1 /*
2 * This file is part of mpv.
3 *
4 * mpv is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either
7 * version 2.1 of the License, or (at your option) any later version.
8 *
9 * mpv is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU Lesser General Public License for more details.
13 *
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with mpv. If not, see <http://www.gnu.org/licenses/>.
16 */
17
18 #include <assert.h>
19 #include <string.h>
20
21 #include <libavcodec/avcodec.h>
22 #include <libavutil/imgutils.h>
23 #include <libavutil/pixfmt.h>
24 #include <libavutil/pixdesc.h>
25
26 #include "config.h"
27
28 #include "video/img_format.h"
29 #include "video/mp_image.h"
30 #include "video/fmt-conversion.h"
31
32 struct mp_imgfmt_entry {
33 const char *name;
34 // Valid if flags!=0.
35 // This can be incomplete, and missing fields are filled in:
36 // - sets num_planes and bpp[], derived from comps[] (rounds to bytes)
37 // - sets MP_IMGFLAG_GRAY, derived from comps[]
38 // - sets MP_IMGFLAG_ALPHA, derived from comps[]
39 // - sets align_x/y if 0, derived from chroma shift
40 // - sets xs[]/ys[] always, derived from num_planes/chroma_shift
41 // - sets MP_IMGFLAG_HAS_COMPS|MP_IMGFLAG_NE if num_planes>0
42 // - sets MP_IMGFLAG_TYPE_UINT if no other type set
43 // - sets id to mp_imgfmt_list[] implied format
44 struct mp_imgfmt_desc desc;
45 };
46
47 #define FRINGE_GBRP(def, dname, b) \
48 [def - IMGFMT_CUST_BASE] = { \
49 .name = dname, \
50 .desc = { .flags = MP_IMGFLAG_COLOR_RGB, \
51 .comps = { {2, 0, 8, (b) - 8}, {0, 0, 8, (b) - 8}, \
52 {1, 0, 8, (b) - 8}, }, }}
53
54 #define FLOAT_YUV(def, dname, xs, ys, a) \
55 [def - IMGFMT_CUST_BASE] = { \
56 .name = dname, \
57 .desc = { .flags = MP_IMGFLAG_COLOR_YUV | MP_IMGFLAG_TYPE_FLOAT, \
58 .chroma_xs = xs, .chroma_ys = ys, \
59 .comps = { {0, 0, 32}, {1, 0, 32}, {2, 0, 32}, \
60 {3 * (a), 0, 32 * (a)} }, }}
61
62 static const struct mp_imgfmt_entry mp_imgfmt_list[] = {
63 // not in ffmpeg
64 [IMGFMT_VDPAU_OUTPUT - IMGFMT_CUST_BASE] = {
65 .name = "vdpau_output",
66 .desc = {
67 .flags = MP_IMGFLAG_NE | MP_IMGFLAG_RGB | MP_IMGFLAG_HWACCEL,
68 },
69 },
70 [IMGFMT_RGB30 - IMGFMT_CUST_BASE] = {
71 .name = "rgb30",
72 .desc = {
73 .flags = MP_IMGFLAG_RGB,
74 .comps = { {0, 20, 10}, {0, 10, 10}, {0, 0, 10} },
75 },
76 },
77 [IMGFMT_YAP8 - IMGFMT_CUST_BASE] = {
78 .name = "yap8",
79 .desc = {
80 .flags = MP_IMGFLAG_COLOR_YUV,
81 .comps = { {0, 0, 8}, {0}, {0}, {1, 0, 8} },
82 },
83 },
84 [IMGFMT_YAP16 - IMGFMT_CUST_BASE] = {
85 .name = "yap16",
86 .desc = {
87 .flags = MP_IMGFLAG_COLOR_YUV,
88 .comps = { {0, 0, 16}, {0}, {0}, {1, 0, 16} },
89 },
90 },
91 [IMGFMT_Y1 - IMGFMT_CUST_BASE] = {
92 .name = "y1",
93 .desc = {
94 .flags = MP_IMGFLAG_COLOR_RGB,
95 .comps = { {0, 0, 8, -7} },
96 },
97 },
98 [IMGFMT_YAPF - IMGFMT_CUST_BASE] = {
99 .name = "grayaf32", // try to mimic ffmpeg naming convention
100 .desc = {
101 .flags = MP_IMGFLAG_COLOR_YUV | MP_IMGFLAG_TYPE_FLOAT,
102 .comps = { {0, 0, 32}, {0}, {0}, {1, 0, 32} },
103 },
104 },
105 FLOAT_YUV(IMGFMT_444PF, "yuv444pf", 0, 0, 0),
106 FLOAT_YUV(IMGFMT_444APF, "yuva444pf", 0, 0, 1),
107 FLOAT_YUV(IMGFMT_420PF, "yuv420pf", 1, 1, 0),
108 FLOAT_YUV(IMGFMT_420APF, "yuva420pf", 1, 1, 1),
109 FLOAT_YUV(IMGFMT_422PF, "yuv422pf", 1, 0, 0),
110 FLOAT_YUV(IMGFMT_422APF, "yuva422pf", 1, 0, 1),
111 FLOAT_YUV(IMGFMT_440PF, "yuv440pf", 0, 1, 0),
112 FLOAT_YUV(IMGFMT_440APF, "yuva440pf", 0, 1, 1),
113 FLOAT_YUV(IMGFMT_410PF, "yuv410pf", 2, 2, 0),
114 FLOAT_YUV(IMGFMT_410APF, "yuva410pf", 2, 2, 1),
115 FLOAT_YUV(IMGFMT_411PF, "yuv411pf", 2, 0, 0),
116 FLOAT_YUV(IMGFMT_411APF, "yuva411pf", 2, 0, 1),
117 FRINGE_GBRP(IMGFMT_GBRP1, "gbrp1", 1),
118 FRINGE_GBRP(IMGFMT_GBRP2, "gbrp2", 2),
119 FRINGE_GBRP(IMGFMT_GBRP3, "gbrp3", 3),
120 FRINGE_GBRP(IMGFMT_GBRP4, "gbrp4", 4),
121 FRINGE_GBRP(IMGFMT_GBRP5, "gbrp5", 5),
122 FRINGE_GBRP(IMGFMT_GBRP6, "gbrp6", 6),
123 // in FFmpeg, but FFmpeg names have an annoying "_vld" suffix
124 [IMGFMT_VIDEOTOOLBOX - IMGFMT_CUST_BASE] = {
125 .name = "videotoolbox",
126 },
127 [IMGFMT_VAAPI - IMGFMT_CUST_BASE] = {
128 .name = "vaapi",
129 },
130 };
131
get_mp_desc(int imgfmt)132 static const struct mp_imgfmt_entry *get_mp_desc(int imgfmt)
133 {
134 if (imgfmt < IMGFMT_CUST_BASE)
135 return NULL;
136 int index = imgfmt - IMGFMT_CUST_BASE;
137 if (index >= MP_ARRAY_SIZE(mp_imgfmt_list))
138 return NULL;
139 const struct mp_imgfmt_entry *e = &mp_imgfmt_list[index];
140 return e->name ? e : NULL;
141 }
142
mp_imgfmt_name_list(void)143 char **mp_imgfmt_name_list(void)
144 {
145 int count = IMGFMT_END - IMGFMT_START;
146 char **list = talloc_zero_array(NULL, char *, count + 1);
147 int num = 0;
148 for (int n = IMGFMT_START; n < IMGFMT_END; n++) {
149 const char *name = mp_imgfmt_to_name(n);
150 if (strcmp(name, "unknown") != 0)
151 list[num++] = talloc_strdup(list, name);
152 }
153 return list;
154 }
155
mp_imgfmt_from_name(bstr name)156 int mp_imgfmt_from_name(bstr name)
157 {
158 if (bstr_equals0(name, "none"))
159 return 0;
160 for (int n = 0; n < MP_ARRAY_SIZE(mp_imgfmt_list); n++) {
161 const struct mp_imgfmt_entry *p = &mp_imgfmt_list[n];
162 if (p->name && bstr_equals0(name, p->name))
163 return IMGFMT_CUST_BASE + n;
164 }
165 return pixfmt2imgfmt(av_get_pix_fmt(mp_tprintf(80, "%.*s", BSTR_P(name))));
166 }
167
mp_imgfmt_to_name_buf(char * buf,size_t buf_size,int fmt)168 char *mp_imgfmt_to_name_buf(char *buf, size_t buf_size, int fmt)
169 {
170 const struct mp_imgfmt_entry *p = get_mp_desc(fmt);
171 const char *name = p ? p->name : NULL;
172 if (!name) {
173 const AVPixFmtDescriptor *pixdesc = av_pix_fmt_desc_get(imgfmt2pixfmt(fmt));
174 if (pixdesc)
175 name = pixdesc->name;
176 }
177 if (!name)
178 name = "unknown";
179 snprintf(buf, buf_size, "%s", name);
180 int len = strlen(buf);
181 if (len > 2 && buf[len - 2] == MP_SELECT_LE_BE('l', 'b') && buf[len - 1] == 'e')
182 buf[len - 2] = '\0';
183 return buf;
184 }
185
fill_pixdesc_layout(struct mp_imgfmt_desc * desc,enum AVPixelFormat fmt,const AVPixFmtDescriptor * pd)186 static void fill_pixdesc_layout(struct mp_imgfmt_desc *desc,
187 enum AVPixelFormat fmt,
188 const AVPixFmtDescriptor *pd)
189 {
190 if (pd->flags & AV_PIX_FMT_FLAG_PAL ||
191 pd->flags & AV_PIX_FMT_FLAG_HWACCEL)
192 goto fail;
193
194 bool has_alpha = pd->flags & AV_PIX_FMT_FLAG_ALPHA;
195 if (pd->nb_components != 1 + has_alpha &&
196 pd->nb_components != 3 + has_alpha)
197 goto fail;
198
199 // Very convenient: we assume we're always on little endian, and FFmpeg
200 // explicitly marks big endian formats => don't need to guess whether a
201 // format is little endian, or not affected by byte order.
202 bool is_be = pd->flags & AV_PIX_FMT_FLAG_BE;
203 bool is_ne = MP_SELECT_LE_BE(false, true) == is_be;
204
205 // Packed sub-sampled YUV is very... special.
206 bool is_packed_ss_yuv = pd->log2_chroma_w && !pd->log2_chroma_h &&
207 pd->comp[1].plane == 0 && pd->comp[2].plane == 0 &&
208 pd->nb_components == 3;
209
210 if (is_packed_ss_yuv)
211 desc->bpp[0] = pd->comp[1].step * 8;
212
213 // Determine if there are any byte overlaps => relevant for determining
214 // access unit for endian, since pixdesc does not expose this, and assumes
215 // a weird model where you do separate memory fetches for each component.
216 bool any_shared_bytes = !!(pd->flags & AV_PIX_FMT_FLAG_BITSTREAM);
217 for (int c = 0; c < pd->nb_components; c++) {
218 for (int i = 0; i < c; i++) {
219 const AVComponentDescriptor *d1 = &pd->comp[c];
220 const AVComponentDescriptor *d2 = &pd->comp[i];
221 if (d1->plane == d2->plane) {
222 if (d1->offset + (d1->depth + 7) / 8u > d2->offset &&
223 d2->offset + (d2->depth + 7) / 8u > d1->offset)
224 any_shared_bytes = true;
225 }
226 }
227 }
228
229 int el_bits = (pd->flags & AV_PIX_FMT_FLAG_BITSTREAM) ? 1 : 8;
230 for (int c = 0; c < pd->nb_components; c++) {
231 const AVComponentDescriptor *d = &pd->comp[c];
232 if (d->plane >= MP_MAX_PLANES)
233 goto fail;
234
235 desc->num_planes = MPMAX(desc->num_planes, d->plane + 1);
236
237 int plane_bits = desc->bpp[d->plane];
238 int c_bits = d->step * el_bits;
239
240 // The first component wins, because either all components result in
241 // the same value, or luma wins (luma always comes before chroma).
242 if (plane_bits) {
243 if (c_bits > plane_bits)
244 goto fail; // inconsistent
245 } else {
246 desc->bpp[d->plane] = plane_bits = c_bits;
247 }
248
249 int shift = d->shift;
250 // What the fuck: for some inexplicable reason, MONOB uses shift=7
251 // in pixdesc, which is basically out of bounds. Pixdesc bug?
252 // Make it behave like MONOW. (No, the bit-order is not different.)
253 if (fmt == AV_PIX_FMT_MONOBLACK)
254 shift = 0;
255
256 int offset = d->offset * el_bits;
257 // The pixdesc logic for reading and endian swapping is as follows
258 // (reverse engineered from av_read_image_line2()):
259 // - determine a word size that will include the component fully;
260 // this includes the "active" bits and the amount "shifted" away
261 // (for example shift=7/depth=18 => 32 bit word reading [31:0])
262 // - the same format can use different word sizes (e.g. bgr565: the R
263 // component at offset 0 is read as 8 bit; BG is read as 16 bits)
264 // - if BE flag is set, swap the word before proceeding
265 // - extract via shift and mask derived by depth
266 int word = mp_round_next_power_of_2(MPMAX(d->depth + shift, 8));
267 // The purpose of this is unknown. It's an absurdity fished out of
268 // av_read_image_line2()'s implementation. It seems technically
269 // unnecessary, and provides no information. On the other hand, it
270 // compensates for seemingly bogus packed integer pixdescs; this
271 // is "why" some formats use d->offset = -1.
272 if (is_be && el_bits == 8 && word == 8)
273 offset += 8;
274 // Pixdesc's model sometimes requires accesses with varying word-sizes,
275 // as seen in bgr565 and other formats. Also, it makes you read some
276 // formats with multiple endian-dependent accesses, where accessing a
277 // larger unit would make more sense. (Consider X2RGB10BE, for which
278 // pixdesc wants you to perform 3 * 2 byte accesses, and swap each of
279 // the read 16 bit words. What you really want is to swap the entire 4
280 // byte thing, and then extract the components with bit shifts).
281 // This is complete bullshit, so we transform it into word swaps before
282 // further processing. Care needs to be taken to not change formats like
283 // P010 or YA16 (prefer component accesses for them; P010 isn't even
284 // representable, because endian_shift is for all planes).
285 // As a heuristic, assume that if any components share a byte, the whole
286 // pixel is read as a single memory access and endian swapped at once.
287 int access_size = 8;
288 if (plane_bits > 8) {
289 if (any_shared_bytes) {
290 access_size = plane_bits;
291 if (is_be && word != access_size) {
292 // Before: offset = 8*byte_offset (with word bits of data)
293 // After: offset = bit_offset into swapped endian_size word
294 offset = access_size - word - offset;
295 }
296 } else {
297 access_size = word;
298 }
299 }
300 int endian_size = (access_size && !is_ne) ? access_size : 8;
301 int endian_shift = mp_log2(endian_size) - 3;
302 if (!MP_IS_POWER_OF_2(endian_size) || endian_shift < 0 || endian_shift > 3)
303 goto fail;
304 if (desc->endian_shift && desc->endian_shift != endian_shift)
305 goto fail;
306 desc->endian_shift = endian_shift;
307
308 // We always use bit offsets; this doesn't lose any information,
309 // and pixdesc is merely more redundant.
310 offset += shift;
311 if (offset < 0 || offset >= (1 << 6))
312 goto fail;
313 if (offset + d->depth > plane_bits)
314 goto fail;
315 if (d->depth < 0 || d->depth >= (1 << 6))
316 goto fail;
317 desc->comps[c] = (struct mp_imgfmt_comp_desc){
318 .plane = d->plane,
319 .offset = offset,
320 .size = d->depth,
321 };
322 }
323
324 for (int p = 0; p < desc->num_planes; p++) {
325 if (!desc->bpp[p])
326 goto fail; // plane doesn't exist
327 }
328
329 // What the fuck: this is probably a pixdesc bug, so fix it.
330 if (fmt == AV_PIX_FMT_RGB8) {
331 desc->comps[2] = (struct mp_imgfmt_comp_desc){0, 0, 2};
332 desc->comps[1] = (struct mp_imgfmt_comp_desc){0, 2, 3};
333 desc->comps[0] = (struct mp_imgfmt_comp_desc){0, 5, 3};
334 }
335
336 // Overlap test. If any shared bits are happening, this is not a format we
337 // can represent (or it's something like Bayer: components in the same bits,
338 // but different alternating lines).
339 bool any_shared_bits = false;
340 for (int c = 0; c < pd->nb_components; c++) {
341 for (int i = 0; i < c; i++) {
342 struct mp_imgfmt_comp_desc *c1 = &desc->comps[c];
343 struct mp_imgfmt_comp_desc *c2 = &desc->comps[i];
344 if (c1->plane == c2->plane) {
345 if (c1->offset + c1->size > c2->offset &&
346 c2->offset + c2->size > c1->offset)
347 any_shared_bits = true;
348 }
349 }
350 }
351
352 if (any_shared_bits) {
353 for (int c = 0; c < pd->nb_components; c++)
354 desc->comps[c] = (struct mp_imgfmt_comp_desc){0};
355 }
356
357 // Many important formats have padding within an access word. For example
358 // yuv420p10 has the upper 6 bit cleared to 0; P010 has the lower 6 bits
359 // cleared to 0. Pixdesc cannot represent that these bits are 0. There are
360 // other formats where padding is not guaranteed to be 0, but they are
361 // described in the same way.
362 // Apply a heuristic that is supposed to identify formats which use
363 // guaranteed 0 padding. This could fail, but nobody said this pixdesc crap
364 // is robust.
365 for (int c = 0; c < pd->nb_components; c++) {
366 struct mp_imgfmt_comp_desc *cd = &desc->comps[c];
367 // Note: rgb444 would defeat our heuristic if we checked only per comp.
368 // also, exclude "bitstream" formats due to monow/monob
369 int fsize = MP_ALIGN_UP(cd->size, 8);
370 if (!any_shared_bytes && el_bits == 8 && fsize != cd->size &&
371 fsize - cd->size <= (1 << 3))
372 {
373 if (!(cd->offset % 8u)) {
374 cd->pad = -(fsize - cd->size);
375 cd->size = fsize;
376 } else if (!((cd->offset + cd->size) % 8u)) {
377 cd->pad = fsize - cd->size;
378 cd->size = fsize;
379 cd->offset = MP_ALIGN_DOWN(cd->offset, 8);
380 }
381 }
382 }
383
384 // The alpha component always has ID 4 (index 3) in our representation, so
385 // move the alpha component to there.
386 if (has_alpha && pd->nb_components < 4) {
387 desc->comps[3] = desc->comps[pd->nb_components - 1];
388 desc->comps[pd->nb_components - 1] = (struct mp_imgfmt_comp_desc){0};
389 }
390
391 if (is_packed_ss_yuv) {
392 desc->flags |= MP_IMGFLAG_PACKED_SS_YUV;
393 desc->bpp[0] /= 1 << pd->log2_chroma_w;
394 } else if (!any_shared_bits) {
395 desc->flags |= MP_IMGFLAG_HAS_COMPS;
396 }
397
398 return;
399
400 fail:
401 for (int n = 0; n < 4; n++)
402 desc->comps[n] = (struct mp_imgfmt_comp_desc){0};
403 // Average bit size fallback.
404 desc->num_planes = av_pix_fmt_count_planes(fmt);
405 for (int p = 0; p < desc->num_planes; p++) {
406 int ls = av_image_get_linesize(fmt, 256, p);
407 desc->bpp[p] = ls > 0 ? ls * 8 / 256 : 0;
408 }
409 }
410
mp_imgfmt_get_desc_from_pixdesc(int mpfmt,struct mp_imgfmt_desc * out)411 static bool mp_imgfmt_get_desc_from_pixdesc(int mpfmt, struct mp_imgfmt_desc *out)
412 {
413 enum AVPixelFormat fmt = imgfmt2pixfmt(mpfmt);
414 const AVPixFmtDescriptor *pd = av_pix_fmt_desc_get(fmt);
415 if (!pd || pd->nb_components > 4)
416 return false;
417
418 struct mp_imgfmt_desc desc = {
419 .id = mpfmt,
420 .chroma_xs = pd->log2_chroma_w,
421 .chroma_ys = pd->log2_chroma_h,
422 };
423
424 if (pd->flags & AV_PIX_FMT_FLAG_ALPHA)
425 desc.flags |= MP_IMGFLAG_ALPHA;
426
427 if (pd->flags & AV_PIX_FMT_FLAG_HWACCEL)
428 desc.flags |= MP_IMGFLAG_TYPE_HW;
429
430 // Pixdesc does not provide a flag for XYZ, so this is the best we can do.
431 if (strncmp(pd->name, "xyz", 3) == 0) {
432 desc.flags |= MP_IMGFLAG_COLOR_XYZ;
433 } else if (pd->flags & AV_PIX_FMT_FLAG_RGB) {
434 desc.flags |= MP_IMGFLAG_COLOR_RGB;
435 } else if (fmt == AV_PIX_FMT_MONOBLACK || fmt == AV_PIX_FMT_MONOWHITE) {
436 desc.flags |= MP_IMGFLAG_COLOR_RGB;
437 } else if (fmt == AV_PIX_FMT_PAL8) {
438 desc.flags |= MP_IMGFLAG_COLOR_RGB | MP_IMGFLAG_TYPE_PAL8;
439 }
440
441 if (pd->flags & AV_PIX_FMT_FLAG_FLOAT)
442 desc.flags |= MP_IMGFLAG_TYPE_FLOAT;
443
444 // Educated guess.
445 if (!(desc.flags & MP_IMGFLAG_COLOR_MASK) &&
446 !(desc.flags & MP_IMGFLAG_TYPE_HW))
447 desc.flags |= MP_IMGFLAG_COLOR_YUV;
448
449 desc.align_x = 1 << desc.chroma_xs;
450 desc.align_y = 1 << desc.chroma_ys;
451
452 fill_pixdesc_layout(&desc, fmt, pd);
453
454 if (desc.flags & (MP_IMGFLAG_HAS_COMPS | MP_IMGFLAG_PACKED_SS_YUV)) {
455 if (!(desc.flags & MP_IMGFLAG_TYPE_MASK))
456 desc.flags |= MP_IMGFLAG_TYPE_UINT;
457 }
458
459 if (desc.bpp[0] % 8u && (pd->flags & AV_PIX_FMT_FLAG_BITSTREAM))
460 desc.align_x = 8 / desc.bpp[0]; // expect power of 2
461
462 // Very heuristical.
463 bool is_ne = !desc.endian_shift;
464 bool need_endian = (desc.comps[0].size % 8u && desc.bpp[0] > 8) ||
465 desc.comps[0].size > 8;
466
467 if (need_endian) {
468 bool is_le = MP_SELECT_LE_BE(is_ne, !is_ne);
469 desc.flags |= is_le ? MP_IMGFLAG_LE : MP_IMGFLAG_BE;
470 } else {
471 desc.flags |= MP_IMGFLAG_LE | MP_IMGFLAG_BE;
472 }
473
474 *out = desc;
475 return true;
476 }
477
mp_imgfmt_get_packed_yuv_locations(int imgfmt,uint8_t * luma_offsets)478 bool mp_imgfmt_get_packed_yuv_locations(int imgfmt, uint8_t *luma_offsets)
479 {
480 struct mp_imgfmt_desc desc = mp_imgfmt_get_desc(imgfmt);
481 if (!(desc.flags & MP_IMGFLAG_PACKED_SS_YUV))
482 return false;
483
484 assert(desc.num_planes == 1);
485
486 // Guess at which positions the additional luma samples are. We iterate
487 // starting with the first byte, and then put a luma sample at places
488 // not covered by other luma/chroma.
489 // Pixdesc does not and can not provide this information. This heuristic
490 // may fail in certain cases. What a load of bullshit, right?
491 int lsize = desc.comps[0].size;
492 int cur_offset = 0;
493 for (int lsample = 1; lsample < (1 << desc.chroma_xs); lsample++) {
494 while (1) {
495 if (cur_offset + lsize > desc.bpp[0] * desc.align_x)
496 return false;
497 bool free = true;
498 for (int c = 0; c < 3; c++) {
499 struct mp_imgfmt_comp_desc *cd = &desc.comps[c];
500 if (!cd->size)
501 continue;
502 if (cd->offset + cd->size > cur_offset &&
503 cur_offset + lsize > cd->offset)
504 {
505 free = false;
506 break;
507 }
508 }
509 if (free)
510 break;
511 cur_offset += lsize;
512 }
513 luma_offsets[lsample] = cur_offset;
514 cur_offset += lsize;
515 }
516
517 luma_offsets[0] = desc.comps[0].offset;
518 return true;
519 }
520
get_native_desc(int mpfmt,struct mp_imgfmt_desc * desc)521 static bool get_native_desc(int mpfmt, struct mp_imgfmt_desc *desc)
522 {
523 const struct mp_imgfmt_entry *p = get_mp_desc(mpfmt);
524 if (!p || !p->desc.flags)
525 return false;
526
527 *desc = p->desc;
528
529 // Fill in some fields mp_imgfmt_entry.desc is not required to set.
530
531 desc->id = mpfmt;
532
533 for (int n = 0; n < MP_NUM_COMPONENTS; n++) {
534 struct mp_imgfmt_comp_desc *cd = &desc->comps[n];
535 if (cd->size)
536 desc->num_planes = MPMAX(desc->num_planes, cd->plane + 1);
537 desc->bpp[cd->plane] =
538 MPMAX(desc->bpp[cd->plane], MP_ALIGN_UP(cd->offset + cd->size, 8));
539 }
540
541 if (!desc->align_x && !desc->align_y) {
542 desc->align_x = 1 << desc->chroma_xs;
543 desc->align_y = 1 << desc->chroma_ys;
544 }
545
546 if (desc->num_planes)
547 desc->flags |= MP_IMGFLAG_HAS_COMPS | MP_IMGFLAG_NE;
548
549 if (!(desc->flags & MP_IMGFLAG_TYPE_MASK))
550 desc->flags |= MP_IMGFLAG_TYPE_UINT;
551
552 return true;
553 }
554
mp_imgfmt_desc_get_num_comps(struct mp_imgfmt_desc * desc)555 int mp_imgfmt_desc_get_num_comps(struct mp_imgfmt_desc *desc)
556 {
557 int flags = desc->flags;
558 if (!(flags & MP_IMGFLAG_COLOR_MASK))
559 return 0;
560 return 3 + (flags & MP_IMGFLAG_GRAY ? -2 : 0) + !!(flags & MP_IMGFLAG_ALPHA);
561 }
562
mp_imgfmt_get_desc(int mpfmt)563 struct mp_imgfmt_desc mp_imgfmt_get_desc(int mpfmt)
564 {
565 struct mp_imgfmt_desc desc;
566
567 if (!get_native_desc(mpfmt, &desc) &&
568 !mp_imgfmt_get_desc_from_pixdesc(mpfmt, &desc))
569 return (struct mp_imgfmt_desc){0};
570
571 for (int p = 0; p < desc.num_planes; p++) {
572 desc.xs[p] = (p == 1 || p == 2) ? desc.chroma_xs : 0;
573 desc.ys[p] = (p == 1 || p == 2) ? desc.chroma_ys : 0;
574 }
575
576 bool is_ba = desc.num_planes > 0;
577 for (int p = 0; p < desc.num_planes; p++)
578 is_ba = !(desc.bpp[p] % 8u);
579
580 if (is_ba)
581 desc.flags |= MP_IMGFLAG_BYTE_ALIGNED;
582
583 if (desc.flags & MP_IMGFLAG_HAS_COMPS) {
584 if (desc.comps[3].size)
585 desc.flags |= MP_IMGFLAG_ALPHA;
586
587 // Assuming all colors are (CCC+[A]) or (C+[A]), the latter being gray.
588 if (!desc.comps[1].size)
589 desc.flags |= MP_IMGFLAG_GRAY;
590
591 bool bb = true;
592 for (int n = 0; n < MP_NUM_COMPONENTS; n++) {
593 if (desc.comps[n].offset % 8u || desc.comps[n].size % 8u)
594 bb = false;
595 }
596 if (bb)
597 desc.flags |= MP_IMGFLAG_BYTES;
598 }
599
600 if ((desc.flags & (MP_IMGFLAG_YUV | MP_IMGFLAG_RGB))
601 && (desc.flags & MP_IMGFLAG_HAS_COMPS)
602 && (desc.flags & MP_IMGFLAG_BYTES)
603 && ((desc.flags & MP_IMGFLAG_TYPE_MASK) == MP_IMGFLAG_TYPE_UINT))
604 {
605 int cnt = mp_imgfmt_desc_get_num_comps(&desc);
606 bool same_depth = true;
607 for (int p = 0; p < desc.num_planes; p++)
608 same_depth &= desc.bpp[p] == desc.bpp[0];
609 if (same_depth && cnt == desc.num_planes) {
610 if (desc.flags & MP_IMGFLAG_YUV) {
611 desc.flags |= MP_IMGFLAG_YUV_P;
612 } else {
613 desc.flags |= MP_IMGFLAG_RGB_P;
614 }
615 }
616 if (cnt == 3 && desc.num_planes == 2 &&
617 desc.bpp[1] == desc.bpp[0] * 2 &&
618 (desc.flags & MP_IMGFLAG_YUV))
619 {
620
621 desc.flags |= MP_IMGFLAG_YUV_NV;
622 }
623 }
624
625 return desc;
626 }
627
validate_regular_imgfmt(const struct mp_regular_imgfmt * fmt)628 static bool validate_regular_imgfmt(const struct mp_regular_imgfmt *fmt)
629 {
630 bool present[MP_NUM_COMPONENTS] = {0};
631 int n_comp = 0;
632
633 for (int n = 0; n < fmt->num_planes; n++) {
634 const struct mp_regular_imgfmt_plane *plane = &fmt->planes[n];
635 n_comp += plane->num_components;
636 if (n_comp > MP_NUM_COMPONENTS)
637 return false;
638 if (!plane->num_components)
639 return false; // no empty planes in between allowed
640
641 bool pad_only = true;
642 int chroma_luma = 0; // luma: 1, chroma: 2, both: 3
643 for (int i = 0; i < plane->num_components; i++) {
644 int comp = plane->components[i];
645 if (comp > MP_NUM_COMPONENTS)
646 return false;
647 if (comp == 0)
648 continue;
649 pad_only = false;
650 if (present[comp - 1])
651 return false; // no duplicates
652 present[comp - 1] = true;
653 chroma_luma |= (comp == 2 || comp == 3) ? 2 : 1;
654 }
655 if (pad_only)
656 return false; // no planes with only padding allowed
657 if ((fmt->chroma_xs > 0 || fmt->chroma_ys > 0) && chroma_luma == 3)
658 return false; // separate chroma/luma planes required
659 }
660
661 if (!(present[0] || present[3]) || // at least component 1 or alpha needed
662 (present[1] && !present[0]) || // component 2 requires component 1
663 (present[2] && !present[1])) // component 3 requires component 2
664 return false;
665
666 return true;
667 }
668
get_forced_csp_from_flags(int flags)669 static enum mp_csp get_forced_csp_from_flags(int flags)
670 {
671 if (flags & MP_IMGFLAG_COLOR_XYZ)
672 return MP_CSP_XYZ;
673
674 if (flags & MP_IMGFLAG_COLOR_RGB)
675 return MP_CSP_RGB;
676
677 return MP_CSP_AUTO;
678 }
679
mp_imgfmt_get_forced_csp(int imgfmt)680 enum mp_csp mp_imgfmt_get_forced_csp(int imgfmt)
681 {
682 return get_forced_csp_from_flags(mp_imgfmt_get_desc(imgfmt).flags);
683 }
684
get_component_type_from_flags(int flags)685 static enum mp_component_type get_component_type_from_flags(int flags)
686 {
687 if (flags & MP_IMGFLAG_TYPE_UINT)
688 return MP_COMPONENT_TYPE_UINT;
689
690 if (flags & MP_IMGFLAG_TYPE_FLOAT)
691 return MP_COMPONENT_TYPE_FLOAT;
692
693 return MP_COMPONENT_TYPE_UNKNOWN;
694 }
695
mp_imgfmt_get_component_type(int imgfmt)696 enum mp_component_type mp_imgfmt_get_component_type(int imgfmt)
697 {
698 return get_component_type_from_flags(mp_imgfmt_get_desc(imgfmt).flags);
699 }
700
mp_find_other_endian(int imgfmt)701 int mp_find_other_endian(int imgfmt)
702 {
703 return pixfmt2imgfmt(av_pix_fmt_swap_endianness(imgfmt2pixfmt(imgfmt)));
704 }
705
mp_get_regular_imgfmt(struct mp_regular_imgfmt * dst,int imgfmt)706 bool mp_get_regular_imgfmt(struct mp_regular_imgfmt *dst, int imgfmt)
707 {
708 struct mp_regular_imgfmt res = {0};
709
710 struct mp_imgfmt_desc desc = mp_imgfmt_get_desc(imgfmt);
711 if (!desc.num_planes)
712 return false;
713 res.num_planes = desc.num_planes;
714
715 if (desc.endian_shift || !(desc.flags & MP_IMGFLAG_HAS_COMPS))
716 return false;
717
718 res.component_type = get_component_type_from_flags(desc.flags);
719 if (!res.component_type)
720 return false;
721
722 struct mp_imgfmt_comp_desc *comp0 = &desc.comps[0];
723 if (comp0->size < 1 || comp0->size > 64 || (comp0->size % 8u))
724 return false;
725
726 res.component_size = comp0->size / 8u;
727 res.component_pad = comp0->pad;
728
729 for (int n = 0; n < res.num_planes; n++) {
730 if (desc.bpp[n] % comp0->size)
731 return false;
732 res.planes[n].num_components = desc.bpp[n] / comp0->size;
733 }
734
735 for (int n = 0; n < MP_NUM_COMPONENTS; n++) {
736 struct mp_imgfmt_comp_desc *comp = &desc.comps[n];
737 if (!comp->size)
738 continue;
739
740 struct mp_regular_imgfmt_plane *plane = &res.planes[comp->plane];
741
742 res.num_planes = MPMAX(res.num_planes, comp->plane + 1);
743
744 // We support uniform depth only.
745 if (comp->size != comp0->size || comp->pad != comp0->pad)
746 return false;
747
748 // Size-aligned only.
749 int pos = comp->offset / comp->size;
750 if (comp->offset != pos * comp->size || pos >= MP_NUM_COMPONENTS)
751 return false;
752
753 if (plane->components[pos])
754 return false;
755 plane->components[pos] = n + 1;
756 }
757
758 res.chroma_xs = desc.chroma_xs;
759 res.chroma_ys = desc.chroma_ys;
760
761 res.forced_csp = get_forced_csp_from_flags(desc.flags);
762
763 if (!validate_regular_imgfmt(&res))
764 return false;
765
766 *dst = res;
767 return true;
768 }
769
regular_imgfmt_equals(struct mp_regular_imgfmt * a,struct mp_regular_imgfmt * b)770 static bool regular_imgfmt_equals(struct mp_regular_imgfmt *a,
771 struct mp_regular_imgfmt *b)
772 {
773 if (a->component_type != b->component_type ||
774 a->component_size != b->component_size ||
775 a->num_planes != b->num_planes ||
776 a->component_pad != b->component_pad ||
777 a->forced_csp != b->forced_csp ||
778 a->chroma_xs != b->chroma_xs ||
779 a->chroma_ys != b->chroma_ys)
780 return false;
781
782 for (int n = 0; n < a->num_planes; n++) {
783 int num_comps = a->planes[n].num_components;
784 if (num_comps != b->planes[n].num_components)
785 return false;
786 for (int i = 0; i < num_comps; i++) {
787 if (a->planes[n].components[i] != b->planes[n].components[i])
788 return false;
789 }
790 }
791
792 return true;
793 }
794
795 // Find a format that matches this one exactly.
mp_find_regular_imgfmt(struct mp_regular_imgfmt * src)796 int mp_find_regular_imgfmt(struct mp_regular_imgfmt *src)
797 {
798 for (int n = IMGFMT_START + 1; n < IMGFMT_END; n++) {
799 struct mp_regular_imgfmt f;
800 if (mp_get_regular_imgfmt(&f, n) && regular_imgfmt_equals(src, &f))
801 return n;
802 }
803 return 0;
804 }
805
806 // Compare the dst image formats, and return the one which can carry more data
807 // (e.g. higher depth, more color components, lower chroma subsampling, etc.),
808 // with respect to what is required to keep most of the src format.
809 // Returns the imgfmt, or 0 on error.
mp_imgfmt_select_best(int dst1,int dst2,int src)810 int mp_imgfmt_select_best(int dst1, int dst2, int src)
811 {
812 enum AVPixelFormat dst1pxf = imgfmt2pixfmt(dst1);
813 enum AVPixelFormat dst2pxf = imgfmt2pixfmt(dst2);
814 enum AVPixelFormat srcpxf = imgfmt2pixfmt(src);
815 enum AVPixelFormat dstlist[] = {dst1pxf, dst2pxf, AV_PIX_FMT_NONE};
816 return pixfmt2imgfmt(avcodec_find_best_pix_fmt_of_list(dstlist, srcpxf, 1, 0));
817 }
818
819 // Same as mp_imgfmt_select_best(), but with a list of dst formats.
mp_imgfmt_select_best_list(int * dst,int num_dst,int src)820 int mp_imgfmt_select_best_list(int *dst, int num_dst, int src)
821 {
822 int best = 0;
823 for (int n = 0; n < num_dst; n++)
824 best = best ? mp_imgfmt_select_best(best, dst[n], src) : dst[n];
825 return best;
826 }
827