1 /*****************************************************************************
2 * algo_phosphor.c : Phosphor algorithm for the VLC deinterlacer
3 *****************************************************************************
4 * Copyright (C) 2011 VLC authors and VideoLAN
7 * Author: Juha Jeronen <juha.jeronen@jyu.fi>
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 *****************************************************************************/
28 #ifdef CAN_COMPILE_MMXEXT
35 #include <vlc_common.h>
37 #include <vlc_picture.h>
38 #include <vlc_filter.h>
40 #include "deinterlace.h" /* filter_sys_t */
41 #include "helpers.h" /* ComposeFrame() */
43 #include "algo_phosphor.h"
45 /*****************************************************************************
47 *****************************************************************************/
50 * Internal helper function: dims (darkens) the given field
51 * of the given picture.
53 * This is used for simulating CRT light output decay in RenderPhosphor().
55 * The strength "1" is recommended. It's a matter of taste,
56 * so it's parametrized.
58 * Note on chroma formats:
59 * - If input is 4:2:2, all planes are processed.
60 * - If input is 4:2:0, only the luma plane is processed, because both fields
61 * have the same chroma. This will distort colours, especially for high
62 * filter strengths, especially for pixels whose U and/or V values are
63 * far away from the origin (which is at 128 in uint8 format).
65 * @param p_dst Input/output picture. Will be modified in-place.
66 * @param i_field Darken which field? 0 = top, 1 = bottom.
67 * @param i_strength Strength of effect: 1, 2 or 3 (division by 2, 4 or 8).
68 * @see RenderPhosphor()
71 static void DarkenField( picture_t *p_dst,
72 const int i_field, const int i_strength,
75 assert( p_dst != NULL );
76 assert( i_field == 0 || i_field == 1 );
77 assert( i_strength >= 1 && i_strength <= 3 );
79 /* Bitwise ANDing with this clears the i_strength highest bits
81 const uint8_t remove_high_u8 = 0xFF >> i_strength;
82 const uint64_t remove_high_u64 = remove_high_u8 *
83 INT64_C(0x0101010101010101);
87 For luma, the operation is just a shift + bitwise AND, so we vectorize
88 even in the C version.
90 There is an MMX version too, because it performs about twice faster.
92 int i_plane = Y_PLANE;
93 uint8_t *p_out, *p_out_end;
94 int w = p_dst->p[i_plane].i_visible_pitch;
95 p_out = p_dst->p[i_plane].p_pixels;
96 p_out_end = p_out + p_dst->p[i_plane].i_pitch
97 * p_dst->p[i_plane].i_visible_lines;
99 /* skip first line for bottom field */
101 p_out += p_dst->p[i_plane].i_pitch;
103 int wm8 = w % 8; /* remainder */
104 int w8 = w - wm8; /* part of width that is divisible by 8 */
105 for( ; p_out < p_out_end ; p_out += 2*p_dst->p[i_plane].i_pitch )
107 uint64_t *po = (uint64_t *)p_out;
110 for( ; x < w8; x += 8, ++po )
111 (*po) = ( ((*po) >> i_strength) & remove_high_u64 );
113 /* handle the width remainder */
114 uint8_t *po_temp = (uint8_t *)po;
115 for( ; x < w; ++x, ++po_temp )
116 (*po_temp) = ( ((*po_temp) >> i_strength) & remove_high_u8 );
119 /* Process chroma if the field chromas are independent.
121 The origin (black) is at YUV = (0, 128, 128) in the uint8 format.
122 The chroma processing is a bit more complicated than luma,
123 and needs MMX for vectorization.
127 for( i_plane++ /* luma already handled*/;
128 i_plane < p_dst->i_planes;
131 int w = p_dst->p[i_plane].i_visible_pitch;
132 p_out = p_dst->p[i_plane].p_pixels;
133 p_out_end = p_out + p_dst->p[i_plane].i_pitch
134 * p_dst->p[i_plane].i_visible_lines;
136 /* skip first line for bottom field */
138 p_out += p_dst->p[i_plane].i_pitch;
140 for( ; p_out < p_out_end ; p_out += 2*p_dst->p[i_plane].i_pitch )
142 /* Handle the width remainder */
144 for( int x = 0; x < w; ++x, ++po )
145 (*po) = 128 + ( ((*po) - 128) / (1 << i_strength) );
147 } /* for i_plane... */
148 } /* if process_chroma */
151 #ifdef CAN_COMPILE_MMXEXT
153 static void DarkenFieldMMX( picture_t *p_dst,
154 const int i_field, const int i_strength,
155 bool process_chroma )
157 assert( p_dst != NULL );
158 assert( i_field == 0 || i_field == 1 );
159 assert( i_strength >= 1 && i_strength <= 3 );
161 uint64_t i_strength_u64 = i_strength; /* needs to know number of bits */
162 const uint8_t remove_high_u8 = 0xFF >> i_strength;
163 const uint64_t remove_high_u64 = remove_high_u8 *
164 INT64_C(0x0101010101010101);
166 int i_plane = Y_PLANE;
167 uint8_t *p_out, *p_out_end;
168 int w = p_dst->p[i_plane].i_visible_pitch;
169 p_out = p_dst->p[i_plane].p_pixels;
170 p_out_end = p_out + p_dst->p[i_plane].i_pitch
171 * p_dst->p[i_plane].i_visible_lines;
173 /* skip first line for bottom field */
175 p_out += p_dst->p[i_plane].i_pitch;
177 int wm8 = w % 8; /* remainder */
178 int w8 = w - wm8; /* part of width that is divisible by 8 */
179 for( ; p_out < p_out_end ; p_out += 2*p_dst->p[i_plane].i_pitch )
181 uint64_t *po = (uint64_t *)p_out;
184 movq_m2r( i_strength_u64, mm1 );
185 movq_m2r( remove_high_u64, mm2 );
186 for( ; x < w8; x += 8 )
188 movq_m2r( (*po), mm0 );
190 psrlq_r2r( mm1, mm0 );
191 pand_r2r( mm2, mm0 );
193 movq_r2m( mm0, (*po++) );
196 /* handle the width remainder */
197 uint8_t *po_temp = (uint8_t *)po;
198 for( ; x < w; ++x, ++po_temp )
199 (*po_temp) = ( ((*po_temp) >> i_strength) & remove_high_u8 );
202 /* Process chroma if the field chromas are independent.
204 The origin (black) is at YUV = (0, 128, 128) in the uint8 format.
205 The chroma processing is a bit more complicated than luma,
206 and needs MMX for vectorization.
210 for( i_plane++ /* luma already handled */;
211 i_plane < p_dst->i_planes;
214 int w = p_dst->p[i_plane].i_visible_pitch;
215 int wm8 = w % 8; /* remainder */
216 int w8 = w - wm8; /* part of width that is divisible by 8 */
218 p_out = p_dst->p[i_plane].p_pixels;
219 p_out_end = p_out + p_dst->p[i_plane].i_pitch
220 * p_dst->p[i_plane].i_visible_lines;
222 /* skip first line for bottom field */
224 p_out += p_dst->p[i_plane].i_pitch;
226 for( ; p_out < p_out_end ; p_out += 2*p_dst->p[i_plane].i_pitch )
230 /* See also easy-to-read C version below. */
231 static const mmx_t b128 = { .uq = 0x8080808080808080ULL };
232 movq_m2r( b128, mm5 );
233 movq_m2r( i_strength_u64, mm6 );
234 movq_m2r( remove_high_u64, mm7 );
236 uint64_t *po8 = (uint64_t *)p_out;
237 for( ; x < w8; x += 8 )
239 movq_m2r( (*po8), mm0 );
241 movq_r2r( mm5, mm2 ); /* 128 */
242 movq_r2r( mm0, mm1 ); /* copy of data */
243 psubusb_r2r( mm2, mm1 ); /* mm1 = max(data - 128, 0) */
244 psubusb_r2r( mm0, mm2 ); /* mm2 = max(128 - data, 0) */
247 psrlq_r2r( mm6, mm1 );
248 psrlq_r2r( mm6, mm2 );
249 pand_r2r( mm7, mm1 );
250 pand_r2r( mm7, mm2 );
252 /* collect results from pos./neg. parts */
253 psubb_r2r( mm2, mm1 );
254 paddb_r2r( mm5, mm1 );
256 movq_r2m( mm1, (*po8++) );
259 /* C version - handle the width remainder */
261 for( ; x < w; ++x, ++po )
262 (*po) = 128 + ( ((*po) - 128) / (1 << i_strength) );
264 } /* for i_plane... */
265 } /* if process_chroma */
271 /*****************************************************************************
273 *****************************************************************************/
275 /* See header for function doc. */
276 int RenderPhosphor( filter_t *p_filter,
278 int i_order, int i_field )
280 assert( p_filter != NULL );
281 assert( p_dst != NULL );
282 assert( i_order >= 0 && i_order <= 2 ); /* 2 = soft field repeat */
283 assert( i_field == 0 || i_field == 1 );
285 filter_sys_t *p_sys = p_filter->p_sys;
287 /* Last two input frames */
288 picture_t *p_in = p_sys->pp_history[HISTORY_SIZE-1];
289 picture_t *p_old = p_sys->pp_history[HISTORY_SIZE-2];
291 /* Use the same input picture as "old" at the first frame after startup */
295 /* If the history mechanism has failed, we can't do anything. */
299 assert( p_old != NULL );
300 assert( p_in != NULL );
302 /* Decide sources for top & bottom fields of output. */
303 picture_t *p_in_top = p_in;
304 picture_t *p_in_bottom = p_in;
305 /* For the first output field this frame,
306 grab "old" field from previous frame. */
309 if( i_field == 0 ) /* rendering top field */
311 else /* i_field == 1, rendering bottom field */
315 compose_chroma_t cc = CC_ALTLINE;
316 if( 2 * p_sys->chroma->p[1].h.num == p_sys->chroma->p[1].h.den &&
317 2 * p_sys->chroma->p[2].h.num == p_sys->chroma->p[2].h.den )
319 /* Only 420 like chroma */
320 switch( p_sys->phosphor.i_chroma_for_420 )
328 else /* i_field == 1 */
329 cc = CC_SOURCE_BOTTOM;
338 /* The above are the only possibilities, if there are no bugs. */
343 ComposeFrame( p_filter, p_dst, p_in_top, p_in_bottom, cc, p_filter->fmt_in.video.i_chroma == VLC_CODEC_YV12 );
345 /* Simulate phosphor light output decay for the old field.
347 The dimmer can also be switched off in the configuration, but that is
348 more of a technical curiosity or an educational toy for advanced users
349 than a useful deinterlacer mode (although it does make telecined
350 material look slightly better than without any filtering).
352 In most use cases the dimmer is used.
354 if( p_sys->phosphor.i_dimmer_strength > 0 )
356 #ifdef CAN_COMPILE_MMXEXT
357 if( vlc_CPU_MMXEXT() )
358 DarkenFieldMMX( p_dst, !i_field, p_sys->phosphor.i_dimmer_strength,
359 p_sys->chroma->p[1].h.num == p_sys->chroma->p[1].h.den &&
360 p_sys->chroma->p[2].h.num == p_sys->chroma->p[2].h.den );
363 DarkenField( 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 );