src/i965_encoder_utils.c

   1 /*
   2  * Copyright © 2011 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the
   6  * "Software"), to deal in the Software without restriction, including
   7  * without limitation the rights to use, copy, modify, merge, publish,
   8  * distribute, sub license, and/or sell copies of the Software, and to
   9  * permit persons to whom the Software is furnished to do so, subject to
  10  * the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the
  13  * next paragraph) shall be included in all copies or substantial portions
  14  * of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  17  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  18  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  19  * IN NO EVENT SHALL PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR
  20  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  21  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  22  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  */
  25
  26 #include <stdlib.h>
  27 #include <assert.h>
  28
  29 #include <va/va.h>
  30 #include <va/va_enc_h264.h>
  31 #include <va/va_enc_mpeg2.h>
  32
  33 #include "i965_encoder_utils.h"
  34
  35 #define BITSTREAM_ALLOCATE_STEPPING     4096
  36
  37 #define NAL_REF_IDC_NONE        0
  38 #define NAL_REF_IDC_LOW         1
  39 #define NAL_REF_IDC_MEDIUM      2
  40 #define NAL_REF_IDC_HIGH        3
  41
  42 #define NAL_NON_IDR             1
  43 #define NAL_IDR                 5
  44 #define NAL_SPS                 7
  45 #define NAL_PPS                 8
  46 #define NAL_SEI                 6
  47
  48 #define SLICE_TYPE_P            0
  49 #define SLICE_TYPE_B            1
  50 #define SLICE_TYPE_I            2
  51
  52 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type) || SLICE_TYPE_I == (type - 5))
  53 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type) || SLICE_TYPE_P == (type - 5))
  54 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type) || SLICE_TYPE_B == (type - 5))
  55
  56 #define ENTROPY_MODE_CAVLC      0
  57 #define ENTROPY_MODE_CABAC      1
  58
  59 #define PROFILE_IDC_BASELINE    66
  60 #define PROFILE_IDC_MAIN        77
  61 #define PROFILE_IDC_HIGH        100
  62
  63 struct __avc_bitstream {
  64     unsigned int *buffer;
  65     int bit_offset;
  66     int max_size_in_dword;
  67 };
  68
  69 typedef struct __avc_bitstream avc_bitstream;
  70
  71 static unsigned int
  72 swap32(unsigned int val)
  73 {
  74     unsigned char *pval = (unsigned char *)&val;
  75
  76     return ((pval[0] << 24)     |
  77             (pval[1] << 16)     |
  78             (pval[2] << 8)      |
  79             (pval[3] << 0));
  80 }
  81
  82 static void
  83 avc_bitstream_start(avc_bitstream *bs)
  84 {
  85     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
  86     bs->buffer = calloc(bs->max_size_in_dword * sizeof(int), 1);
  87     bs->bit_offset = 0;
  88 }
  89
  90 static void
  91 avc_bitstream_end(avc_bitstream *bs)
  92 {
  93     int pos = (bs->bit_offset >> 5);
  94     int bit_offset = (bs->bit_offset & 0x1f);
  95     int bit_left = 32 - bit_offset;
  96
  97     if (bit_offset) {
  98         bs->buffer[pos] = swap32((bs->buffer[pos] << bit_left));
  99     }
 100
 101     // free(bs->buffer);
 102 }
 103
 104 static void
 105 avc_bitstream_put_ui(avc_bitstream *bs, unsigned int val, int size_in_bits)
 106 {
 107     int pos = (bs->bit_offset >> 5);
 108     int bit_offset = (bs->bit_offset & 0x1f);
 109     int bit_left = 32 - bit_offset;
 110
 111     if (!size_in_bits)
 112         return;
 113
 114     if (size_in_bits < 32)
 115         val &= (( 1 << size_in_bits) - 1);
 116
 117     bs->bit_offset += size_in_bits;
 118
 119     if (bit_left > size_in_bits) {
 120         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 121     } else {
 122         size_in_bits -= bit_left;
 123         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 124         bs->buffer[pos] = swap32(bs->buffer[pos]);
 125
 126         if (pos + 1 == bs->max_size_in_dword) {
 127             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 128             bs->buffer = realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 129         }
 130
 131         bs->buffer[pos + 1] = val;
 132     }
 133 }
 134
 135 static void
 136 avc_bitstream_put_ue(avc_bitstream *bs, unsigned int val)
 137 {
 138     int size_in_bits = 0;
 139     int tmp_val = ++val;
 140
 141     while (tmp_val) {
 142         tmp_val >>= 1;
 143         size_in_bits++;
 144     }
 145
 146     avc_bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 147     avc_bitstream_put_ui(bs, val, size_in_bits);
 148 }
 149
 150 static void
 151 avc_bitstream_put_se(avc_bitstream *bs, int val)
 152 {
 153     unsigned int new_val;
 154
 155     if (val <= 0)
 156         new_val = -2 * val;
 157     else
 158         new_val = 2 * val - 1;
 159
 160     avc_bitstream_put_ue(bs, new_val);
 161 }
 162
 163 static void
 164 avc_bitstream_byte_aligning(avc_bitstream *bs, int bit)
 165 {
 166     int bit_offset = (bs->bit_offset & 0x7);
 167     int bit_left = 8 - bit_offset;
 168     int new_val;
 169
 170     if (!bit_offset)
 171         return;
 172
 173     assert(bit == 0 || bit == 1);
 174
 175     if (bit)
 176         new_val = (1 << bit_left) - 1;
 177     else
 178         new_val = 0;
 179
 180     avc_bitstream_put_ui(bs, new_val, bit_left);
 181 }
 182 static void avc_rbsp_trailing_bits(avc_bitstream *bs)
 183 {
 184     avc_bitstream_put_ui(bs, 1, 1);
 185     avc_bitstream_byte_aligning(bs, 0);
 186 }
 187 static void nal_start_code_prefix(avc_bitstream *bs)
 188 {
 189     avc_bitstream_put_ui(bs, 0x00000001, 32);
 190 }
 191
 192 static void nal_header(avc_bitstream *bs, int nal_ref_idc, int nal_unit_type)
 193 {
 194     avc_bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 195     avc_bitstream_put_ui(bs, nal_ref_idc, 2);
 196     avc_bitstream_put_ui(bs, nal_unit_type, 5);
 197 }
 198
 199 static void
 200 slice_header(avc_bitstream *bs,
 201              VAEncSequenceParameterBufferH264 *sps_param,
 202              VAEncPictureParameterBufferH264 *pic_param,
 203              VAEncSliceParameterBufferH264 *slice_param)
 204 {
 205     int first_mb_in_slice = slice_param->macroblock_address;
 206
 207     avc_bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 208     avc_bitstream_put_ue(bs, slice_param->slice_type);  /* slice_type */
 209     avc_bitstream_put_ue(bs, slice_param->pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 210     avc_bitstream_put_ui(bs, pic_param->frame_num, sps_param->seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 211
 212     /* frame_mbs_only_flag == 1 */
 213     if (!sps_param->seq_fields.bits.frame_mbs_only_flag) {
 214         /* FIXME: */
 215         assert(0);
 216     }
 217
 218     if (pic_param->pic_fields.bits.idr_pic_flag)
 219         avc_bitstream_put_ue(bs, slice_param->idr_pic_id);              /* idr_pic_id: 0 */
 220
 221     if (sps_param->seq_fields.bits.pic_order_cnt_type == 0) {
 222         avc_bitstream_put_ui(bs, pic_param->CurrPic.TopFieldOrderCnt, sps_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 223         /* pic_order_present_flag == 0 */
 224     } else {
 225         /* FIXME: */
 226         assert(0);
 227     }
 228
 229     /* redundant_pic_cnt_present_flag == 0 */
 230
 231     /* slice type */
 232     if (IS_P_SLICE(slice_param->slice_type)) {
 233         avc_bitstream_put_ui(bs, 0, 1);            /* num_ref_idx_active_override_flag: 0 */
 234
 235         /* ref_pic_list_reordering */
 236         avc_bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 237     } else if (IS_B_SLICE(slice_param->slice_type)) {
 238         avc_bitstream_put_ui(bs, slice_param->direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 239         avc_bitstream_put_ui(bs, 0, 1);            /* num_ref_idx_active_override_flag: 0 */
 240
 241         /* ref_pic_list_reordering */
 242         avc_bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 243         avc_bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 244     }
 245
 246     if ((pic_param->pic_fields.bits.weighted_pred_flag &&
 247          IS_P_SLICE(slice_param->slice_type)) ||
 248         ((pic_param->pic_fields.bits.weighted_bipred_idc == 1) &&
 249          IS_B_SLICE(slice_param->slice_type))) {
 250         /* FIXME: fill weight/offset table */
 251         assert(0);
 252     }
 253
 254     /* dec_ref_pic_marking */
 255     if (pic_param->pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 256         unsigned char no_output_of_prior_pics_flag = 0;
 257         unsigned char long_term_reference_flag = 0;
 258         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 259
 260         if (pic_param->pic_fields.bits.idr_pic_flag) {
 261             avc_bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 262             avc_bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 263         } else {
 264             avc_bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 265         }
 266     }
 267
 268     if (pic_param->pic_fields.bits.entropy_coding_mode_flag &&
 269         !IS_I_SLICE(slice_param->slice_type))
 270         avc_bitstream_put_ue(bs, slice_param->cabac_init_idc);               /* cabac_init_idc: 0 */
 271
 272     avc_bitstream_put_se(bs, slice_param->slice_qp_delta);                   /* slice_qp_delta: 0 */
 273
 274     /* ignore for SP/SI */
 275
 276     if (pic_param->pic_fields.bits.deblocking_filter_control_present_flag) {
 277         avc_bitstream_put_ue(bs, slice_param->disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 278
 279         if (slice_param->disable_deblocking_filter_idc != 1) {
 280             avc_bitstream_put_se(bs, slice_param->slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 281             avc_bitstream_put_se(bs, slice_param->slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 282         }
 283     }
 284
 285     if (pic_param->pic_fields.bits.entropy_coding_mode_flag) {
 286         avc_bitstream_byte_aligning(bs, 1);
 287     }
 288 }
 289
 290 int
 291 build_avc_slice_header(VAEncSequenceParameterBufferH264 *sps_param,
 292                        VAEncPictureParameterBufferH264 *pic_param,
 293                        VAEncSliceParameterBufferH264 *slice_param,
 294                        unsigned char **slice_header_buffer)
 295 {
 296     avc_bitstream bs;
 297     int is_idr = !!pic_param->pic_fields.bits.idr_pic_flag;
 298
 299     avc_bitstream_start(&bs);
 300     nal_start_code_prefix(&bs);
 301
 302     if (IS_I_SLICE(slice_param->slice_type)) {
 303         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 304     } else if (IS_P_SLICE(slice_param->slice_type)) {
 305         nal_header(&bs, NAL_REF_IDC_MEDIUM, is_idr ? NAL_IDR : NAL_NON_IDR);
 306     } else {
 307         assert(IS_B_SLICE(slice_param->slice_type));
 308         nal_header(&bs, NAL_REF_IDC_NONE, is_idr ? NAL_IDR : NAL_NON_IDR);
 309     }
 310
 311     slice_header(&bs, sps_param, pic_param, slice_param);
 312
 313     avc_bitstream_end(&bs);
 314     *slice_header_buffer = (unsigned char *)bs.buffer;
 315
 316     return bs.bit_offset;
 317 }
 318
 319 int
 320 build_avc_sei_buffering_period(int cpb_removal_length,
 321                                unsigned int init_cpb_removal_delay,
 322                                unsigned int init_cpb_removal_delay_offset,
 323                                unsigned char **sei_buffer)
 324 {
 325     unsigned char *byte_buf;
 326     int byte_size, i;
 327
 328     avc_bitstream nal_bs;
 329     avc_bitstream sei_bs;
 330
 331     avc_bitstream_start(&sei_bs);
 332     avc_bitstream_put_ue(&sei_bs, 0);       /*seq_parameter_set_id*/
 333     avc_bitstream_put_ui(&sei_bs, init_cpb_removal_delay, cpb_removal_length);
 334     avc_bitstream_put_ui(&sei_bs, init_cpb_removal_delay_offset, cpb_removal_length);
 335     if ( sei_bs.bit_offset & 0x7) {
 336         avc_bitstream_put_ui(&sei_bs, 1, 1);
 337     }
 338     avc_bitstream_end(&sei_bs);
 339     byte_size = (sei_bs.bit_offset + 7) / 8;
 340
 341     avc_bitstream_start(&nal_bs);
 342     nal_start_code_prefix(&nal_bs);
 343     nal_header(&nal_bs, NAL_REF_IDC_NONE, NAL_SEI);
 344
 345     avc_bitstream_put_ui(&nal_bs, 0, 8);
 346     avc_bitstream_put_ui(&nal_bs, byte_size, 8);
 347
 348     byte_buf = (unsigned char *)sei_bs.buffer;
 349     for(i = 0; i < byte_size; i++) {
 350         avc_bitstream_put_ui(&nal_bs, byte_buf[i], 8);
 351     }
 352     free(byte_buf);
 353
 354     avc_rbsp_trailing_bits(&nal_bs);
 355     avc_bitstream_end(&nal_bs);
 356
 357     *sei_buffer = (unsigned char *)nal_bs.buffer;
 358
 359     return nal_bs.bit_offset;
 360 }
 361
 362 int
 363 build_avc_sei_pic_timing(unsigned int cpb_removal_length, unsigned int cpb_removal_delay,
 364                          unsigned int dpb_output_length, unsigned int dpb_output_delay,
 365                          unsigned char **sei_buffer)
 366 {
 367     unsigned char *byte_buf;
 368     int byte_size, i;
 369
 370     avc_bitstream nal_bs;
 371     avc_bitstream sei_bs;
 372
 373     avc_bitstream_start(&sei_bs);
 374     avc_bitstream_put_ui(&sei_bs, cpb_removal_delay, cpb_removal_length);
 375     avc_bitstream_put_ui(&sei_bs, dpb_output_delay, dpb_output_length);
 376     if ( sei_bs.bit_offset & 0x7) {
 377         avc_bitstream_put_ui(&sei_bs, 1, 1);
 378     }
 379     avc_bitstream_end(&sei_bs);
 380     byte_size = (sei_bs.bit_offset + 7) / 8;
 381
 382     avc_bitstream_start(&nal_bs);
 383     nal_start_code_prefix(&nal_bs);
 384     nal_header(&nal_bs, NAL_REF_IDC_NONE, NAL_SEI);
 385
 386     avc_bitstream_put_ui(&nal_bs, 0x01, 8);
 387     avc_bitstream_put_ui(&nal_bs, byte_size, 8);
 388
 389     byte_buf = (unsigned char *)sei_bs.buffer;
 390     for(i = 0; i < byte_size; i++) {
 391         avc_bitstream_put_ui(&nal_bs, byte_buf[i], 8);
 392     }
 393     free(byte_buf);
 394
 395     avc_rbsp_trailing_bits(&nal_bs);
 396     avc_bitstream_end(&nal_bs);
 397
 398     *sei_buffer = (unsigned char *)nal_bs.buffer;
 399
 400     return nal_bs.bit_offset;
 401 }
 402
 403
 404 int
 405 build_avc_sei_buffer_timing(unsigned int init_cpb_removal_length,
 406                                 unsigned int init_cpb_removal_delay,
 407                                 unsigned int init_cpb_removal_delay_offset,
 408                                 unsigned int cpb_removal_length,
 409                                 unsigned int cpb_removal_delay,
 410                                 unsigned int dpb_output_length,
 411                                 unsigned int dpb_output_delay,
 412                                 unsigned char **sei_buffer)
 413 {
 414     unsigned char *byte_buf;
 415     int bp_byte_size, i, pic_byte_size;
 416
 417     avc_bitstream nal_bs;
 418     avc_bitstream sei_bp_bs, sei_pic_bs;
 419
 420     avc_bitstream_start(&sei_bp_bs);
 421     avc_bitstream_put_ue(&sei_bp_bs, 0);       /*seq_parameter_set_id*/
 422     avc_bitstream_put_ui(&sei_bp_bs, init_cpb_removal_delay, cpb_removal_length);
 423     avc_bitstream_put_ui(&sei_bp_bs, init_cpb_removal_delay_offset, cpb_removal_length);
 424     if ( sei_bp_bs.bit_offset & 0x7) {
 425         avc_bitstream_put_ui(&sei_bp_bs, 1, 1);
 426     }
 427     avc_bitstream_end(&sei_bp_bs);
 428     bp_byte_size = (sei_bp_bs.bit_offset + 7) / 8;
 429
 430     avc_bitstream_start(&sei_pic_bs);
 431     avc_bitstream_put_ui(&sei_pic_bs, cpb_removal_delay, cpb_removal_length);
 432     avc_bitstream_put_ui(&sei_pic_bs, dpb_output_delay, dpb_output_length);
 433     if ( sei_pic_bs.bit_offset & 0x7) {
 434         avc_bitstream_put_ui(&sei_pic_bs, 1, 1);
 435     }
 436     avc_bitstream_end(&sei_pic_bs);
 437     pic_byte_size = (sei_pic_bs.bit_offset + 7) / 8;
 438
 439     avc_bitstream_start(&nal_bs);
 440     nal_start_code_prefix(&nal_bs);
 441     nal_header(&nal_bs, NAL_REF_IDC_NONE, NAL_SEI);
 442
 443         /* Write the SEI buffer period data */
 444     avc_bitstream_put_ui(&nal_bs, 0, 8);
 445     avc_bitstream_put_ui(&nal_bs, bp_byte_size, 8);
 446
 447     byte_buf = (unsigned char *)sei_bp_bs.buffer;
 448     for(i = 0; i < bp_byte_size; i++) {
 449         avc_bitstream_put_ui(&nal_bs, byte_buf[i], 8);
 450     }
 451     free(byte_buf);
 452         /* write the SEI timing data */
 453     avc_bitstream_put_ui(&nal_bs, 0x01, 8);
 454     avc_bitstream_put_ui(&nal_bs, pic_byte_size, 8);
 455
 456     byte_buf = (unsigned char *)sei_pic_bs.buffer;
 457     for(i = 0; i < pic_byte_size; i++) {
 458         avc_bitstream_put_ui(&nal_bs, byte_buf[i], 8);
 459     }
 460     free(byte_buf);
 461
 462     avc_rbsp_trailing_bits(&nal_bs);
 463     avc_bitstream_end(&nal_bs);
 464
 465     *sei_buffer = (unsigned char *)nal_bs.buffer;
 466
 467     return nal_bs.bit_offset;
 468 }
 469
 470 int
 471 build_mpeg2_slice_header(VAEncSequenceParameterBufferMPEG2 *sps_param,
 472                          VAEncPictureParameterBufferMPEG2 *pic_param,
 473                          VAEncSliceParameterBufferMPEG2 *slice_param,
 474                          unsigned char **slice_header_buffer)
 475 {
 476     avc_bitstream bs;
 477
 478     avc_bitstream_start(&bs);
 479     avc_bitstream_end(&bs);
 480     *slice_header_buffer = (unsigned char *)bs.buffer;
 481
 482     return bs.bit_offset;
 483 }