video_chroma: added I420_ABGR32 support (mostly for opengl), some clean up as well
[vlc.git] / modules / video_chroma / i420_rgb.h
1 /*****************************************************************************
2  * i420_rgb.h : YUV to bitmap RGB conversion module for vlc
3  *****************************************************************************
4  * Copyright (C) 2000, 2004 the VideoLAN team
5  * $Id$
6  *
7  * Authors: Samuel Hocevar <sam@zoy.org>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 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 General Public License for more details.
18  *
19  * You should have received a copy of the GNU General Public License
20  * along with this program; if not, write to the Free Software
21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
22  *****************************************************************************/
23
24 /** Number of entries in RGB palette/colormap */
25 #define CMAP_RGB2_SIZE 256
26
27 /**
28  * chroma_sys_t: chroma method descriptor
29
30  * This structure is part of the chroma transformation descriptor, it
31  * describes the yuv2rgb specific properties.
32  */
33 struct chroma_sys_t
34 {
35     uint8_t  *p_buffer;
36     int *p_offset;
37
38 #ifdef MODULE_NAME_IS_i420_rgb
39     /**< Pre-calculated conversion tables */
40     void *p_base;                      /**< base for all conversion tables */
41     uint8_t   *p_rgb8;                 /**< RGB 8 bits table */
42     uint16_t  *p_rgb16;                /**< RGB 16 bits table */
43     uint32_t  *p_rgb32;                /**< RGB 32 bits table */
44
45     /**< To get RGB value for palette entry i, use (p_rgb_r[i], p_rgb_g[i],
46        p_rgb_b[i]). Note these are 16 bits per pixel. For 8bpp entries,
47        shift right 8 bits.
48     */
49     uint16_t  p_rgb_r[CMAP_RGB2_SIZE];  /**< Red values of palette */
50     uint16_t  p_rgb_g[CMAP_RGB2_SIZE];  /**< Green values of palette */
51     uint16_t  p_rgb_b[CMAP_RGB2_SIZE];  /**< Blue values of palette */
52 #endif
53 };
54
55 /*****************************************************************************
56  * Prototypes
57  *****************************************************************************/
58 #ifdef MODULE_NAME_IS_i420_rgb
59 void E_(I420_RGB8)         ( vout_thread_t *, picture_t *, picture_t * );
60 void E_(I420_RGB16_dither) ( vout_thread_t *, picture_t *, picture_t * );
61 void E_(I420_RGB16)        ( vout_thread_t *, picture_t *, picture_t * );
62 void E_(I420_RGB32)        ( vout_thread_t *, picture_t *, picture_t * );
63 #else // if defined(MODULE_NAME_IS_i420_rgb_mmx)
64 void E_(I420_R5G5B5)       ( vout_thread_t *, picture_t *, picture_t * );
65 void E_(I420_R5G6B5)       ( vout_thread_t *, picture_t *, picture_t * );
66 void E_(I420_A8R8G8B8)     ( vout_thread_t *, picture_t *, picture_t * );
67 void E_(I420_B8G8R8A8)     ( vout_thread_t *, picture_t *, picture_t * );
68 void E_(I420_A8B8G8R8)     ( vout_thread_t *, picture_t *, picture_t * );
69 #endif
70
71 /*****************************************************************************
72  * CONVERT_*_PIXEL: pixel conversion macros
73  *****************************************************************************
74  * These conversion routines are used by YUV conversion functions.
75  * conversion are made from p_y, p_u, p_v, which are modified, to p_buffer,
76  * which is also modified. CONVERT_4YUV_PIXEL is used for 8bpp dithering,
77  * CONVERT_4YUV_PIXEL_SCALE does the same but also scales the output.
78  *****************************************************************************/
79 #define CONVERT_Y_PIXEL( BPP )                                                \
80     /* Only Y sample is present */                                            \
81     p_ybase = p_yuv + *p_y++;                                                 \
82     *p_buffer++ = p_ybase[RED_OFFSET-((V_RED_COEF*128)>>SHIFT) + i_red] |     \
83         p_ybase[GREEN_OFFSET-(((U_GREEN_COEF+V_GREEN_COEF)*128)>>SHIFT)       \
84         + i_green ] | p_ybase[BLUE_OFFSET-((U_BLUE_COEF*128)>>SHIFT) + i_blue];
85
86 #define CONVERT_YUV_PIXEL( BPP )                                              \
87     /* Y, U and V samples are present */                                      \
88     i_uval =    *p_u++;                                                       \
89     i_vval =    *p_v++;                                                       \
90     i_red =     (V_RED_COEF * i_vval) >> SHIFT;                               \
91     i_green =   (U_GREEN_COEF * i_uval + V_GREEN_COEF * i_vval) >> SHIFT;     \
92     i_blue =    (U_BLUE_COEF * i_uval) >> SHIFT;                              \
93     CONVERT_Y_PIXEL( BPP )                                                    \
94
95 #define CONVERT_Y_PIXEL_DITHER( BPP )                                         \
96     /* Only Y sample is present */                                            \
97     p_ybase = p_yuv + *p_y++;                                                 \
98     *p_buffer++ = p_ybase[RED_OFFSET-((V_RED_COEF*128+p_dither[i_real_y])>>SHIFT) + i_red] |     \
99         p_ybase[GREEN_OFFSET-(((U_GREEN_COEF+V_GREEN_COEF)*128+p_dither[i_real_y])>>SHIFT)       \
100         + i_green ] | p_ybase[BLUE_OFFSET-((U_BLUE_COEF*128+p_dither[i_real_y])>>SHIFT) + i_blue];
101
102 #define CONVERT_YUV_PIXEL_DITHER( BPP )                                       \
103     /* Y, U and V samples are present */                                      \
104     i_uval =    *p_u++;                                                       \
105     i_vval =    *p_v++;                                                       \
106     i_red =     (V_RED_COEF * i_vval) >> SHIFT;                               \
107     i_green =   (U_GREEN_COEF * i_uval + V_GREEN_COEF * i_vval) >> SHIFT;     \
108     i_blue =    (U_BLUE_COEF * i_uval) >> SHIFT;                              \
109     CONVERT_Y_PIXEL_DITHER( BPP )                                             \
110
111 #define CONVERT_4YUV_PIXEL( CHROMA )                                          \
112     *p_pic++ = p_lookup[                                                      \
113         (((*p_y++ + dither10[i_real_y]) >> 4) << 7)                           \
114       + ((*p_u + dither20[i_real_y]) >> 5) * 9                                \
115       + ((*p_v + dither20[i_real_y]) >> 5) ];                                 \
116     *p_pic++ = p_lookup[                                                      \
117         (((*p_y++ + dither11[i_real_y]) >> 4) << 7)                           \
118       + ((*p_u++ + dither21[i_real_y]) >> 5) * 9                              \
119       + ((*p_v++ + dither21[i_real_y]) >> 5) ];                               \
120     *p_pic++ = p_lookup[                                                      \
121         (((*p_y++ + dither12[i_real_y]) >> 4) << 7)                           \
122       + ((*p_u + dither22[i_real_y]) >> 5) * 9                                \
123       + ((*p_v + dither22[i_real_y]) >> 5) ];                                 \
124     *p_pic++ = p_lookup[                                                      \
125         (((*p_y++ + dither13[i_real_y]) >> 4) << 7)                           \
126       + ((*p_u++ + dither23[i_real_y]) >> 5) * 9                              \
127       + ((*p_v++ + dither23[i_real_y]) >> 5) ];                               \
128
129 #define CONVERT_4YUV_PIXEL_SCALE( CHROMA )                                    \
130     *p_pic++ = p_lookup[                                                      \
131         ( ((*p_y + dither10[i_real_y]) >> 4) << 7)                            \
132         + ((*p_u + dither20[i_real_y]) >> 5) * 9                              \
133         + ((*p_v + dither20[i_real_y]) >> 5) ];                               \
134     p_y += *p_offset++;                                                       \
135     p_u += *p_offset;                                                         \
136     p_v += *p_offset++;                                                       \
137     *p_pic++ = p_lookup[                                                      \
138         ( ((*p_y + dither11[i_real_y]) >> 4) << 7)                            \
139         + ((*p_u + dither21[i_real_y]) >> 5) * 9                              \
140         + ((*p_v + dither21[i_real_y]) >> 5) ];                               \
141     p_y += *p_offset++;                                                       \
142     p_u += *p_offset;                                                         \
143     p_v += *p_offset++;                                                       \
144     *p_pic++ = p_lookup[                                                      \
145         ( ((*p_y + dither12[i_real_y]) >> 4) << 7)                            \
146         + ((*p_u + dither22[i_real_y]) >> 5) * 9                              \
147         + ((*p_v + dither22[i_real_y]) >> 5) ];                               \
148     p_y += *p_offset++;                                                       \
149     p_u += *p_offset;                                                         \
150     p_v += *p_offset++;                                                       \
151     *p_pic++ = p_lookup[                                                      \
152         ( ((*p_y + dither13[i_real_y]) >> 4) << 7)                            \
153         + ((*p_u + dither23[i_real_y]) >> 5) * 9                              \
154         + ((*p_v + dither23[i_real_y]) >> 5) ];                               \
155     p_y += *p_offset++;                                                       \
156     p_u += *p_offset;                                                         \
157     p_v += *p_offset++;                                                       \
158
159 /*****************************************************************************
160  * SCALE_WIDTH: scale a line horizontally
161  *****************************************************************************
162  * This macro scales a line using rendering buffer and offset array. It works
163  * for 1, 2 and 4 Bpp.
164  *****************************************************************************/
165 #define SCALE_WIDTH                                                           \
166     if( b_hscale )                                                            \
167     {                                                                         \
168         /* Horizontal scaling, conversion has been done to buffer.            \
169          * Rewind buffer and offset, then copy and scale line */              \
170         p_buffer = p_buffer_start;                                            \
171         p_offset = p_offset_start;                                            \
172         for( i_x = p_vout->output.i_width / 16; i_x--; )                      \
173         {                                                                     \
174             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
175             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
176             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
177             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
178             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
179             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
180             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
181             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
182             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
183             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
184             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
185             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
186             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
187             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
188             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
189             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
190         }                                                                     \
191         for( i_x = p_vout->output.i_width & 15; i_x--; )                      \
192         {                                                                     \
193             *p_pic++ = *p_buffer;   p_buffer += *p_offset++;                  \
194         }                                                                     \
195         p_pic = (void*)((uint8_t*)p_pic + i_right_margin );                   \
196     }                                                                         \
197     else                                                                      \
198     {                                                                         \
199         /* No scaling, conversion has been done directly in picture memory.   \
200          * Increment of picture pointer to end of line is still needed */     \
201         p_pic = (void*)((uint8_t*)p_pic + p_dest->p->i_pitch );               \
202     }                                                                         \
203
204 /*****************************************************************************
205  * SCALE_WIDTH_DITHER: scale a line horizontally for dithered 8 bpp
206  *****************************************************************************
207  * This macro scales a line using an offset array.
208  *****************************************************************************/
209 #define SCALE_WIDTH_DITHER( CHROMA )                                          \
210     if( b_hscale )                                                            \
211     {                                                                         \
212         /* Horizontal scaling - we can't use a buffer due to dithering */     \
213         p_offset = p_offset_start;                                            \
214         for( i_x = p_vout->output.i_width / 16; i_x--; )                      \
215         {                                                                     \
216             CONVERT_4YUV_PIXEL_SCALE( CHROMA )                                \
217             CONVERT_4YUV_PIXEL_SCALE( CHROMA )                                \
218             CONVERT_4YUV_PIXEL_SCALE( CHROMA )                                \
219             CONVERT_4YUV_PIXEL_SCALE( CHROMA )                                \
220         }                                                                     \
221     }                                                                         \
222     else                                                                      \
223     {                                                                         \
224         for( i_x = p_vout->render.i_width / 16; i_x--;  )                     \
225         {                                                                     \
226             CONVERT_4YUV_PIXEL( CHROMA )                                      \
227             CONVERT_4YUV_PIXEL( CHROMA )                                      \
228             CONVERT_4YUV_PIXEL( CHROMA )                                      \
229             CONVERT_4YUV_PIXEL( CHROMA )                                      \
230         }                                                                     \
231     }                                                                         \
232     /* Increment of picture pointer to end of line is still needed */         \
233     p_pic = (void*)((uint8_t*)p_pic + i_right_margin );                       \
234                                                                               \
235     /* Increment the Y coordinate in the matrix, modulo 4 */                  \
236     i_real_y = (i_real_y + 1) & 0x3;                                          \
237
238 /*****************************************************************************
239  * SCALE_HEIGHT: handle vertical scaling
240  *****************************************************************************
241  * This macro handle vertical scaling for a picture. CHROMA may be 420, 422 or
242  * 444 for RGB conversion, or 400 for gray conversion. It works for 1, 2, 3
243  * and 4 Bpp.
244  *****************************************************************************/
245 #define SCALE_HEIGHT( CHROMA, BPP )                                           \
246     /* If line is odd, rewind 4:2:0 U and V samples */                        \
247     if( ((CHROMA == 420) || (CHROMA == 422)) && !(i_y & 0x1) )                \
248     {                                                                         \
249         p_u -= i_chroma_width;                                                \
250         p_v -= i_chroma_width;                                                \
251     }                                                                         \
252                                                                               \
253     /*                                                                        \
254      * Handle vertical scaling. The current line can be copied or next one    \
255      * can be ignored.                                                        \
256      */                                                                       \
257     switch( i_vscale )                                                        \
258     {                                                                         \
259     case -1:                             /* vertical scaling factor is < 1 */ \
260         while( (i_scale_count -= p_vout->output.i_height) > 0 )               \
261         {                                                                     \
262             /* Height reduction: skip next source line */                     \
263             p_y += p_vout->render.i_width;                                    \
264             i_y++;                                                            \
265             if( (CHROMA == 420) || (CHROMA == 422) )                          \
266             {                                                                 \
267                 if( i_y & 0x1 )                                               \
268                 {                                                             \
269                     p_u += i_chroma_width;                                    \
270                     p_v += i_chroma_width;                                    \
271                 }                                                             \
272             }                                                                 \
273             else if( CHROMA == 444 )                                          \
274             {                                                                 \
275                 p_u += p_vout->render.i_width;                                \
276                 p_v += p_vout->render.i_width;                                \
277             }                                                                 \
278         }                                                                     \
279         i_scale_count += p_vout->render.i_height;                             \
280         break;                                                                \
281     case 1:                              /* vertical scaling factor is > 1 */ \
282         while( (i_scale_count -= p_vout->render.i_height) > 0 )               \
283         {                                                                     \
284             /* Height increment: copy previous picture line */                \
285             p_vout->p_libvlc->pf_memcpy( p_pic, p_pic_start,                     \
286                                       p_vout->output.i_width * BPP );         \
287             p_pic = (void*)((uint8_t*)p_pic + p_dest->p->i_pitch );           \
288         }                                                                     \
289         i_scale_count += p_vout->output.i_height;                             \
290         break;                                                                \
291     }                                                                         \
292
293 /*****************************************************************************
294  * SCALE_HEIGHT_DITHER: handle vertical scaling for dithered 8 bpp
295  *****************************************************************************
296  * This macro handles vertical scaling for a picture. CHROMA may be 420,
297  * 422 or 444 for RGB conversion, or 400 for gray conversion.
298  *****************************************************************************/
299 #define SCALE_HEIGHT_DITHER( CHROMA )                                         \
300                                                                               \
301     /* If line is odd, rewind 4:2:0 U and V samples */                        \
302     if( ((CHROMA == 420) || (CHROMA == 422)) && !(i_y & 0x1) )                \
303     {                                                                         \
304         p_u -= i_chroma_width;                                                \
305         p_v -= i_chroma_width;                                                \
306     }                                                                         \
307                                                                               \
308     /*                                                                        \
309      * Handle vertical scaling. The current line can be copied or next one    \
310      * can be ignored.                                                        \
311      */                                                                       \
312                                                                               \
313     switch( i_vscale )                                                        \
314     {                                                                         \
315     case -1:                             /* vertical scaling factor is < 1 */ \
316         while( (i_scale_count -= p_vout->output.i_height) > 0 )               \
317         {                                                                     \
318             /* Height reduction: skip next source line */                     \
319             p_y += p_vout->render.i_width;                                    \
320             i_y++;                                                            \
321             if( (CHROMA == 420) || (CHROMA == 422) )                          \
322             {                                                                 \
323                 if( i_y & 0x1 )                                               \
324                 {                                                             \
325                     p_u += i_chroma_width;                                    \
326                     p_v += i_chroma_width;                                    \
327                 }                                                             \
328             }                                                                 \
329             else if( CHROMA == 444 )                                          \
330             {                                                                 \
331                 p_u += p_vout->render.i_width;                                \
332                 p_v += p_vout->render.i_width;                                \
333             }                                                                 \
334         }                                                                     \
335         i_scale_count += p_vout->render.i_height;                             \
336         break;                                                                \
337     case 1:                              /* vertical scaling factor is > 1 */ \
338         while( (i_scale_count -= p_vout->render.i_height) > 0 )               \
339         {                                                                     \
340             p_y -= p_vout->render.i_width;                                    \
341             p_u -= i_chroma_width;                                            \
342             p_v -= i_chroma_width;                                            \
343             SCALE_WIDTH_DITHER( CHROMA );                                     \
344         }                                                                     \
345         i_scale_count += p_vout->output.i_height;                             \
346         break;                                                                \
347     }                                                                         \
348