FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/dnxhdenc.c
Date: 2024-07-26 21:54:09
Exec Total Coverage
Lines: 679 799 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(MpegEncContext *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->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->idsp.perm_type != FF_IDCT_PERM_NONE)
173 ✗ ff_block_permute(block, ctx->idsp.idct_permutation,
174 scantable, last_non_zero);
175
176 ✗ return last_non_zero;
177 }
178
179 3945600 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
180 int n, int qscale, int *overflow)
181 {
182 3945600 const uint8_t *scantable= ctx->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->idsp.perm_type != FF_IDCT_PERM_NONE)
204 ✗ ff_block_permute(block, ctx->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 int qscale, i;
270 73 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
271 73 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
272
273
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l, ctx->m.avctx->qmax + 1) ||
274
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c, ctx->m.avctx->qmax + 1) ||
275
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l16, ctx->m.avctx->qmax + 1) ||
276
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c16, ctx->m.avctx->qmax + 1))
277 ✗ return AVERROR(ENOMEM);
278
279
2/2
✓ Branch 0 taken 66 times.
✓ Branch 1 taken 7 times.
 73 if (ctx->bit_depth == 8) {
280
2/2
✓ Branch 0 taken 4158 times.
✓ Branch 1 taken 66 times.
 4224 for (i = 1; i < 64; i++) {
281 4158 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
282 4158 weight_matrix[j] = ctx->cid_table->luma_weight[i];
283 }
284 66 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
285 weight_matrix, ctx->intra_quant_bias, 1,
286 66 ctx->m.avctx->qmax, 1);
287
2/2
✓ Branch 0 taken 4158 times.
✓ Branch 1 taken 66 times.
 4224 for (i = 1; i < 64; i++) {
288 4158 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
289 4158 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
290 }
291 66 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
292 weight_matrix, ctx->intra_quant_bias, 1,
293 66 ctx->m.avctx->qmax, 1);
294
295
2/2
✓ Branch 0 taken 42184 times.
✓ Branch 1 taken 66 times.
 42250 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
296
2/2
✓ Branch 0 taken 2699776 times.
✓ Branch 1 taken 42184 times.
 2741960 for (i = 0; i < 64; i++) {
297 2699776 ctx->qmatrix_l[qscale][i] <<= 2;
298 2699776 ctx->qmatrix_c[qscale][i] <<= 2;
299 2699776 ctx->qmatrix_l16[qscale][0][i] <<= 2;
300 2699776 ctx->qmatrix_l16[qscale][1][i] <<= 2;
301 2699776 ctx->qmatrix_c16[qscale][0][i] <<= 2;
302 2699776 ctx->qmatrix_c16[qscale][1][i] <<= 2;
303 }
304 }
305 } else {
306 // 10-bit
307
2/2
✓ Branch 0 taken 56 times.
✓ Branch 1 taken 7 times.
 63 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
308
2/2
✓ Branch 0 taken 3528 times.
✓ Branch 1 taken 56 times.
 3584 for (i = 1; i < 64; i++) {
309 3528 int j = ff_zigzag_direct[i];
310
311 /* The quantization formula from the VC-3 standard is:
312 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
313 * (qscale * weight_table[i]))
314 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
315 * The s factor compensates scaling of DCT coefficients done by
316 * the DCT routines, and therefore is not present in standard.
317 * It's 8 for 8-bit samples and 4 for 10-bit ones.
318 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
319 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
320 * (qscale * weight_table[i])
321 * For 10-bit samples, p / s == 2 */
322 3528 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
323 3528 (qscale * luma_weight_table[i]);
324 3528 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
325 3528 (qscale * chroma_weight_table[i]);
326 }
327 }
328 }
329
330 73 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
331 73 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
332 73 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
333 73 ctx->m.q_intra_matrix = ctx->qmatrix_l;
334
335 73 return 0;
336 }
337
338 73 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
339 {
340
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_rc, (ctx->m.avctx->qmax + 1) * ctx->m.mb_num))
341 ✗ return AVERROR(ENOMEM);
342
343
2/2
✓ Branch 0 taken 58 times.
✓ Branch 1 taken 15 times.
 73 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) {
344
1/2
✓ Branch 1 taken 58 times.
✗ Branch 2 not taken.
 58 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp, ctx->m.mb_num) ||
345
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 58 times.
 58 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp_tmp, ctx->m.mb_num))
346 ✗ return AVERROR(ENOMEM);
347 }
348 73 ctx->frame_bits = (ctx->coding_unit_size -
349 73 ctx->data_offset - 4 - ctx->min_padding) * 8;
350 73 ctx->qscale = 1;
351 73 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
352 73 return 0;
353 }
354
355 73 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
356 {
357 73 DNXHDEncContext *ctx = avctx->priv_data;
358 int i, ret;
359
360
2/3
✓ Branch 0 taken 66 times.
✓ Branch 1 taken 7 times.
✗ Branch 2 not taken.
 73 switch (avctx->pix_fmt) {
361 66 case AV_PIX_FMT_YUV422P:
362 66 ctx->bit_depth = 8;
363 66 break;
364 7 case AV_PIX_FMT_YUV422P10:
365 case AV_PIX_FMT_YUV444P10:
366 case AV_PIX_FMT_GBRP10:
367 7 ctx->bit_depth = 10;
368 7 break;
369 }
370
371
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 &&
372 ✗ avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||
373
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 ||
374
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 avctx->pix_fmt == AV_PIX_FMT_GBRP10))) {
375 ✗ av_log(avctx, AV_LOG_ERROR,
376 "pixel format is incompatible with DNxHD profile\n");
377 ✗ return AVERROR(EINVAL);
378 }
379
380
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) {
381 ✗ av_log(avctx, AV_LOG_ERROR,
382 "pixel format is incompatible with DNxHR HQX profile\n");
383 ✗ return AVERROR(EINVAL);
384 }
385
386
2/2
✓ Branch 0 taken 61 times.
✓ Branch 1 taken 12 times.
 73 if ((ctx->profile == AV_PROFILE_DNXHR_LB ||
387
2/2
✓ Branch 0 taken 53 times.
✓ Branch 1 taken 8 times.
 61 ctx->profile == AV_PROFILE_DNXHR_SQ ||
388
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) {
389 ✗ av_log(avctx, AV_LOG_ERROR,
390 "pixel format is incompatible with DNxHR LB/SQ/HQ profile\n");
391 ✗ return AVERROR(EINVAL);
392 }
393
394 73 ctx->is_444 = ctx->profile == AV_PROFILE_DNXHR_444;
395 73 avctx->profile = ctx->profile;
396 73 ctx->cid = ff_dnxhd_find_cid(avctx, ctx->bit_depth);
397
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (!ctx->cid) {
398 ✗ av_log(avctx, AV_LOG_ERROR,
399 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
400 ✗ ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
401 ✗ return AVERROR(EINVAL);
402 }
403 73 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
404
405
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)
406 40 avctx->codec_tag = MKTAG('A','V','d','h');
407
408
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) {
409 ✗ av_log(avctx, AV_LOG_ERROR,
410 "Input dimensions too small, input must be at least 256x120\n");
411 ✗ return AVERROR(EINVAL);
412 }
413
414 73 ctx->cid_table = ff_dnxhd_get_cid_table(ctx->cid);
415
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 av_assert0(ctx->cid_table);
416
417 73 ctx->m.avctx = avctx;
418 73 ctx->m.mb_intra = 1;
419 73 ctx->m.h263_aic = 1;
420
421 73 avctx->bits_per_raw_sample = ctx->bit_depth;
422
423 73 ff_blockdsp_init(&ctx->bdsp);
424 73 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
425 73 ff_mpv_idct_init(&ctx->m);
426 73 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
427 73 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
428 73 ff_dct_encode_init(&ctx->m);
429
430
2/2
✓ Branch 0 taken 40 times.
✓ Branch 1 taken 33 times.
 73 if (ctx->profile != AV_PROFILE_DNXHD)
431 40 ff_videodsp_init(&ctx->m.vdsp, ctx->bit_depth);
432
433
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) {
434 ✗ ctx->m.dct_quantize = dnxhd_10bit_dct_quantize_444;
435 ✗ ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
436 ✗ ctx->block_width_l2 = 4;
437
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 66 times.
 73 } else if (ctx->bit_depth == 10) {
438 7 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
439 7 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
440 7 ctx->block_width_l2 = 4;
441 } else {
442 66 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
443 66 ctx->block_width_l2 = 3;
444 }
445
446 73 ff_dnxhdenc_init(ctx);
447
448 73 ctx->m.mb_height = (avctx->height + 15) / 16;
449 73 ctx->m.mb_width = (avctx->width + 15) / 16;
450
451
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 62 times.
 73 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
452 11 ctx->interlaced = 1;
453 11 ctx->m.mb_height /= 2;
454 }
455
456
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) {
457 ✗ av_log(avctx, AV_LOG_ERROR,
458 "Interlaced encoding is not supported for DNxHR profiles.\n");
459 ✗ return AVERROR(EINVAL);
460 }
461
462 73 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
463
464
2/2
✓ Branch 0 taken 40 times.
✓ Branch 1 taken 33 times.
 73 if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
465 40 ctx->frame_size = ff_dnxhd_get_hr_frame_size(ctx->cid,
466 avctx->width, avctx->height);
467
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 40 times.
 40 av_assert0(ctx->frame_size >= 0);
468 40 ctx->coding_unit_size = ctx->frame_size;
469 } else {
470 33 ctx->frame_size = ctx->cid_table->frame_size;
471 33 ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
472 }
473
474
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 61 times.
 73 if (ctx->m.mb_height > 68)
475 12 ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);
476 else
477 61 ctx->data_offset = 0x280;
478
479 // XXX tune lbias/cbias
480
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
481 ✗ return ret;
482
483 /* Avid Nitris hardware decoder requires a minimum amount of padding
484 * in the coding unit payload */
485
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (ctx->nitris_compat)
486 ✗ ctx->min_padding = 1600;
487
488
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 if ((ret = dnxhd_init_vlc(ctx)) < 0)
489 ✗ return ret;
490
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 if ((ret = dnxhd_init_rc(ctx)) < 0)
491 ✗ return ret;
492
493
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->slice_size, ctx->m.mb_height) ||
494
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->slice_offs, ctx->m.mb_height) ||
495
1/2
✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_bits, ctx->m.mb_num) ||
496
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 73 times.
 73 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_qscale, ctx->m.mb_num))
497 ✗ return AVERROR(ENOMEM);
498
499
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (avctx->active_thread_type == FF_THREAD_SLICE) {
500 ✗ if (avctx->thread_count > MAX_THREADS) {
501 ✗ av_log(avctx, AV_LOG_ERROR, "too many threads\n");
502 ✗ return AVERROR(EINVAL);
503 }
504 }
505
506
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (avctx->qmax <= 1) {
507 ✗ av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
508 ✗ return AVERROR(EINVAL);
509 }
510
511 73 ctx->thread[0] = ctx;
512
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (avctx->active_thread_type == FF_THREAD_SLICE) {
513 ✗ for (i = 1; i < avctx->thread_count; i++) {
514 ✗ ctx->thread[i] = av_memdup(ctx, sizeof(DNXHDEncContext));
515 ✗ if (!ctx->thread[i])
516 ✗ return AVERROR(ENOMEM);
517 }
518 }
519
520 73 return 0;
521 }
522
523 380 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
524 {
525 380 DNXHDEncContext *ctx = avctx->priv_data;
526
527 380 memset(buf, 0, ctx->data_offset);
528
529 // * write prefix */
530 380 AV_WB16(buf + 0x02, ctx->data_offset);
531
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)
532 200 buf[4] = 0x03;
533 else
534 180 buf[4] = 0x01;
535
536
2/2
✓ Branch 0 taken 110 times.
✓ Branch 1 taken 270 times.
 380 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
537 380 buf[6] = 0x80; // crc flag off
538 380 buf[7] = 0xa0; // reserved
539 380 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
540 380 AV_WB16(buf + 0x1a, avctx->width); // SPL
541 380 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
542
543
2/2
✓ Branch 0 taken 55 times.
✓ Branch 1 taken 325 times.
 380 buf[0x21] = ctx->bit_depth == 10 ? 0x58 : 0x38;
544 380 buf[0x22] = 0x88 + (ctx->interlaced << 2);
545 380 AV_WB32(buf + 0x28, ctx->cid); // CID
546
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);
547
548 380 buf[0x5f] = 0x01; // UDL
549
550 380 buf[0x167] = 0x02; // reserved
551 380 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
552 380 AV_WB16(buf + 0x16c, ctx->m.mb_height); // Ns
553 380 buf[0x16f] = 0x10; // reserved
554
555 380 ctx->msip = buf + 0x170;
556 380 return 0;
557 }
558
559 26918560 static av_always_inline void dnxhd_encode_dc(PutBitContext *pb, DNXHDEncContext *ctx, int diff)
560 {
561 int nbits;
562
2/2
✓ Branch 0 taken 12317800 times.
✓ Branch 1 taken 14600760 times.
 26918560 if (diff < 0) {
563 12317800 nbits = av_log2_16bit(-2 * diff);
564 12317800 diff--;
565 } else {
566 14600760 nbits = av_log2_16bit(2 * diff);
567 }
568 26918560 put_bits(pb, ctx->cid_table->dc_bits[nbits] + nbits,
569 26918560 (ctx->cid_table->dc_codes[nbits] << nbits) +
570 26918560 av_zero_extend(diff, nbits));
571 26918560 }
572
573 static av_always_inline
574 26918560 void dnxhd_encode_block(PutBitContext *pb, DNXHDEncContext *ctx,
575 int16_t *block, int last_index, int n)
576 {
577 26918560 int last_non_zero = 0;
578 int slevel, i, j;
579
580 26918560 dnxhd_encode_dc(pb, ctx, block[0] - ctx->m.last_dc[n]);
581 26918560 ctx->m.last_dc[n] = block[0];
582
583
2/2
✓ Branch 0 taken 296815535 times.
✓ Branch 1 taken 26918560 times.
 323734095 for (i = 1; i <= last_index; i++) {
584 296815535 j = ctx->m.intra_scantable.permutated[i];
585 296815535 slevel = block[j];
586
2/2
✓ Branch 0 taken 111020901 times.
✓ Branch 1 taken 185794634 times.
 296815535 if (slevel) {
587 111020901 int run_level = i - last_non_zero - 1;
588 111020901 int rlevel = slevel * (1 << 1) | !!run_level;
589 111020901 put_bits(pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
590
2/2
✓ Branch 0 taken 36001883 times.
✓ Branch 1 taken 75019018 times.
 111020901 if (run_level)
591 36001883 put_bits(pb, ctx->run_bits[run_level],
592 36001883 ctx->run_codes[run_level]);
593 111020901 last_non_zero = i;
594 }
595 }
596 26918560 put_bits(pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
597 26918560 }
598
599 static av_always_inline
600 3024000 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
601 int qscale, int last_index)
602 {
603 const uint8_t *weight_matrix;
604 int level;
605 int i;
606
607
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3024000 times.
 3024000 if (ctx->is_444) {
608 ✗ weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight
609 ✗ : ctx->cid_table->chroma_weight;
610 } else {
611 3024000 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
612
2/2
✓ Branch 0 taken 1512000 times.
✓ Branch 1 taken 1512000 times.
 3024000 : ctx->cid_table->luma_weight;
613 }
614
615
2/2
✓ Branch 0 taken 32773315 times.
✓ Branch 1 taken 3024000 times.
 35797315 for (i = 1; i <= last_index; i++) {
616 32773315 int j = ctx->m.intra_scantable.permutated[i];
617 32773315 level = block[j];
618
2/2
✓ Branch 0 taken 16774641 times.
✓ Branch 1 taken 15998674 times.
 32773315 if (level) {
619
2/2
✓ Branch 0 taken 8458215 times.
✓ Branch 1 taken 8316426 times.
 16774641 if (level < 0) {
620 8458215 level = (1 - 2 * level) * qscale * weight_matrix[i];
621
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8458215 times.
 8458215 if (ctx->bit_depth == 10) {
622 ✗ if (weight_matrix[i] != 8)
623 ✗ level += 8;
624 ✗ level >>= 4;
625 } else {
626
2/2
✓ Branch 0 taken 6479585 times.
✓ Branch 1 taken 1978630 times.
 8458215 if (weight_matrix[i] != 32)
627 6479585 level += 32;
628 8458215 level >>= 6;
629 }
630 8458215 level = -level;
631 } else {
632 8316426 level = (2 * level + 1) * qscale * weight_matrix[i];
633
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8316426 times.
 8316426 if (ctx->bit_depth == 10) {
634 ✗ if (weight_matrix[i] != 8)
635 ✗ level += 8;
636 ✗ level >>= 4;
637 } else {
638
2/2
✓ Branch 0 taken 6813970 times.
✓ Branch 1 taken 1502456 times.
 8316426 if (weight_matrix[i] != 32)
639 6813970 level += 32;
640 8316426 level >>= 6;
641 }
642 }
643 16774641 block[j] = level;
644 }
645 }
646 3024000 }
647
648 3024000 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
649 {
650 3024000 int score = 0;
651 int i;
652
2/2
✓ Branch 0 taken 193536000 times.
✓ Branch 1 taken 3024000 times.
 196560000 for (i = 0; i < 64; i++)
653 193536000 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
654 3024000 return score;
655 }
656
657 static av_always_inline
658 51881432 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
659 {
660 51881432 int last_non_zero = 0;
661 51881432 int bits = 0;
662 int i, j, level;
663
2/2
✓ Branch 0 taken 463138205 times.
✓ Branch 1 taken 51881432 times.
 515019637 for (i = 1; i <= last_index; i++) {
664 463138205 j = ctx->m.intra_scantable.permutated[i];
665 463138205 level = block[j];
666
2/2
✓ Branch 0 taken 191817206 times.
✓ Branch 1 taken 271320999 times.
 463138205 if (level) {
667 191817206 int run_level = i - last_non_zero - 1;
668 191817206 bits += ctx->vlc_bits[level * (1 << 1) |
669 191817206 !!run_level] + ctx->run_bits[run_level];
670 191817206 last_non_zero = i;
671 }
672 }
673 51881432 return bits;
674 }
675
676 static av_always_inline
677 9849999 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
678 {
679 9849999 const int bs = ctx->block_width_l2;
680 9849999 const int bw = 1 << bs;
681 9849999 int dct_y_offset = ctx->dct_y_offset;
682 9849999 int dct_uv_offset = ctx->dct_uv_offset;
683 9849999 int linesize = ctx->m.linesize;
684 9849999 int uvlinesize = ctx->m.uvlinesize;
685 9849999 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
686 9849999 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
687 9849999 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
688 9849999 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
689 9849999 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
690 9849999 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
691 9849999 PixblockDSPContext *pdsp = &ctx->m.pdsp;
692 9849999 VideoDSPContext *vdsp = &ctx->m.vdsp;
693
694
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.avctx->width ||
695
2/2
✓ Branch 0 taken 18064 times.
✓ Branch 1 taken 7185088 times.
 7203152 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
696 22511 int y_w = ctx->m.avctx->width - (mb_x << 4);
697 22511 int y_h = ctx->m.avctx->height - (mb_y << 4);
698 22511 int uv_w = (y_w + 1) / 2;
699 22511 int uv_h = y_h;
700 22511 linesize = 16;
701 22511 uvlinesize = 8;
702
703 22511 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
704 linesize, ctx->m.linesize,
705 linesize, 16,
706 0, 0, y_w, y_h);
707 22511 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
708 uvlinesize, ctx->m.uvlinesize,
709 uvlinesize, 16,
710 0, 0, uv_w, uv_h);
711 22511 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
712 uvlinesize, ctx->m.uvlinesize,
713 uvlinesize, 16,
714 0, 0, uv_w, uv_h);
715
716 22511 dct_y_offset = bw * linesize;
717 22511 dct_uv_offset = bw * uvlinesize;
718 22511 ptr_y = &ctx->edge_buf_y[0];
719 22511 ptr_u = &ctx->edge_buf_uv[0][0];
720 22511 ptr_v = &ctx->edge_buf_uv[1][0];
721
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.avctx->width ||
722 ✗ (mb_y << 4) + 16 > ctx->m.avctx->height)) {
723 ✗ int y_w = ctx->m.avctx->width - (mb_x << 4);
724 ✗ int y_h = ctx->m.avctx->height - (mb_y << 4);
725 ✗ int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;
726 ✗ int uv_h = y_h;
727 ✗ linesize = 32;
728 ✗ uvlinesize = 16 + 16 * ctx->is_444;
729
730 ✗ vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
731 linesize, ctx->m.linesize,
732 linesize / 2, 16,
733 0, 0, y_w, y_h);
734 ✗ vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
735 uvlinesize, ctx->m.uvlinesize,
736 uvlinesize / 2, 16,
737 0, 0, uv_w, uv_h);
738 ✗ vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
739 uvlinesize, ctx->m.uvlinesize,
740 uvlinesize / 2, 16,
741 0, 0, uv_w, uv_h);
742
743 ✗ dct_y_offset = bw * linesize / 2;
744 ✗ dct_uv_offset = bw * uvlinesize / 2;
745 ✗ ptr_y = &ctx->edge_buf_y[0];
746 ✗ ptr_u = &ctx->edge_buf_uv[0][0];
747 ✗ ptr_v = &ctx->edge_buf_uv[1][0];
748 }
749
750
1/2
✓ Branch 0 taken 9849999 times.
✗ Branch 1 not taken.
 9849999 if (!ctx->is_444) {
751 9849999 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
752 9849999 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
753 9849999 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
754 9849999 pdsp->get_pixels(ctx->blocks[3], ptr_v, uvlinesize);
755
756
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.mb_height && ctx->m.avctx->height == 1080) {
757
2/2
✓ Branch 0 taken 35280 times.
✓ Branch 1 taken 21280 times.
 56560 if (ctx->interlaced) {
758 35280 ctx->get_pixels_8x4_sym(ctx->blocks[4],
759 ptr_y + dct_y_offset,
760 linesize);
761 35280 ctx->get_pixels_8x4_sym(ctx->blocks[5],
762 35280 ptr_y + dct_y_offset + bw,
763 linesize);
764 35280 ctx->get_pixels_8x4_sym(ctx->blocks[6],
765 ptr_u + dct_uv_offset,
766 uvlinesize);
767 35280 ctx->get_pixels_8x4_sym(ctx->blocks[7],
768 ptr_v + dct_uv_offset,
769 uvlinesize);
770 } else {
771 21280 ctx->bdsp.clear_block(ctx->blocks[4]);
772 21280 ctx->bdsp.clear_block(ctx->blocks[5]);
773 21280 ctx->bdsp.clear_block(ctx->blocks[6]);
774 21280 ctx->bdsp.clear_block(ctx->blocks[7]);
775 }
776 } else {
777 9793439 pdsp->get_pixels(ctx->blocks[4],
778 ptr_y + dct_y_offset, linesize);
779 9793439 pdsp->get_pixels(ctx->blocks[5],
780 9793439 ptr_y + dct_y_offset + bw, linesize);
781 9793439 pdsp->get_pixels(ctx->blocks[6],
782 ptr_u + dct_uv_offset, uvlinesize);
783 9793439 pdsp->get_pixels(ctx->blocks[7],
784 ptr_v + dct_uv_offset, uvlinesize);
785 }
786 } else {
787 ✗ pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
788 ✗ pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
789 ✗ pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);
790 ✗ pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);
791
792 ✗ pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
793 ✗ pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);
794 ✗ pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);
795 ✗ pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);
796
797 ✗ pdsp->get_pixels(ctx->blocks[4], ptr_v, uvlinesize);
798 ✗ pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);
799 ✗ pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);
800 ✗ pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);
801 }
802 9849999 }
803
804 static av_always_inline
805 78799992 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
806 {
807 int x;
808
809
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 78799992 times.
 78799992 if (ctx->is_444) {
810 ✗ x = (i >> 1) % 3;
811 } else {
812 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
813 78799992 x = component[i];
814 }
815 78799992 return x;
816 }
817
818 41557 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
819 int jobnr, int threadnr)
820 {
821 41557 DNXHDEncContext *ctx = avctx->priv_data;
822 41557 int mb_y = jobnr, mb_x;
823 41557 int qscale = ctx->qscale;
824 41557 LOCAL_ALIGNED_16(int16_t, block, [64]);
825 41557 ctx = ctx->thread[threadnr];
826
827 41557 ctx->m.last_dc[0] =
828 41557 ctx->m.last_dc[1] =
829 41557 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
830
831
2/2
✓ Branch 0 taken 6485179 times.
✓ Branch 1 taken 41557 times.
 6526736 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
832 6485179 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
833 6485179 int ssd = 0;
834 6485179 int ac_bits = 0;
835 6485179 int dc_bits = 0;
836 int i;
837
838 6485179 dnxhd_get_blocks(ctx, mb_x, mb_y);
839
840
2/2
✓ Branch 0 taken 51881432 times.
✓ Branch 1 taken 6485179 times.
 58366611 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
841 51881432 int16_t *src_block = ctx->blocks[i];
842 int overflow, nbits, diff, last_index;
843 51881432 int n = dnxhd_switch_matrix(ctx, i);
844
845 51881432 memcpy(block, src_block, 64 * sizeof(*block));
846 51881432 last_index = ctx->m.dct_quantize(&ctx->m, block,
847
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),
848 qscale, &overflow);
849 51881432 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
850
851 51881432 diff = block[0] - ctx->m.last_dc[n];
852
2/2
✓ Branch 0 taken 23766663 times.
✓ Branch 1 taken 28114769 times.
 51881432 if (diff < 0)
853 23766663 nbits = av_log2_16bit(-2 * diff);
854 else
855 28114769 nbits = av_log2_16bit(2 * diff);
856
857 av_assert1(nbits < ctx->bit_depth + 4);
858 51881432 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
859
860 51881432 ctx->m.last_dc[n] = block[0];
861
862
2/2
✓ Branch 0 taken 3024000 times.
✓ Branch 1 taken 48857432 times.
 51881432 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
863 3024000 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
864 3024000 ctx->m.idsp.idct(block);
865 3024000 ssd += dnxhd_ssd_block(block, src_block);
866 }
867 }
868 6485179 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd = ssd;
869 6485179 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
870 6485179 (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];
871 }
872 41557 return 0;
873 }
874
875 20765 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
876 int jobnr, int threadnr)
877 {
878 20765 DNXHDEncContext *ctx = avctx->priv_data;
879 20765 PutBitContext pb0, *const pb = &pb0;
880 20765 int mb_y = jobnr, mb_x;
881 20765 ctx = ctx->thread[threadnr];
882 20765 init_put_bits(pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
883 20765 ctx->slice_size[jobnr]);
884
885 20765 ctx->m.last_dc[0] =
886 20765 ctx->m.last_dc[1] =
887 20765 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
888
2/2
✓ Branch 0 taken 3364820 times.
✓ Branch 1 taken 20765 times.
 3385585 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
889 3364820 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
890 3364820 int qscale = ctx->mb_qscale[mb];
891 int i;
892
893 3364820 put_bits(pb, 11, qscale);
894 3364820 put_bits(pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
895
896 3364820 dnxhd_get_blocks(ctx, mb_x, mb_y);
897
898
2/2
✓ Branch 0 taken 26918560 times.
✓ Branch 1 taken 3364820 times.
 30283380 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
899 26918560 int16_t *block = ctx->blocks[i];
900 26918560 int overflow, n = dnxhd_switch_matrix(ctx, i);
901 26918560 int last_index = ctx->m.dct_quantize(&ctx->m, block,
902
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),
903 qscale, &overflow);
904
905 26918560 dnxhd_encode_block(pb, ctx, block, last_index, n);
906 }
907 }
908 20765 flush_put_bits(pb);
909 20765 memset(put_bits_ptr(pb), 0, put_bytes_left(pb, 0));
910 20765 return 0;
911 }
912
913 380 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
914 {
915 int mb_y, mb_x;
916 380 int offset = 0;
917
2/2
✓ Branch 0 taken 20765 times.
✓ Branch 1 taken 380 times.
 21145 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
918 int thread_size;
919 20765 ctx->slice_offs[mb_y] = offset;
920 20765 ctx->slice_size[mb_y] = 0;
921
2/2
✓ Branch 0 taken 3364820 times.
✓ Branch 1 taken 20765 times.
 3385585 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
922 3364820 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
923 3364820 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
924 }
925 20765 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31U) & ~31U;
926 20765 ctx->slice_size[mb_y] >>= 3;
927 20765 thread_size = ctx->slice_size[mb_y];
928 20765 offset += thread_size;
929 }
930 380 }
931
932 11520 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
933 int jobnr, int threadnr)
934 {
935 11520 DNXHDEncContext *ctx = avctx->priv_data;
936 11520 int mb_y = jobnr, mb_x, x, y;
937
2/2
✓ Branch 0 taken 220 times.
✓ Branch 1 taken 11300 times.
 11740 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
938
2/2
✓ Branch 0 taken 135 times.
✓ Branch 1 taken 85 times.
 220 ((avctx->height >> ctx->interlaced) & 0xF);
939
940 11520 ctx = ctx->thread[threadnr];
941
2/2
✓ Branch 0 taken 10955 times.
✓ Branch 1 taken 565 times.
 11520 if (ctx->bit_depth == 8) {
942 10955 const uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
943
2/2
✓ Branch 0 taken 1919700 times.
✓ Branch 1 taken 10955 times.
 1930655 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
944 1919700 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
945 int sum;
946 int varc;
947
948
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) {
949 1901720 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
950 1901720 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
951 } else {
952 17980 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
953 17980 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
954 17980 sum = varc = 0;
955
2/2
✓ Branch 0 taken 226195 times.
✓ Branch 1 taken 17980 times.
 244175 for (y = 0; y < bh; y++) {
956
2/2
✓ Branch 0 taken 3573835 times.
✓ Branch 1 taken 226195 times.
 3800030 for (x = 0; x < bw; x++) {
957 3573835 uint8_t val = pix[x + y * ctx->m.linesize];
958 3573835 sum += val;
959 3573835 varc += val * val;
960 }
961 }
962 }
963 1919700 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
964
965 1919700 ctx->mb_cmp[mb].value = varc;
966 1919700 ctx->mb_cmp[mb].mb = mb;
967 }
968 } else { // 10-bit
969 565 const int linesize = ctx->m.linesize >> 1;
970
2/2
✓ Branch 0 taken 58800 times.
✓ Branch 1 taken 565 times.
 59365 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
971 58800 const uint16_t *pix = (const uint16_t *)ctx->thread[0]->src[0] +
972 58800 ((mb_y << 4) * linesize) + (mb_x << 4);
973 58800 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
974 58800 int sum = 0;
975 58800 int sqsum = 0;
976 58800 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
977 58800 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
978 int mean, sqmean;
979 int i, j;
980 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
981
2/2
✓ Branch 0 taken 936000 times.
✓ Branch 1 taken 58800 times.
 994800 for (i = 0; i < bh; ++i) {
982
2/2
✓ Branch 0 taken 14976000 times.
✓ Branch 1 taken 936000 times.
 15912000 for (j = 0; j < bw; ++j) {
983 // Turn 16-bit pixels into 10-bit ones.
984 14976000 const int sample = (unsigned) pix[j] >> 6;
985 14976000 sum += sample;
986 14976000 sqsum += sample * sample;
987 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
988 }
989 936000 pix += linesize;
990 }
991 58800 mean = sum >> 8; // 16*16 == 2^8
992 58800 sqmean = sqsum >> 8;
993 58800 ctx->mb_cmp[mb].value = sqmean - mean * mean;
994 58800 ctx->mb_cmp[mb].mb = mb;
995 }
996 }
997 11520 return 0;
998 }
999
1000 15 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
1001 {
1002 int lambda, up_step, down_step;
1003 15 int last_lower = INT_MAX, last_higher = 0;
1004 int x, y, q;
1005
1006
2/2
✓ Branch 0 taken 105 times.
✓ Branch 1 taken 15 times.
 120 for (q = 1; q < avctx->qmax; q++) {
1007 105 ctx->qscale = q;
1008 105 avctx->execute2(avctx, dnxhd_calc_bits_thread,
1009 NULL, NULL, ctx->m.mb_height);
1010 }
1011 15 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
1012 15 lambda = ctx->lambda;
1013
1014 70 for (;;) {
1015 85 int bits = 0;
1016 85 int end = 0;
1017
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 80 times.
 85 if (lambda == last_higher) {
1018 5 lambda++;
1019 5 end = 1; // need to set final qscales/bits
1020 }
1021
2/2
✓ Branch 0 taken 3809 times.
✓ Branch 1 taken 53 times.
 3862 for (y = 0; y < ctx->m.mb_height; y++) {
1022
2/2
✓ Branch 0 taken 304720 times.
✓ Branch 1 taken 3809 times.
 308529 for (x = 0; x < ctx->m.mb_width; x++) {
1023 304720 unsigned min = UINT_MAX;
1024 304720 int qscale = 1;
1025 304720 int mb = y * ctx->m.mb_width + x;
1026 304720 int rc = 0;
1027
2/2
✓ Branch 0 taken 2133040 times.
✓ Branch 1 taken 304720 times.
 2437760 for (q = 1; q < avctx->qmax; q++) {
1028 2133040 int i = (q*ctx->m.mb_num) + mb;
1029 2133040 unsigned score = ctx->mb_rc[i].bits * lambda +
1030 2133040 ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
1031
2/2
✓ Branch 0 taken 493989 times.
✓ Branch 1 taken 1639051 times.
 2133040 if (score < min) {
1032 493989 min = score;
1033 493989 qscale = q;
1034 493989 rc = i;
1035 }
1036 }
1037 304720 bits += ctx->mb_rc[rc].bits;
1038 304720 ctx->mb_qscale[mb] = qscale;
1039 304720 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1040 }
1041 3809 bits = (bits + 31) & ~31; // padding
1042
2/2
✓ Branch 0 taken 32 times.
✓ Branch 1 taken 3777 times.
 3809 if (bits > ctx->frame_bits)
1043 32 break;
1044 }
1045
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 80 times.
 85 if (end) {
1046
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 5 times.
 5 if (bits > ctx->frame_bits)
1047 ✗ return AVERROR(EINVAL);
1048 5 break;
1049 }
1050
2/2
✓ Branch 0 taken 48 times.
✓ Branch 1 taken 32 times.
 80 if (bits < ctx->frame_bits) {
1051 48 last_lower = FFMIN(lambda, last_lower);
1052
2/2
✓ Branch 0 taken 25 times.
✓ Branch 1 taken 23 times.
 48 if (last_higher != 0)
1053 25 lambda = (lambda+last_higher)>>1;
1054 else
1055 23 lambda -= down_step;
1056 48 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
1057 48 up_step = 1<<LAMBDA_FRAC_BITS;
1058 48 lambda = FFMAX(1, lambda);
1059
2/2
✓ Branch 0 taken 10 times.
✓ Branch 1 taken 38 times.
 48 if (lambda == last_lower)
1060 10 break;
1061 } else {
1062 32 last_higher = FFMAX(lambda, last_higher);
1063
1/2
✓ Branch 0 taken 32 times.
✗ Branch 1 not taken.
 32 if (last_lower != INT_MAX)
1064 32 lambda = (lambda+last_lower)>>1;
1065 ✗ else if ((int64_t)lambda + up_step > INT_MAX)
1066 ✗ return AVERROR(EINVAL);
1067 else
1068 ✗ lambda += up_step;
1069 32 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
1070 32 down_step = 1<<LAMBDA_FRAC_BITS;
1071 }
1072 }
1073 15 ctx->lambda = lambda;
1074 15 return 0;
1075 }
1076
1077 365 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
1078 {
1079 365 int bits = 0;
1080 365 int up_step = 1;
1081 365 int down_step = 1;
1082 365 int last_higher = 0;
1083 365 int last_lower = INT_MAX;
1084 int qscale;
1085 int x, y;
1086
1087 365 qscale = ctx->qscale;
1088 for (;;) {
1089 691 bits = 0;
1090 691 ctx->qscale = qscale;
1091 // XXX avoid recalculating bits
1092 691 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
1093 NULL, NULL, ctx->m.mb_height);
1094
2/2
✓ Branch 0 taken 34838 times.
✓ Branch 1 taken 393 times.
 35231 for (y = 0; y < ctx->m.mb_height; y++) {
1095
2/2
✓ Branch 0 taken 5760121 times.
✓ Branch 1 taken 34838 times.
 5794959 for (x = 0; x < ctx->m.mb_width; x++)
1096 5760121 bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;
1097 34838 bits = (bits+31)&~31; // padding
1098
2/2
✓ Branch 0 taken 298 times.
✓ Branch 1 taken 34540 times.
 34838 if (bits > ctx->frame_bits)
1099 298 break;
1100 }
1101
2/2
✓ Branch 0 taken 393 times.
✓ Branch 1 taken 298 times.
 691 if (bits < ctx->frame_bits) {
1102
2/2
✓ Branch 0 taken 145 times.
✓ Branch 1 taken 248 times.
 393 if (qscale == 1)
1103 145 return 1;
1104
2/2
✓ Branch 0 taken 195 times.
✓ Branch 1 taken 53 times.
 248 if (last_higher == qscale - 1) {
1105 195 qscale = last_higher;
1106 195 break;
1107 }
1108 53 last_lower = FFMIN(qscale, last_lower);
1109
2/2
✓ Branch 0 taken 38 times.
✓ Branch 1 taken 15 times.
 53 if (last_higher != 0)
1110 38 qscale = (qscale + last_higher) >> 1;
1111 else
1112 15 qscale -= down_step++;
1113
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 53 times.
 53 if (qscale < 1)
1114 ✗ qscale = 1;
1115 53 up_step = 1;
1116 } else {
1117
2/2
✓ Branch 0 taken 25 times.
✓ Branch 1 taken 273 times.
 298 if (last_lower == qscale + 1)
1118 25 break;
1119 273 last_higher = FFMAX(qscale, last_higher);
1120
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 268 times.
 273 if (last_lower != INT_MAX)
1121 5 qscale = (qscale + last_lower) >> 1;
1122 else
1123 268 qscale += up_step++;
1124 273 down_step = 1;
1125
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 273 times.
 273 if (qscale >= ctx->m.avctx->qmax)
1126 ✗ return AVERROR(EINVAL);
1127 }
1128 }
1129 220 ctx->qscale = qscale;
1130 220 return 0;
1131 }
1132
1133 #define BUCKET_BITS 8
1134 #define RADIX_PASSES 4
1135 #define NBUCKETS (1 << BUCKET_BITS)
1136
1137 11871000 static inline int get_bucket(int value, int shift)
1138 {
1139 11871000 value >>= shift;
1140 11871000 value &= NBUCKETS - 1;
1141 11871000 return NBUCKETS - 1 - value;
1142 }
1143
1144 220 static void radix_count(const RCCMPEntry *data, int size,
1145 int buckets[RADIX_PASSES][NBUCKETS])
1146 {
1147 int i, j;
1148 220 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1149
2/2
✓ Branch 0 taken 1978500 times.
✓ Branch 1 taken 220 times.
 1978720 for (i = 0; i < size; i++) {
1150 1978500 int v = data[i].value;
1151
2/2
✓ Branch 0 taken 7914000 times.
✓ Branch 1 taken 1978500 times.
 9892500 for (j = 0; j < RADIX_PASSES; j++) {
1152 7914000 buckets[j][get_bucket(v, 0)]++;
1153 7914000 v >>= BUCKET_BITS;
1154 }
1155 av_assert1(!v);
1156 }
1157
2/2
✓ Branch 0 taken 880 times.
✓ Branch 1 taken 220 times.
 1100 for (j = 0; j < RADIX_PASSES; j++) {
1158 880 int offset = size;
1159
2/2
✓ Branch 0 taken 225280 times.
✓ Branch 1 taken 880 times.
 226160 for (i = NBUCKETS - 1; i >= 0; i--)
1160 225280 buckets[j][i] = offset -= buckets[j][i];
1161 av_assert1(!buckets[j][0]);
1162 }
1163 220 }
1164
1165 440 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1166 int size, int buckets[NBUCKETS], int pass)
1167 {
1168 440 int shift = pass * BUCKET_BITS;
1169 int i;
1170
2/2
✓ Branch 0 taken 3957000 times.
✓ Branch 1 taken 440 times.
 3957440 for (i = 0; i < size; i++) {
1171 3957000 int v = get_bucket(data[i].value, shift);
1172 3957000 int pos = buckets[v]++;
1173 3957000 dst[pos] = data[i];
1174 }
1175 440 }
1176
1177 220 static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)
1178 {
1179 int buckets[RADIX_PASSES][NBUCKETS];
1180 220 radix_count(data, size, buckets);
1181 220 radix_sort_pass(tmp, data, size, buckets[0], 0);
1182 220 radix_sort_pass(data, tmp, size, buckets[1], 1);
1183
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]) {
1184 ✗ radix_sort_pass(tmp, data, size, buckets[2], 2);
1185 ✗ radix_sort_pass(data, tmp, size, buckets[3], 3);
1186 }
1187 220 }
1188
1189 365 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1190 {
1191 365 int max_bits = 0;
1192 int ret, x, y;
1193
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 365 times.
 365 if ((ret = dnxhd_find_qscale(ctx)) < 0)
1194 ✗ return ret;
1195
2/2
✓ Branch 0 taken 20090 times.
✓ Branch 1 taken 365 times.
 20455 for (y = 0; y < ctx->m.mb_height; y++) {
1196
2/2
✓ Branch 0 taken 3310820 times.
✓ Branch 1 taken 20090 times.
 3330910 for (x = 0; x < ctx->m.mb_width; x++) {
1197 3310820 int mb = y * ctx->m.mb_width + x;
1198 3310820 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1199 int delta_bits;
1200 3310820 ctx->mb_qscale[mb] = ctx->qscale;
1201 3310820 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1202 3310820 max_bits += ctx->mb_rc[rc].bits;
1203 if (!RC_VARIANCE) {
1204 delta_bits = ctx->mb_rc[rc].bits -
1205 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1206 ctx->mb_cmp[mb].mb = mb;
1207 ctx->mb_cmp[mb].value =
1208 delta_bits ? ((ctx->mb_rc[rc].ssd -
1209 ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /
1210 delta_bits
1211 : INT_MIN; // avoid increasing qscale
1212 }
1213 }
1214 20090 max_bits += 31; // worst padding
1215 }
1216
2/2
✓ Branch 0 taken 220 times.
✓ Branch 1 taken 145 times.
 365 if (!ret) {
1217 if (RC_VARIANCE)
1218 220 avctx->execute2(avctx, dnxhd_mb_var_thread,
1219 NULL, NULL, ctx->m.mb_height);
1220 220 radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.mb_num);
1221 222 retry:
1222
4/4
✓ Branch 0 taken 647231 times.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 647011 times.
✓ Branch 3 taken 220 times.
 647233 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1223 647011 int mb = ctx->mb_cmp[x].mb;
1224 647011 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1225 647011 max_bits -= ctx->mb_rc[rc].bits -
1226 647011 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1227
1/2
✓ Branch 0 taken 647011 times.
✗ Branch 1 not taken.
 647011 if (ctx->mb_qscale[mb] < 255)
1228 647011 ctx->mb_qscale[mb]++;
1229 647011 ctx->mb_bits[mb] = ctx->mb_rc[rc + ctx->m.mb_num].bits;
1230 }
1231
1232
2/2
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 220 times.
 222 if (max_bits > ctx->frame_bits)
1233 2 goto retry;
1234 }
1235 365 return 0;
1236 }
1237
1238 325 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1239 {
1240 int i;
1241
1242
2/2
✓ Branch 0 taken 325 times.
✓ Branch 1 taken 325 times.
 650 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1243 325 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1244 325 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1245 325 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1246 325 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1247 }
1248
1249
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 325 times.
 325 ctx->cur_field = (frame->flags & AV_FRAME_FLAG_INTERLACED) &&
1250 ✗ !(frame->flags & AV_FRAME_FLAG_TOP_FIELD_FIRST);
1251 325 }
1252
1253 325 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1254 const AVFrame *frame, int *got_packet)
1255 {
1256 325 DNXHDEncContext *ctx = avctx->priv_data;
1257 325 int first_field = 1;
1258 int offset, i, ret;
1259 uint8_t *buf;
1260
1261
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)
1262 ✗ return ret;
1263 325 buf = pkt->data;
1264
1265 325 dnxhd_load_picture(ctx, frame);
1266
1267 380 encode_coding_unit:
1268
2/2
✓ Branch 0 taken 1140 times.
✓ Branch 1 taken 380 times.
 1520 for (i = 0; i < 3; i++) {
1269 1140 ctx->src[i] = frame->data[i];
1270
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)
1271 165 ctx->src[i] += frame->linesize[i];
1272 }
1273
1274 380 dnxhd_write_header(avctx, buf);
1275
1276
2/2
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 365 times.
 380 if (avctx->mb_decision == FF_MB_DECISION_RD)
1277 15 ret = dnxhd_encode_rdo(avctx, ctx);
1278 else
1279 365 ret = dnxhd_encode_fast(avctx, ctx);
1280
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 380 times.
 380 if (ret < 0) {
1281 ✗ av_log(avctx, AV_LOG_ERROR,
1282 "picture could not fit ratecontrol constraints, increase qmax\n");
1283 ✗ return ret;
1284 }
1285
1286 380 dnxhd_setup_threads_slices(ctx);
1287
1288 380 offset = 0;
1289
2/2
✓ Branch 0 taken 20765 times.
✓ Branch 1 taken 380 times.
 21145 for (i = 0; i < ctx->m.mb_height; i++) {
1290 20765 AV_WB32(ctx->msip + i * 4, offset);
1291 20765 offset += ctx->slice_size[i];
1292 av_assert1(!(ctx->slice_size[i] & 3));
1293 }
1294
1295 380 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1296
1297 av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1298 380 memset(buf + ctx->data_offset + offset, 0,
1299 380 ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1300
1301 380 AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1302
1303
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) {
1304 55 first_field = 0;
1305 55 ctx->cur_field ^= 1;
1306 55 buf += ctx->coding_unit_size;
1307 55 goto encode_coding_unit;
1308 }
1309
1310 325 ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1311
1312 325 *got_packet = 1;
1313 325 return 0;
1314 }
1315
1316 73 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1317 {
1318 73 DNXHDEncContext *ctx = avctx->priv_data;
1319 int i;
1320
1321 73 av_freep(&ctx->orig_vlc_codes);
1322 73 av_freep(&ctx->orig_vlc_bits);
1323 73 av_freep(&ctx->run_codes);
1324 73 av_freep(&ctx->run_bits);
1325
1326 73 av_freep(&ctx->mb_bits);
1327 73 av_freep(&ctx->mb_qscale);
1328 73 av_freep(&ctx->mb_rc);
1329 73 av_freep(&ctx->mb_cmp);
1330 73 av_freep(&ctx->mb_cmp_tmp);
1331 73 av_freep(&ctx->slice_size);
1332 73 av_freep(&ctx->slice_offs);
1333
1334 73 av_freep(&ctx->qmatrix_c);
1335 73 av_freep(&ctx->qmatrix_l);
1336 73 av_freep(&ctx->qmatrix_c16);
1337 73 av_freep(&ctx->qmatrix_l16);
1338
1339
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 73 times.
 73 if (ctx->thread[1]) {
1340 ✗ for (i = 1; i < avctx->thread_count; i++)
1341 ✗ av_freep(&ctx->thread[i]);
1342 }
1343
1344 73 return 0;
1345 }
1346
1347 static const FFCodecDefault dnxhd_defaults[] = {
1348 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1349 { NULL },
1350 };
1351
1352 const FFCodec ff_dnxhd_encoder = {
1353 .p.name = "dnxhd",
1354 CODEC_LONG_NAME("VC3/DNxHD"),
1355 .p.type = AVMEDIA_TYPE_VIDEO,
1356 .p.id = AV_CODEC_ID_DNXHD,
1357 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS |
1358 AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE,
1359 .priv_data_size = sizeof(DNXHDEncContext),
1360 .init = dnxhd_encode_init,
1361 FF_CODEC_ENCODE_CB(dnxhd_encode_picture),
1362 .close = dnxhd_encode_end,
1363 .p.pix_fmts = (const enum AVPixelFormat[]) {
1364 AV_PIX_FMT_YUV422P,
1365 AV_PIX_FMT_YUV422P10,
1366 AV_PIX_FMT_YUV444P10,
1367 AV_PIX_FMT_GBRP10,
1368 AV_PIX_FMT_NONE
1369 },
1370 .p.priv_class = &dnxhd_class,
1371 .defaults = dnxhd_defaults,
1372 .p.profiles = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),
1373 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
1374 };
1375
1376 73 void ff_dnxhdenc_init(DNXHDEncContext *ctx)
1377 {
1378 #if ARCH_X86
1379 73 ff_dnxhdenc_init_x86(ctx);
1380 #endif
1381 73 }
1382