FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/dnxhdenc.c
Date: 2025-05-15 14:57:22
Exec Total Coverage
Lines: 678 798 85.0%
Functions: 30 31 96.8%
Branches: 303 415 73.0%
Line Branch Exec Source
1 /*
2 * VC3/DNxHD encoder
3 * Copyright (c) 2007 Baptiste Coudurier <baptiste dot coudurier at smartjog dot com>
4 * Copyright (c) 2011 MirriAd Ltd
5 *
6 * VC-3 encoder funded by the British Broadcasting Corporation
7 * 10 bit support added by MirriAd Ltd, Joseph Artsimovich <joseph@mirriad.com>
8 *
9 * This file is part of FFmpeg.
10 *
11 * FFmpeg is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public
13 * License as published by the Free Software Foundation; either
14 * version 2.1 of the License, or (at your option) any later version.
15 *
16 * FFmpeg is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
20 *
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with FFmpeg; if not, write to the Free Software
23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 */
25
26 #include "libavutil/attributes.h"
27 #include "libavutil/internal.h"
28 #include "libavutil/mem.h"
29 #include "libavutil/mem_internal.h"
30 #include "libavutil/opt.h"
31
32 #include "avcodec.h"
33 #include "blockdsp.h"
34 #include "codec_internal.h"
35 #include "encode.h"
36 #include "fdctdsp.h"
37 #include "mathops.h"
38 #include "mpegvideo.h"
39 #include "mpegvideoenc.h"
40 #include "pixblockdsp.h"
41 #include "packet_internal.h"
42 #include "profiles.h"
43 #include "dnxhdenc.h"
44
45 // The largest value that will not lead to overflow for 10-bit samples.
46 #define DNX10BIT_QMAT_SHIFT 18
47 #define RC_VARIANCE 1 // use variance or ssd for fast rc
48 #define LAMBDA_FRAC_BITS 10
49
50 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
51 static const AVOption options[] = {
52 { "nitris_compat", "encode with Avid Nitris compatibility",
53 offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
54 { "ibias", "intra quant bias",
55 offsetof(DNXHDEncContext, intra_quant_bias), AV_OPT_TYPE_INT,
56 { .i64 = 0 }, INT_MIN, INT_MAX, VE },
57 { "profile", NULL, offsetof(DNXHDEncContext, profile), AV_OPT_TYPE_INT,
58 { .i64 = AV_PROFILE_DNXHD },
59 AV_PROFILE_DNXHD, AV_PROFILE_DNXHR_444, VE, .unit = "profile" },
60 { "dnxhd", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_PROFILE_DNXHD },
61 0, 0, VE, .unit = "profile" },
62 { "dnxhr_444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_PROFILE_DNXHR_444 },
63 0, 0, VE, .unit = "profile" },
64 { "dnxhr_hqx", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_PROFILE_DNXHR_HQX },
65 0, 0, VE, .unit = "profile" },
66 { "dnxhr_hq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_PROFILE_DNXHR_HQ },
67 0, 0, VE, .unit = "profile" },
68 { "dnxhr_sq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_PROFILE_DNXHR_SQ },
69 0, 0, VE, .unit = "profile" },
70 { "dnxhr_lb", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_PROFILE_DNXHR_LB },
71 0, 0, VE, .unit = "profile" },
72 { NULL }
73 };
74
75 static const AVClass dnxhd_class = {
76 .class_name = "dnxhd",
77 .item_name = av_default_item_name,
78 .option = options,
79 .version = LIBAVUTIL_VERSION_INT,
80 };
81
82 97920 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *restrict block,
83 const uint8_t *pixels,
84 ptrdiff_t line_size)
85 {
86 int i;
87
2/2
✓ Branch 0 taken 391680 times.
✓ Branch 1 taken 97920 times.
 489600 for (i = 0; i < 4; i++) {
88 391680 block[0] = pixels[0];
89 391680 block[1] = pixels[1];
90 391680 block[2] = pixels[2];
91 391680 block[3] = pixels[3];
92 391680 block[4] = pixels[4];
93 391680 block[5] = pixels[5];
94 391680 block[6] = pixels[6];
95 391680 block[7] = pixels[7];
96 391680 pixels += line_size;
97 391680 block += 8;
98 }
99 97920 memcpy(block, block - 8, sizeof(*block) * 8);
100 97920 memcpy(block + 8, block - 16, sizeof(*block) * 8);
101 97920 memcpy(block + 16, block - 24, sizeof(*block) * 8);
102 97920 memcpy(block + 24, block - 32, sizeof(*block) * 8);
103 97920 }
104
105 static av_always_inline
106 43200 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *restrict block,
107 const uint8_t *pixels,
108 ptrdiff_t line_size)
109 {
110 43200 memcpy(block + 0 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
111 43200 memcpy(block + 7 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
112 43200 memcpy(block + 1 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
113 43200 memcpy(block + 6 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
114 43200 memcpy(block + 2 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
115 43200 memcpy(block + 5 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
116 43200 memcpy(block + 3 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
117 43200 memcpy(block + 4 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
118 43200 }
119
120 ✗ static int dnxhd_10bit_dct_quantize_444(MPVEncContext *ctx, int16_t *block,
121 int n, int qscale, int *overflow)
122 {
123 int i, j, level, last_non_zero, start_i;
124 const int *qmat;
125 ✗ const uint8_t *scantable = ctx->c.intra_scantable.scantable;
126 int bias;
127 ✗ int max = 0;
128 unsigned int threshold1, threshold2;
129
130 ✗ ctx->fdsp.fdct(block);
131
132 ✗ block[0] = (block[0] + 2) >> 2;
133 ✗ start_i = 1;
134 ✗ last_non_zero = 0;
135 ✗ qmat = n < 4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
136 ✗ bias= ctx->intra_quant_bias * (1 << (16 - 8));
137 ✗ threshold1 = (1 << 16) - bias - 1;
138 ✗ threshold2 = (threshold1 << 1);
139
140 ✗ for (i = 63; i >= start_i; i--) {
141 ✗ j = scantable[i];
142 ✗ level = block[j] * qmat[j];
143
144 ✗ if (((unsigned)(level + threshold1)) > threshold2) {
145 ✗ last_non_zero = i;
146 ✗ break;
147 } else{
148 ✗ block[j]=0;
149 }
150 }
151
152 ✗ for (i = start_i; i <= last_non_zero; i++) {
153 ✗ j = scantable[i];
154 ✗ level = block[j] * qmat[j];
155
156 ✗ if (((unsigned)(level + threshold1)) > threshold2) {
157 ✗ if (level > 0) {
158 ✗ level = (bias + level) >> 16;
159 ✗ block[j] = level;
160 } else{
161 ✗ level = (bias - level) >> 16;
162 ✗ block[j] = -level;
163 }
164 ✗ max |= level;
165 } else {
166 ✗ block[j] = 0;
167 }
168 }
169 ✗ *overflow = ctx->max_qcoeff < max; //overflow might have happened
170
171 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
172 ✗ if (ctx->c.idsp.perm_type != FF_IDCT_PERM_NONE)
173 ✗ ff_block_permute(block, ctx->c.idsp.idct_permutation,
174 scantable, last_non_zero);
175
176 ✗ return last_non_zero;
177 }
178
179 3945600 static int dnxhd_10bit_dct_quantize(MPVEncContext *ctx, int16_t *block,
180 int n, int qscale, int *overflow)
181 {
182 3945600 const uint8_t *scantable = ctx->c.intra_scantable.scantable;
183
2/2
✓ Branch 0 taken 1972800 times.
✓ Branch 1 taken 1972800 times.
 3945600 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
184 3945600 int last_non_zero = 0;
185 int i;
186
187 3945600 ctx->fdsp.fdct(block);
188
189 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
190 3945600 block[0] = (block[0] + 2) >> 2;
191
192
2/2
✓ Branch 0 taken 248572800 times.
✓ Branch 1 taken 3945600 times.
 252518400 for (i = 1; i < 64; ++i) {
193 248572800 int j = scantable[i];
194 248572800 int sign = FF_SIGNBIT(block[j]);
195 248572800 int level = (block[j] ^ sign) - sign;
196 248572800 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
197 248572800 block[j] = (level ^ sign) - sign;
198
2/2
✓ Branch 0 taken 32824223 times.
✓ Branch 1 taken 215748577 times.
 248572800 if (level)
199 32824223 last_non_zero = i;
200 }
201
202 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
203
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3945600 times.
 3945600 if (ctx->c.idsp.perm_type != FF_IDCT_PERM_NONE)
204 ✗ ff_block_permute(block, ctx->c.idsp.idct_permutation,
205 scantable, last_non_zero);
206
207 3945600 return last_non_zero;
208 }
209
210 73 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
211 {
212 int i, j, level, run;
213 73 int max_level = 1 << (ctx->bit_depth + 2);
214
215
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->orig_vlc_codes, max_level * 4) ||
216
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->orig_vlc_bits, max_level * 4) ||
217
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !(ctx->run_codes = av_mallocz(63 * 2)) ||
218
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 !(ctx->run_bits = av_mallocz(63)))
219 ✗ return AVERROR(ENOMEM);
220 73 ctx->vlc_codes = ctx->orig_vlc_codes + max_level * 2;
221 73 ctx->vlc_bits = ctx->orig_vlc_bits + max_level * 2;
222
2/2
✓ Branch 0 taken 192512 times.
✓ Branch 1 taken 73 times.
 192585 for (level = -max_level; level < max_level; level++) {
223
2/2
✓ Branch 0 taken 385024 times.
✓ Branch 1 taken 192512 times.
 577536 for (run = 0; run < 2; run++) {
224 385024 int index = level * (1 << 1) | run;
225 385024 int sign, offset = 0, alevel = level;
226
227 385024 MASK_ABS(sign, alevel);
228
2/2
✓ Branch 0 taken 366190 times.
✓ Branch 1 taken 18834 times.
 385024 if (alevel > 64) {
229 366190 offset = (alevel - 1) >> 6;
230 366190 alevel -= offset << 6;
231 }
232
2/2
✓ Branch 0 taken 65136947 times.
✓ Branch 1 taken 73 times.
 65137020 for (j = 0; j < 257; j++) {
233
4/4
✓ Branch 0 taken 1197496 times.
✓ Branch 1 taken 63939451 times.
✓ Branch 2 taken 1163352 times.
✓ Branch 3 taken 34144 times.
 65136947 if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&
234
5/6
✓ Branch 0 taken 549285 times.
✓ Branch 1 taken 614067 times.
✓ Branch 2 taken 549285 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 390917 times.
✓ Branch 5 taken 192512 times.
 1197496 (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&
235
3/4
✓ Branch 0 taken 192439 times.
✓ Branch 1 taken 198478 times.
✓ Branch 2 taken 192439 times.
✗ Branch 3 not taken.
 390917 (!run || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {
236 av_assert1(!ctx->vlc_codes[index]);
237
2/2
✓ Branch 0 taken 384878 times.
✓ Branch 1 taken 73 times.
 384951 if (alevel) {
238 384878 ctx->vlc_codes[index] =
239 384878 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
240 384878 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
241 } else {
242 73 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
243 73 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
244 }
245 384951 break;
246 }
247 }
248
3/4
✓ Branch 0 taken 384878 times.
✓ Branch 1 taken 146 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 384878 times.
 385024 av_assert0(!alevel || j < 257);
249
2/2
✓ Branch 0 taken 366190 times.
✓ Branch 1 taken 18834 times.
 385024 if (offset) {
250 366190 ctx->vlc_codes[index] =
251 366190 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
252 366190 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
253 }
254 }
255 }
256
2/2
✓ Branch 0 taken 4526 times.
✓ Branch 1 taken 73 times.
 4599 for (i = 0; i < 62; i++) {
257 4526 int run = ctx->cid_table->run[i];
258
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 4526 times.
 4526 av_assert0(run < 63);
259 4526 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
260 4526 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
261 }
262 73 return 0;
263 }
264
265 73 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
266 {
267 // init first elem to 1 to avoid div by 0 in convert_matrix
268 73 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
269 73 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
270 73 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
271
272
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l, ctx->m.c.avctx->qmax + 1) ||
273
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c, ctx->m.c.avctx->qmax + 1) ||
274
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l16, ctx->m.c.avctx->qmax + 1) ||
275
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c16, ctx->m.c.avctx->qmax + 1))
276 ✗ return AVERROR(ENOMEM);
277
278
2/2
✓ Branch 0 taken 66 times.
✓ Branch 1 taken 7 times.
 73 if (ctx->bit_depth == 8) {
279
2/2
✓ Branch 0 taken 4158 times.
✓ Branch 1 taken 66 times.
 4224 for (int i = 1; i < 64; i++) {
280 4158 int j = ctx->m.c.idsp.idct_permutation[ff_zigzag_direct[i]];
281 4158 weight_matrix[j] = ctx->cid_table->luma_weight[i];
282 }
283 66 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
284 weight_matrix, ctx->intra_quant_bias, 1,
285 66 ctx->m.c.avctx->qmax, 1);
286
2/2
✓ Branch 0 taken 4158 times.
✓ Branch 1 taken 66 times.
 4224 for (int i = 1; i < 64; i++) {
287 4158 int j = ctx->m.c.idsp.idct_permutation[ff_zigzag_direct[i]];
288 4158 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
289 }
290 66 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
291 weight_matrix, ctx->intra_quant_bias, 1,
292 66 ctx->m.c.avctx->qmax, 1);
293
294
2/2
✓ Branch 0 taken 42184 times.
✓ Branch 1 taken 66 times.
 42250 for (int qscale = 1; qscale <= ctx->m.c.avctx->qmax; qscale++) {
295
2/2
✓ Branch 0 taken 2699776 times.
✓ Branch 1 taken 42184 times.
 2741960 for (int i = 0; i < 64; i++) {
296 2699776 ctx->qmatrix_l[qscale][i] <<= 2;
297 2699776 ctx->qmatrix_c[qscale][i] <<= 2;
298 2699776 ctx->qmatrix_l16[qscale][0][i] <<= 2;
299 2699776 ctx->qmatrix_l16[qscale][1][i] <<= 2;
300 2699776 ctx->qmatrix_c16[qscale][0][i] <<= 2;
301 2699776 ctx->qmatrix_c16[qscale][1][i] <<= 2;
302 }
303 }
304 } else {
305 // 10-bit
306
2/2
✓ Branch 0 taken 56 times.
✓ Branch 1 taken 7 times.
 63 for (int qscale = 1; qscale <= ctx->m.c.avctx->qmax; qscale++) {
307
2/2
✓ Branch 0 taken 3528 times.
✓ Branch 1 taken 56 times.
 3584 for (int i = 1; i < 64; i++) {
308 3528 int j = ff_zigzag_direct[i];
309
310 /* The quantization formula from the VC-3 standard is:
311 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
312 * (qscale * weight_table[i]))
313 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
314 * The s factor compensates scaling of DCT coefficients done by
315 * the DCT routines, and therefore is not present in standard.
316 * It's 8 for 8-bit samples and 4 for 10-bit ones.
317 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
318 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
319 * (qscale * weight_table[i])
320 * For 10-bit samples, p / s == 2 */
321 3528 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
322 3528 (qscale * luma_weight_table[i]);
323 3528 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
324 3528 (qscale * chroma_weight_table[i]);
325 }
326 }
327 }
328
329 73 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
330 73 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
331 73 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
332 73 ctx->m.q_intra_matrix = ctx->qmatrix_l;
333
334 73 return 0;
335 }
336
337 73 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
338 {
339
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_rc, (ctx->m.c.avctx->qmax + 1) * ctx->m.c.mb_num))
340 ✗ return AVERROR(ENOMEM);
341
342
2/2
✓ Branch 0 taken 58 times.
✓ Branch 1 taken 15 times.
 73 if (ctx->m.c.avctx->mb_decision != FF_MB_DECISION_RD) {
343
1/2
✓ Branch 1 taken 58 times.
✗ Branch 2 not taken.
 58 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp, ctx->m.c.mb_num) ||
344
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 58 times.
 58 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp_tmp, ctx->m.c.mb_num))
345 ✗ return AVERROR(ENOMEM);
346 }
347 73 ctx->frame_bits = (ctx->coding_unit_size -
348 73 ctx->data_offset - 4 - ctx->min_padding) * 8;
349 73 ctx->qscale = 1;
350 73 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
351 73 return 0;
352 }
353
354 73 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
355 {
356 73 DNXHDEncContext *ctx = avctx->priv_data;
357 int i, ret;
358
359
2/3
✓ Branch 0 taken 66 times.
✓ Branch 1 taken 7 times.
✗ Branch 2 not taken.
 73 switch (avctx->pix_fmt) {
360 66 case AV_PIX_FMT_YUV422P:
361 66 ctx->bit_depth = 8;
362 66 break;
363 7 case AV_PIX_FMT_YUV422P10:
364 case AV_PIX_FMT_YUV444P10:
365 case AV_PIX_FMT_GBRP10:
366 7 ctx->bit_depth = 10;
367 7 break;
368 }
369
370
1/4
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
 73 if ((ctx->profile == AV_PROFILE_DNXHR_444 && (avctx->pix_fmt != AV_PIX_FMT_YUV444P10 &&
371 ✗ avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||
372
2/4
✓ Branch 0 taken 73 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
✗ Branch 3 not taken.
 73 (ctx->profile != AV_PROFILE_DNXHR_444 && (avctx->pix_fmt == AV_PIX_FMT_YUV444P10 ||
373
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 avctx->pix_fmt == AV_PIX_FMT_GBRP10))) {
374 ✗ av_log(avctx, AV_LOG_ERROR,
375 "pixel format is incompatible with DNxHD profile\n");
376 ✗ return AVERROR(EINVAL);
377 }
378
379
1/4
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
 73 if (ctx->profile == AV_PROFILE_DNXHR_HQX && avctx->pix_fmt != AV_PIX_FMT_YUV422P10) {
380 ✗ av_log(avctx, AV_LOG_ERROR,
381 "pixel format is incompatible with DNxHR HQX profile\n");
382 ✗ return AVERROR(EINVAL);
383 }
384
385
2/2
✓ Branch 0 taken 61 times.
✓ Branch 1 taken 12 times.
 73 if ((ctx->profile == AV_PROFILE_DNXHR_LB ||
386
2/2
✓ Branch 0 taken 53 times.
✓ Branch 1 taken 8 times.
 61 ctx->profile == AV_PROFILE_DNXHR_SQ ||
387
3/4
✓ Branch 0 taken 20 times.
✓ Branch 1 taken 33 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 40 times.
 73 ctx->profile == AV_PROFILE_DNXHR_HQ) && avctx->pix_fmt != AV_PIX_FMT_YUV422P) {
388 ✗ av_log(avctx, AV_LOG_ERROR,
389 "pixel format is incompatible with DNxHR LB/SQ/HQ profile\n");
390 ✗ return AVERROR(EINVAL);
391 }
392
393 73 ctx->is_444 = ctx->profile == AV_PROFILE_DNXHR_444;
394 73 avctx->profile = ctx->profile;
395 73 ctx->cid = ff_dnxhd_find_cid(avctx, ctx->bit_depth);
396
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (!ctx->cid) {
397 ✗ av_log(avctx, AV_LOG_ERROR,
398 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
399 ✗ ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
400 ✗ return AVERROR(EINVAL);
401 }
402 73 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
403
404
3/4
✓ Branch 0 taken 40 times.
✓ Branch 1 taken 33 times.
✓ Branch 2 taken 40 times.
✗ Branch 3 not taken.
 73 if (ctx->cid >= 1270 && ctx->cid <= 1274)
405 40 avctx->codec_tag = MKTAG('A','V','d','h');
406
407
2/4
✓ Branch 0 taken 73 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 73 times.
 73 if (avctx->width < 256 || avctx->height < 120) {
408 ✗ av_log(avctx, AV_LOG_ERROR,
409 "Input dimensions too small, input must be at least 256x120\n");
410 ✗ return AVERROR(EINVAL);
411 }
412
413 73 ctx->cid_table = ff_dnxhd_get_cid_table(ctx->cid);
414
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 av_assert0(ctx->cid_table);
415
416 73 ctx->m.c.avctx = avctx;
417 73 ctx->m.c.mb_intra = 1;
418 73 ctx->m.c.h263_aic = 1;
419
420 73 avctx->bits_per_raw_sample = ctx->bit_depth;
421
422 73 ff_blockdsp_init(&ctx->m.c.bdsp);
423 73 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
424 73 ff_mpv_idct_init(&ctx->m.c);
425 73 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
426 73 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
427 73 ff_dct_encode_init(&ctx->m);
428
429
2/2
✓ Branch 0 taken 40 times.
✓ Branch 1 taken 33 times.
 73 if (ctx->profile != AV_PROFILE_DNXHD)
430 40 ff_videodsp_init(&ctx->m.c.vdsp, ctx->bit_depth);
431
432
2/4
✓ Branch 0 taken 73 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 73 times.
 73 if (ctx->is_444 || ctx->profile == AV_PROFILE_DNXHR_HQX) {
433 ✗ ctx->m.dct_quantize = dnxhd_10bit_dct_quantize_444;
434 ✗ ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
435 ✗ ctx->block_width_l2 = 4;
436
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 66 times.
 73 } else if (ctx->bit_depth == 10) {
437 7 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
438 7 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
439 7 ctx->block_width_l2 = 4;
440 } else {
441 66 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
442 66 ctx->block_width_l2 = 3;
443 }
444
445 73 ff_dnxhdenc_init(ctx);
446
447 73 ctx->m.c.mb_height = (avctx->height + 15) / 16;
448 73 ctx->m.c.mb_width = (avctx->width + 15) / 16;
449
450
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 62 times.
 73 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
451 11 ctx->interlaced = 1;
452 11 ctx->m.c.mb_height /= 2;
453 }
454
455
3/4
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 62 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 11 times.
 73 if (ctx->interlaced && ctx->profile != AV_PROFILE_DNXHD) {
456 ✗ av_log(avctx, AV_LOG_ERROR,
457 "Interlaced encoding is not supported for DNxHR profiles.\n");
458 ✗ return AVERROR(EINVAL);
459 }
460
461 73 ctx->m.c.mb_num = ctx->m.c.mb_height * ctx->m.c.mb_width;
462
463
2/2
✓ Branch 0 taken 40 times.
✓ Branch 1 taken 33 times.
 73 if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
464 40 ctx->frame_size = ff_dnxhd_get_hr_frame_size(ctx->cid,
465 avctx->width, avctx->height);
466
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 40 times.
 40 av_assert0(ctx->frame_size >= 0);
467 40 ctx->coding_unit_size = ctx->frame_size;
468 } else {
469 33 ctx->frame_size = ctx->cid_table->frame_size;
470 33 ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
471 }
472
473
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 61 times.
 73 if (ctx->m.c.mb_height > 68)
474 12 ctx->data_offset = 0x170 + (ctx->m.c.mb_height << 2);
475 else
476 61 ctx->data_offset = 0x280;
477
478 // XXX tune lbias/cbias
479
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
480 ✗ return ret;
481
482 /* Avid Nitris hardware decoder requires a minimum amount of padding
483 * in the coding unit payload */
484
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (ctx->nitris_compat)
485 ✗ ctx->min_padding = 1600;
486
487
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 if ((ret = dnxhd_init_vlc(ctx)) < 0)
488 ✗ return ret;
489
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 if ((ret = dnxhd_init_rc(ctx)) < 0)
490 ✗ return ret;
491
492
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->slice_size, ctx->m.c.mb_height) ||
493
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->slice_offs, ctx->m.c.mb_height) ||
494
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_bits, ctx->m.c.mb_num) ||
495
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_qscale, ctx->m.c.mb_num))
496 ✗ return AVERROR(ENOMEM);
497
498
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (avctx->active_thread_type == FF_THREAD_SLICE) {
499 ✗ if (avctx->thread_count > MAX_THREADS) {
500 ✗ av_log(avctx, AV_LOG_ERROR, "too many threads\n");
501 ✗ return AVERROR(EINVAL);
502 }
503 }
504
505
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (avctx->qmax <= 1) {
506 ✗ av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
507 ✗ return AVERROR(EINVAL);
508 }
509
510 73 ctx->thread[0] = ctx;
511
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (avctx->active_thread_type == FF_THREAD_SLICE) {
512 ✗ for (i = 1; i < avctx->thread_count; i++) {
513 ✗ ctx->thread[i] = av_memdup(ctx, sizeof(DNXHDEncContext));
514 ✗ if (!ctx->thread[i])
515 ✗ return AVERROR(ENOMEM);
516 }
517 }
518
519 73 return 0;
520 }
521
522 380 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
523 {
524 380 DNXHDEncContext *ctx = avctx->priv_data;
525
526 380 memset(buf, 0, ctx->data_offset);
527
528 // * write prefix */
529 380 AV_WB16(buf + 0x02, ctx->data_offset);
530
3/4
✓ Branch 0 taken 200 times.
✓ Branch 1 taken 180 times.
✓ Branch 2 taken 200 times.
✗ Branch 3 not taken.
 380 if (ctx->cid >= 1270 && ctx->cid <= 1274)
531 200 buf[4] = 0x03;
532 else
533 180 buf[4] = 0x01;
534
535
2/2
✓ Branch 0 taken 110 times.
✓ Branch 1 taken 270 times.
 380 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
536 380 buf[6] = 0x80; // crc flag off
537 380 buf[7] = 0xa0; // reserved
538 380 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
539 380 AV_WB16(buf + 0x1a, avctx->width); // SPL
540 380 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
541
542
2/2
✓ Branch 0 taken 55 times.
✓ Branch 1 taken 325 times.
 380 buf[0x21] = ctx->bit_depth == 10 ? 0x58 : 0x38;
543 380 buf[0x22] = 0x88 + (ctx->interlaced << 2);
544 380 AV_WB32(buf + 0x28, ctx->cid); // CID
545
2/2
✓ Branch 0 taken 270 times.
✓ Branch 1 taken 110 times.
 380 buf[0x2c] = (!ctx->interlaced << 7) | (ctx->is_444 << 6) | (avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
546
547 380 buf[0x5f] = 0x01; // UDL
548
549 380 buf[0x167] = 0x02; // reserved
550 380 AV_WB16(buf + 0x16a, ctx->m.c.mb_height * 4 + 4); // MSIPS
551 380 AV_WB16(buf + 0x16c, ctx->m.c.mb_height); // Ns
552 380 buf[0x16f] = 0x10; // reserved
553
554 380 ctx->msip = buf + 0x170;
555 380 return 0;
556 }
557
558 26918560 static av_always_inline void dnxhd_encode_dc(PutBitContext *pb, DNXHDEncContext *ctx, int diff)
559 {
560 int nbits;
561
2/2
✓ Branch 0 taken 12317800 times.
✓ Branch 1 taken 14600760 times.
 26918560 if (diff < 0) {
562 12317800 nbits = av_log2_16bit(-2 * diff);
563 12317800 diff--;
564 } else {
565 14600760 nbits = av_log2_16bit(2 * diff);
566 }
567 26918560 put_bits(pb, ctx->cid_table->dc_bits[nbits] + nbits,
568 26918560 (ctx->cid_table->dc_codes[nbits] << nbits) +
569 26918560 av_zero_extend(diff, nbits));
570 26918560 }
571
572 static av_always_inline
573 26918560 void dnxhd_encode_block(PutBitContext *pb, DNXHDEncContext *ctx,
574 int16_t *block, int last_index, int n)
575 {
576 26918560 int last_non_zero = 0;
577 int slevel, i, j;
578
579 26918560 dnxhd_encode_dc(pb, ctx, block[0] - ctx->m.c.last_dc[n]);
580 26918560 ctx->m.c.last_dc[n] = block[0];
581
582
2/2
✓ Branch 0 taken 296815535 times.
✓ Branch 1 taken 26918560 times.
 323734095 for (i = 1; i <= last_index; i++) {
583 296815535 j = ctx->m.c.intra_scantable.permutated[i];
584 296815535 slevel = block[j];
585
2/2
✓ Branch 0 taken 111020901 times.
✓ Branch 1 taken 185794634 times.
 296815535 if (slevel) {
586 111020901 int run_level = i - last_non_zero - 1;
587 111020901 int rlevel = slevel * (1 << 1) | !!run_level;
588 111020901 put_bits(pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
589
2/2
✓ Branch 0 taken 36001883 times.
✓ Branch 1 taken 75019018 times.
 111020901 if (run_level)
590 36001883 put_bits(pb, ctx->run_bits[run_level],
591 36001883 ctx->run_codes[run_level]);
592 111020901 last_non_zero = i;
593 }
594 }
595 26918560 put_bits(pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
596 26918560 }
597
598 static av_always_inline
599 3024000 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
600 int qscale, int last_index)
601 {
602 const uint8_t *weight_matrix;
603 int level;
604 int i;
605
606
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3024000 times.
 3024000 if (ctx->is_444) {
607 ✗ weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight
608 ✗ : ctx->cid_table->chroma_weight;
609 } else {
610 3024000 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
611
2/2
✓ Branch 0 taken 1512000 times.
✓ Branch 1 taken 1512000 times.
 3024000 : ctx->cid_table->luma_weight;
612 }
613
614
2/2
✓ Branch 0 taken 32773315 times.
✓ Branch 1 taken 3024000 times.
 35797315 for (i = 1; i <= last_index; i++) {
615 32773315 int j = ctx->m.c.intra_scantable.permutated[i];
616 32773315 level = block[j];
617
2/2
✓ Branch 0 taken 16774641 times.
✓ Branch 1 taken 15998674 times.
 32773315 if (level) {
618
2/2
✓ Branch 0 taken 8458215 times.
✓ Branch 1 taken 8316426 times.
 16774641 if (level < 0) {
619 8458215 level = (1 - 2 * level) * qscale * weight_matrix[i];
620
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8458215 times.
 8458215 if (ctx->bit_depth == 10) {
621 ✗ if (weight_matrix[i] != 8)
622 ✗ level += 8;
623 ✗ level >>= 4;
624 } else {
625
2/2
✓ Branch 0 taken 6479585 times.
✓ Branch 1 taken 1978630 times.
 8458215 if (weight_matrix[i] != 32)
626 6479585 level += 32;
627 8458215 level >>= 6;
628 }
629 8458215 level = -level;
630 } else {
631 8316426 level = (2 * level + 1) * qscale * weight_matrix[i];
632
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8316426 times.
 8316426 if (ctx->bit_depth == 10) {
633 ✗ if (weight_matrix[i] != 8)
634 ✗ level += 8;
635 ✗ level >>= 4;
636 } else {
637
2/2
✓ Branch 0 taken 6813970 times.
✓ Branch 1 taken 1502456 times.
 8316426 if (weight_matrix[i] != 32)
638 6813970 level += 32;
639 8316426 level >>= 6;
640 }
641 }
642 16774641 block[j] = level;
643 }
644 }
645 3024000 }
646
647 3024000 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
648 {
649 3024000 int score = 0;
650 int i;
651
2/2
✓ Branch 0 taken 193536000 times.
✓ Branch 1 taken 3024000 times.
 196560000 for (i = 0; i < 64; i++)
652 193536000 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
653 3024000 return score;
654 }
655
656 static av_always_inline
657 51881432 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
658 {
659 51881432 int last_non_zero = 0;
660 51881432 int bits = 0;
661 int i, j, level;
662
2/2
✓ Branch 0 taken 463138205 times.
✓ Branch 1 taken 51881432 times.
 515019637 for (i = 1; i <= last_index; i++) {
663 463138205 j = ctx->m.c.intra_scantable.permutated[i];
664 463138205 level = block[j];
665
2/2
✓ Branch 0 taken 191817206 times.
✓ Branch 1 taken 271320999 times.
 463138205 if (level) {
666 191817206 int run_level = i - last_non_zero - 1;
667 191817206 bits += ctx->vlc_bits[level * (1 << 1) |
668 191817206 !!run_level] + ctx->run_bits[run_level];
669 191817206 last_non_zero = i;
670 }
671 }
672 51881432 return bits;
673 }
674
675 static av_always_inline
676 9849999 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
677 {
678 9849999 const int bs = ctx->block_width_l2;
679 9849999 const int bw = 1 << bs;
680 9849999 int dct_y_offset = ctx->dct_y_offset;
681 9849999 int dct_uv_offset = ctx->dct_uv_offset;
682 9849999 int linesize = ctx->m.c.linesize;
683 9849999 int uvlinesize = ctx->m.c.uvlinesize;
684 9849999 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
685 9849999 ((mb_y << 4) * ctx->m.c.linesize) + (mb_x << bs + 1);
686 9849999 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
687 9849999 ((mb_y << 4) * ctx->m.c.uvlinesize) + (mb_x << bs + ctx->is_444);
688 9849999 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
689 9849999 ((mb_y << 4) * ctx->m.c.uvlinesize) + (mb_x << bs + ctx->is_444);
690 9849999 PixblockDSPContext *pdsp = &ctx->m.pdsp;
691 9849999 VideoDSPContext *vdsp = &ctx->m.c.vdsp;
692
693
6/6
✓ Branch 0 taken 9356799 times.
✓ Branch 1 taken 493200 times.
✓ Branch 2 taken 7207599 times.
✓ Branch 3 taken 2149200 times.
✓ Branch 4 taken 7203152 times.
✓ Branch 5 taken 4447 times.
 9849999 if (ctx->bit_depth != 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.c.avctx->width ||
694
2/2
✓ Branch 0 taken 18064 times.
✓ Branch 1 taken 7185088 times.
 7203152 (mb_y << 4) + 16 > ctx->m.c.avctx->height)) {
695 22511 int y_w = ctx->m.c.avctx->width - (mb_x << 4);
696 22511 int y_h = ctx->m.c.avctx->height - (mb_y << 4);
697 22511 int uv_w = (y_w + 1) / 2;
698 22511 int uv_h = y_h;
699 22511 linesize = 16;
700 22511 uvlinesize = 8;
701
702 22511 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
703 linesize, ctx->m.c.linesize,
704 linesize, 16,
705 0, 0, y_w, y_h);
706 22511 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
707 uvlinesize, ctx->m.c.uvlinesize,
708 uvlinesize, 16,
709 0, 0, uv_w, uv_h);
710 22511 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
711 uvlinesize, ctx->m.c.uvlinesize,
712 uvlinesize, 16,
713 0, 0, uv_w, uv_h);
714
715 22511 dct_y_offset = bw * linesize;
716 22511 dct_uv_offset = bw * uvlinesize;
717 22511 ptr_y = &ctx->edge_buf_y[0];
718 22511 ptr_u = &ctx->edge_buf_uv[0][0];
719 22511 ptr_v = &ctx->edge_buf_uv[1][0];
720
3/6
✓ Branch 0 taken 493200 times.
✓ Branch 1 taken 9334288 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 493200 times.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
 9827488 } else if (ctx->bit_depth == 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.c.avctx->width ||
721 ✗ (mb_y << 4) + 16 > ctx->m.c.avctx->height)) {
722 ✗ int y_w = ctx->m.c.avctx->width - (mb_x << 4);
723 ✗ int y_h = ctx->m.c.avctx->height - (mb_y << 4);
724 ✗ int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;
725 ✗ int uv_h = y_h;
726 ✗ linesize = 32;
727 ✗ uvlinesize = 16 + 16 * ctx->is_444;
728
729 ✗ vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
730 linesize, ctx->m.c.linesize,
731 linesize / 2, 16,
732 0, 0, y_w, y_h);
733 ✗ vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
734 uvlinesize, ctx->m.c.uvlinesize,
735 uvlinesize / 2, 16,
736 0, 0, uv_w, uv_h);
737 ✗ vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
738 uvlinesize, ctx->m.c.uvlinesize,
739 uvlinesize / 2, 16,
740 0, 0, uv_w, uv_h);
741
742 ✗ dct_y_offset = bw * linesize / 2;
743 ✗ dct_uv_offset = bw * uvlinesize / 2;
744 ✗ ptr_y = &ctx->edge_buf_y[0];
745 ✗ ptr_u = &ctx->edge_buf_uv[0][0];
746 ✗ ptr_v = &ctx->edge_buf_uv[1][0];
747 }
748
749
1/2
✓ Branch 0 taken 9849999 times.
✗ Branch 1 not taken.
 9849999 if (!ctx->is_444) {
750 9849999 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
751 9849999 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
752 9849999 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
753 9849999 pdsp->get_pixels(ctx->blocks[3], ptr_v, uvlinesize);
754
755
4/4
✓ Branch 0 taken 131663 times.
✓ Branch 1 taken 9718336 times.
✓ Branch 2 taken 56560 times.
✓ Branch 3 taken 75103 times.
 9849999 if (mb_y + 1 == ctx->m.c.mb_height && ctx->m.c.avctx->height == 1080) {
756
2/2
✓ Branch 0 taken 35280 times.
✓ Branch 1 taken 21280 times.
 56560 if (ctx->interlaced) {
757 35280 ctx->get_pixels_8x4_sym(ctx->blocks[4],
758 ptr_y + dct_y_offset,
759 linesize);
760 35280 ctx->get_pixels_8x4_sym(ctx->blocks[5],
761 35280 ptr_y + dct_y_offset + bw,
762 linesize);
763 35280 ctx->get_pixels_8x4_sym(ctx->blocks[6],
764 ptr_u + dct_uv_offset,
765 uvlinesize);
766 35280 ctx->get_pixels_8x4_sym(ctx->blocks[7],
767 ptr_v + dct_uv_offset,
768 uvlinesize);
769 } else {
770 21280 ctx->m.c.bdsp.clear_block(ctx->blocks[4]);
771 21280 ctx->m.c.bdsp.clear_block(ctx->blocks[5]);
772 21280 ctx->m.c.bdsp.clear_block(ctx->blocks[6]);
773 21280 ctx->m.c.bdsp.clear_block(ctx->blocks[7]);
774 }
775 } else {
776 9793439 pdsp->get_pixels(ctx->blocks[4],
777 ptr_y + dct_y_offset, linesize);
778 9793439 pdsp->get_pixels(ctx->blocks[5],
779 9793439 ptr_y + dct_y_offset + bw, linesize);
780 9793439 pdsp->get_pixels(ctx->blocks[6],
781 ptr_u + dct_uv_offset, uvlinesize);
782 9793439 pdsp->get_pixels(ctx->blocks[7],
783 ptr_v + dct_uv_offset, uvlinesize);
784 }
785 } else {
786 ✗ pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
787 ✗ pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
788 ✗ pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);
789 ✗ pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);
790
791 ✗ pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
792 ✗ pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);
793 ✗ pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);
794 ✗ pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);
795
796 ✗ pdsp->get_pixels(ctx->blocks[4], ptr_v, uvlinesize);
797 ✗ pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);
798 ✗ pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);
799 ✗ pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);
800 }
801 9849999 }
802
803 static av_always_inline
804 78799992 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
805 {
806 int x;
807
808
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 78799992 times.
 78799992 if (ctx->is_444) {
809 ✗ x = (i >> 1) % 3;
810 } else {
811 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
812 78799992 x = component[i];
813 }
814 78799992 return x;
815 }
816
817 41557 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
818 int jobnr, int threadnr)
819 {
820 41557 DNXHDEncContext *ctx = avctx->priv_data;
821 41557 int mb_y = jobnr;
822 41557 int qscale = ctx->qscale;
823 41557 LOCAL_ALIGNED_16(int16_t, block, [64]);
824 41557 ctx = ctx->thread[threadnr];
825
826 41557 ctx->m.c.last_dc[0] =
827 41557 ctx->m.c.last_dc[1] =
828 41557 ctx->m.c.last_dc[2] = 1 << (ctx->bit_depth + 2);
829
830
2/2
✓ Branch 0 taken 6485179 times.
✓ Branch 1 taken 41557 times.
 6526736 for (int mb_x = 0; mb_x < ctx->m.c.mb_width; mb_x++) {
831 6485179 unsigned mb = mb_y * ctx->m.c.mb_width + mb_x;
832 6485179 int ssd = 0;
833 6485179 int ac_bits = 0;
834 6485179 int dc_bits = 0;
835 int i;
836
837 6485179 dnxhd_get_blocks(ctx, mb_x, mb_y);
838
839
2/2
✓ Branch 0 taken 51881432 times.
✓ Branch 1 taken 6485179 times.
 58366611 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
840 51881432 int16_t *src_block = ctx->blocks[i];
841 int overflow, nbits, diff, last_index;
842 51881432 int n = dnxhd_switch_matrix(ctx, i);
843
844 51881432 memcpy(block, src_block, 64 * sizeof(*block));
845 51881432 last_index = ctx->m.dct_quantize(&ctx->m, block,
846
1/4
✗ Branch 0 not taken.
✓ Branch 1 taken 51881432 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
 51881432 ctx->is_444 ? 4 * (n > 0): 4 & (2*i),
847 qscale, &overflow);
848 51881432 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
849
850 51881432 diff = block[0] - ctx->m.c.last_dc[n];
851
2/2
✓ Branch 0 taken 23766663 times.
✓ Branch 1 taken 28114769 times.
 51881432 if (diff < 0)
852 23766663 nbits = av_log2_16bit(-2 * diff);
853 else
854 28114769 nbits = av_log2_16bit(2 * diff);
855
856 av_assert1(nbits < ctx->bit_depth + 4);
857 51881432 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
858
859 51881432 ctx->m.c.last_dc[n] = block[0];
860
861
2/2
✓ Branch 0 taken 3024000 times.
✓ Branch 1 taken 48857432 times.
 51881432 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
862 3024000 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
863 3024000 ctx->m.c.idsp.idct(block);
864 3024000 ssd += dnxhd_ssd_block(block, src_block);
865 }
866 }
867 6485179 ctx->mb_rc[(qscale * ctx->m.c.mb_num) + mb].ssd = ssd;
868 6485179 ctx->mb_rc[(qscale * ctx->m.c.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
869 6485179 (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];
870 }
871 41557 return 0;
872 }
873
874 20765 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
875 int jobnr, int threadnr)
876 {
877 20765 DNXHDEncContext *ctx = avctx->priv_data;
878 20765 PutBitContext pb0, *const pb = &pb0;
879 20765 int mb_y = jobnr;
880 20765 ctx = ctx->thread[threadnr];
881 20765 init_put_bits(pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
882 20765 ctx->slice_size[jobnr]);
883
884 20765 ctx->m.c.last_dc[0] =
885 20765 ctx->m.c.last_dc[1] =
886 20765 ctx->m.c.last_dc[2] = 1 << (ctx->bit_depth + 2);
887
2/2
✓ Branch 0 taken 3364820 times.
✓ Branch 1 taken 20765 times.
 3385585 for (int mb_x = 0; mb_x < ctx->m.c.mb_width; mb_x++) {
888 3364820 unsigned mb = mb_y * ctx->m.c.mb_width + mb_x;
889 3364820 int qscale = ctx->mb_qscale[mb];
890 int i;
891
892 3364820 put_bits(pb, 11, qscale);
893 3364820 put_bits(pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
894
895 3364820 dnxhd_get_blocks(ctx, mb_x, mb_y);
896
897
2/2
✓ Branch 0 taken 26918560 times.
✓ Branch 1 taken 3364820 times.
 30283380 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
898 26918560 int16_t *block = ctx->blocks[i];
899 26918560 int overflow, n = dnxhd_switch_matrix(ctx, i);
900 26918560 int last_index = ctx->m.dct_quantize(&ctx->m, block,
901
1/4
✗ Branch 0 not taken.
✓ Branch 1 taken 26918560 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
 26918560 ctx->is_444 ? (((i >> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),
902 qscale, &overflow);
903
904 26918560 dnxhd_encode_block(pb, ctx, block, last_index, n);
905 }
906 }
907 20765 flush_put_bits(pb);
908 20765 memset(put_bits_ptr(pb), 0, put_bytes_left(pb, 0));
909 20765 return 0;
910 }
911
912 380 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
913 {
914
2/2
✓ Branch 0 taken 20765 times.
✓ Branch 1 taken 380 times.
 21145 for (int mb_y = 0, offset = 0; mb_y < ctx->m.c.mb_height; mb_y++) {
915 int thread_size;
916 20765 ctx->slice_offs[mb_y] = offset;
917 20765 ctx->slice_size[mb_y] = 0;
918
2/2
✓ Branch 0 taken 3364820 times.
✓ Branch 1 taken 20765 times.
 3385585 for (int mb_x = 0; mb_x < ctx->m.c.mb_width; mb_x++) {
919 3364820 unsigned mb = mb_y * ctx->m.c.mb_width + mb_x;
920 3364820 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
921 }
922 20765 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31U) & ~31U;
923 20765 ctx->slice_size[mb_y] >>= 3;
924 20765 thread_size = ctx->slice_size[mb_y];
925 20765 offset += thread_size;
926 }
927 380 }
928
929 11520 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
930 int jobnr, int threadnr)
931 {
932 11520 DNXHDEncContext *ctx = avctx->priv_data;
933 11520 int mb_y = jobnr, x, y;
934
2/2
✓ Branch 0 taken 220 times.
✓ Branch 1 taken 11300 times.
 11740 int partial_last_row = (mb_y == ctx->m.c.mb_height - 1) &&
935
2/2
✓ Branch 0 taken 135 times.
✓ Branch 1 taken 85 times.
 220 ((avctx->height >> ctx->interlaced) & 0xF);
936
937 11520 ctx = ctx->thread[threadnr];
938
2/2
✓ Branch 0 taken 10955 times.
✓ Branch 1 taken 565 times.
 11520 if (ctx->bit_depth == 8) {
939 10955 const uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.c.linesize);
940
2/2
✓ Branch 0 taken 1919700 times.
✓ Branch 1 taken 10955 times.
 1930655 for (int mb_x = 0; mb_x < ctx->m.c.mb_width; ++mb_x, pix += 16) {
941 1919700 unsigned mb = mb_y * ctx->m.c.mb_width + mb_x;
942 int sum;
943 int varc;
944
945
6/6
✓ Branch 0 taken 1908265 times.
✓ Branch 1 taken 11435 times.
✓ Branch 2 taken 1907880 times.
✓ Branch 3 taken 385 times.
✓ Branch 4 taken 1901720 times.
✓ Branch 5 taken 6160 times.
 1919700 if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {
946 1901720 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.c.linesize);
947 1901720 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.c.linesize);
948 } else {
949 17980 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
950 17980 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
951 17980 sum = varc = 0;
952
2/2
✓ Branch 0 taken 226195 times.
✓ Branch 1 taken 17980 times.
 244175 for (y = 0; y < bh; y++) {
953
2/2
✓ Branch 0 taken 3573835 times.
✓ Branch 1 taken 226195 times.
 3800030 for (x = 0; x < bw; x++) {
954 3573835 uint8_t val = pix[x + y * ctx->m.c.linesize];
955 3573835 sum += val;
956 3573835 varc += val * val;
957 }
958 }
959 }
960 1919700 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
961
962 1919700 ctx->mb_cmp[mb].value = varc;
963 1919700 ctx->mb_cmp[mb].mb = mb;
964 }
965 } else { // 10-bit
966 565 const int linesize = ctx->m.c.linesize >> 1;
967
2/2
✓ Branch 0 taken 58800 times.
✓ Branch 1 taken 565 times.
 59365 for (int mb_x = 0; mb_x < ctx->m.c.mb_width; ++mb_x) {
968 58800 const uint16_t *pix = (const uint16_t *)ctx->thread[0]->src[0] +
969 58800 ((mb_y << 4) * linesize) + (mb_x << 4);
970 58800 unsigned mb = mb_y * ctx->m.c.mb_width + mb_x;
971 58800 int sum = 0;
972 58800 int sqsum = 0;
973 58800 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
974 58800 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
975 int mean, sqmean;
976 int i, j;
977 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
978
2/2
✓ Branch 0 taken 936000 times.
✓ Branch 1 taken 58800 times.
 994800 for (i = 0; i < bh; ++i) {
979
2/2
✓ Branch 0 taken 14976000 times.
✓ Branch 1 taken 936000 times.
 15912000 for (j = 0; j < bw; ++j) {
980 // Turn 16-bit pixels into 10-bit ones.
981 14976000 const int sample = (unsigned) pix[j] >> 6;
982 14976000 sum += sample;
983 14976000 sqsum += sample * sample;
984 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
985 }
986 936000 pix += linesize;
987 }
988 58800 mean = sum >> 8; // 16*16 == 2^8
989 58800 sqmean = sqsum >> 8;
990 58800 ctx->mb_cmp[mb].value = sqmean - mean * mean;
991 58800 ctx->mb_cmp[mb].mb = mb;
992 }
993 }
994 11520 return 0;
995 }
996
997 15 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
998 {
999 int lambda, up_step, down_step;
1000 15 int last_lower = INT_MAX, last_higher = 0;
1001
1002
2/2
✓ Branch 0 taken 105 times.
✓ Branch 1 taken 15 times.
 120 for (int q = 1; q < avctx->qmax; q++) {
1003 105 ctx->qscale = q;
1004 105 avctx->execute2(avctx, dnxhd_calc_bits_thread,
1005 NULL, NULL, ctx->m.c.mb_height);
1006 }
1007 15 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
1008 15 lambda = ctx->lambda;
1009
1010 70 for (;;) {
1011 85 int bits = 0;
1012 85 int end = 0;
1013
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 80 times.
 85 if (lambda == last_higher) {
1014 5 lambda++;
1015 5 end = 1; // need to set final qscales/bits
1016 }
1017
2/2
✓ Branch 0 taken 3809 times.
✓ Branch 1 taken 53 times.
 3862 for (int y = 0; y < ctx->m.c.mb_height; y++) {
1018
2/2
✓ Branch 0 taken 304720 times.
✓ Branch 1 taken 3809 times.
 308529 for (int x = 0; x < ctx->m.c.mb_width; x++) {
1019 304720 unsigned min = UINT_MAX;
1020 304720 int qscale = 1;
1021 304720 int mb = y * ctx->m.c.mb_width + x;
1022 304720 int rc = 0;
1023
2/2
✓ Branch 0 taken 2133040 times.
✓ Branch 1 taken 304720 times.
 2437760 for (int q = 1; q < avctx->qmax; q++) {
1024 2133040 int i = (q*ctx->m.c.mb_num) + mb;
1025 2133040 unsigned score = ctx->mb_rc[i].bits * lambda +
1026 2133040 ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
1027
2/2
✓ Branch 0 taken 493989 times.
✓ Branch 1 taken 1639051 times.
 2133040 if (score < min) {
1028 493989 min = score;
1029 493989 qscale = q;
1030 493989 rc = i;
1031 }
1032 }
1033 304720 bits += ctx->mb_rc[rc].bits;
1034 304720 ctx->mb_qscale[mb] = qscale;
1035 304720 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1036 }
1037 3809 bits = (bits + 31) & ~31; // padding
1038
2/2
✓ Branch 0 taken 32 times.
✓ Branch 1 taken 3777 times.
 3809 if (bits > ctx->frame_bits)
1039 32 break;
1040 }
1041
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 80 times.
 85 if (end) {
1042
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 5 times.
 5 if (bits > ctx->frame_bits)
1043 ✗ return AVERROR(EINVAL);
1044 5 break;
1045 }
1046
2/2
✓ Branch 0 taken 48 times.
✓ Branch 1 taken 32 times.
 80 if (bits < ctx->frame_bits) {
1047 48 last_lower = FFMIN(lambda, last_lower);
1048
2/2
✓ Branch 0 taken 25 times.
✓ Branch 1 taken 23 times.
 48 if (last_higher != 0)
1049 25 lambda = (lambda+last_higher)>>1;
1050 else
1051 23 lambda -= down_step;
1052 48 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
1053 48 up_step = 1<<LAMBDA_FRAC_BITS;
1054 48 lambda = FFMAX(1, lambda);
1055
2/2
✓ Branch 0 taken 10 times.
✓ Branch 1 taken 38 times.
 48 if (lambda == last_lower)
1056 10 break;
1057 } else {
1058 32 last_higher = FFMAX(lambda, last_higher);
1059
1/2
✓ Branch 0 taken 32 times.
✗ Branch 1 not taken.
 32 if (last_lower != INT_MAX)
1060 32 lambda = (lambda+last_lower)>>1;
1061 ✗ else if ((int64_t)lambda + up_step > INT_MAX)
1062 ✗ return AVERROR(EINVAL);
1063 else
1064 ✗ lambda += up_step;
1065 32 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
1066 32 down_step = 1<<LAMBDA_FRAC_BITS;
1067 }
1068 }
1069 15 ctx->lambda = lambda;
1070 15 return 0;
1071 }
1072
1073 365 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
1074 {
1075 365 int bits = 0;
1076 365 int up_step = 1;
1077 365 int down_step = 1;
1078 365 int last_higher = 0;
1079 365 int last_lower = INT_MAX;
1080 int qscale;
1081
1082 365 qscale = ctx->qscale;
1083 for (;;) {
1084 691 bits = 0;
1085 691 ctx->qscale = qscale;
1086 // XXX avoid recalculating bits
1087 691 ctx->m.c.avctx->execute2(ctx->m.c.avctx, dnxhd_calc_bits_thread,
1088 NULL, NULL, ctx->m.c.mb_height);
1089
2/2
✓ Branch 0 taken 34838 times.
✓ Branch 1 taken 393 times.
 35231 for (int y = 0; y < ctx->m.c.mb_height; y++) {
1090
2/2
✓ Branch 0 taken 5760121 times.
✓ Branch 1 taken 34838 times.
 5794959 for (int x = 0; x < ctx->m.c.mb_width; x++)
1091 5760121 bits += ctx->mb_rc[(qscale*ctx->m.c.mb_num) + (y*ctx->m.c.mb_width+x)].bits;
1092 34838 bits = (bits+31)&~31; // padding
1093
2/2
✓ Branch 0 taken 298 times.
✓ Branch 1 taken 34540 times.
 34838 if (bits > ctx->frame_bits)
1094 298 break;
1095 }
1096
2/2
✓ Branch 0 taken 393 times.
✓ Branch 1 taken 298 times.
 691 if (bits < ctx->frame_bits) {
1097
2/2
✓ Branch 0 taken 145 times.
✓ Branch 1 taken 248 times.
 393 if (qscale == 1)
1098 145 return 1;
1099
2/2
✓ Branch 0 taken 195 times.
✓ Branch 1 taken 53 times.
 248 if (last_higher == qscale - 1) {
1100 195 qscale = last_higher;
1101 195 break;
1102 }
1103 53 last_lower = FFMIN(qscale, last_lower);
1104
2/2
✓ Branch 0 taken 38 times.
✓ Branch 1 taken 15 times.
 53 if (last_higher != 0)
1105 38 qscale = (qscale + last_higher) >> 1;
1106 else
1107 15 qscale -= down_step++;
1108
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 53 times.
 53 if (qscale < 1)
1109 ✗ qscale = 1;
1110 53 up_step = 1;
1111 } else {
1112
2/2
✓ Branch 0 taken 25 times.
✓ Branch 1 taken 273 times.
 298 if (last_lower == qscale + 1)
1113 25 break;
1114 273 last_higher = FFMAX(qscale, last_higher);
1115
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 268 times.
 273 if (last_lower != INT_MAX)
1116 5 qscale = (qscale + last_lower) >> 1;
1117 else
1118 268 qscale += up_step++;
1119 273 down_step = 1;
1120
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 273 times.
 273 if (qscale >= ctx->m.c.avctx->qmax)
1121 ✗ return AVERROR(EINVAL);
1122 }
1123 }
1124 220 ctx->qscale = qscale;
1125 220 return 0;
1126 }
1127
1128 #define BUCKET_BITS 8
1129 #define RADIX_PASSES 4
1130 #define NBUCKETS (1 << BUCKET_BITS)
1131
1132 11871000 static inline int get_bucket(int value, int shift)
1133 {
1134 11871000 value >>= shift;
1135 11871000 value &= NBUCKETS - 1;
1136 11871000 return NBUCKETS - 1 - value;
1137 }
1138
1139 220 static void radix_count(const RCCMPEntry *data, int size,
1140 int buckets[RADIX_PASSES][NBUCKETS])
1141 {
1142 int i, j;
1143 220 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1144
2/2
✓ Branch 0 taken 1978500 times.
✓ Branch 1 taken 220 times.
 1978720 for (i = 0; i < size; i++) {
1145 1978500 int v = data[i].value;
1146
2/2
✓ Branch 0 taken 7914000 times.
✓ Branch 1 taken 1978500 times.
 9892500 for (j = 0; j < RADIX_PASSES; j++) {
1147 7914000 buckets[j][get_bucket(v, 0)]++;
1148 7914000 v >>= BUCKET_BITS;
1149 }
1150 av_assert1(!v);
1151 }
1152
2/2
✓ Branch 0 taken 880 times.
✓ Branch 1 taken 220 times.
 1100 for (j = 0; j < RADIX_PASSES; j++) {
1153 880 int offset = size;
1154
2/2
✓ Branch 0 taken 225280 times.
✓ Branch 1 taken 880 times.
 226160 for (i = NBUCKETS - 1; i >= 0; i--)
1155 225280 buckets[j][i] = offset -= buckets[j][i];
1156 av_assert1(!buckets[j][0]);
1157 }
1158 220 }
1159
1160 440 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1161 int size, int buckets[NBUCKETS], int pass)
1162 {
1163 440 int shift = pass * BUCKET_BITS;
1164 int i;
1165
2/2
✓ Branch 0 taken 3957000 times.
✓ Branch 1 taken 440 times.
 3957440 for (i = 0; i < size; i++) {
1166 3957000 int v = get_bucket(data[i].value, shift);
1167 3957000 int pos = buckets[v]++;
1168 3957000 dst[pos] = data[i];
1169 }
1170 440 }
1171
1172 220 static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)
1173 {
1174 int buckets[RADIX_PASSES][NBUCKETS];
1175 220 radix_count(data, size, buckets);
1176 220 radix_sort_pass(tmp, data, size, buckets[0], 0);
1177 220 radix_sort_pass(data, tmp, size, buckets[1], 1);
1178
2/4
✓ Branch 0 taken 220 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 220 times.
 220 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
1179 ✗ radix_sort_pass(tmp, data, size, buckets[2], 2);
1180 ✗ radix_sort_pass(data, tmp, size, buckets[3], 3);
1181 }
1182 220 }
1183
1184 365 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1185 {
1186 365 int max_bits = 0;
1187 int ret;
1188
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 365 times.
 365 if ((ret = dnxhd_find_qscale(ctx)) < 0)
1189 ✗ return ret;
1190
2/2
✓ Branch 0 taken 20090 times.
✓ Branch 1 taken 365 times.
 20455 for (int y = 0; y < ctx->m.c.mb_height; y++) {
1191
2/2
✓ Branch 0 taken 3310820 times.
✓ Branch 1 taken 20090 times.
 3330910 for (int x = 0; x < ctx->m.c.mb_width; x++) {
1192 3310820 int mb = y * ctx->m.c.mb_width + x;
1193 3310820 int rc = (ctx->qscale * ctx->m.c.mb_num ) + mb;
1194 int delta_bits;
1195 3310820 ctx->mb_qscale[mb] = ctx->qscale;
1196 3310820 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1197 3310820 max_bits += ctx->mb_rc[rc].bits;
1198 if (!RC_VARIANCE) {
1199 delta_bits = ctx->mb_rc[rc].bits -
1200 ctx->mb_rc[rc + ctx->m.c.mb_num].bits;
1201 ctx->mb_cmp[mb].mb = mb;
1202 ctx->mb_cmp[mb].value =
1203 delta_bits ? ((ctx->mb_rc[rc].ssd -
1204 ctx->mb_rc[rc + ctx->m.c.mb_num].ssd) * 100) /
1205 delta_bits
1206 : INT_MIN; // avoid increasing qscale
1207 }
1208 }
1209 20090 max_bits += 31; // worst padding
1210 }
1211
2/2
✓ Branch 0 taken 220 times.
✓ Branch 1 taken 145 times.
 365 if (!ret) {
1212 if (RC_VARIANCE)
1213 220 avctx->execute2(avctx, dnxhd_mb_var_thread,
1214 NULL, NULL, ctx->m.c.mb_height);
1215 220 radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.c.mb_num);
1216 222 retry:
1217
4/4
✓ Branch 0 taken 647231 times.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 647011 times.
✓ Branch 3 taken 220 times.
 647233 for (int x = 0; x < ctx->m.c.mb_num && max_bits > ctx->frame_bits; x++) {
1218 647011 int mb = ctx->mb_cmp[x].mb;
1219 647011 int rc = (ctx->qscale * ctx->m.c.mb_num ) + mb;
1220 647011 max_bits -= ctx->mb_rc[rc].bits -
1221 647011 ctx->mb_rc[rc + ctx->m.c.mb_num].bits;
1222
1/2
✓ Branch 0 taken 647011 times.
✗ Branch 1 not taken.
 647011 if (ctx->mb_qscale[mb] < 255)
1223 647011 ctx->mb_qscale[mb]++;
1224 647011 ctx->mb_bits[mb] = ctx->mb_rc[rc + ctx->m.c.mb_num].bits;
1225 }
1226
1227
2/2
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 220 times.
 222 if (max_bits > ctx->frame_bits)
1228 2 goto retry;
1229 }
1230 365 return 0;
1231 }
1232
1233 325 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1234 {
1235
2/2
✓ Branch 0 taken 325 times.
✓ Branch 1 taken 325 times.
 650 for (int i = 0; i < ctx->m.c.avctx->thread_count; i++) {
1236 325 ctx->thread[i]->m.c.linesize = frame->linesize[0] << ctx->interlaced;
1237 325 ctx->thread[i]->m.c.uvlinesize = frame->linesize[1] << ctx->interlaced;
1238 325 ctx->thread[i]->dct_y_offset = ctx->m.c.linesize *8;
1239 325 ctx->thread[i]->dct_uv_offset = ctx->m.c.uvlinesize*8;
1240 }
1241
1242
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 325 times.
 325 ctx->cur_field = (frame->flags & AV_FRAME_FLAG_INTERLACED) &&
1243 ✗ !(frame->flags & AV_FRAME_FLAG_TOP_FIELD_FIRST);
1244 325 }
1245
1246 325 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1247 const AVFrame *frame, int *got_packet)
1248 {
1249 325 DNXHDEncContext *ctx = avctx->priv_data;
1250 325 int first_field = 1;
1251 int offset, i, ret;
1252 uint8_t *buf;
1253
1254
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 325 times.
 325 if ((ret = ff_get_encode_buffer(avctx, pkt, ctx->frame_size, 0)) < 0)
1255 ✗ return ret;
1256 325 buf = pkt->data;
1257
1258 325 dnxhd_load_picture(ctx, frame);
1259
1260 380 encode_coding_unit:
1261
2/2
✓ Branch 0 taken 1140 times.
✓ Branch 1 taken 380 times.
 1520 for (i = 0; i < 3; i++) {
1262 1140 ctx->src[i] = frame->data[i];
1263
4/4
✓ Branch 0 taken 330 times.
✓ Branch 1 taken 810 times.
✓ Branch 2 taken 165 times.
✓ Branch 3 taken 165 times.
 1140 if (ctx->interlaced && ctx->cur_field)
1264 165 ctx->src[i] += frame->linesize[i];
1265 }
1266
1267 380 dnxhd_write_header(avctx, buf);
1268
1269
2/2
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 365 times.
 380 if (avctx->mb_decision == FF_MB_DECISION_RD)
1270 15 ret = dnxhd_encode_rdo(avctx, ctx);
1271 else
1272 365 ret = dnxhd_encode_fast(avctx, ctx);
1273
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 380 times.
 380 if (ret < 0) {
1274 ✗ av_log(avctx, AV_LOG_ERROR,
1275 "picture could not fit ratecontrol constraints, increase qmax\n");
1276 ✗ return ret;
1277 }
1278
1279 380 dnxhd_setup_threads_slices(ctx);
1280
1281 380 offset = 0;
1282
2/2
✓ Branch 0 taken 20765 times.
✓ Branch 1 taken 380 times.
 21145 for (i = 0; i < ctx->m.c.mb_height; i++) {
1283 20765 AV_WB32(ctx->msip + i * 4, offset);
1284 20765 offset += ctx->slice_size[i];
1285 av_assert1(!(ctx->slice_size[i] & 3));
1286 }
1287
1288 380 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.c.mb_height);
1289
1290 av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1291 380 memset(buf + ctx->data_offset + offset, 0,
1292 380 ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1293
1294 380 AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1295
1296
4/4
✓ Branch 0 taken 110 times.
✓ Branch 1 taken 270 times.
✓ Branch 2 taken 55 times.
✓ Branch 3 taken 55 times.
 380 if (ctx->interlaced && first_field) {
1297 55 first_field = 0;
1298 55 ctx->cur_field ^= 1;
1299 55 buf += ctx->coding_unit_size;
1300 55 goto encode_coding_unit;
1301 }
1302
1303 325 ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1304
1305 325 *got_packet = 1;
1306 325 return 0;
1307 }
1308
1309 73 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1310 {
1311 73 DNXHDEncContext *ctx = avctx->priv_data;
1312 int i;
1313
1314 73 av_freep(&ctx->orig_vlc_codes);
1315 73 av_freep(&ctx->orig_vlc_bits);
1316 73 av_freep(&ctx->run_codes);
1317 73 av_freep(&ctx->run_bits);
1318
1319 73 av_freep(&ctx->mb_bits);
1320 73 av_freep(&ctx->mb_qscale);
1321 73 av_freep(&ctx->mb_rc);
1322 73 av_freep(&ctx->mb_cmp);
1323 73 av_freep(&ctx->mb_cmp_tmp);
1324 73 av_freep(&ctx->slice_size);
1325 73 av_freep(&ctx->slice_offs);
1326
1327 73 av_freep(&ctx->qmatrix_c);
1328 73 av_freep(&ctx->qmatrix_l);
1329 73 av_freep(&ctx->qmatrix_c16);
1330 73 av_freep(&ctx->qmatrix_l16);
1331
1332
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (ctx->thread[1]) {
1333 ✗ for (i = 1; i < avctx->thread_count; i++)
1334 ✗ av_freep(&ctx->thread[i]);
1335 }
1336
1337 73 return 0;
1338 }
1339
1340 static const FFCodecDefault dnxhd_defaults[] = {
1341 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1342 { NULL },
1343 };
1344
1345 const FFCodec ff_dnxhd_encoder = {
1346 .p.name = "dnxhd",
1347 CODEC_LONG_NAME("VC3/DNxHD"),
1348 .p.type = AVMEDIA_TYPE_VIDEO,
1349 .p.id = AV_CODEC_ID_DNXHD,
1350 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS |
1351 AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE,
1352 .priv_data_size = sizeof(DNXHDEncContext),
1353 .init = dnxhd_encode_init,
1354 FF_CODEC_ENCODE_CB(dnxhd_encode_picture),
1355 .close = dnxhd_encode_end,
1356 CODEC_PIXFMTS(AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV422P10,
1357 AV_PIX_FMT_YUV444P10, AV_PIX_FMT_GBRP10),
1358 .color_ranges = AVCOL_RANGE_MPEG,
1359 .p.priv_class = &dnxhd_class,
1360 .defaults = dnxhd_defaults,
1361 .p.profiles = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),
1362 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
1363 };
1364
1365 73 void ff_dnxhdenc_init(DNXHDEncContext *ctx)
1366 {
1367 #if ARCH_X86
1368 73 ff_dnxhdenc_init_x86(ctx);
1369 #endif
1370 73 }
1371