demux/playlist: xspf: only use text-elements inside current tag
[vlc.git] / modules / video_filter / sepia.c
blob2a00130fbfe19beefd163fd8c86ca584ac0d561c
1 /*****************************************************************************
2 * sepia.c : Sepia video plugin for vlc
3 *****************************************************************************
4 * Copyright (C) 2010 VLC authors and VideoLAN
5 * $Id$
7 * Authors: Branko Kokanovic <branko.kokanovic@gmail.com>
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.
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.
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 *****************************************************************************/
24 /*****************************************************************************
25 * Preamble
26 *****************************************************************************/
28 #ifdef HAVE_CONFIG_H
29 # include "config.h"
30 #endif
32 #include <vlc_common.h>
33 #include <vlc_plugin.h>
34 #include <vlc_filter.h>
35 #include <vlc_picture.h>
36 #include <vlc_cpu.h>
37 #include <vlc_atomic.h>
39 #include <assert.h>
40 #include "filter_picture.h"
42 /*****************************************************************************
43 * Local prototypes
44 *****************************************************************************/
45 static int Create ( vlc_object_t * );
46 static void Destroy ( vlc_object_t * );
48 static void RVSepia( picture_t *, picture_t *, int );
49 static void PlanarI420Sepia( picture_t *, picture_t *, int);
50 static void PackedYUVSepia( picture_t *, picture_t *, int);
51 static picture_t *Filter( filter_t *, picture_t * );
52 static const char *const ppsz_filter_options[] = {
53 "intensity", NULL
56 /*****************************************************************************
57 * Module descriptor
58 *****************************************************************************/
59 #define SEPIA_INTENSITY_TEXT N_("Sepia intensity")
60 #define SEPIA_INTENSITY_LONGTEXT N_("Intensity of sepia effect" )
62 #define CFG_PREFIX "sepia-"
64 vlc_module_begin ()
65 set_description( N_("Sepia video filter") )
66 set_shortname( N_("Sepia" ) )
67 set_help( N_("Gives video a warmer tone by applying sepia effect") )
68 set_category( CAT_VIDEO )
69 set_subcategory( SUBCAT_VIDEO_VFILTER )
70 set_capability( "video filter", 0 )
71 add_integer_with_range( CFG_PREFIX "intensity", 120, 0, 255,
72 SEPIA_INTENSITY_TEXT, SEPIA_INTENSITY_LONGTEXT,
73 false )
74 set_callbacks( Create, Destroy )
75 vlc_module_end ()
77 /*****************************************************************************
78 * callback prototypes
79 *****************************************************************************/
80 static int FilterCallback( vlc_object_t *, char const *,
81 vlc_value_t, vlc_value_t, void * );
83 typedef void (*SepiaFunction)( picture_t *, picture_t *, int );
85 static const struct
87 vlc_fourcc_t i_chroma;
88 SepiaFunction pf_sepia;
89 } p_sepia_cfg[] = {
90 { VLC_CODEC_I420, PlanarI420Sepia },
91 { VLC_CODEC_RGB24, RVSepia },
92 { VLC_CODEC_RGB32, RVSepia },
93 { VLC_CODEC_UYVY, PackedYUVSepia },
94 { VLC_CODEC_VYUY, PackedYUVSepia },
95 { VLC_CODEC_YUYV, PackedYUVSepia },
96 { VLC_CODEC_YVYU, PackedYUVSepia },
97 { 0, NULL }
100 /*****************************************************************************
101 * filter_sys_t: adjust filter method descriptor
102 *****************************************************************************/
103 struct filter_sys_t
105 SepiaFunction pf_sepia;
106 atomic_int i_intensity;
109 /*****************************************************************************
110 * Create: allocates Sepia video thread output method
111 *****************************************************************************
112 * This function allocates and initializes a Sepia vout method.
113 *****************************************************************************/
114 static int Create( vlc_object_t *p_this )
116 filter_t *p_filter = (filter_t *)p_this;
117 filter_sys_t *p_sys;
119 /* Allocate structure */
120 p_sys = p_filter->p_sys = malloc( sizeof( filter_sys_t ) );
121 if( p_filter->p_sys == NULL )
122 return VLC_ENOMEM;
124 p_sys->pf_sepia = NULL;
126 for( int i = 0; p_sepia_cfg[i].i_chroma != 0; i++ )
128 if( p_sepia_cfg[i].i_chroma != p_filter->fmt_in.video.i_chroma )
129 continue;
130 p_sys->pf_sepia = p_sepia_cfg[i].pf_sepia;
133 if( p_sys->pf_sepia == NULL )
135 msg_Err( p_filter, "Unsupported input chroma (%4.4s)",
136 (char*)&(p_filter->fmt_in.video.i_chroma) );
137 free( p_sys );
138 return VLC_EGENERIC;
141 config_ChainParse( p_filter, CFG_PREFIX, ppsz_filter_options,
142 p_filter->p_cfg );
143 atomic_init( &p_sys->i_intensity,
144 var_CreateGetIntegerCommand( p_filter, CFG_PREFIX "intensity" ) );
145 var_AddCallback( p_filter, CFG_PREFIX "intensity", FilterCallback, NULL );
147 p_filter->pf_video_filter = Filter;
149 return VLC_SUCCESS;
152 /*****************************************************************************
153 * Destroy: destroy sepia video thread output method
154 *****************************************************************************
155 * Terminate an output method
156 *****************************************************************************/
157 static void Destroy( vlc_object_t *p_this )
159 filter_t *p_filter = (filter_t *)p_this;
161 var_DelCallback( p_filter, CFG_PREFIX "intensity", FilterCallback, NULL );
163 free( p_filter->p_sys );
166 /*****************************************************************************
167 * Render: displays previously rendered output
168 *****************************************************************************
169 * This function send the currently rendered image to sepia image, waits
170 * until it is displayed and switch the two rendering buffers, preparing next
171 * frame.
172 *****************************************************************************/
173 static picture_t *Filter( filter_t *p_filter, picture_t *p_pic )
175 picture_t *p_outpic;
177 if( !p_pic ) return NULL;
179 filter_sys_t *p_sys = p_filter->p_sys;
180 int intensity = atomic_load( &p_sys->i_intensity );
182 p_outpic = filter_NewPicture( p_filter );
183 if( !p_outpic )
185 msg_Warn( p_filter, "can't get output picture" );
186 picture_Release( p_pic );
187 return NULL;
190 p_sys->pf_sepia( p_pic, p_outpic, intensity );
192 return CopyInfoAndRelease( p_outpic, p_pic );
195 #if defined(CAN_COMPILE_SSE2)
196 /*****************************************************************************
197 * Sepia8ySSE2
198 *****************************************************************************
199 * This function applies sepia effect to eight bytes of yellow using SSE4.1
200 * instructions. It copies those 8 bytes to 128b register and fills the gaps
201 * with zeroes and following operations are made with word-operating instructs.
202 *****************************************************************************/
203 VLC_SSE
204 static inline void Sepia8ySSE2(uint8_t * dst, const uint8_t * src,
205 int i_intensity_spread)
207 __asm__ volatile (
208 // y = y - y / 4 + i_intensity / 4
209 "movq (%1), %%xmm1\n"
210 "punpcklbw %%xmm7, %%xmm1\n"
211 "movq (%1), %%xmm2\n" // store bytes as words with 0s in between
212 "punpcklbw %%xmm7, %%xmm2\n"
213 "movd %2, %%xmm3\n"
214 "pshufd $0, %%xmm3, %%xmm3\n"
215 "psrlw $2, %%xmm2\n" // rotate right 2
216 "psubusb %%xmm1, %%xmm2\n" // subtract
217 "psrlw $2, %%xmm3\n"
218 "paddsb %%xmm1, %%xmm3\n" // add
219 "packuswb %%xmm2, %%xmm1\n" // pack back to bytes
220 "movq %%xmm1, (%0) \n" // load to dest
222 :"r" (dst), "r"(src), "r"(i_intensity_spread)
223 :"memory", "xmm1", "xmm2", "xmm3");
226 VLC_SSE
227 static void PlanarI420SepiaSSE( picture_t *p_pic, picture_t *p_outpic,
228 int i_intensity )
230 /* prepared values to copy for U and V channels */
231 const uint8_t filling_const_8u = 128 - i_intensity / 6;
232 const uint8_t filling_const_8v = 128 + i_intensity / 14;
233 /* prepared value for faster broadcasting in xmm register */
234 int i_intensity_spread = 0x10001 * (uint8_t) i_intensity;
236 __asm__ volatile(
237 "pxor %%xmm7, %%xmm7\n"
238 ::: "xmm7");
240 /* iterate for every two visible line in the frame */
241 for (int y = 0; y < p_pic->p[Y_PLANE].i_visible_lines - 1; y += 2)
243 const int i_dy_line1_start = y * p_outpic->p[Y_PLANE].i_pitch;
244 const int i_dy_line2_start = (y + 1) * p_outpic->p[Y_PLANE].i_pitch;
245 const int i_du_line_start = (y / 2) * p_outpic->p[U_PLANE].i_pitch;
246 const int i_dv_line_start = (y / 2) * p_outpic->p[V_PLANE].i_pitch;
247 int x = 0;
248 /* iterate for every visible line in the frame (eight values at once) */
249 for ( ; x < p_pic->p[Y_PLANE].i_visible_pitch - 15; x += 16 )
251 /* Compute yellow channel values with asm function */
252 Sepia8ySSE2(&p_outpic->p[Y_PLANE].p_pixels[i_dy_line1_start + x],
253 &p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x],
254 i_intensity_spread );
255 Sepia8ySSE2(&p_outpic->p[Y_PLANE].p_pixels[i_dy_line2_start + x],
256 &p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x],
257 i_intensity_spread );
258 Sepia8ySSE2(&p_outpic->p[Y_PLANE].p_pixels[i_dy_line1_start + x + 8],
259 &p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x + 8],
260 i_intensity_spread );
261 Sepia8ySSE2(&p_outpic->p[Y_PLANE].p_pixels[i_dy_line2_start + x + 8],
262 &p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x + 8],
263 i_intensity_spread );
264 /* Copy precomputed values to destination memory location */
265 memset(&p_outpic->p[U_PLANE].p_pixels[i_du_line_start + (x / 2)],
266 filling_const_8u, 8 );
267 memset(&p_outpic->p[V_PLANE].p_pixels[i_dv_line_start + (x / 2)],
268 filling_const_8v, 8 );
270 /* Completing the job, the cycle above takes really big chunks, so
271 this makes sure the job will be done completely */
272 for ( ; x < p_pic->p[Y_PLANE].i_visible_pitch - 1; x += 2 )
274 // y = y - y/4 {to prevent overflow} + intensity / 4
275 p_outpic->p[Y_PLANE].p_pixels[i_dy_line1_start + x] =
276 p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x] -
277 (p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x] >> 2) +
278 (i_intensity >> 2);
279 p_outpic->p[Y_PLANE].p_pixels[i_dy_line1_start + x + 1] =
280 p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x + 1] -
281 (p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x + 1] >> 2) +
282 (i_intensity >> 2);
283 p_outpic->p[Y_PLANE].p_pixels[i_dy_line2_start + x] =
284 p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x] -
285 (p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x] >> 2) +
286 (i_intensity >> 2);
287 p_outpic->p[Y_PLANE].p_pixels[i_dy_line2_start + x + 1] =
288 p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x + 1] -
289 (p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x + 1] >> 2) +
290 (i_intensity >> 2);
291 // u = 128 {half => B&W} - intensity / 6
292 p_outpic->p[U_PLANE].p_pixels[i_du_line_start + (x / 2)] =
293 filling_const_8u;
294 // v = 128 {half => B&W} + intensity / 14
295 p_outpic->p[V_PLANE].p_pixels[i_dv_line_start + (x / 2)] =
296 filling_const_8v;
300 #endif
302 /*****************************************************************************
303 * PlanarI420Sepia: Applies sepia to one frame of the planar I420 video
304 *****************************************************************************
305 * This function applies sepia effect to one frame of the video by iterating
306 * through video lines. We iterate for every two lines and for every two pixels
307 * in line to calculate new sepia values for four y components as well for u
308 * and v components.
309 *****************************************************************************/
310 static void PlanarI420Sepia( picture_t *p_pic, picture_t *p_outpic,
311 int i_intensity )
313 #if defined(CAN_COMPILE_SSE2)
314 if (vlc_CPU_SSE2())
315 return PlanarI420SepiaSSE( p_pic, p_outpic, i_intensity );
316 #endif
318 // prepared values to copy for U and V channels
319 const uint8_t filling_const_8u = 128 - i_intensity / 6;
320 const uint8_t filling_const_8v = 128 + i_intensity / 14;
322 /* iterate for every two visible line in the frame */
323 for( int y = 0; y < p_pic->p[Y_PLANE].i_visible_lines - 1; y += 2 )
325 const int i_dy_line1_start = y * p_outpic->p[Y_PLANE].i_pitch;
326 const int i_dy_line2_start = ( y + 1 ) * p_outpic->p[Y_PLANE].i_pitch;
327 const int i_du_line_start = (y/2) * p_outpic->p[U_PLANE].i_pitch;
328 const int i_dv_line_start = (y/2) * p_outpic->p[V_PLANE].i_pitch;
329 // to prevent sigsegv if one pic is smaller (theoretically)
330 int i_picture_size_limit = p_pic->p[Y_PLANE].i_visible_pitch
331 < p_outpic->p[Y_PLANE].i_visible_pitch
332 ? (p_pic->p[Y_PLANE].i_visible_pitch - 1) :
333 (p_outpic->p[Y_PLANE].i_visible_pitch - 1);
334 /* iterate for every two visible line in the frame */
335 for( int x = 0; x < i_picture_size_limit; x += 2 )
337 // y = y - y/4 {to prevent overflow} + intensity / 4
338 p_outpic->p[Y_PLANE].p_pixels[i_dy_line1_start + x] =
339 p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x] -
340 (p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x] >> 2) +
341 (i_intensity >> 2);
342 p_outpic->p[Y_PLANE].p_pixels[i_dy_line1_start + x + 1] =
343 p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x + 1] -
344 (p_pic->p[Y_PLANE].p_pixels[i_dy_line1_start + x + 1] >> 2) +
345 (i_intensity >> 2);
346 p_outpic->p[Y_PLANE].p_pixels[i_dy_line2_start + x] =
347 p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x] -
348 (p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x] >> 2) +
349 (i_intensity >> 2);
350 p_outpic->p[Y_PLANE].p_pixels[i_dy_line2_start + x + 1] =
351 p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x + 1] -
352 (p_pic->p[Y_PLANE].p_pixels[i_dy_line2_start + x + 1] >> 2) +
353 (i_intensity >> 2);
354 // u = 128 {half => B&W} - intensity / 6
355 p_outpic->p[U_PLANE].p_pixels[i_du_line_start + (x / 2)] =
356 filling_const_8u;
357 // v = 128 {half => B&W} + intensity / 14
358 p_outpic->p[V_PLANE].p_pixels[i_dv_line_start + (x / 2)] =
359 filling_const_8v;
364 /*****************************************************************************
365 * PackedYUVSepia: Applies sepia to one frame of the packed YUV video
366 *****************************************************************************
367 * This function applies sepia effext to one frame of the video by iterating
368 * through video lines. In every pass, we calculate new values for pixels
369 * (UYVY, VYUY, YUYV and YVYU formats are supported)
370 *****************************************************************************/
371 static void PackedYUVSepia( picture_t *p_pic, picture_t *p_outpic,
372 int i_intensity )
374 uint8_t *p_in, *p_in_end, *p_line_end, *p_out;
375 int i_yindex = 1, i_uindex = 2, i_vindex = 0;
377 GetPackedYuvOffsets( p_outpic->format.i_chroma,
378 &i_yindex, &i_uindex, &i_vindex );
380 // prepared values to copy for U and V channels
381 const uint8_t filling_const_8u = 128 - i_intensity / 6;
382 const uint8_t filling_const_8v = 128 + i_intensity / 14;
384 p_in = p_pic->p[0].p_pixels;
385 p_in_end = p_in + p_pic->p[0].i_visible_lines
386 * p_pic->p[0].i_pitch;
387 p_out = p_outpic->p[0].p_pixels;
390 while( p_in < p_in_end )
392 p_line_end = p_in + p_pic->p[0].i_visible_pitch;
393 while( p_in < p_line_end )
395 /* calculate new, sepia values */
396 p_out[i_yindex] =
397 p_in[i_yindex] - (p_in[i_yindex] >> 2) + (i_intensity >> 2);
398 p_out[i_yindex + 2] =
399 p_in[i_yindex + 2] - (p_in[i_yindex + 2] >> 2)
400 + (i_intensity >> 2);
401 p_out[i_uindex] = filling_const_8u;
402 p_out[i_vindex] = filling_const_8v;
403 p_in += 4;
404 p_out += 4;
406 p_in += p_pic->p[0].i_pitch - p_pic->p[0].i_visible_pitch;
407 p_out += p_outpic->p[0].i_pitch
408 - p_outpic->p[0].i_visible_pitch;
413 /*****************************************************************************
414 * RVSepia: Applies sepia to one frame of the RV24/RV32 video
415 *****************************************************************************
416 * This function applies sepia effect to one frame of the video by iterating
417 * through video lines and calculating new values for every byte in chunks of
418 * 3 (RV24) or 4 (RV32) bytes.
419 *****************************************************************************/
420 static void RVSepia( picture_t *p_pic, picture_t *p_outpic, int i_intensity )
422 #define SCALEBITS 10
423 #define ONE_HALF (1 << (SCALEBITS - 1))
424 #define FIX(x) ((int) ((x) * (1<<SCALEBITS) + 0.5))
425 uint8_t *p_in, *p_in_end, *p_line_end, *p_out;
426 bool b_isRV32 = p_pic->format.i_chroma == VLC_CODEC_RGB32;
427 int i_rindex = 0, i_gindex = 1, i_bindex = 2;
429 GetPackedRgbIndexes( &p_outpic->format, &i_rindex, &i_gindex, &i_bindex );
431 p_in = p_pic->p[0].p_pixels;
432 p_in_end = p_in + p_pic->p[0].i_visible_lines
433 * p_pic->p[0].i_pitch;
434 p_out = p_outpic->p[0].p_pixels;
436 /* Precompute values constant for this certain i_intensity, using the same
437 * formula as YUV functions above */
438 uint8_t r_intensity = (( FIX( 1.40200 * 255.0 / 224.0 ) * (i_intensity * 14)
439 + ONE_HALF )) >> SCALEBITS;
440 uint8_t g_intensity = (( - FIX(0.34414*255.0/224.0) * ( - i_intensity / 6 )
441 - FIX( 0.71414 * 255.0 / 224.0) * ( i_intensity * 14 )
442 + ONE_HALF )) >> SCALEBITS;
443 uint8_t b_intensity = (( FIX( 1.77200 * 255.0 / 224.0) * ( - i_intensity / 6 )
444 + ONE_HALF )) >> SCALEBITS;
446 while (p_in < p_in_end)
448 p_line_end = p_in + p_pic->p[0].i_visible_pitch;
449 while (p_in < p_line_end)
451 /* do sepia: this calculation is based on the formula to calculate
452 * YUV->RGB and RGB->YUV (in filter_picture.h) mode and that
453 * y = y - y/4 + intensity/4 . As Y is the only channel that changes
454 * through the whole image. After that, precomputed values are added
455 * for each RGB channel and saved in the output image.
456 * FIXME: needs cleanup */
457 uint8_t i_y = ((( 66 * p_in[i_rindex] + 129 * p_in[i_gindex] + 25
458 * p_in[i_bindex] + 128 ) >> 8 ) * FIX(255.0/219.0))
459 - (((( 66 * p_in[i_rindex] + 129 * p_in[i_gindex] + 25
460 * p_in[i_bindex] + 128 ) >> 8 )
461 * FIX( 255.0 / 219.0 )) >> 2 ) + ( i_intensity >> 2 );
462 p_out[i_rindex] = vlc_uint8(i_y + r_intensity);
463 p_out[i_gindex] = vlc_uint8(i_y + g_intensity);
464 p_out[i_bindex] = vlc_uint8(i_y + b_intensity);
465 p_in += 3;
466 p_out += 3;
467 /* for rv32 we take 4 chunks at the time */
468 if (b_isRV32) {
469 /* alpha channel stays the same */
470 *p_out++ = *p_in++;
474 p_in += p_pic->p[0].i_pitch - p_pic->p[0].i_visible_pitch;
475 p_out += p_outpic->p[0].i_pitch
476 - p_outpic->p[0].i_visible_pitch;
478 #undef SCALEBITS
479 #undef ONE_HALF
480 #undef FIX
483 static int FilterCallback ( vlc_object_t *p_this, char const *psz_var,
484 vlc_value_t oldval, vlc_value_t newval,
485 void *p_data )
487 VLC_UNUSED(psz_var); VLC_UNUSED(oldval); VLC_UNUSED(p_data);
488 filter_t *p_filter = (filter_t*)p_this;
489 filter_sys_t *p_sys = p_filter->p_sys;
491 atomic_store( &p_sys->i_intensity, newval.i_int );
492 return VLC_SUCCESS;