************************************************************
   5  * Copyright (C) 1999, 2000 VideoLAN
   6  * $Id: video_yuvmmx.c,v 1.7 2001/03/21 13:42:34 sam Exp $
   7  *
   8  * Authors: Samuel Hocevar <sam@zoy.org>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 02111-1307, USA.
  24  *****************************************************************************/
  25
  26 #define MODULE_NAME yuvmmx
  27 #include "modules_inner.h"
  28
  29 /*****************************************************************************
  30  * Preamble
  31  *****************************************************************************/
  32 #include "defs.h"
  33
  34 #include <math.h>                                            /* exp(), pow() */
  35 #include <errno.h>                                                 /* ENOMEM */
  36 #include <stdlib.h>                                                /* free() */
  37 #include <string.h>                                            /* strerror() */
  38
  39 #include "config.h"
  40 #include "common.h"
  41 #include "threads.h"
  42 #include "mtime.h"
  43 #include "tests.h"
  44
  45 #include "modules.h"
  46
  47 #include "video.h"
  48 #include "video_output.h"
  49
  50 #include "video_common.h"
  51
  52 #include "intf_msg.h"
  53
  54 static int     yuv_Probe      ( probedata_t *p_data );
  55 static int     yuv_Init       ( vout_thread_t *p_vout );
  56 static int     yuv_Reset      ( vout_thread_t *p_vout );
  57 static void    yuv_End        ( vout_thread_t *p_vout );
  58
  59 static void    SetYUV         ( vout_thread_t *p_vout );
  60
  61 /*****************************************************************************
  62  * Functions exported as capabilities. They are declared as static so that
  63  * we don't pollute the namespace too much.
  64  *****************************************************************************/
  65 void _M( yuv_getfunctions )( function_list_t * p_function_list )
  66 {
  67     p_function_list->pf_probe = yuv_Probe;
  68     p_function_list->functions.yuv.pf_init = yuv_Init;
  69     p_function_list->functions.yuv.pf_reset = yuv_Reset;
  70     p_function_list->functions.yuv.pf_end = yuv_End;
  71 }
  72
  73 /*****************************************************************************
  74  * yuv_Probe: tests probe the audio device and return a score
  75  *****************************************************************************
  76  * This function tries to open the DSP and returns a score to the plugin
  77  * manager so that it can choose the most appropriate one.
  78  *****************************************************************************/
  79 static int yuv_Probe( probedata_t *p_data )
  80 {
  81     /* Test for MMX support in the CPU */
  82     if( TestCPU( CPU_CAPABILITY_MMX ) )
  83     {
  84         if( TestMethod( YUV_METHOD_VAR, "yuvmmx" ) )
  85         {
  86             return( 999 );
  87         }
  88         else
  89         {
  90             return( 100 );
  91         }
  92     }
  93     else
  94     {
  95         return( 0 );
  96     }
  97 }
  98
  99 /*****************************************************************************
 100  * yuv_Init: allocate and initialize translations tables
 101  *****************************************************************************
 102  * This function will allocate memory to store translation tables, depending
 103  * of the screen depth.
 104  *****************************************************************************/
 105 static int yuv_Init( vout_thread_t *p_vout )
 106 {
 107     size_t      tables_size;                        /* tables size, in bytes */
 108
 109     /* Computes tables size for 8bbp only */
 110     if( p_vout->i_bytes_per_pixel == 1 )
 111     {
 112         tables_size = sizeof( u8 )
 113                 * (p_vout->b_grayscale ? GRAY_TABLE_SIZE : PALETTE_TABLE_SIZE);
 114
 115         /* Allocate memory */
 116         p_vout->yuv.p_base = malloc( tables_size );
 117         if( p_vout->yuv.p_base == NULL )
 118         {
 119             intf_ErrMsg("error: %s", strerror(ENOMEM));
 120             return( 1 );
 121         }
 122     }
 123     else
 124     {
 125         p_vout->yuv.p_base = NULL;
 126     }
 127
 128     /* Allocate memory for conversion buffer and offset array */
 129     p_vout->yuv.p_buffer = malloc( VOUT_MAX_WIDTH * p_vout->i_bytes_per_pixel );
 130     if( p_vout->yuv.p_buffer == NULL )
 131     {
 132         intf_ErrMsg("error: %s", strerror(ENOMEM));
 133         free( p_vout->yuv.p_base );
 134         return( 1 );
 135     }
 136     p_vout->yuv.p_offset = malloc( p_vout->i_width * sizeof( int ) );
 137     if( p_vout->yuv.p_offset == NULL )
 138     {
 139         intf_ErrMsg("error: %s", strerror(ENOMEM));
 140         free( p_vout->yuv.p_base );
 141         free( p_vout->yuv.p_buffer );
 142         return( 1 );
 143     }
 144
 145     /* Initialize tables */
 146     SetYUV( p_vout );
 147     return( 0 );
 148 }
 149
 150 /*****************************************************************************
 151  * yuv_End: destroy translations tables
 152  *****************************************************************************
 153  * Free memory allocated by yuv_CCreate.
 154  *****************************************************************************/
 155 static void yuv_End( vout_thread_t *p_vout )
 156 {
 157     free( p_vout->yuv.p_base );
 158     free( p_vout->yuv.p_buffer );
 159     free( p_vout->yuv.p_offset );
 160 }
 161
 162 /*****************************************************************************
 163  * yuv_Reset: re-initialize translations tables
 164  *****************************************************************************
 165  * This function will initialize the tables allocated by vout_CreateTables and
 166  * set functions pointers.
 167  *****************************************************************************/
 168 static int yuv_Reset( vout_thread_t *p_vout )
 169 {
 170     yuv_End( p_vout );
 171     return( yuv_Init( p_vout ) );
 172 }
 173
 174 /*****************************************************************************
 175  * SetYUV: compute tables and set function pointers
 176  *****************************************************************************/
 177 static void SetYUV( vout_thread_t *p_vout )
 178 {
 179     int         i_index;                                  /* index in tables */
 180
 181     /*
 182      * Set pointers and build YUV tables
 183      */
 184     if( p_vout->b_grayscale )
 185     {
 186         /* Grayscale: build gray table */
 187         if( p_vout->i_bytes_per_pixel == 1 )
 188         {
 189             u16 bright[256], transp[256];
 190
 191             for( i_index = 0; i_index < 256; i_index++)
 192             {
 193                 bright[ i_index ] = i_index << 8;
 194                 transp[ i_index ] = 0;
 195             }
 196             /* the colors have been allocated, we can set the palette */
 197             p_vout->pf_setpalette( p_vout, bright, bright, bright, transp );
 198             p_vout->i_white_pixel = 0xff;
 199             p_vout->i_black_pixel = 0x00;
 200             p_vout->i_gray_pixel = 0x44;
 201             p_vout->i_blue_pixel = 0x3b;
 202         }
 203     }
 204     else
 205     {
 206         /* Color: build red, green and blue tables */
 207         if( p_vout->i_bytes_per_pixel == 1 )
 208         {
 209             #define RGB_MIN 0
 210             #define RGB_MAX 255
 211             #define CLIP( x ) ( ((x < 0) ? 0 : (x > 255) ? 255 : x) << 8 )
 212             #define SHIFT 20
 213             #define U_GREEN_COEF    ((int)(-0.391 * (1<<SHIFT) / 1.164))
 214             #define U_BLUE_COEF     ((int)(2.018 * (1<<SHIFT) / 1.164))
 215             #define V_RED_COEF      ((int)(1.596 * (1<<SHIFT) / 1.164))
 216             #define V_GREEN_COEF    ((int)(-0.813 * (1<<SHIFT) / 1.164))
 217
 218             int y,u,v;
 219             int r,g,b;
 220             int uvr, uvg, uvb;
 221             int i = 0, j = 0;
 222             u16 red[256], green[256], blue[256], transp[256];
 223             unsigned char lookup[PALETTE_TABLE_SIZE];
 224
 225             p_vout->yuv.yuv.p_rgb8 = (u8 *)p_vout->yuv.p_base;
 226
 227             /* this loop calculates the intersection of an YUV box
 228              * and the RGB cube. */
 229             for ( y = 0; y <= 256; y += 16 )
 230             {
 231                 for ( u = 0; u <= 256; u += 32 )
 232                 for ( v = 0; v <= 256; v += 32 )
 233                 {
 234                     uvr = (V_RED_COEF*(v-128)) >> SHIFT;
 235                     uvg = (U_GREEN_COEF*(u-128) + V_GREEN_COEF*(v-128)) >> SHIFT;
 236                     uvb = (U_BLUE_COEF*(u-128)) >> SHIFT;
 237                     r = y + uvr;
 238                     g = y + uvg;
 239                     b = y + uvb;
 240
 241                     if( r >= RGB_MIN && g >= RGB_MIN && b >= RGB_MIN
 242                             && r <= RGB_MAX && g <= RGB_MAX && b <= RGB_MAX )
 243                     {
 244                         /* this one should never happen unless someone fscked up my code */
 245                         if(j == 256) { intf_ErrMsg( "vout error: no colors left to build palette" ); break; }
 246
 247                         /* clip the colors */
 248                         red[j] = CLIP( r );
 249                         green[j] = CLIP( g );
 250                         blue[j] = CLIP( b );
 251                         transp[j] = 0;
 252
 253                         /* allocate color */
 254                         lookup[i] = 1;
 255                         p_vout->yuv.yuv.p_rgb8[i++] = j;
 256                         j++;
 257                     }
 258                     else
 259                     {
 260                         lookup[i] = 0;
 261                         p_vout->yuv.yuv.p_rgb8[i++] = 0;
 262                     }
 263                 }
 264                 i += 128-81;
 265             }
 266
 267             /* the colors have been allocated, we can set the palette */
 268             /* there will eventually be a way to know which colors
 269              * couldn't be allocated and try to find a replacement */
 270             p_vout->pf_setpalette( p_vout, red, green, blue, transp );
 271
 272             p_vout->i_white_pixel = 0xff;
 273             p_vout->i_black_pixel = 0x00;
 274             p_vout->i_gray_pixel = 0x44;
 275             p_vout->i_blue_pixel = 0x3b;
 276
 277             i = 0;
 278             /* this loop allocates colors that got outside
 279              * the RGB cube */
 280             for ( y = 0; y <= 256; y += 16 )
 281             {
 282                 for ( u = 0; u <= 256; u += 32 )
 283                 {
 284                     for ( v = 0; v <= 256; v += 32 )
 285                     {
 286                         int u2, v2;
 287                         int dist, mindist = 100000000;
 288
 289                         if( lookup[i] || y==0)
 290                         {
 291                             i++;
 292                             continue;
 293                         }
 294
 295                         /* heavy. yeah. */
 296                         for( u2 = 0; u2 <= 256; u2 += 32 )
 297                         for( v2 = 0; v2 <= 256; v2 += 32 )
 298                         {
 299                             j = ((y>>4)<<7) + (u2>>5)*9 + (v2>>5);
 300                             dist = (u-u2)*(u-u2) + (v-v2)*(v-v2);
 301                             if( lookup[j] )
 302                             /* find the nearest color */
 303                             if( dist < mindist )
 304                             {
 305                                 p_vout->yuv.yuv.p_rgb8[i] = p_vout->yuv.yuv.p_rgb8[j];
 306                                 mindist = dist;
 307                             }
 308                             j -= 128;
 309                             if( lookup[j] )
 310                             /* find the nearest color */
 311                             if( dist + 128 < mindist )
 312                             {
 313                                 p_vout->yuv.yuv.p_rgb8[i] = p_vout->yuv.yuv.p_rgb8[j];
 314                                 mindist = dist + 128;
 315                             }
 316                         }
 317                         i++;
 318                     }
 319                 }
 320                 i += 128-81;
 321             }
 322         }
 323     }
 324
 325     /*
 326      * Set functions pointers
 327      */
 328     if( p_vout->b_YCbr)
 329     {
 330         switch( p_vout->i_bytes_per_pixel)
 331         {
 332         case 1:
 333             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420YCbr8;
 334             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertYUV422YCbr8;
 335             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertYUV444YCbr8;
 336             break;
 337
 338         case 2:
 339             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420YCbr16;
 340             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertYUV422YCbr16;
 341             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertYUV444YCbr16;
 342             break;
 343
 344         case 3:
 345             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420YCbr24;
 346             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertYUV422YCbr24;
 347             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertYUV444YCbr24;
 348             break;
 349
 350         case 4:
 351             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420YCbr32;
 352             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertYUV422YCbr32;
 353             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertYUV444YCbr32;
 354             break;
 355         }
 356     }
 357     else if( p_vout->b_grayscale )
 358     {
 359         /* Grayscale */
 360         switch( p_vout->i_bytes_per_pixel )
 361         {
 362         case 1:
 363             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertY4Gray8;
 364             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertY4Gray8;
 365             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertY4Gray8;
 366             break;
 367         case 2:
 368             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertY4Gray16;
 369             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertY4Gray16;
 370             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertY4Gray16;
 371             break;
 372         case 3:
 373             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420RGB24;
 374             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertY4Gray24;
 375             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertY4Gray24;
 376             break;
 377         case 4:
 378             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420RGB32;
 379             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertY4Gray32;
 380             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertY4Gray32;
 381             break;
 382         }
 383     }
 384     else
 385     {
 386         /* Color */
 387         switch( p_vout->i_bytes_per_pixel )
 388         {
 389         case 1:
 390             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420RGB8;
 391             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertYUV422RGB8;
 392             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertYUV444RGB8;
 393             break;
 394         case 2:
 395             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420RGB16;
 396             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertYUV422RGB16;
 397             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertYUV444RGB16;
 398             break;
 399         case 3:
 400             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420RGB24;
 401             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertYUV422RGB24;
 402             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertYUV444RGB24;
 403             break;
 404         case 4:
 405             p_vout->yuv.pf_yuv420 = (vout_yuv_convert_t *) ConvertYUV420RGB32;
 406             p_vout->yuv.pf_yuv422 = (vout_yuv_convert_t *) ConvertYUV422RGB32;
 407             p_vout->yuv.pf_yuv444 = (vout_yuv_convert_t *) ConvertYUV444RGB32;
 408             break;
 409         }
 410     }
 411 }
 412
 413 /*****************************************************************************
 414  * SetOffset: build offset array for conversion functions
 415  *****************************************************************************
 416  * This function will build an offset array used in later conversion functions.
 417  * It will also set horizontal and vertical scaling indicators. If b_double
 418  * is set, the p_offset structure has interleaved Y and U/V offsets.
 419  *****************************************************************************/
 420 void SetOffset( int i_width, int i_height, int i_pic_width, int i_pic_height,
 421                 boolean_t *pb_h_scaling, int *pi_v_scaling,
 422                 int *p_offset, boolean_t b_double )
 423 {
 424     int i_x;                                    /* x position in destination */
 425     int i_scale_count;                                     /* modulo counter */
 426
 427     /*
 428      * Prepare horizontal offset array
 429      */
 430     if( i_pic_width - i_width == 0 )
 431     {
 432         /* No horizontal scaling: YUV conversion is done directly to picture */
 433         *pb_h_scaling = 0;
 434     }
 435     else if( i_pic_width - i_width > 0 )
 436     {
 437         /* Prepare scaling array for horizontal extension */
 438         *pb_h_scaling =  1;
 439         i_scale_count =  i_pic_width;
 440         for( i_x = i_width; i_x--; )
 441         {
 442             while( (i_scale_count -= i_width) > 0 )
 443             {
 444                 *p_offset++ = 0;
 445             }
 446             *p_offset++ = 1;
 447             i_scale_count += i_pic_width;
 448         }
 449     }
 450     else /* if( i_pic_width - i_width < 0 ) */
 451     {
 452         /* Prepare scaling array for horizontal reduction */
 453         *pb_h_scaling =  1;
 454         i_scale_count =  i_width;
 455         for( i_x = i_pic_width; i_x--; )
 456         {
 457             *p_offset = 1;
 458             while( (i_scale_count -= i_pic_width) > 0 )
 459             {
 460                 *p_offset += 1;
 461             }
 462             p_offset++;
 463             i_scale_count += i_width;
 464         }
 465     }
 466
 467     /*
 468      * Set vertical scaling indicator
 469      */
 470     if( i_pic_height - i_height == 0 )
 471     {
 472         *pi_v_scaling = 0;
 473     }
 474     else if( i_pic_height - i_height > 0 )
 475     {
 476         *pi_v_scaling = 1;
 477     }
 478     else /* if( i_pic_height - i_height < 0 ) */
 479     {
 480         *pi_v_scaling = -1;
 481     }
 482 }
 483