GCC Code Coverage Report
Directory: ../../../ffmpeg/ Exec Total Coverage
File: src/libavcodec/dnxhdenc.c Lines: 679 803 84.6 %
Date: 2021-01-21 21:11:50 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_internal.h"
29
#include "libavutil/opt.h"
30
31
#include "avcodec.h"
32
#include "blockdsp.h"
33
#include "fdctdsp.h"
34
#include "internal.h"
35
#include "mpegvideo.h"
36
#include "pixblockdsp.h"
37
#include "packet_internal.h"
38
#include "profiles.h"
39
#include "dnxhdenc.h"
40
41
// The largest value that will not lead to overflow for 10-bit samples.
42
#define DNX10BIT_QMAT_SHIFT 18
43
#define RC_VARIANCE 1 // use variance or ssd for fast rc
44
#define LAMBDA_FRAC_BITS 10
45
46
#define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
47
static const AVOption options[] = {
48
    { "nitris_compat", "encode with Avid Nitris compatibility",
49
        offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
50
    { "ibias", "intra quant bias",
51
        offsetof(DNXHDEncContext, intra_quant_bias), AV_OPT_TYPE_INT,
52
        { .i64 = 0 }, INT_MIN, INT_MAX, VE },
53
    { "profile",       NULL, offsetof(DNXHDEncContext, profile), AV_OPT_TYPE_INT,
54
        { .i64 = FF_PROFILE_DNXHD },
55
        FF_PROFILE_DNXHD, FF_PROFILE_DNXHR_444, VE, "profile" },
56
    { "dnxhd",     NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHD },
57
        0, 0, VE, "profile" },
58
    { "dnxhr_444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_444 },
59
        0, 0, VE, "profile" },
60
    { "dnxhr_hqx", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQX },
61
        0, 0, VE, "profile" },
62
    { "dnxhr_hq",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQ },
63
        0, 0, VE, "profile" },
64
    { "dnxhr_sq",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_SQ },
65
        0, 0, VE, "profile" },
66
    { "dnxhr_lb",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_LB },
67
        0, 0, VE, "profile" },
68
    { NULL }
69
};
70
71
static const AVClass dnxhd_class = {
72
    .class_name = "dnxhd",
73
    .item_name  = av_default_item_name,
74
    .option     = options,
75
    .version    = LIBAVUTIL_VERSION_INT,
76
};
77
78
97920
static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,
79
                                          const uint8_t *pixels,
80
                                          ptrdiff_t line_size)
81
{
82
    int i;
83
489600
    for (i = 0; i < 4; i++) {
84
391680
        block[0] = pixels[0];
85
391680
        block[1] = pixels[1];
86
391680
        block[2] = pixels[2];
87
391680
        block[3] = pixels[3];
88
391680
        block[4] = pixels[4];
89
391680
        block[5] = pixels[5];
90
391680
        block[6] = pixels[6];
91
391680
        block[7] = pixels[7];
92
391680
        pixels  += line_size;
93
391680
        block   += 8;
94
    }
95
97920
    memcpy(block,      block -  8, sizeof(*block) * 8);
96
97920
    memcpy(block +  8, block - 16, sizeof(*block) * 8);
97
97920
    memcpy(block + 16, block - 24, sizeof(*block) * 8);
98
97920
    memcpy(block + 24, block - 32, sizeof(*block) * 8);
99
97920
}
100
101
static av_always_inline
102
43200
void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,
103
                                    const uint8_t *pixels,
104
                                    ptrdiff_t line_size)
105
{
106
43200
    memcpy(block + 0 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
107
43200
    memcpy(block + 7 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
108
43200
    memcpy(block + 1 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
109
43200
    memcpy(block + 6 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
110
43200
    memcpy(block + 2 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
111
43200
    memcpy(block + 5 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
112
43200
    memcpy(block + 3 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
113
43200
    memcpy(block + 4 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
114
43200
}
115
116
static int dnxhd_10bit_dct_quantize_444(MpegEncContext *ctx, int16_t *block,
117
                                        int n, int qscale, int *overflow)
118
{
119
    int i, j, level, last_non_zero, start_i;
120
    const int *qmat;
121
    const uint8_t *scantable= ctx->intra_scantable.scantable;
122
    int bias;
123
    int max = 0;
124
    unsigned int threshold1, threshold2;
125
126
    ctx->fdsp.fdct(block);
127
128
    block[0] = (block[0] + 2) >> 2;
129
    start_i = 1;
130
    last_non_zero = 0;
131
    qmat = n < 4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
132
    bias= ctx->intra_quant_bias * (1 << (16 - 8));
133
    threshold1 = (1 << 16) - bias - 1;
134
    threshold2 = (threshold1 << 1);
135
136
    for (i = 63; i >= start_i; i--) {
137
        j = scantable[i];
138
        level = block[j] * qmat[j];
139
140
        if (((unsigned)(level + threshold1)) > threshold2) {
141
            last_non_zero = i;
142
            break;
143
        } else{
144
            block[j]=0;
145
        }
146
    }
147
148
    for (i = start_i; i <= last_non_zero; i++) {
149
        j = scantable[i];
150
        level = block[j] * qmat[j];
151
152
        if (((unsigned)(level + threshold1)) > threshold2) {
153
            if (level > 0) {
154
                level = (bias + level) >> 16;
155
                block[j] = level;
156
            } else{
157
                level = (bias - level) >> 16;
158
                block[j] = -level;
159
            }
160
            max |= level;
161
        } else {
162
            block[j] = 0;
163
        }
164
    }
165
    *overflow = ctx->max_qcoeff < max; //overflow might have happened
166
167
    /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
168
    if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
169
        ff_block_permute(block, ctx->idsp.idct_permutation,
170
                         scantable, last_non_zero);
171
172
    return last_non_zero;
173
}
174
175
3945600
static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
176
                                    int n, int qscale, int *overflow)
177
{
178
3945600
    const uint8_t *scantable= ctx->intra_scantable.scantable;
179
3945600
    const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
180
3945600
    int last_non_zero = 0;
181
    int i;
182
183
3945600
    ctx->fdsp.fdct(block);
184
185
    // Divide by 4 with rounding, to compensate scaling of DCT coefficients
186
3945600
    block[0] = (block[0] + 2) >> 2;
187
188
252518400
    for (i = 1; i < 64; ++i) {
189
248572800
        int j = scantable[i];
190
248572800
        int sign = FF_SIGNBIT(block[j]);
191
248572800
        int level = (block[j] ^ sign) - sign;
192
248572800
        level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
193
248572800
        block[j] = (level ^ sign) - sign;
194
248572800
        if (level)
195
32824223
            last_non_zero = i;
196
    }
197
198
    /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
199
3945600
    if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
200
        ff_block_permute(block, ctx->idsp.idct_permutation,
201
                         scantable, last_non_zero);
202
203
3945600
    return last_non_zero;
204
}
205
206
73
static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
207
{
208
    int i, j, level, run;
209
73
    int max_level = 1 << (ctx->bit_depth + 2);
210
211
73
    if (!FF_ALLOCZ_TYPED_ARRAY(ctx->orig_vlc_codes, max_level * 4) ||
212
73
        !FF_ALLOCZ_TYPED_ARRAY(ctx->orig_vlc_bits,  max_level * 4) ||
213
73
        !(ctx->run_codes = av_mallocz(63 * 2))                     ||
214
73
        !(ctx->run_bits  = av_mallocz(63)))
215
        return AVERROR(ENOMEM);
216
73
    ctx->vlc_codes = ctx->orig_vlc_codes + max_level * 2;
217
73
    ctx->vlc_bits  = ctx->orig_vlc_bits + max_level * 2;
218
192585
    for (level = -max_level; level < max_level; level++) {
219
577536
        for (run = 0; run < 2; run++) {
220
385024
            int index = level * (1 << 1) | run;
221
385024
            int sign, offset = 0, alevel = level;
222
223
385024
            MASK_ABS(sign, alevel);
224
385024
            if (alevel > 64) {
225
366190
                offset  = (alevel - 1) >> 6;
226
366190
                alevel -= offset << 6;
227
            }
228
65137020
            for (j = 0; j < 257; j++) {
229

65136947
                if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&
230

1197496
                    (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&
231

390917
                    (!run    || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {
232
                    av_assert1(!ctx->vlc_codes[index]);
233
384951
                    if (alevel) {
234
384878
                        ctx->vlc_codes[index] =
235
384878
                            (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
236
384878
                        ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
237
                    } else {
238
73
                        ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
239
73
                        ctx->vlc_bits[index]  = ctx->cid_table->ac_bits[j];
240
                    }
241
384951
                    break;
242
                }
243
            }
244

385024
            av_assert0(!alevel || j < 257);
245
385024
            if (offset) {
246
366190
                ctx->vlc_codes[index] =
247
366190
                    (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
248
366190
                ctx->vlc_bits[index] += ctx->cid_table->index_bits;
249
            }
250
        }
251
    }
252
4599
    for (i = 0; i < 62; i++) {
253
4526
        int run = ctx->cid_table->run[i];
254
4526
        av_assert0(run < 63);
255
4526
        ctx->run_codes[run] = ctx->cid_table->run_codes[i];
256
4526
        ctx->run_bits[run]  = ctx->cid_table->run_bits[i];
257
    }
258
73
    return 0;
259
}
260
261
73
static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
262
{
263
    // init first elem to 1 to avoid div by 0 in convert_matrix
264
73
    uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
265
    int qscale, i;
266
73
    const uint8_t *luma_weight_table   = ctx->cid_table->luma_weight;
267
73
    const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
268
269
73
    if (!FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l,   ctx->m.avctx->qmax + 1) ||
270
73
        !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c,   ctx->m.avctx->qmax + 1) ||
271
73
        !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l16, ctx->m.avctx->qmax + 1) ||
272
73
        !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c16, ctx->m.avctx->qmax + 1))
273
        return AVERROR(ENOMEM);
274
275
73
    if (ctx->bit_depth == 8) {
276
4224
        for (i = 1; i < 64; i++) {
277
4158
            int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
278
4158
            weight_matrix[j] = ctx->cid_table->luma_weight[i];
279
        }
280
66
        ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
281
                          weight_matrix, ctx->intra_quant_bias, 1,
282
66
                          ctx->m.avctx->qmax, 1);
283
4224
        for (i = 1; i < 64; i++) {
284
4158
            int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
285
4158
            weight_matrix[j] = ctx->cid_table->chroma_weight[i];
286
        }
287
66
        ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
288
                          weight_matrix, ctx->intra_quant_bias, 1,
289
66
                          ctx->m.avctx->qmax, 1);
290
291
42250
        for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
292
2741960
            for (i = 0; i < 64; i++) {
293
2699776
                ctx->qmatrix_l[qscale][i]      <<= 2;
294
2699776
                ctx->qmatrix_c[qscale][i]      <<= 2;
295
2699776
                ctx->qmatrix_l16[qscale][0][i] <<= 2;
296
2699776
                ctx->qmatrix_l16[qscale][1][i] <<= 2;
297
2699776
                ctx->qmatrix_c16[qscale][0][i] <<= 2;
298
2699776
                ctx->qmatrix_c16[qscale][1][i] <<= 2;
299
            }
300
        }
301
    } else {
302
        // 10-bit
303
63
        for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
304
3584
            for (i = 1; i < 64; i++) {
305
3528
                int j = ff_zigzag_direct[i];
306
307
                /* The quantization formula from the VC-3 standard is:
308
                 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
309
                 *             (qscale * weight_table[i]))
310
                 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
311
                 * The s factor compensates scaling of DCT coefficients done by
312
                 * the DCT routines, and therefore is not present in standard.
313
                 * It's 8 for 8-bit samples and 4 for 10-bit ones.
314
                 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
315
                 *     ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
316
                 *     (qscale * weight_table[i])
317
                 * For 10-bit samples, p / s == 2 */
318
3528
                ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
319
3528
                                            (qscale * luma_weight_table[i]);
320
3528
                ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
321
3528
                                            (qscale * chroma_weight_table[i]);
322
            }
323
        }
324
    }
325
326
73
    ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
327
73
    ctx->m.q_chroma_intra_matrix   = ctx->qmatrix_c;
328
73
    ctx->m.q_intra_matrix16        = ctx->qmatrix_l16;
329
73
    ctx->m.q_intra_matrix          = ctx->qmatrix_l;
330
331
73
    return 0;
332
}
333
334
73
static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
335
{
336
73
    if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_rc, (ctx->m.avctx->qmax + 1) * ctx->m.mb_num))
337
        return AVERROR(ENOMEM);
338
339
73
    if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) {
340
58
        if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp,     ctx->m.mb_num) ||
341
58
            !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp_tmp, ctx->m.mb_num))
342
            return AVERROR(ENOMEM);
343
    }
344
73
    ctx->frame_bits = (ctx->coding_unit_size -
345
73
                       ctx->data_offset - 4 - ctx->min_padding) * 8;
346
73
    ctx->qscale = 1;
347
73
    ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
348
73
    return 0;
349
}
350
351
73
static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
352
{
353
73
    DNXHDEncContext *ctx = avctx->priv_data;
354
    int i, index, ret;
355
356
73
    switch (avctx->pix_fmt) {
357
66
    case AV_PIX_FMT_YUV422P:
358
66
        ctx->bit_depth = 8;
359
66
        break;
360
7
    case AV_PIX_FMT_YUV422P10:
361
    case AV_PIX_FMT_YUV444P10:
362
    case AV_PIX_FMT_GBRP10:
363
7
        ctx->bit_depth = 10;
364
7
        break;
365
    default:
366
        av_log(avctx, AV_LOG_ERROR,
367
               "pixel format is incompatible with DNxHD\n");
368
        return AVERROR(EINVAL);
369
    }
370
371

73
    if ((ctx->profile == FF_PROFILE_DNXHR_444 && (avctx->pix_fmt != AV_PIX_FMT_YUV444P10 &&
372
                                                  avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||
373

73
        (ctx->profile != FF_PROFILE_DNXHR_444 && (avctx->pix_fmt == AV_PIX_FMT_YUV444P10 ||
374
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

73
    if (ctx->profile == FF_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
73
    if ((ctx->profile == FF_PROFILE_DNXHR_LB ||
387
61
         ctx->profile == FF_PROFILE_DNXHR_SQ ||
388

73
         ctx->profile == FF_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 == FF_PROFILE_DNXHR_444;
395
73
    avctx->profile = ctx->profile;
396
73
    ctx->cid = ff_dnxhd_find_cid(avctx, ctx->bit_depth);
397
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

73
    if (ctx->cid >= 1270 && ctx->cid <= 1274)
406
40
        avctx->codec_tag = MKTAG('A','V','d','h');
407
408

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
    index = ff_dnxhd_get_cid_table(ctx->cid);
415
73
    av_assert0(index >= 0);
416
417
73
    ctx->cid_table = &ff_dnxhd_cid_table[index];
418
419
73
    ctx->m.avctx    = avctx;
420
73
    ctx->m.mb_intra = 1;
421
73
    ctx->m.h263_aic = 1;
422
423
73
    avctx->bits_per_raw_sample = ctx->bit_depth;
424
425
73
    ff_blockdsp_init(&ctx->bdsp, avctx);
426
73
    ff_fdctdsp_init(&ctx->m.fdsp, avctx);
427
73
    ff_mpv_idct_init(&ctx->m);
428
73
    ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
429
73
    ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
430
73
    ff_dct_encode_init(&ctx->m);
431
432
73
    if (ctx->profile != FF_PROFILE_DNXHD)
433
40
        ff_videodsp_init(&ctx->m.vdsp, ctx->bit_depth);
434
435
73
    if (!ctx->m.dct_quantize)
436
        ctx->m.dct_quantize = ff_dct_quantize_c;
437
438

73
    if (ctx->is_444 || ctx->profile == FF_PROFILE_DNXHR_HQX) {
439
        ctx->m.dct_quantize     = dnxhd_10bit_dct_quantize_444;
440
        ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
441
        ctx->block_width_l2     = 4;
442
73
    } else if (ctx->bit_depth == 10) {
443
7
        ctx->m.dct_quantize     = dnxhd_10bit_dct_quantize;
444
7
        ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
445
7
        ctx->block_width_l2     = 4;
446
    } else {
447
66
        ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
448
66
        ctx->block_width_l2     = 3;
449
    }
450
451
    if (ARCH_X86)
452
73
        ff_dnxhdenc_init_x86(ctx);
453
454
73
    ctx->m.mb_height = (avctx->height + 15) / 16;
455
73
    ctx->m.mb_width  = (avctx->width  + 15) / 16;
456
457
73
    if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
458
11
        ctx->interlaced   = 1;
459
11
        ctx->m.mb_height /= 2;
460
    }
461
462

73
    if (ctx->interlaced && ctx->profile != FF_PROFILE_DNXHD) {
463
        av_log(avctx, AV_LOG_ERROR,
464
               "Interlaced encoding is not supported for DNxHR profiles.\n");
465
        return AVERROR(EINVAL);
466
    }
467
468
73
    ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
469
470
73
    if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
471
40
        ctx->frame_size = avpriv_dnxhd_get_hr_frame_size(ctx->cid,
472
                                                     avctx->width, avctx->height);
473
40
        av_assert0(ctx->frame_size >= 0);
474
40
        ctx->coding_unit_size = ctx->frame_size;
475
    } else {
476
33
        ctx->frame_size = ctx->cid_table->frame_size;
477
33
        ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
478
    }
479
480
73
    if (ctx->m.mb_height > 68)
481
12
        ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);
482
    else
483
61
        ctx->data_offset = 0x280;
484
485
    // XXX tune lbias/cbias
486
73
    if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
487
        return ret;
488
489
    /* Avid Nitris hardware decoder requires a minimum amount of padding
490
     * in the coding unit payload */
491
73
    if (ctx->nitris_compat)
492
        ctx->min_padding = 1600;
493
494
73
    if ((ret = dnxhd_init_vlc(ctx)) < 0)
495
        return ret;
496
73
    if ((ret = dnxhd_init_rc(ctx)) < 0)
497
        return ret;
498
499
73
    if (!FF_ALLOCZ_TYPED_ARRAY(ctx->slice_size, ctx->m.mb_height) ||
500
73
        !FF_ALLOCZ_TYPED_ARRAY(ctx->slice_offs, ctx->m.mb_height) ||
501
73
        !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_bits,    ctx->m.mb_num)    ||
502
73
        !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_qscale,  ctx->m.mb_num))
503
        return AVERROR(ENOMEM);
504
#if FF_API_CODED_FRAME
505
FF_DISABLE_DEPRECATION_WARNINGS
506
73
    avctx->coded_frame->key_frame = 1;
507
73
    avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
508
FF_ENABLE_DEPRECATION_WARNINGS
509
#endif
510
511
73
    if (avctx->active_thread_type == FF_THREAD_SLICE) {
512
        if (avctx->thread_count > MAX_THREADS) {
513
            av_log(avctx, AV_LOG_ERROR, "too many threads\n");
514
            return AVERROR(EINVAL);
515
        }
516
    }
517
518
73
    if (avctx->qmax <= 1) {
519
        av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
520
        return AVERROR(EINVAL);
521
    }
522
523
73
    ctx->thread[0] = ctx;
524
73
    if (avctx->active_thread_type == FF_THREAD_SLICE) {
525
        for (i = 1; i < avctx->thread_count; i++) {
526
            ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
527
            if (!ctx->thread[i])
528
                return AVERROR(ENOMEM);
529
            memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
530
        }
531
    }
532
533
73
    return 0;
534
}
535
536
380
static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
537
{
538
380
    DNXHDEncContext *ctx = avctx->priv_data;
539
540
380
    memset(buf, 0, ctx->data_offset);
541
542
    // * write prefix */
543
380
    AV_WB16(buf + 0x02, ctx->data_offset);
544

380
    if (ctx->cid >= 1270 && ctx->cid <= 1274)
545
200
        buf[4] = 0x03;
546
    else
547
180
        buf[4] = 0x01;
548
549
380
    buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
550
380
    buf[6] = 0x80; // crc flag off
551
380
    buf[7] = 0xa0; // reserved
552
380
    AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
553
380
    AV_WB16(buf + 0x1a, avctx->width);  // SPL
554
380
    AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
555
556
380
    buf[0x21] = ctx->bit_depth == 10 ? 0x58 : 0x38;
557
380
    buf[0x22] = 0x88 + (ctx->interlaced << 2);
558
380
    AV_WB32(buf + 0x28, ctx->cid); // CID
559
380
    buf[0x2c] = (!ctx->interlaced << 7) | (ctx->is_444 << 6) | (avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
560
561
380
    buf[0x5f] = 0x01; // UDL
562
563
380
    buf[0x167] = 0x02; // reserved
564
380
    AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
565
380
    AV_WB16(buf + 0x16c, ctx->m.mb_height); // Ns
566
380
    buf[0x16f] = 0x10; // reserved
567
568
380
    ctx->msip = buf + 0x170;
569
380
    return 0;
570
}
571
572
26918560
static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
573
{
574
    int nbits;
575
26918560
    if (diff < 0) {
576
12317800
        nbits = av_log2_16bit(-2 * diff);
577
12317800
        diff--;
578
    } else {
579
14600760
        nbits = av_log2_16bit(2 * diff);
580
    }
581
26918560
    put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
582
26918560
             (ctx->cid_table->dc_codes[nbits] << nbits) +
583
26918560
             av_mod_uintp2(diff, nbits));
584
26918560
}
585
586
static av_always_inline
587
26918560
void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
588
                        int last_index, int n)
589
{
590
26918560
    int last_non_zero = 0;
591
    int slevel, i, j;
592
593
26918560
    dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
594
26918560
    ctx->m.last_dc[n] = block[0];
595
596
323734349
    for (i = 1; i <= last_index; i++) {
597
296815789
        j = ctx->m.intra_scantable.permutated[i];
598
296815789
        slevel = block[j];
599
296815789
        if (slevel) {
600
111021078
            int run_level = i - last_non_zero - 1;
601
111021078
            int rlevel = slevel * (1 << 1) | !!run_level;
602
111021078
            put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
603
111021078
            if (run_level)
604
36001894
                put_bits(&ctx->m.pb, ctx->run_bits[run_level],
605
36001894
                         ctx->run_codes[run_level]);
606
111021078
            last_non_zero = i;
607
        }
608
    }
609
26918560
    put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
610
26918560
}
611
612
static av_always_inline
613
3024000
void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
614
                        int qscale, int last_index)
615
{
616
    const uint8_t *weight_matrix;
617
    int level;
618
    int i;
619
620
3024000
    if (ctx->is_444) {
621
        weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight
622
                                      : ctx->cid_table->chroma_weight;
623
    } else {
624
3024000
        weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
625
3024000
                                : ctx->cid_table->luma_weight;
626
    }
627
628
35797315
    for (i = 1; i <= last_index; i++) {
629
32773315
        int j = ctx->m.intra_scantable.permutated[i];
630
32773315
        level = block[j];
631
32773315
        if (level) {
632
16774641
            if (level < 0) {
633
8458215
                level = (1 - 2 * level) * qscale * weight_matrix[i];
634
8458215
                if (ctx->bit_depth == 10) {
635
                    if (weight_matrix[i] != 8)
636
                        level += 8;
637
                    level >>= 4;
638
                } else {
639
8458215
                    if (weight_matrix[i] != 32)
640
6479585
                        level += 32;
641
8458215
                    level >>= 6;
642
                }
643
8458215
                level = -level;
644
            } else {
645
8316426
                level = (2 * level + 1) * qscale * weight_matrix[i];
646
8316426
                if (ctx->bit_depth == 10) {
647
                    if (weight_matrix[i] != 8)
648
                        level += 8;
649
                    level >>= 4;
650
                } else {
651
8316426
                    if (weight_matrix[i] != 32)
652
6813970
                        level += 32;
653
8316426
                    level >>= 6;
654
                }
655
            }
656
16774641
            block[j] = level;
657
        }
658
    }
659
3024000
}
660
661
3024000
static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
662
{
663
3024000
    int score = 0;
664
    int i;
665
196560000
    for (i = 0; i < 64; i++)
666
193536000
        score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
667
3024000
    return score;
668
}
669
670
static av_always_inline
671
51881432
int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
672
{
673
51881432
    int last_non_zero = 0;
674
51881432
    int bits = 0;
675
    int i, j, level;
676
515019637
    for (i = 1; i <= last_index; i++) {
677
463138205
        j = ctx->m.intra_scantable.permutated[i];
678
463138205
        level = block[j];
679
463138205
        if (level) {
680
191817206
            int run_level = i - last_non_zero - 1;
681
191817206
            bits += ctx->vlc_bits[level * (1 << 1) |
682
191817206
                    !!run_level] + ctx->run_bits[run_level];
683
191817206
            last_non_zero = i;
684
        }
685
    }
686
51881432
    return bits;
687
}
688
689
static av_always_inline
690
9849999
void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
691
{
692
9849999
    const int bs = ctx->block_width_l2;
693
9849999
    const int bw = 1 << bs;
694
9849999
    int dct_y_offset = ctx->dct_y_offset;
695
9849999
    int dct_uv_offset = ctx->dct_uv_offset;
696
9849999
    int linesize = ctx->m.linesize;
697
9849999
    int uvlinesize = ctx->m.uvlinesize;
698
9849999
    const uint8_t *ptr_y = ctx->thread[0]->src[0] +
699
9849999
                           ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
700
9849999
    const uint8_t *ptr_u = ctx->thread[0]->src[1] +
701
9849999
                           ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
702
9849999
    const uint8_t *ptr_v = ctx->thread[0]->src[2] +
703
9849999
                           ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
704
9849999
    PixblockDSPContext *pdsp = &ctx->m.pdsp;
705
9849999
    VideoDSPContext *vdsp = &ctx->m.vdsp;
706
707

9849999
    if (ctx->bit_depth != 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
708
7203152
                                                           (mb_y << 4) + 16 > ctx->m.avctx->height)) {
709
22511
        int y_w = ctx->m.avctx->width  - (mb_x << 4);
710
22511
        int y_h = ctx->m.avctx->height - (mb_y << 4);
711
22511
        int uv_w = (y_w + 1) / 2;
712
22511
        int uv_h = y_h;
713
22511
        linesize = 16;
714
22511
        uvlinesize = 8;
715
716
22511
        vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
717
                               linesize, ctx->m.linesize,
718
                               linesize, 16,
719
                               0, 0, y_w, y_h);
720
22511
        vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
721
                               uvlinesize, ctx->m.uvlinesize,
722
                               uvlinesize, 16,
723
                               0, 0, uv_w, uv_h);
724
22511
        vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
725
                               uvlinesize, ctx->m.uvlinesize,
726
                               uvlinesize, 16,
727
                               0, 0, uv_w, uv_h);
728
729
22511
        dct_y_offset =  bw * linesize;
730
22511
        dct_uv_offset = bw * uvlinesize;
731
22511
        ptr_y = &ctx->edge_buf_y[0];
732
22511
        ptr_u = &ctx->edge_buf_uv[0][0];
733
22511
        ptr_v = &ctx->edge_buf_uv[1][0];
734

9827488
    } else if (ctx->bit_depth == 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
735
                                                                  (mb_y << 4) + 16 > ctx->m.avctx->height)) {
736
        int y_w = ctx->m.avctx->width  - (mb_x << 4);
737
        int y_h = ctx->m.avctx->height - (mb_y << 4);
738
        int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;
739
        int uv_h = y_h;
740
        linesize = 32;
741
        uvlinesize = 16 + 16 * ctx->is_444;
742
743
        vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
744
                               linesize, ctx->m.linesize,
745
                               linesize / 2, 16,
746
                               0, 0, y_w, y_h);
747
        vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
748
                               uvlinesize, ctx->m.uvlinesize,
749
                               uvlinesize / 2, 16,
750
                               0, 0, uv_w, uv_h);
751
        vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
752
                               uvlinesize, ctx->m.uvlinesize,
753
                               uvlinesize / 2, 16,
754
                               0, 0, uv_w, uv_h);
755
756
        dct_y_offset =  bw * linesize / 2;
757
        dct_uv_offset = bw * uvlinesize / 2;
758
        ptr_y = &ctx->edge_buf_y[0];
759
        ptr_u = &ctx->edge_buf_uv[0][0];
760
        ptr_v = &ctx->edge_buf_uv[1][0];
761
    }
762
763
9849999
    if (!ctx->is_444) {
764
9849999
        pdsp->get_pixels(ctx->blocks[0], ptr_y,      linesize);
765
9849999
        pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
766
9849999
        pdsp->get_pixels(ctx->blocks[2], ptr_u,      uvlinesize);
767
9849999
        pdsp->get_pixels(ctx->blocks[3], ptr_v,      uvlinesize);
768
769

9849999
        if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
770
56560
            if (ctx->interlaced) {
771
35280
                ctx->get_pixels_8x4_sym(ctx->blocks[4],
772
                                        ptr_y + dct_y_offset,
773
                                        linesize);
774
35280
                ctx->get_pixels_8x4_sym(ctx->blocks[5],
775
35280
                                        ptr_y + dct_y_offset + bw,
776
                                        linesize);
777
35280
                ctx->get_pixels_8x4_sym(ctx->blocks[6],
778
                                        ptr_u + dct_uv_offset,
779
                                        uvlinesize);
780
35280
                ctx->get_pixels_8x4_sym(ctx->blocks[7],
781
                                        ptr_v + dct_uv_offset,
782
                                        uvlinesize);
783
            } else {
784
21280
                ctx->bdsp.clear_block(ctx->blocks[4]);
785
21280
                ctx->bdsp.clear_block(ctx->blocks[5]);
786
21280
                ctx->bdsp.clear_block(ctx->blocks[6]);
787
21280
                ctx->bdsp.clear_block(ctx->blocks[7]);
788
            }
789
        } else {
790
9793439
            pdsp->get_pixels(ctx->blocks[4],
791
                             ptr_y + dct_y_offset, linesize);
792
9793439
            pdsp->get_pixels(ctx->blocks[5],
793
9793439
                             ptr_y + dct_y_offset + bw, linesize);
794
9793439
            pdsp->get_pixels(ctx->blocks[6],
795
                             ptr_u + dct_uv_offset, uvlinesize);
796
9793439
            pdsp->get_pixels(ctx->blocks[7],
797
                             ptr_v + dct_uv_offset, uvlinesize);
798
        }
799
    } else {
800
        pdsp->get_pixels(ctx->blocks[0], ptr_y,      linesize);
801
        pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
802
        pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);
803
        pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);
804
805
        pdsp->get_pixels(ctx->blocks[2], ptr_u,      uvlinesize);
806
        pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);
807
        pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);
808
        pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);
809
810
        pdsp->get_pixels(ctx->blocks[4], ptr_v,      uvlinesize);
811
        pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);
812
        pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);
813
        pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);
814
    }
815
9849999
}
816
817
static av_always_inline
818
78799992
int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
819
{
820
    int x;
821
822
78799992
    if (ctx->is_444) {
823
        x = (i >> 1) % 3;
824
    } else {
825
        const static uint8_t component[8]={0,0,1,2,0,0,1,2};
826
78799992
        x = component[i];
827
    }
828
78799992
    return x;
829
}
830
831
41557
static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
832
                                  int jobnr, int threadnr)
833
{
834
41557
    DNXHDEncContext *ctx = avctx->priv_data;
835
41557
    int mb_y = jobnr, mb_x;
836
41557
    int qscale = ctx->qscale;
837
41557
    LOCAL_ALIGNED_16(int16_t, block, [64]);
838
41557
    ctx = ctx->thread[threadnr];
839
840
41557
    ctx->m.last_dc[0] =
841
41557
    ctx->m.last_dc[1] =
842
41557
    ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
843
844
6526736
    for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
845
6485179
        unsigned mb = mb_y * ctx->m.mb_width + mb_x;
846
6485179
        int ssd     = 0;
847
6485179
        int ac_bits = 0;
848
6485179
        int dc_bits = 0;
849
        int i;
850
851
6485179
        dnxhd_get_blocks(ctx, mb_x, mb_y);
852
853
58366611
        for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
854
51881432
            int16_t *src_block = ctx->blocks[i];
855
            int overflow, nbits, diff, last_index;
856
51881432
            int n = dnxhd_switch_matrix(ctx, i);
857
858
51881432
            memcpy(block, src_block, 64 * sizeof(*block));
859
51881432
            last_index = ctx->m.dct_quantize(&ctx->m, block,
860

51881432
                                             ctx->is_444 ? 4 * (n > 0): 4 & (2*i),
861
                                             qscale, &overflow);
862
51881432
            ac_bits   += dnxhd_calc_ac_bits(ctx, block, last_index);
863
864
51881432
            diff = block[0] - ctx->m.last_dc[n];
865
51881432
            if (diff < 0)
866
23766663
                nbits = av_log2_16bit(-2 * diff);
867
            else
868
28114769
                nbits = av_log2_16bit(2 * diff);
869
870
            av_assert1(nbits < ctx->bit_depth + 4);
871
51881432
            dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
872
873
51881432
            ctx->m.last_dc[n] = block[0];
874
875
51881432
            if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
876
3024000
                dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
877
3024000
                ctx->m.idsp.idct(block);
878
3024000
                ssd += dnxhd_ssd_block(block, src_block);
879
            }
880
        }
881
6485179
        ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd  = ssd;
882
6485179
        ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
883
6485179
                                     (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];
884
    }
885
41557
    return 0;
886
}
887
888
20765
static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
889
                               int jobnr, int threadnr)
890
{
891
20765
    DNXHDEncContext *ctx = avctx->priv_data;
892
20765
    int mb_y = jobnr, mb_x;
893
20765
    ctx = ctx->thread[threadnr];
894
20765
    init_put_bits(&ctx->m.pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
895
20765
                  ctx->slice_size[jobnr]);
896
897
20765
    ctx->m.last_dc[0] =
898
20765
    ctx->m.last_dc[1] =
899
20765
    ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
900
3385585
    for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
901
3364820
        unsigned mb = mb_y * ctx->m.mb_width + mb_x;
902
3364820
        int qscale = ctx->mb_qscale[mb];
903
        int i;
904
905
3364820
        put_bits(&ctx->m.pb, 11, qscale);
906
3364820
        put_bits(&ctx->m.pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
907
908
3364820
        dnxhd_get_blocks(ctx, mb_x, mb_y);
909
910
30283380
        for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
911
26918560
            int16_t *block = ctx->blocks[i];
912
26918560
            int overflow, n = dnxhd_switch_matrix(ctx, i);
913
26918560
            int last_index = ctx->m.dct_quantize(&ctx->m, block,
914

26918560
                                                 ctx->is_444 ? (((i >> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),
915
                                                 qscale, &overflow);
916
917
26918560
            dnxhd_encode_block(ctx, block, last_index, n);
918
        }
919
    }
920
20765
    if (put_bits_count(&ctx->m.pb) & 31)
921
20137
        put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
922
20765
    flush_put_bits(&ctx->m.pb);
923
20765
    return 0;
924
}
925
926
380
static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
927
{
928
    int mb_y, mb_x;
929
380
    int offset = 0;
930
21145
    for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
931
        int thread_size;
932
20765
        ctx->slice_offs[mb_y] = offset;
933
20765
        ctx->slice_size[mb_y] = 0;
934
3385585
        for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
935
3364820
            unsigned mb = mb_y * ctx->m.mb_width + mb_x;
936
3364820
            ctx->slice_size[mb_y] += ctx->mb_bits[mb];
937
        }
938
20765
        ctx->slice_size[mb_y]   = (ctx->slice_size[mb_y] + 31) & ~31;
939
20765
        ctx->slice_size[mb_y] >>= 3;
940
20765
        thread_size = ctx->slice_size[mb_y];
941
20765
        offset += thread_size;
942
    }
943
380
}
944
945
11520
static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
946
                               int jobnr, int threadnr)
947
{
948
11520
    DNXHDEncContext *ctx = avctx->priv_data;
949
11520
    int mb_y = jobnr, mb_x, x, y;
950
11740
    int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
951
220
                           ((avctx->height >> ctx->interlaced) & 0xF);
952
953
11520
    ctx = ctx->thread[threadnr];
954
11520
    if (ctx->bit_depth == 8) {
955
10955
        uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
956
1930655
        for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
957
1919700
            unsigned mb = mb_y * ctx->m.mb_width + mb_x;
958
            int sum;
959
            int varc;
960
961

1919700
            if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {
962
1901720
                sum  = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
963
1901720
                varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
964
            } else {
965
17980
                int bw = FFMIN(avctx->width - 16 * mb_x, 16);
966
17980
                int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
967
17980
                sum = varc = 0;
968
244175
                for (y = 0; y < bh; y++) {
969
3800030
                    for (x = 0; x < bw; x++) {
970
3573835
                        uint8_t val = pix[x + y * ctx->m.linesize];
971
3573835
                        sum  += val;
972
3573835
                        varc += val * val;
973
                    }
974
                }
975
            }
976
1919700
            varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
977
978
1919700
            ctx->mb_cmp[mb].value = varc;
979
1919700
            ctx->mb_cmp[mb].mb    = mb;
980
        }
981
    } else { // 10-bit
982
565
        const int linesize = ctx->m.linesize >> 1;
983
59365
        for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
984
58800
            uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
985
58800
                            ((mb_y << 4) * linesize) + (mb_x << 4);
986
58800
            unsigned mb  = mb_y * ctx->m.mb_width + mb_x;
987
58800
            int sum = 0;
988
58800
            int sqsum = 0;
989
58800
            int bw = FFMIN(avctx->width - 16 * mb_x, 16);
990
58800
            int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
991
            int mean, sqmean;
992
            int i, j;
993
            // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
994
994800
            for (i = 0; i < bh; ++i) {
995
15912000
                for (j = 0; j < bw; ++j) {
996
                    // Turn 16-bit pixels into 10-bit ones.
997
14976000
                    const int sample = (unsigned) pix[j] >> 6;
998
14976000
                    sum   += sample;
999
14976000
                    sqsum += sample * sample;
1000
                    // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
1001
                }
1002
936000
                pix += linesize;
1003
            }
1004
58800
            mean = sum >> 8; // 16*16 == 2^8
1005
58800
            sqmean = sqsum >> 8;
1006
58800
            ctx->mb_cmp[mb].value = sqmean - mean * mean;
1007
58800
            ctx->mb_cmp[mb].mb    = mb;
1008
        }
1009
    }
1010
11520
    return 0;
1011
}
1012
1013
15
static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
1014
{
1015
    int lambda, up_step, down_step;
1016
15
    int last_lower = INT_MAX, last_higher = 0;
1017
    int x, y, q;
1018
1019
120
    for (q = 1; q < avctx->qmax; q++) {
1020
105
        ctx->qscale = q;
1021
105
        avctx->execute2(avctx, dnxhd_calc_bits_thread,
1022
                        NULL, NULL, ctx->m.mb_height);
1023
    }
1024
15
    up_step = down_step = 2 << LAMBDA_FRAC_BITS;
1025
15
    lambda  = ctx->lambda;
1026
1027
70
    for (;;) {
1028
85
        int bits = 0;
1029
85
        int end  = 0;
1030
85
        if (lambda == last_higher) {
1031
5
            lambda++;
1032
5
            end = 1; // need to set final qscales/bits
1033
        }
1034
3862
        for (y = 0; y < ctx->m.mb_height; y++) {
1035
308529
            for (x = 0; x < ctx->m.mb_width; x++) {
1036
304720
                unsigned min = UINT_MAX;
1037
304720
                int qscale = 1;
1038
304720
                int mb     = y * ctx->m.mb_width + x;
1039
304720
                int rc = 0;
1040
2437760
                for (q = 1; q < avctx->qmax; q++) {
1041
2133040
                    int i = (q*ctx->m.mb_num) + mb;
1042
2133040
                    unsigned score = ctx->mb_rc[i].bits * lambda +
1043
2133040
                                     ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
1044
2133040
                    if (score < min) {
1045
493989
                        min    = score;
1046
493989
                        qscale = q;
1047
493989
                        rc = i;
1048
                    }
1049
                }
1050
304720
                bits += ctx->mb_rc[rc].bits;
1051
304720
                ctx->mb_qscale[mb] = qscale;
1052
304720
                ctx->mb_bits[mb]   = ctx->mb_rc[rc].bits;
1053
            }
1054
3809
            bits = (bits + 31) & ~31; // padding
1055
3809
            if (bits > ctx->frame_bits)
1056
32
                break;
1057
        }
1058
85
        if (end) {
1059
5
            if (bits > ctx->frame_bits)
1060
                return AVERROR(EINVAL);
1061
5
            break;
1062
        }
1063
80
        if (bits < ctx->frame_bits) {
1064
48
            last_lower = FFMIN(lambda, last_lower);
1065
48
            if (last_higher != 0)
1066
25
                lambda = (lambda+last_higher)>>1;
1067
            else
1068
23
                lambda -= down_step;
1069
48
            down_step = FFMIN((int64_t)down_step*5, INT_MAX);
1070
48
            up_step = 1<<LAMBDA_FRAC_BITS;
1071
48
            lambda = FFMAX(1, lambda);
1072
48
            if (lambda == last_lower)
1073
10
                break;
1074
        } else {
1075
32
            last_higher = FFMAX(lambda, last_higher);
1076
32
            if (last_lower != INT_MAX)
1077
32
                lambda = (lambda+last_lower)>>1;
1078
            else if ((int64_t)lambda + up_step > INT_MAX)
1079
                return AVERROR(EINVAL);
1080
            else
1081
                lambda += up_step;
1082
32
            up_step = FFMIN((int64_t)up_step*5, INT_MAX);
1083
32
            down_step = 1<<LAMBDA_FRAC_BITS;
1084
        }
1085
    }
1086
15
    ctx->lambda = lambda;
1087
15
    return 0;
1088
}
1089
1090
365
static int dnxhd_find_qscale(DNXHDEncContext *ctx)
1091
{
1092
365
    int bits = 0;
1093
365
    int up_step = 1;
1094
365
    int down_step = 1;
1095
365
    int last_higher = 0;
1096
365
    int last_lower = INT_MAX;
1097
    int qscale;
1098
    int x, y;
1099
1100
365
    qscale = ctx->qscale;
1101
    for (;;) {
1102
691
        bits = 0;
1103
691
        ctx->qscale = qscale;
1104
        // XXX avoid recalculating bits
1105
691
        ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
1106
                               NULL, NULL, ctx->m.mb_height);
1107
35231
        for (y = 0; y < ctx->m.mb_height; y++) {
1108
5794959
            for (x = 0; x < ctx->m.mb_width; x++)
1109
5760121
                bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;
1110
34838
            bits = (bits+31)&~31; // padding
1111
34838
            if (bits > ctx->frame_bits)
1112
298
                break;
1113
        }
1114
691
        if (bits < ctx->frame_bits) {
1115
393
            if (qscale == 1)
1116
145
                return 1;
1117
248
            if (last_higher == qscale - 1) {
1118
195
                qscale = last_higher;
1119
195
                break;
1120
            }
1121
53
            last_lower = FFMIN(qscale, last_lower);
1122
53
            if (last_higher != 0)
1123
38
                qscale = (qscale + last_higher) >> 1;
1124
            else
1125
15
                qscale -= down_step++;
1126
53
            if (qscale < 1)
1127
                qscale = 1;
1128
53
            up_step = 1;
1129
        } else {
1130
298
            if (last_lower == qscale + 1)
1131
25
                break;
1132
273
            last_higher = FFMAX(qscale, last_higher);
1133
273
            if (last_lower != INT_MAX)
1134
5
                qscale = (qscale + last_lower) >> 1;
1135
            else
1136
268
                qscale += up_step++;
1137
273
            down_step = 1;
1138
273
            if (qscale >= ctx->m.avctx->qmax)
1139
                return AVERROR(EINVAL);
1140
        }
1141
    }
1142
220
    ctx->qscale = qscale;
1143
220
    return 0;
1144
}
1145
1146
#define BUCKET_BITS 8
1147
#define RADIX_PASSES 4
1148
#define NBUCKETS (1 << BUCKET_BITS)
1149
1150
11871000
static inline int get_bucket(int value, int shift)
1151
{
1152
11871000
    value >>= shift;
1153
11871000
    value  &= NBUCKETS - 1;
1154
11871000
    return NBUCKETS - 1 - value;
1155
}
1156
1157
220
static void radix_count(const RCCMPEntry *data, int size,
1158
                        int buckets[RADIX_PASSES][NBUCKETS])
1159
{
1160
    int i, j;
1161
220
    memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1162
1978720
    for (i = 0; i < size; i++) {
1163
1978500
        int v = data[i].value;
1164
9892500
        for (j = 0; j < RADIX_PASSES; j++) {
1165
7914000
            buckets[j][get_bucket(v, 0)]++;
1166
7914000
            v >>= BUCKET_BITS;
1167
        }
1168
        av_assert1(!v);
1169
    }
1170
1100
    for (j = 0; j < RADIX_PASSES; j++) {
1171
880
        int offset = size;
1172
226160
        for (i = NBUCKETS - 1; i >= 0; i--)
1173
225280
            buckets[j][i] = offset -= buckets[j][i];
1174
        av_assert1(!buckets[j][0]);
1175
    }
1176
220
}
1177
1178
440
static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1179
                            int size, int buckets[NBUCKETS], int pass)
1180
{
1181
440
    int shift = pass * BUCKET_BITS;
1182
    int i;
1183
3957440
    for (i = 0; i < size; i++) {
1184
3957000
        int v   = get_bucket(data[i].value, shift);
1185
3957000
        int pos = buckets[v]++;
1186
3957000
        dst[pos] = data[i];
1187
    }
1188
440
}
1189
1190
220
static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)
1191
{
1192
    int buckets[RADIX_PASSES][NBUCKETS];
1193
220
    radix_count(data, size, buckets);
1194
220
    radix_sort_pass(tmp, data, size, buckets[0], 0);
1195
220
    radix_sort_pass(data, tmp, size, buckets[1], 1);
1196

220
    if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
1197
        radix_sort_pass(tmp, data, size, buckets[2], 2);
1198
        radix_sort_pass(data, tmp, size, buckets[3], 3);
1199
    }
1200
220
}
1201
1202
365
static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1203
{
1204
365
    int max_bits = 0;
1205
    int ret, x, y;
1206
365
    if ((ret = dnxhd_find_qscale(ctx)) < 0)
1207
        return ret;
1208
20455
    for (y = 0; y < ctx->m.mb_height; y++) {
1209
3330910
        for (x = 0; x < ctx->m.mb_width; x++) {
1210
3310820
            int mb = y * ctx->m.mb_width + x;
1211
3310820
            int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1212
            int delta_bits;
1213
3310820
            ctx->mb_qscale[mb] = ctx->qscale;
1214
3310820
            ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1215
3310820
            max_bits += ctx->mb_rc[rc].bits;
1216
            if (!RC_VARIANCE) {
1217
                delta_bits = ctx->mb_rc[rc].bits -
1218
                             ctx->mb_rc[rc + ctx->m.mb_num].bits;
1219
                ctx->mb_cmp[mb].mb = mb;
1220
                ctx->mb_cmp[mb].value =
1221
                    delta_bits ? ((ctx->mb_rc[rc].ssd -
1222
                                   ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /
1223
                                  delta_bits
1224
                               : INT_MIN; // avoid increasing qscale
1225
            }
1226
        }
1227
20090
        max_bits += 31; // worst padding
1228
    }
1229
365
    if (!ret) {
1230
        if (RC_VARIANCE)
1231
220
            avctx->execute2(avctx, dnxhd_mb_var_thread,
1232
                            NULL, NULL, ctx->m.mb_height);
1233
220
        radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.mb_num);
1234

647025
        for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1235
646805
            int mb = ctx->mb_cmp[x].mb;
1236
646805
            int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1237
646805
            max_bits -= ctx->mb_rc[rc].bits -
1238
646805
                        ctx->mb_rc[rc + ctx->m.mb_num].bits;
1239
646805
            ctx->mb_qscale[mb] = ctx->qscale + 1;
1240
646805
            ctx->mb_bits[mb]   = ctx->mb_rc[rc + ctx->m.mb_num].bits;
1241
        }
1242
    }
1243
365
    return 0;
1244
}
1245
1246
325
static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1247
{
1248
    int i;
1249
1250
650
    for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1251
325
        ctx->thread[i]->m.linesize    = frame->linesize[0] << ctx->interlaced;
1252
325
        ctx->thread[i]->m.uvlinesize  = frame->linesize[1] << ctx->interlaced;
1253
325
        ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;
1254
325
        ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1255
    }
1256
1257
#if FF_API_CODED_FRAME
1258
FF_DISABLE_DEPRECATION_WARNINGS
1259
325
    ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1260
FF_ENABLE_DEPRECATION_WARNINGS
1261
#endif
1262

325
    ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1263
325
}
1264
1265
325
static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1266
                                const AVFrame *frame, int *got_packet)
1267
{
1268
325
    DNXHDEncContext *ctx = avctx->priv_data;
1269
325
    int first_field = 1;
1270
    int offset, i, ret;
1271
    uint8_t *buf;
1272
1273
325
    if ((ret = ff_alloc_packet2(avctx, pkt, ctx->frame_size, 0)) < 0)
1274
        return ret;
1275
325
    buf = pkt->data;
1276
1277
325
    dnxhd_load_picture(ctx, frame);
1278
1279
380
encode_coding_unit:
1280
1520
    for (i = 0; i < 3; i++) {
1281
1140
        ctx->src[i] = frame->data[i];
1282

1140
        if (ctx->interlaced && ctx->cur_field)
1283
165
            ctx->src[i] += frame->linesize[i];
1284
    }
1285
1286
380
    dnxhd_write_header(avctx, buf);
1287
1288
380
    if (avctx->mb_decision == FF_MB_DECISION_RD)
1289
15
        ret = dnxhd_encode_rdo(avctx, ctx);
1290
    else
1291
365
        ret = dnxhd_encode_fast(avctx, ctx);
1292
380
    if (ret < 0) {
1293
        av_log(avctx, AV_LOG_ERROR,
1294
               "picture could not fit ratecontrol constraints, increase qmax\n");
1295
        return ret;
1296
    }
1297
1298
380
    dnxhd_setup_threads_slices(ctx);
1299
1300
380
    offset = 0;
1301
21145
    for (i = 0; i < ctx->m.mb_height; i++) {
1302
20765
        AV_WB32(ctx->msip + i * 4, offset);
1303
20765
        offset += ctx->slice_size[i];
1304
        av_assert1(!(ctx->slice_size[i] & 3));
1305
    }
1306
1307
380
    avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1308
1309
    av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1310
380
    memset(buf + ctx->data_offset + offset, 0,
1311
380
           ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1312
1313
380
    AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1314
1315

380
    if (ctx->interlaced && first_field) {
1316
55
        first_field     = 0;
1317
55
        ctx->cur_field ^= 1;
1318
55
        buf            += ctx->coding_unit_size;
1319
55
        goto encode_coding_unit;
1320
    }
1321
1322
#if FF_API_CODED_FRAME
1323
FF_DISABLE_DEPRECATION_WARNINGS
1324
325
    avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1325
FF_ENABLE_DEPRECATION_WARNINGS
1326
#endif
1327
1328
325
    ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1329
1330
325
    pkt->flags |= AV_PKT_FLAG_KEY;
1331
325
    *got_packet = 1;
1332
325
    return 0;
1333
}
1334
1335
73
static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1336
{
1337
73
    DNXHDEncContext *ctx = avctx->priv_data;
1338
    int i;
1339
1340
73
    av_freep(&ctx->orig_vlc_codes);
1341
73
    av_freep(&ctx->orig_vlc_bits);
1342
73
    av_freep(&ctx->run_codes);
1343
73
    av_freep(&ctx->run_bits);
1344
1345
73
    av_freep(&ctx->mb_bits);
1346
73
    av_freep(&ctx->mb_qscale);
1347
73
    av_freep(&ctx->mb_rc);
1348
73
    av_freep(&ctx->mb_cmp);
1349
73
    av_freep(&ctx->mb_cmp_tmp);
1350
73
    av_freep(&ctx->slice_size);
1351
73
    av_freep(&ctx->slice_offs);
1352
1353
73
    av_freep(&ctx->qmatrix_c);
1354
73
    av_freep(&ctx->qmatrix_l);
1355
73
    av_freep(&ctx->qmatrix_c16);
1356
73
    av_freep(&ctx->qmatrix_l16);
1357
1358
73
    if (avctx->active_thread_type == FF_THREAD_SLICE) {
1359
        for (i = 1; i < avctx->thread_count; i++)
1360
            av_freep(&ctx->thread[i]);
1361
    }
1362
1363
73
    return 0;
1364
}
1365
1366
static const AVCodecDefault dnxhd_defaults[] = {
1367
    { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1368
    { NULL },
1369
};
1370
1371
AVCodec ff_dnxhd_encoder = {
1372
    .name           = "dnxhd",
1373
    .long_name      = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1374
    .type           = AVMEDIA_TYPE_VIDEO,
1375
    .id             = AV_CODEC_ID_DNXHD,
1376
    .priv_data_size = sizeof(DNXHDEncContext),
1377
    .init           = dnxhd_encode_init,
1378
    .encode2        = dnxhd_encode_picture,
1379
    .close          = dnxhd_encode_end,
1380
    .capabilities   = AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_FRAME_THREADS,
1381
    .caps_internal  = FF_CODEC_CAP_INIT_CLEANUP,
1382
    .pix_fmts       = (const enum AVPixelFormat[]) {
1383
        AV_PIX_FMT_YUV422P,
1384
        AV_PIX_FMT_YUV422P10,
1385
        AV_PIX_FMT_YUV444P10,
1386
        AV_PIX_FMT_GBRP10,
1387
        AV_PIX_FMT_NONE
1388
    },
1389
    .priv_class     = &dnxhd_class,
1390
    .defaults       = dnxhd_defaults,
1391
    .profiles       = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),
1392
};