test/encode/avcenc.c

   1 /*
   2  * Copyright (c) 2012 Intel Corporation. All Rights Reserved.
   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  * Simple AVC encoder based on libVA.
  26  *
  27  * Usage:
  28  * ./avcenc <width> <height> <input file> <output file> [qp]
  29  */
  30
  31 #include "sysdeps.h"
  32 #include <stdio.h>
  33 #include <string.h>
  34 #include <stdlib.h>
  35 #include <getopt.h>
  36 #include <unistd.h>
  37
  38 #include <sys/time.h>
  39 #include <sys/types.h>
  40 #include <sys/stat.h>
  41 #include <fcntl.h>
  42 #include <assert.h>
  43 #include <time.h>
  44
  45 #include <pthread.h>
  46
  47 #include <va/va.h>
  48 #include <va/va_enc_h264.h>
  49 #include "va_display.h"
  50
  51 #define NAL_REF_IDC_NONE        0
  52 #define NAL_REF_IDC_LOW         1
  53 #define NAL_REF_IDC_MEDIUM      2
  54 #define NAL_REF_IDC_HIGH        3
  55
  56 #define NAL_NON_IDR             1
  57 #define NAL_IDR                 5
  58 #define NAL_SPS                 7
  59 #define NAL_PPS                 8
  60 #define NAL_SEI                 6
  61
  62 #define SLICE_TYPE_P            0
  63 #define SLICE_TYPE_B            1
  64 #define SLICE_TYPE_I            2
  65
  66 #define FRAME_IDR 7
  67
  68 #define ENTROPY_MODE_CAVLC      0
  69 #define ENTROPY_MODE_CABAC      1
  70
  71 #define PROFILE_IDC_BASELINE    66
  72 #define PROFILE_IDC_MAIN        77
  73 #define PROFILE_IDC_HIGH        100
  74
  75 #define CHECK_VASTATUS(va_status,func)                                  \
  76     if (va_status != VA_STATUS_SUCCESS) {                               \
  77         fprintf(stderr,"%s:%s (%d) failed,exit\n", __func__, func, __LINE__); \
  78         exit(1);                                                        \
  79     }
  80
  81 static VADisplay va_dpy;
  82
  83 static int picture_width, picture_width_in_mbs;
  84 static int picture_height, picture_height_in_mbs;
  85 static int frame_size;
  86 static unsigned char *newImageBuffer = 0;
  87
  88 static int qp_value = 26;
  89
  90 static int intra_period = 30;
  91 static int frame_bit_rate = -1;
  92 static int frame_rate = 30;
  93 static int ip_period = 1;
  94
  95 static VAEntrypoint select_entrypoint = VAEntrypointEncSlice;
  96
  97 #define MAX_SLICES      32
  98
  99
 100 static  unsigned int MaxFrameNum = (1<<12);
 101 static  unsigned int Log2MaxFrameNum = 12;
 102 static  unsigned int Log2MaxPicOrderCntLsb = 8;
 103
 104 static const struct option longopts[] = {
 105     {"qp", required_argument, 0, 1},
 106     {"fb", required_argument, 0, 2},
 107     {"mode", required_argument, 0, 3},
 108     {"low-power", no_argument, 0, 4},
 109     { NULL, 0, NULL, 0}
 110 };
 111
 112 static int
 113 build_packed_pic_buffer(unsigned char **header_buffer);
 114
 115 static int
 116 build_packed_seq_buffer(unsigned char **header_buffer);
 117
 118 static int
 119 build_packed_sei_pic_timing(unsigned int cpb_removal_length,
 120                                 unsigned int dpb_output_length,
 121                                 unsigned char **sei_buffer);
 122
 123 static int
 124 build_packed_idr_sei_buffer_timing(unsigned int init_cpb_removal_delay_length,
 125                                 unsigned int cpb_removal_length,
 126                                 unsigned int dpb_output_length,
 127                                 unsigned char **sei_buffer);
 128
 129 struct upload_thread_param
 130 {
 131     FILE *yuv_fp;
 132     VASurfaceID surface_id;
 133 };
 134
 135 static void
 136 upload_yuv_to_surface(FILE *yuv_fp, VASurfaceID surface_id);
 137
 138 static struct {
 139     VAProfile profile;
 140     int constraint_set_flag;
 141     VAEncSequenceParameterBufferH264 seq_param;
 142     VAEncPictureParameterBufferH264 pic_param;
 143     VAEncSliceParameterBufferH264 slice_param[MAX_SLICES];
 144     VAContextID context_id;
 145     VAConfigID config_id;
 146     VABufferID seq_param_buf_id;                /* Sequence level parameter */
 147     VABufferID pic_param_buf_id;                /* Picture level parameter */
 148     VABufferID slice_param_buf_id[MAX_SLICES];  /* Slice level parameter, multil slices */
 149     VABufferID codedbuf_buf_id;                 /* Output buffer, compressed data */
 150     VABufferID packed_seq_header_param_buf_id;
 151     VABufferID packed_seq_buf_id;
 152     VABufferID packed_pic_header_param_buf_id;
 153     VABufferID packed_pic_buf_id;
 154     VABufferID packed_sei_header_param_buf_id;   /* the SEI buffer */
 155     VABufferID packed_sei_buf_id;
 156     VABufferID misc_parameter_hrd_buf_id;
 157
 158     int num_slices;
 159     int codedbuf_i_size;
 160     int codedbuf_pb_size;
 161     int current_input_surface;
 162     int rate_control_method;
 163     struct upload_thread_param upload_thread_param;
 164     pthread_t upload_thread_id;
 165     int upload_thread_value;
 166     int i_initial_cpb_removal_delay;
 167     int i_initial_cpb_removal_delay_offset;
 168     int i_initial_cpb_removal_delay_length;
 169     int i_cpb_removal_delay;
 170     int i_cpb_removal_delay_length;
 171     int i_dpb_output_delay_length;
 172     int time_offset_length;
 173
 174     unsigned long long idr_frame_num;
 175     unsigned long long prev_idr_cpb_removal;
 176     unsigned long long current_idr_cpb_removal;
 177     unsigned long long current_cpb_removal;
 178     /* This is relative to the current_cpb_removal */
 179     unsigned int current_dpb_removal_delta;
 180 } avcenc_context;
 181
 182 static  VAPictureH264 ReferenceFrames[16], RefPicList0[32], RefPicList1[32];
 183
 184 static void create_encode_pipe()
 185 {
 186     VAEntrypoint entrypoints[5];
 187     int num_entrypoints,slice_entrypoint;
 188     VAConfigAttrib attrib[2];
 189     int major_ver, minor_ver;
 190     VAStatus va_status;
 191
 192     va_dpy = va_open_display();
 193     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 194     CHECK_VASTATUS(va_status, "vaInitialize");
 195
 196     vaQueryConfigEntrypoints(va_dpy, avcenc_context.profile, entrypoints,
 197                              &num_entrypoints);
 198
 199     for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 200         if (entrypoints[slice_entrypoint] == select_entrypoint)
 201             break;
 202     }
 203
 204     if (slice_entrypoint == num_entrypoints) {
 205         /* not find Slice entry point */
 206         assert(0);
 207     }
 208
 209     /* find out the format for the render target, and rate control mode */
 210     attrib[0].type = VAConfigAttribRTFormat;
 211     attrib[1].type = VAConfigAttribRateControl;
 212     vaGetConfigAttributes(va_dpy, avcenc_context.profile, select_entrypoint,
 213                           &attrib[0], 2);
 214
 215     if ((attrib[0].value & VA_RT_FORMAT_YUV420) == 0) {
 216         /* not find desired YUV420 RT format */
 217         assert(0);
 218     }
 219
 220     if ((attrib[1].value & avcenc_context.rate_control_method) == 0) {
 221         /* Can't find matched RC mode */
 222         printf("Can't find the desired RC mode, exit\n");
 223         assert(0);
 224     }
 225
 226     attrib[0].value = VA_RT_FORMAT_YUV420; /* set to desired RT format */
 227     attrib[1].value = avcenc_context.rate_control_method; /* set to desired RC mode */
 228
 229     va_status = vaCreateConfig(va_dpy, avcenc_context.profile, select_entrypoint,
 230                                &attrib[0], 2,&avcenc_context.config_id);
 231     CHECK_VASTATUS(va_status, "vaCreateConfig");
 232
 233     /* Create a context for this decode pipe */
 234     va_status = vaCreateContext(va_dpy, avcenc_context.config_id,
 235                                 picture_width, picture_height,
 236                                 VA_PROGRESSIVE,
 237                                 0, 0,
 238                                 &avcenc_context.context_id);
 239     CHECK_VASTATUS(va_status, "vaCreateContext");
 240 }
 241
 242 static void destory_encode_pipe()
 243 {
 244     vaDestroyContext(va_dpy,avcenc_context.context_id);
 245     vaDestroyConfig(va_dpy,avcenc_context.config_id);
 246     vaTerminate(va_dpy);
 247     va_close_display(va_dpy);
 248 }
 249
 250 /***************************************************
 251  *
 252  *  The encode pipe resource define
 253  *
 254  ***************************************************/
 255 #define SID_INPUT_PICTURE_0                     0
 256 #define SID_INPUT_PICTURE_1                     1
 257 #define SID_REFERENCE_PICTURE_L0                2
 258 #define SID_REFERENCE_PICTURE_L1                3
 259 #define SID_RECON_PICTURE                       4
 260 #define SID_NUMBER                              SID_RECON_PICTURE + 1
 261
 262 #define SURFACE_NUM 16 /* 16 surfaces for reference */
 263
 264 static  VASurfaceID surface_ids[SID_NUMBER];
 265 static  VASurfaceID ref_surface[SURFACE_NUM];
 266 static  int use_slot[SURFACE_NUM];
 267
 268 static  unsigned long long current_frame_display = 0;
 269 static  unsigned long long current_IDR_display = 0;
 270
 271 static  VAPictureH264 CurrentCurrPic;
 272
 273 #define current_slot (current_frame_display % SURFACE_NUM)
 274
 275 static int frame_number;
 276 static unsigned long long enc_frame_number;
 277 static int current_frame_type;
 278 static int current_frame_num;
 279 static unsigned int current_poc;
 280
 281 static  unsigned int num_ref_frames = 2;
 282 static  unsigned int numShortTerm = 0;
 283 /***************************************************/
 284
 285 static int get_free_slot()
 286 {
 287     int i, index = -1;
 288
 289     for (i = 0; i < SURFACE_NUM; i++) {
 290         if (use_slot[i] == 0) {
 291             index = i;
 292             break;
 293         }
 294     }
 295     if (index < 0) {
 296         printf("WARNING: No free slot to store the reconstructed frame \n");
 297         index = SURFACE_NUM - 1;
 298     }
 299     return index;
 300 }
 301
 302 static void *
 303 upload_thread_function(void *data)
 304 {
 305     struct upload_thread_param *param = data;
 306
 307     upload_yuv_to_surface(param->yuv_fp, param->surface_id);
 308
 309     return NULL;
 310 }
 311
 312 static void alloc_encode_resource(FILE *yuv_fp)
 313 {
 314     VAStatus va_status;
 315
 316     // Create surface
 317     va_status = vaCreateSurfaces(
 318         va_dpy,
 319         VA_RT_FORMAT_YUV420, picture_width, picture_height,
 320         surface_ids, SID_NUMBER,
 321         NULL, 0
 322     );
 323
 324     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 325
 326     // Create surface
 327     va_status = vaCreateSurfaces(
 328         va_dpy,
 329         VA_RT_FORMAT_YUV420, picture_width, picture_height,
 330         ref_surface, SURFACE_NUM,
 331         NULL, 0
 332     );
 333
 334     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 335
 336
 337     newImageBuffer = (unsigned char *)malloc(frame_size);
 338
 339     /* firstly upload YUV data to SID_INPUT_PICTURE_1 */
 340     avcenc_context.upload_thread_param.yuv_fp = yuv_fp;
 341     avcenc_context.upload_thread_param.surface_id = surface_ids[SID_INPUT_PICTURE_1];
 342
 343     avcenc_context.upload_thread_value = pthread_create(&avcenc_context.upload_thread_id,
 344                                                         NULL,
 345                                                         upload_thread_function,
 346                                                         (void*)&avcenc_context.upload_thread_param);
 347 }
 348
 349 static void release_encode_resource()
 350 {
 351     pthread_join(avcenc_context.upload_thread_id, NULL);
 352     free(newImageBuffer);
 353
 354     // Release all the surfaces resource
 355     vaDestroySurfaces(va_dpy, surface_ids, SID_NUMBER);
 356     // Release all the reference surfaces
 357     vaDestroySurfaces(va_dpy, ref_surface, SURFACE_NUM);
 358 }
 359
 360 static void avcenc_update_sei_param(int is_idr)
 361 {
 362         VAEncPackedHeaderParameterBuffer packed_header_param_buffer;
 363         unsigned int length_in_bits;
 364         unsigned char *packed_sei_buffer = NULL;
 365         VAStatus va_status;
 366
 367         if (is_idr)
 368             length_in_bits = build_packed_idr_sei_buffer_timing(
 369                                 avcenc_context.i_initial_cpb_removal_delay_length,
 370                                 avcenc_context.i_cpb_removal_delay_length,
 371                                 avcenc_context.i_dpb_output_delay_length,
 372                                 &packed_sei_buffer);
 373        else
 374             length_in_bits = build_packed_sei_pic_timing(
 375                                 avcenc_context.i_cpb_removal_delay_length,
 376                                 avcenc_context.i_dpb_output_delay_length,
 377                                 &packed_sei_buffer);
 378
 379         packed_header_param_buffer.type = VAEncPackedHeaderH264_SEI;
 380         packed_header_param_buffer.bit_length = length_in_bits;
 381         packed_header_param_buffer.has_emulation_bytes = 0;
 382
 383         va_status = vaCreateBuffer(va_dpy,
 384                                 avcenc_context.context_id,
 385                                 VAEncPackedHeaderParameterBufferType,
 386                                 sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
 387                                 &avcenc_context.packed_sei_header_param_buf_id);
 388         CHECK_VASTATUS(va_status,"vaCreateBuffer");
 389
 390         va_status = vaCreateBuffer(va_dpy,
 391                                 avcenc_context.context_id,
 392                                 VAEncPackedHeaderDataBufferType,
 393                                 (length_in_bits + 7) / 8, 1, packed_sei_buffer,
 394                                 &avcenc_context.packed_sei_buf_id);
 395         CHECK_VASTATUS(va_status,"vaCreateBuffer");
 396         free(packed_sei_buffer);
 397         return;
 398 }
 399
 400 #define partition(ref, field, key, ascending)   \
 401     while (i <= j) {                            \
 402         if (ascending) {                        \
 403             while (ref[i].field < key)          \
 404                 i++;                            \
 405             while (ref[j].field > key)          \
 406                 j--;                            \
 407         } else {                                \
 408             while (ref[i].field > key)          \
 409                 i++;                            \
 410             while (ref[j].field < key)          \
 411                 j--;                            \
 412         }                                       \
 413         if (i <= j) {                           \
 414             tmp = ref[i];                       \
 415             ref[i] = ref[j];                    \
 416             ref[j] = tmp;                       \
 417             i++;                                \
 418             j--;                                \
 419         }                                       \
 420     }                                           \
 421
 422 static void sort_one(VAPictureH264 ref[], int left, int right,
 423                      int ascending, int frame_idx)
 424 {
 425     int i = left, j = right;
 426     unsigned int key;
 427     VAPictureH264 tmp;
 428
 429     if (frame_idx) {
 430         key = ref[(left + right) / 2].frame_idx;
 431         partition(ref, frame_idx, key, ascending);
 432     } else {
 433         key = ref[(left + right) / 2].TopFieldOrderCnt;
 434         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
 435     }
 436
 437     /* recursion */
 438     if (left < j)
 439         sort_one(ref, left, j, ascending, frame_idx);
 440
 441     if (i < right)
 442         sort_one(ref, i, right, ascending, frame_idx);
 443 }
 444
 445 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
 446                      int partition_ascending, int list0_ascending, int list1_ascending)
 447 {
 448     int i = left, j = right;
 449     VAPictureH264 tmp;
 450
 451     if (frame_idx) {
 452         partition(ref, frame_idx, key, partition_ascending);
 453     } else {
 454         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
 455     }
 456
 457     sort_one(ref, left, i-1, list0_ascending, frame_idx);
 458     sort_one(ref, j+1, right, list1_ascending, frame_idx);
 459 }
 460
 461 static int update_RefPicList()
 462 {
 463
 464     if (current_frame_type == SLICE_TYPE_P) {
 465         memcpy(RefPicList0, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
 466         sort_one(RefPicList0, 0, numShortTerm-1, 0, 1);
 467     }
 468
 469     if (current_frame_type == SLICE_TYPE_B) {
 470         memcpy(RefPicList0, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
 471         sort_two(RefPicList0, 0, numShortTerm-1, current_poc, 0,
 472                  1, 0, 1);
 473
 474         memcpy(RefPicList1, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
 475         sort_two(RefPicList1, 0, numShortTerm-1, current_poc, 0,
 476                  0, 1, 0);
 477     }
 478
 479     return 0;
 480 }
 481
 482 static void avcenc_update_picture_parameter(int slice_type, int is_idr)
 483 {
 484     VAEncPictureParameterBufferH264 *pic_param;
 485     VAStatus va_status;
 486     int recon_index;
 487
 488     recon_index = get_free_slot();
 489     // Picture level
 490     pic_param = &avcenc_context.pic_param;
 491
 492     pic_param->CurrPic.picture_id = ref_surface[recon_index];
 493     pic_param->CurrPic.frame_idx = current_frame_num;
 494     pic_param->CurrPic.flags = 0;
 495
 496     pic_param->CurrPic.TopFieldOrderCnt = current_poc;
 497     pic_param->CurrPic.BottomFieldOrderCnt = pic_param->CurrPic.TopFieldOrderCnt;
 498
 499     assert(avcenc_context.codedbuf_buf_id != VA_INVALID_ID);
 500     pic_param->coded_buf = avcenc_context.codedbuf_buf_id;
 501     pic_param->frame_num = current_frame_num;
 502     pic_param->pic_fields.bits.idr_pic_flag = !!is_idr;
 503     pic_param->pic_fields.bits.reference_pic_flag = (slice_type != SLICE_TYPE_B);
 504     CurrentCurrPic = pic_param->CurrPic;
 505
 506     if (slice_type == SLICE_TYPE_P || slice_type == SLICE_TYPE_B)
 507         memset(pic_param->ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
 508
 509     if ((slice_type == SLICE_TYPE_P) || (slice_type == SLICE_TYPE_B)) {
 510         pic_param->ReferenceFrames[0] = RefPicList0[0];
 511     }
 512     if (slice_type == SLICE_TYPE_B) {
 513         pic_param->ReferenceFrames[1] = RefPicList1[0];
 514     }
 515
 516     va_status = vaCreateBuffer(va_dpy,
 517                                avcenc_context.context_id,
 518                                VAEncPictureParameterBufferType,
 519                                sizeof(*pic_param), 1, pic_param,
 520                                &avcenc_context.pic_param_buf_id);
 521     CHECK_VASTATUS(va_status,"vaCreateBuffer");
 522
 523 }
 524
 525 #ifndef VA_FOURCC_I420
 526 #define VA_FOURCC_I420          0x30323449
 527 #endif
 528
 529 static void upload_yuv_to_surface(FILE *yuv_fp, VASurfaceID surface_id)
 530 {
 531     VAImage surface_image;
 532     VAStatus va_status;
 533     void *surface_p = NULL;
 534     unsigned char *y_src, *u_src, *v_src;
 535     unsigned char *y_dst, *u_dst, *v_dst;
 536     int y_size = picture_width * picture_height;
 537     int u_size = (picture_width >> 1) * (picture_height >> 1);
 538     int row, col;
 539     size_t n_items;
 540
 541     do {
 542         n_items = fread(newImageBuffer, frame_size, 1, yuv_fp);
 543     } while (n_items != 1);
 544
 545     va_status = vaDeriveImage(va_dpy, surface_id, &surface_image);
 546     CHECK_VASTATUS(va_status,"vaDeriveImage");
 547
 548     vaMapBuffer(va_dpy, surface_image.buf, &surface_p);
 549     assert(VA_STATUS_SUCCESS == va_status);
 550
 551     y_src = newImageBuffer;
 552     u_src = newImageBuffer + y_size; /* UV offset for NV12 */
 553     v_src = newImageBuffer + y_size + u_size;
 554
 555     y_dst = surface_p + surface_image.offsets[0];
 556     u_dst = surface_p + surface_image.offsets[1]; /* UV offset for NV12 */
 557     v_dst = surface_p + surface_image.offsets[2];
 558
 559     /* Y plane */
 560     for (row = 0; row < surface_image.height; row++) {
 561         memcpy(y_dst, y_src, surface_image.width);
 562         y_dst += surface_image.pitches[0];
 563         y_src += picture_width;
 564     }
 565
 566     if (surface_image.format.fourcc == VA_FOURCC_NV12) { /* UV plane */
 567         for (row = 0; row < surface_image.height / 2; row++) {
 568             for (col = 0; col < surface_image.width / 2; col++) {
 569                 u_dst[col * 2] = u_src[col];
 570                 u_dst[col * 2 + 1] = v_src[col];
 571             }
 572
 573             u_dst += surface_image.pitches[1];
 574             u_src += (picture_width / 2);
 575             v_src += (picture_width / 2);
 576         }
 577     } else if (surface_image.format.fourcc == VA_FOURCC_YV12 ||
 578                surface_image.format.fourcc == VA_FOURCC_I420) {
 579         const int U = surface_image.format.fourcc == VA_FOURCC_I420 ? 1 : 2;
 580         const int V = surface_image.format.fourcc == VA_FOURCC_I420 ? 2 : 1;
 581
 582         u_dst = surface_p + surface_image.offsets[U];
 583         v_dst = surface_p + surface_image.offsets[V];
 584
 585         for (row = 0; row < surface_image.height / 2; row++) {
 586             memcpy(u_dst, u_src, surface_image.width / 2);
 587             memcpy(v_dst, v_src, surface_image.width / 2);
 588             u_dst += surface_image.pitches[U];
 589             v_dst += surface_image.pitches[V];
 590             u_src += (picture_width / 2);
 591             v_src += (picture_width / 2);
 592         }
 593     }
 594
 595     vaUnmapBuffer(va_dpy, surface_image.buf);
 596     vaDestroyImage(va_dpy, surface_image.image_id);
 597 }
 598
 599 static void avcenc_update_slice_parameter(int slice_type)
 600 {
 601     VAEncSliceParameterBufferH264 *slice_param;
 602     VAStatus va_status;
 603     int i;
 604
 605     // Slice level
 606     i = 0;
 607     slice_param = &avcenc_context.slice_param[i];
 608     slice_param->macroblock_address = 0;
 609     slice_param->num_macroblocks = picture_height_in_mbs * picture_width_in_mbs;
 610     slice_param->pic_parameter_set_id = 0;
 611     slice_param->slice_type = slice_type;
 612     slice_param->direct_spatial_mv_pred_flag = 0;
 613     slice_param->num_ref_idx_l0_active_minus1 = 0;      /* FIXME: ??? */
 614     slice_param->num_ref_idx_l1_active_minus1 = 0;
 615     slice_param->cabac_init_idc = 0;
 616     slice_param->slice_qp_delta = 0;
 617     slice_param->disable_deblocking_filter_idc = 0;
 618     slice_param->slice_alpha_c0_offset_div2 = 2;
 619     slice_param->slice_beta_offset_div2 = 2;
 620     slice_param->idr_pic_id = 0;
 621
 622     /* FIXME: fill other fields */
 623     if ((slice_type == SLICE_TYPE_P) || (slice_type == SLICE_TYPE_B)) {
 624         memset(slice_param->RefPicList0, 0xFF, 32 * sizeof(VAPictureH264));
 625         slice_param->RefPicList0[0] = RefPicList0[0];
 626     }
 627
 628     if ((slice_type == SLICE_TYPE_B)) {
 629         memset(slice_param->RefPicList1, 0xFF, 32 * sizeof(VAPictureH264));
 630         slice_param->RefPicList1[0] = RefPicList1[0];
 631     }
 632
 633     va_status = vaCreateBuffer(va_dpy,
 634                                avcenc_context.context_id,
 635                                VAEncSliceParameterBufferType,
 636                                sizeof(*slice_param), 1, slice_param,
 637                                &avcenc_context.slice_param_buf_id[i]);
 638     CHECK_VASTATUS(va_status,"vaCreateBuffer");;
 639     i++;
 640
 641 #if 0
 642     slice_param = &avcenc_context.slice_param[i];
 643     slice_param->macroblock_address = picture_height_in_mbs * picture_width_in_mbs / 2;
 644     slice_param->num_macroblocks = picture_height_in_mbs * picture_width_in_mbs / 2;
 645     slice_param->pic_parameter_set_id = 0;
 646     slice_param->slice_type = slice_type;
 647     slice_param->direct_spatial_mv_pred_flag = 0;
 648     slice_param->num_ref_idx_l0_active_minus1 = 0;      /* FIXME: ??? */
 649     slice_param->num_ref_idx_l1_active_minus1 = 0;
 650     slice_param->cabac_init_idc = 0;
 651     slice_param->slice_qp_delta = 0;
 652     slice_param->disable_deblocking_filter_idc = 0;
 653     slice_param->slice_alpha_c0_offset_div2 = 2;
 654     slice_param->slice_beta_offset_div2 = 2;
 655     slice_param->idr_pic_id = 0;
 656
 657     /* FIXME: fill other fields */
 658
 659     va_status = vaCreateBuffer(va_dpy,
 660                                avcenc_context.context_id,
 661                                VAEncSliceParameterBufferType,
 662                                sizeof(*slice_param), 1, slice_param,
 663                                &avcenc_context.slice_param_buf_id[i]);
 664     CHECK_VASTATUS(va_status,"vaCreateBuffer");;
 665     i++;
 666 #endif
 667
 668     avcenc_context.num_slices = i;
 669 }
 670
 671 static int update_ReferenceFrames(void)
 672 {
 673     int i;
 674     /* B-frame is not used for reference */
 675     if (current_frame_type == SLICE_TYPE_B)
 676         return 0;
 677
 678     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
 679     numShortTerm++;
 680     if (numShortTerm > num_ref_frames)
 681         numShortTerm = num_ref_frames;
 682     for (i=numShortTerm-1; i>0; i--)
 683         ReferenceFrames[i] = ReferenceFrames[i-1];
 684     ReferenceFrames[0] = CurrentCurrPic;
 685
 686     if (current_frame_type != SLICE_TYPE_B)
 687         current_frame_num++;
 688     if (current_frame_num > MaxFrameNum)
 689         current_frame_num = 0;
 690
 691     /* Update the use_slot. Only when the surface is used in reference
 692      * frame list, the use_slot[index] is set
 693      */
 694     for (i = 0; i < SURFACE_NUM; i++) {
 695         int j;
 696         bool found;
 697
 698         found = false;
 699         for (j = 0; j < numShortTerm; j++) {
 700             if (ref_surface[i] == ReferenceFrames[j].picture_id) {
 701                 found = true;
 702                 break;
 703             }
 704         }
 705         if (found)
 706             use_slot[i] = 1;
 707         else
 708             use_slot[i] = 0;
 709     }
 710
 711     return 0;
 712 }
 713
 714 static int begin_picture(FILE *yuv_fp, int frame_num, int display_num, int slice_type, int is_idr)
 715 {
 716     VAStatus va_status;
 717
 718     if (avcenc_context.upload_thread_value != 0) {
 719         fprintf(stderr, "FATAL error!!!\n");
 720         exit(1);
 721     }
 722
 723     pthread_join(avcenc_context.upload_thread_id, NULL);
 724
 725     avcenc_context.upload_thread_value = -1;
 726
 727     if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_0)
 728         avcenc_context.current_input_surface = SID_INPUT_PICTURE_1;
 729     else
 730         avcenc_context.current_input_surface = SID_INPUT_PICTURE_0;
 731
 732     if (is_idr) {
 733         VAEncPackedHeaderParameterBuffer packed_header_param_buffer;
 734         unsigned int length_in_bits;
 735         unsigned char *packed_seq_buffer = NULL, *packed_pic_buffer = NULL;
 736
 737         assert(slice_type == SLICE_TYPE_I);
 738         length_in_bits = build_packed_seq_buffer(&packed_seq_buffer);
 739         packed_header_param_buffer.type = VAEncPackedHeaderSequence;
 740         packed_header_param_buffer.bit_length = length_in_bits;
 741         packed_header_param_buffer.has_emulation_bytes = 0;
 742         va_status = vaCreateBuffer(va_dpy,
 743                                    avcenc_context.context_id,
 744                                    VAEncPackedHeaderParameterBufferType,
 745                                    sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
 746                                    &avcenc_context.packed_seq_header_param_buf_id);
 747         CHECK_VASTATUS(va_status,"vaCreateBuffer");
 748
 749         va_status = vaCreateBuffer(va_dpy,
 750                                    avcenc_context.context_id,
 751                                    VAEncPackedHeaderDataBufferType,
 752                                    (length_in_bits + 7) / 8, 1, packed_seq_buffer,
 753                                    &avcenc_context.packed_seq_buf_id);
 754         CHECK_VASTATUS(va_status,"vaCreateBuffer");
 755
 756         length_in_bits = build_packed_pic_buffer(&packed_pic_buffer);
 757         packed_header_param_buffer.type = VAEncPackedHeaderPicture;
 758         packed_header_param_buffer.bit_length = length_in_bits;
 759         packed_header_param_buffer.has_emulation_bytes = 0;
 760
 761         va_status = vaCreateBuffer(va_dpy,
 762                                    avcenc_context.context_id,
 763                                    VAEncPackedHeaderParameterBufferType,
 764                                    sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
 765                                    &avcenc_context.packed_pic_header_param_buf_id);
 766         CHECK_VASTATUS(va_status,"vaCreateBuffer");
 767
 768         va_status = vaCreateBuffer(va_dpy,
 769                                    avcenc_context.context_id,
 770                                    VAEncPackedHeaderDataBufferType,
 771                                    (length_in_bits + 7) / 8, 1, packed_pic_buffer,
 772                                    &avcenc_context.packed_pic_buf_id);
 773         CHECK_VASTATUS(va_status,"vaCreateBuffer");
 774
 775         free(packed_seq_buffer);
 776         free(packed_pic_buffer);
 777     }
 778
 779     /* sequence parameter set */
 780     VAEncSequenceParameterBufferH264 *seq_param = &avcenc_context.seq_param;
 781     va_status = vaCreateBuffer(va_dpy,
 782                                avcenc_context.context_id,
 783                                VAEncSequenceParameterBufferType,
 784                                sizeof(*seq_param), 1, seq_param,
 785                                &avcenc_context.seq_param_buf_id);
 786     CHECK_VASTATUS(va_status,"vaCreateBuffer");
 787
 788
 789     /* hrd parameter */
 790     VAEncMiscParameterBuffer *misc_param;
 791     VAEncMiscParameterHRD *misc_hrd_param;
 792     vaCreateBuffer(va_dpy,
 793                    avcenc_context.context_id,
 794                    VAEncMiscParameterBufferType,
 795                    sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
 796                    1,
 797                    NULL,
 798                    &avcenc_context.misc_parameter_hrd_buf_id);
 799     CHECK_VASTATUS(va_status, "vaCreateBuffer");
 800
 801     vaMapBuffer(va_dpy,
 802                 avcenc_context.misc_parameter_hrd_buf_id,
 803                 (void **)&misc_param);
 804     misc_param->type = VAEncMiscParameterTypeHRD;
 805     misc_hrd_param = (VAEncMiscParameterHRD *)misc_param->data;
 806
 807     if (frame_bit_rate > 0) {
 808         misc_hrd_param->initial_buffer_fullness = frame_bit_rate * 1000 * 4;
 809         misc_hrd_param->buffer_size = frame_bit_rate * 1000 * 8;
 810     } else {
 811         misc_hrd_param->initial_buffer_fullness = 0;
 812         misc_hrd_param->buffer_size = 0;
 813     }
 814
 815     vaUnmapBuffer(va_dpy, avcenc_context.misc_parameter_hrd_buf_id);
 816
 817     return 0;
 818 }
 819
 820 int avcenc_render_picture()
 821 {
 822     VAStatus va_status;
 823     VABufferID va_buffers[10];
 824     unsigned int num_va_buffers = 0;
 825     int i;
 826
 827     va_buffers[num_va_buffers++] = avcenc_context.seq_param_buf_id;
 828     va_buffers[num_va_buffers++] = avcenc_context.pic_param_buf_id;
 829
 830     if (avcenc_context.packed_seq_header_param_buf_id != VA_INVALID_ID)
 831         va_buffers[num_va_buffers++] = avcenc_context.packed_seq_header_param_buf_id;
 832
 833     if (avcenc_context.packed_seq_buf_id != VA_INVALID_ID)
 834         va_buffers[num_va_buffers++] = avcenc_context.packed_seq_buf_id;
 835
 836     if (avcenc_context.packed_pic_header_param_buf_id != VA_INVALID_ID)
 837         va_buffers[num_va_buffers++] = avcenc_context.packed_pic_header_param_buf_id;
 838
 839     if (avcenc_context.packed_pic_buf_id != VA_INVALID_ID)
 840         va_buffers[num_va_buffers++] = avcenc_context.packed_pic_buf_id;
 841
 842     if (avcenc_context.packed_sei_header_param_buf_id != VA_INVALID_ID)
 843         va_buffers[num_va_buffers++] = avcenc_context.packed_sei_header_param_buf_id;
 844
 845     if (avcenc_context.packed_sei_buf_id != VA_INVALID_ID)
 846         va_buffers[num_va_buffers++] = avcenc_context.packed_sei_buf_id;
 847
 848     if (avcenc_context.misc_parameter_hrd_buf_id != VA_INVALID_ID)
 849         va_buffers[num_va_buffers++] =  avcenc_context.misc_parameter_hrd_buf_id;
 850
 851     va_status = vaBeginPicture(va_dpy,
 852                                avcenc_context.context_id,
 853                                surface_ids[avcenc_context.current_input_surface]);
 854     CHECK_VASTATUS(va_status,"vaBeginPicture");
 855
 856     va_status = vaRenderPicture(va_dpy,
 857                                 avcenc_context.context_id,
 858                                 va_buffers,
 859                                 num_va_buffers);
 860     CHECK_VASTATUS(va_status,"vaRenderPicture");
 861
 862     for(i = 0; i < avcenc_context.num_slices; i++) {
 863         va_status = vaRenderPicture(va_dpy,
 864                                 avcenc_context.context_id,
 865                                 &avcenc_context.slice_param_buf_id[i],
 866                                 1);
 867         CHECK_VASTATUS(va_status,"vaRenderPicture");
 868     }
 869
 870     va_status = vaEndPicture(va_dpy, avcenc_context.context_id);
 871     CHECK_VASTATUS(va_status,"vaEndPicture");
 872
 873     return 0;
 874 }
 875
 876 static int avcenc_destroy_buffers(VABufferID *va_buffers, unsigned int num_va_buffers)
 877 {
 878     VAStatus va_status;
 879     unsigned int i;
 880
 881     for (i = 0; i < num_va_buffers; i++) {
 882         if (va_buffers[i] != VA_INVALID_ID) {
 883             va_status = vaDestroyBuffer(va_dpy, va_buffers[i]);
 884             CHECK_VASTATUS(va_status,"vaDestroyBuffer");
 885             va_buffers[i] = VA_INVALID_ID;
 886         }
 887     }
 888
 889     return 0;
 890 }
 891
 892 static void end_picture()
 893 {
 894
 895     update_ReferenceFrames();
 896     avcenc_destroy_buffers(&avcenc_context.seq_param_buf_id, 1);
 897     avcenc_destroy_buffers(&avcenc_context.pic_param_buf_id, 1);
 898     avcenc_destroy_buffers(&avcenc_context.packed_seq_header_param_buf_id, 1);
 899     avcenc_destroy_buffers(&avcenc_context.packed_seq_buf_id, 1);
 900     avcenc_destroy_buffers(&avcenc_context.packed_pic_header_param_buf_id, 1);
 901     avcenc_destroy_buffers(&avcenc_context.packed_pic_buf_id, 1);
 902     avcenc_destroy_buffers(&avcenc_context.packed_sei_header_param_buf_id, 1);
 903     avcenc_destroy_buffers(&avcenc_context.packed_sei_buf_id, 1);
 904     avcenc_destroy_buffers(&avcenc_context.slice_param_buf_id[0], avcenc_context.num_slices);
 905     avcenc_destroy_buffers(&avcenc_context.codedbuf_buf_id, 1);
 906     avcenc_destroy_buffers(&avcenc_context.misc_parameter_hrd_buf_id, 1);
 907
 908     memset(avcenc_context.slice_param, 0, sizeof(avcenc_context.slice_param));
 909     avcenc_context.num_slices = 0;
 910 }
 911
 912 #define BITSTREAM_ALLOCATE_STEPPING     4096
 913
 914 struct __bitstream {
 915     unsigned int *buffer;
 916     int bit_offset;
 917     int max_size_in_dword;
 918 };
 919
 920 typedef struct __bitstream bitstream;
 921
 922 #if 0
 923 static int
 924 get_coded_bitsteam_length(unsigned char *buffer, int buffer_length)
 925 {
 926     int i;
 927
 928     for (i = 0; i < buffer_length - 3; i++) {
 929         if (!buffer[i] &&
 930             !buffer[i + 1] &&
 931             !buffer[i + 2] &&
 932             !buffer[i + 3])
 933             break;
 934     }
 935
 936     return i;
 937 }
 938 #endif
 939
 940 static unsigned int
 941 va_swap32(unsigned int val)
 942 {
 943     unsigned char *pval = (unsigned char *)&val;
 944
 945     return ((pval[0] << 24)     |
 946             (pval[1] << 16)     |
 947             (pval[2] << 8)      |
 948             (pval[3] << 0));
 949 }
 950
 951 static void
 952 bitstream_start(bitstream *bs)
 953 {
 954     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 955     bs->buffer = calloc(bs->max_size_in_dword * sizeof(int), 1);
 956     assert(bs->buffer);
 957     bs->bit_offset = 0;
 958 }
 959
 960 static void
 961 bitstream_end(bitstream *bs)
 962 {
 963     int pos = (bs->bit_offset >> 5);
 964     int bit_offset = (bs->bit_offset & 0x1f);
 965     int bit_left = 32 - bit_offset;
 966
 967     if (bit_offset) {
 968         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 969     }
 970 }
 971
 972 static void
 973 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 974 {
 975     int pos = (bs->bit_offset >> 5);
 976     int bit_offset = (bs->bit_offset & 0x1f);
 977     int bit_left = 32 - bit_offset;
 978
 979     if (!size_in_bits)
 980         return;
 981
 982     bs->bit_offset += size_in_bits;
 983
 984     if (bit_left > size_in_bits) {
 985         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 986     } else {
 987         size_in_bits -= bit_left;
 988         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 989         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 990
 991         if (pos + 1 == bs->max_size_in_dword) {
 992             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 993             bs->buffer = realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 994             assert(bs->buffer);
 995         }
 996
 997         bs->buffer[pos + 1] = val;
 998     }
 999 }
1000
1001 static void
1002 bitstream_put_ue(bitstream *bs, unsigned int val)
1003 {
1004     int size_in_bits = 0;
1005     int tmp_val = ++val;
1006
1007     while (tmp_val) {
1008         tmp_val >>= 1;
1009         size_in_bits++;
1010     }
1011
1012     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
1013     bitstream_put_ui(bs, val, size_in_bits);
1014 }
1015
1016 static void
1017 bitstream_put_se(bitstream *bs, int val)
1018 {
1019     unsigned int new_val;
1020
1021     if (val <= 0)
1022         new_val = -2 * val;
1023     else
1024         new_val = 2 * val - 1;
1025
1026     bitstream_put_ue(bs, new_val);
1027 }
1028
1029 static void
1030 bitstream_byte_aligning(bitstream *bs, int bit)
1031 {
1032     int bit_offset = (bs->bit_offset & 0x7);
1033     int bit_left = 8 - bit_offset;
1034     int new_val;
1035
1036     if (!bit_offset)
1037         return;
1038
1039     assert(bit == 0 || bit == 1);
1040
1041     if (bit)
1042         new_val = (1 << bit_left) - 1;
1043     else
1044         new_val = 0;
1045
1046     bitstream_put_ui(bs, new_val, bit_left);
1047 }
1048
1049 static void
1050 rbsp_trailing_bits(bitstream *bs)
1051 {
1052     bitstream_put_ui(bs, 1, 1);
1053     bitstream_byte_aligning(bs, 0);
1054 }
1055
1056 static void nal_start_code_prefix(bitstream *bs)
1057 {
1058     bitstream_put_ui(bs, 0x00000001, 32);
1059 }
1060
1061 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
1062 {
1063     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
1064     bitstream_put_ui(bs, nal_ref_idc, 2);
1065     bitstream_put_ui(bs, nal_unit_type, 5);
1066 }
1067
1068 static void sps_rbsp(bitstream *bs)
1069 {
1070     VAEncSequenceParameterBufferH264 *seq_param = &avcenc_context.seq_param;
1071     int profile_idc = PROFILE_IDC_BASELINE;
1072
1073     if (avcenc_context.profile == VAProfileH264High)
1074         profile_idc = PROFILE_IDC_HIGH;
1075     else if (avcenc_context.profile == VAProfileH264Main)
1076         profile_idc = PROFILE_IDC_MAIN;
1077
1078     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
1079     bitstream_put_ui(bs, !!(avcenc_context.constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
1080     bitstream_put_ui(bs, !!(avcenc_context.constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
1081     bitstream_put_ui(bs, !!(avcenc_context.constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
1082     bitstream_put_ui(bs, !!(avcenc_context.constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
1083     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
1084     bitstream_put_ui(bs, seq_param->level_idc, 8);      /* level_idc */
1085     bitstream_put_ue(bs, seq_param->seq_parameter_set_id);      /* seq_parameter_set_id */
1086
1087     if ( profile_idc == PROFILE_IDC_HIGH) {
1088         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
1089         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
1090         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
1091         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
1092         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
1093     }
1094
1095     bitstream_put_ue(bs, seq_param->seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
1096     bitstream_put_ue(bs, seq_param->seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
1097
1098     if (seq_param->seq_fields.bits.pic_order_cnt_type == 0)
1099         bitstream_put_ue(bs, seq_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
1100     else {
1101         assert(0);
1102     }
1103
1104     bitstream_put_ue(bs, seq_param->max_num_ref_frames);        /* num_ref_frames */
1105     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
1106
1107     bitstream_put_ue(bs, seq_param->picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
1108     bitstream_put_ue(bs, seq_param->picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
1109     bitstream_put_ui(bs, seq_param->seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
1110
1111     if (!seq_param->seq_fields.bits.frame_mbs_only_flag) {
1112         assert(0);
1113     }
1114
1115     bitstream_put_ui(bs, seq_param->seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
1116     bitstream_put_ui(bs, seq_param->frame_cropping_flag, 1);            /* frame_cropping_flag */
1117
1118     if (seq_param->frame_cropping_flag) {
1119         bitstream_put_ue(bs, seq_param->frame_crop_left_offset);        /* frame_crop_left_offset */
1120         bitstream_put_ue(bs, seq_param->frame_crop_right_offset);       /* frame_crop_right_offset */
1121         bitstream_put_ue(bs, seq_param->frame_crop_top_offset);         /* frame_crop_top_offset */
1122         bitstream_put_ue(bs, seq_param->frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
1123     }
1124
1125     if ( frame_bit_rate < 0 ) {
1126         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
1127     } else {
1128         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
1129         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
1130         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
1131         bitstream_put_ui(bs, 0, 1); /* video_signal_type_present_flag */
1132         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
1133         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
1134         {
1135             bitstream_put_ui(bs, 1, 32);
1136             bitstream_put_ui(bs, frame_rate * 2, 32);
1137             bitstream_put_ui(bs, 1, 1);
1138         }
1139         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
1140         {
1141             // hrd_parameters
1142             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
1143             bitstream_put_ui(bs, 0, 4); /* bit_rate_scale */
1144             bitstream_put_ui(bs, 2, 4); /* cpb_size_scale */
1145
1146             /* the frame_bit_rate is in kbps */
1147             bitstream_put_ue(bs, (((frame_bit_rate * 1000)>> 6) - 1)); /* bit_rate_value_minus1[0] */
1148             bitstream_put_ue(bs, ((frame_bit_rate * 8000) >> 6) - 1); /* cpb_size_value_minus1[0] */
1149             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
1150
1151             /* initial_cpb_removal_delay_length_minus1 */
1152             bitstream_put_ui(bs,
1153                 (avcenc_context.i_initial_cpb_removal_delay_length - 1), 5);
1154             /* cpb_removal_delay_length_minus1 */
1155             bitstream_put_ui(bs,
1156                 (avcenc_context.i_cpb_removal_delay_length - 1), 5);
1157             /* dpb_output_delay_length_minus1 */
1158             bitstream_put_ui(bs,
1159                 (avcenc_context.i_dpb_output_delay_length - 1), 5);
1160             /* time_offset_length  */
1161             bitstream_put_ui(bs,
1162                 (avcenc_context.time_offset_length - 1), 5);
1163         }
1164         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
1165         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
1166
1167         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
1168         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
1169     }
1170
1171     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
1172 }
1173
1174 #if 0
1175 static void build_nal_sps(FILE *avc_fp)
1176 {
1177     bitstream bs;
1178
1179     bitstream_start(&bs);
1180     nal_start_code_prefix(&bs);
1181     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
1182     sps_rbsp(&bs);
1183     bitstream_end(&bs, avc_fp);
1184 }
1185 #endif
1186
1187 static void pps_rbsp(bitstream *bs)
1188 {
1189     VAEncPictureParameterBufferH264 *pic_param = &avcenc_context.pic_param;
1190
1191     bitstream_put_ue(bs, pic_param->pic_parameter_set_id);      /* pic_parameter_set_id */
1192     bitstream_put_ue(bs, pic_param->seq_parameter_set_id);      /* seq_parameter_set_id */
1193
1194     bitstream_put_ui(bs, pic_param->pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
1195
1196     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
1197
1198     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
1199
1200     bitstream_put_ue(bs, pic_param->num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
1201     bitstream_put_ue(bs, pic_param->num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
1202
1203     bitstream_put_ui(bs, pic_param->pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
1204     bitstream_put_ui(bs, pic_param->pic_fields.bits.weighted_bipred_idc, 2);    /* weighted_bipred_idc: 0 */
1205
1206     bitstream_put_se(bs, pic_param->pic_init_qp - 26);  /* pic_init_qp_minus26 */
1207     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
1208     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
1209
1210     bitstream_put_ui(bs, pic_param->pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
1211     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
1212     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
1213
1214     /* more_rbsp_data */
1215     bitstream_put_ui(bs, pic_param->pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
1216     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
1217     bitstream_put_se(bs, pic_param->second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
1218
1219     rbsp_trailing_bits(bs);
1220 }
1221
1222 #if 0
1223 static void build_nal_pps(FILE *avc_fp)
1224 {
1225     bitstream bs;
1226
1227     bitstream_start(&bs);
1228     nal_start_code_prefix(&bs);
1229     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
1230     pps_rbsp(&bs);
1231     bitstream_end(&bs, avc_fp);
1232 }
1233
1234 static void
1235 build_header(FILE *avc_fp)
1236 {
1237     build_nal_sps(avc_fp);
1238     build_nal_pps(avc_fp);
1239 }
1240 #endif
1241
1242 static int
1243 build_packed_pic_buffer(unsigned char **header_buffer)
1244 {
1245     bitstream bs;
1246
1247     bitstream_start(&bs);
1248     nal_start_code_prefix(&bs);
1249     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
1250     pps_rbsp(&bs);
1251     bitstream_end(&bs);
1252
1253     *header_buffer = (unsigned char *)bs.buffer;
1254     return bs.bit_offset;
1255 }
1256
1257 static int
1258 build_packed_seq_buffer(unsigned char **header_buffer)
1259 {
1260     bitstream bs;
1261
1262     bitstream_start(&bs);
1263     nal_start_code_prefix(&bs);
1264     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
1265     sps_rbsp(&bs);
1266     bitstream_end(&bs);
1267
1268     *header_buffer = (unsigned char *)bs.buffer;
1269     return bs.bit_offset;
1270 }
1271
1272 static int
1273 build_packed_idr_sei_buffer_timing(unsigned int init_cpb_removal_delay_length,
1274                                 unsigned int cpb_removal_length,
1275                                 unsigned int dpb_output_length,
1276                                 unsigned char **sei_buffer)
1277 {
1278     unsigned char *byte_buf;
1279     int bp_byte_size, i, pic_byte_size;
1280     unsigned int cpb_removal_delay;
1281
1282     bitstream nal_bs;
1283     bitstream sei_bp_bs, sei_pic_bs;
1284
1285     bitstream_start(&sei_bp_bs);
1286     bitstream_put_ue(&sei_bp_bs, 0);       /*seq_parameter_set_id*/
1287     /* SEI buffer period info */
1288     /* NALHrdBpPresentFlag == 1 */
1289     bitstream_put_ui(&sei_bp_bs, avcenc_context.i_initial_cpb_removal_delay,
1290                      init_cpb_removal_delay_length);
1291     bitstream_put_ui(&sei_bp_bs, avcenc_context.i_initial_cpb_removal_delay_offset,
1292                      init_cpb_removal_delay_length);
1293     if ( sei_bp_bs.bit_offset & 0x7) {
1294         bitstream_put_ui(&sei_bp_bs, 1, 1);
1295     }
1296     bitstream_end(&sei_bp_bs);
1297     bp_byte_size = (sei_bp_bs.bit_offset + 7) / 8;
1298
1299     /* SEI pic timing info */
1300     bitstream_start(&sei_pic_bs);
1301     /* The info of CPB and DPB delay is controlled by CpbDpbDelaysPresentFlag,
1302      * which is derived as 1 if one of the following conditions is true:
1303      * nal_hrd_parameters_present_flag is present in the bitstream and is equal to 1,
1304      * vcl_hrd_parameters_present_flag is present in the bitstream and is equal to 1,
1305      */
1306     cpb_removal_delay = (avcenc_context.current_cpb_removal - avcenc_context.prev_idr_cpb_removal);
1307     bitstream_put_ui(&sei_pic_bs, cpb_removal_delay, cpb_removal_length);
1308     bitstream_put_ui(&sei_pic_bs, avcenc_context.current_dpb_removal_delta,
1309                      dpb_output_length);
1310     if ( sei_pic_bs.bit_offset & 0x7) {
1311         bitstream_put_ui(&sei_pic_bs, 1, 1);
1312     }
1313     /* The pic_structure_present_flag determines whether the pic_structure
1314      * info is written into the SEI pic timing info.
1315      * Currently it is set to zero.
1316      */
1317     bitstream_end(&sei_pic_bs);
1318     pic_byte_size = (sei_pic_bs.bit_offset + 7) / 8;
1319
1320     bitstream_start(&nal_bs);
1321     nal_start_code_prefix(&nal_bs);
1322     nal_header(&nal_bs, NAL_REF_IDC_NONE, NAL_SEI);
1323
1324         /* Write the SEI buffer period data */
1325     bitstream_put_ui(&nal_bs, 0, 8);
1326     bitstream_put_ui(&nal_bs, bp_byte_size, 8);
1327
1328     byte_buf = (unsigned char *)sei_bp_bs.buffer;
1329     for(i = 0; i < bp_byte_size; i++) {
1330         bitstream_put_ui(&nal_bs, byte_buf[i], 8);
1331     }
1332     free(byte_buf);
1333         /* write the SEI pic timing data */
1334     bitstream_put_ui(&nal_bs, 0x01, 8);
1335     bitstream_put_ui(&nal_bs, pic_byte_size, 8);
1336
1337     byte_buf = (unsigned char *)sei_pic_bs.buffer;
1338     for(i = 0; i < pic_byte_size; i++) {
1339         bitstream_put_ui(&nal_bs, byte_buf[i], 8);
1340     }
1341     free(byte_buf);
1342
1343     rbsp_trailing_bits(&nal_bs);
1344     bitstream_end(&nal_bs);
1345
1346     *sei_buffer = (unsigned char *)nal_bs.buffer;
1347
1348     return nal_bs.bit_offset;
1349 }
1350
1351 static int
1352 build_packed_sei_pic_timing(unsigned int cpb_removal_length,
1353                                 unsigned int dpb_output_length,
1354                                 unsigned char **sei_buffer)
1355 {
1356     unsigned char *byte_buf;
1357     int i, pic_byte_size;
1358     unsigned int cpb_removal_delay;
1359
1360     bitstream nal_bs;
1361     bitstream sei_pic_bs;
1362
1363     bitstream_start(&sei_pic_bs);
1364     /* The info of CPB and DPB delay is controlled by CpbDpbDelaysPresentFlag,
1365      * which is derived as 1 if one of the following conditions is true:
1366      * nal_hrd_parameters_present_flag is present in the bitstream and is equal to 1,
1367      * vcl_hrd_parameters_present_flag is present in the bitstream and is equal to 1,
1368      */
1369     cpb_removal_delay = (avcenc_context.current_cpb_removal - avcenc_context.current_idr_cpb_removal);
1370     bitstream_put_ui(&sei_pic_bs, cpb_removal_delay, cpb_removal_length);
1371     bitstream_put_ui(&sei_pic_bs, avcenc_context.current_dpb_removal_delta,
1372                      dpb_output_length);
1373     if ( sei_pic_bs.bit_offset & 0x7) {
1374         bitstream_put_ui(&sei_pic_bs, 1, 1);
1375     }
1376
1377     /* The pic_structure_present_flag determines whether the pic_structure
1378      * info is written into the SEI pic timing info.
1379      * Currently it is set to zero.
1380      */
1381     bitstream_end(&sei_pic_bs);
1382     pic_byte_size = (sei_pic_bs.bit_offset + 7) / 8;
1383
1384     bitstream_start(&nal_bs);
1385     nal_start_code_prefix(&nal_bs);
1386     nal_header(&nal_bs, NAL_REF_IDC_NONE, NAL_SEI);
1387
1388         /* write the SEI Pic timing data */
1389     bitstream_put_ui(&nal_bs, 0x01, 8);
1390     bitstream_put_ui(&nal_bs, pic_byte_size, 8);
1391
1392     byte_buf = (unsigned char *)sei_pic_bs.buffer;
1393     for(i = 0; i < pic_byte_size; i++) {
1394         bitstream_put_ui(&nal_bs, byte_buf[i], 8);
1395     }
1396     free(byte_buf);
1397
1398     rbsp_trailing_bits(&nal_bs);
1399     bitstream_end(&nal_bs);
1400
1401     *sei_buffer = (unsigned char *)nal_bs.buffer;
1402
1403     return nal_bs.bit_offset;
1404 }
1405
1406 #if 0
1407 static void
1408 slice_header(bitstream *bs, int frame_num, int display_frame, int slice_type, int nal_ref_idc, int is_idr)
1409 {
1410     VAEncSequenceParameterBufferH264 *seq_param = &avcenc_context.seq_param;
1411     VAEncPictureParameterBufferH264 *pic_param = &avcenc_context.pic_param;
1412     int is_cabac = (pic_param->pic_fields.bits.entropy_coding_mode_flag == ENTROPY_MODE_CABAC);
1413
1414     bitstream_put_ue(bs, 0);                   /* first_mb_in_slice: 0 */
1415     bitstream_put_ue(bs, slice_type);          /* slice_type */
1416     bitstream_put_ue(bs, 0);                   /* pic_parameter_set_id: 0 */
1417     bitstream_put_ui(bs, frame_num & 0x0F, seq_param->seq_fields.bits.log2_max_frame_num_minus4 + 4);    /* frame_num */
1418
1419     /* frame_mbs_only_flag == 1 */
1420     if (!seq_param->seq_fields.bits.frame_mbs_only_flag) {
1421         /* FIXME: */
1422         assert(0);
1423     }
1424
1425     if (is_idr)
1426         bitstream_put_ue(bs, 0);                /* idr_pic_id: 0 */
1427
1428     if (seq_param->seq_fields.bits.pic_order_cnt_type == 0) {
1429         bitstream_put_ui(bs, (display_frame*2) & 0x3F, seq_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
1430         /* only support frame */
1431     } else {
1432         /* FIXME: */
1433         assert(0);
1434     }
1435
1436     /* redundant_pic_cnt_present_flag == 0 */
1437
1438     /* slice type */
1439     if (slice_type == SLICE_TYPE_P) {
1440         bitstream_put_ui(bs, 0, 1);            /* num_ref_idx_active_override_flag: 0 */
1441         /* ref_pic_list_reordering */
1442         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
1443     } else if (slice_type == SLICE_TYPE_B) {
1444         bitstream_put_ui(bs, 1, 1);            /* direct_spatial_mv_pred: 1 */
1445         bitstream_put_ui(bs, 0, 1);            /* num_ref_idx_active_override_flag: 0 */
1446         /* ref_pic_list_reordering */
1447         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
1448         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
1449     }
1450
1451     /* weighted_pred_flag == 0 */
1452
1453     /* dec_ref_pic_marking */
1454     if (nal_ref_idc != 0) {
1455         if ( is_idr) {
1456             bitstream_put_ui(bs, 0, 1);            /* no_output_of_prior_pics_flag: 0 */
1457             bitstream_put_ui(bs, 0, 1);            /* long_term_reference_flag: 0 */
1458         } else {
1459             bitstream_put_ui(bs, 0, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
1460         }
1461     }
1462
1463     if (is_cabac && (slice_type != SLICE_TYPE_I))
1464         bitstream_put_ue(bs, 0);               /* cabac_init_idc: 0 */
1465
1466     bitstream_put_se(bs, 0);                   /* slice_qp_delta: 0 */
1467
1468     if (pic_param->pic_fields.bits.deblocking_filter_control_present_flag == 1) {
1469         bitstream_put_ue(bs, 0);               /* disable_deblocking_filter_idc: 0 */
1470         bitstream_put_se(bs, 2);               /* slice_alpha_c0_offset_div2: 2 */
1471         bitstream_put_se(bs, 2);               /* slice_beta_offset_div2: 2 */
1472     }
1473 }
1474
1475 static void
1476 slice_data(bitstream *bs)
1477 {
1478     VACodedBufferSegment *coded_buffer_segment;
1479     unsigned char *coded_mem;
1480     int i, slice_data_length;
1481     VAStatus va_status;
1482     VASurfaceStatus surface_status;
1483
1484     va_status = vaSyncSurface(va_dpy, surface_ids[avcenc_context.current_input_surface]);
1485     CHECK_VASTATUS(va_status,"vaSyncSurface");
1486
1487     surface_status = 0;
1488     va_status = vaQuerySurfaceStatus(va_dpy, surface_ids[avcenc_context.current_input_surface], &surface_status);
1489     CHECK_VASTATUS(va_status,"vaQuerySurfaceStatus");
1490
1491     va_status = vaMapBuffer(va_dpy, avcenc_context.codedbuf_buf_id, (void **)(&coded_buffer_segment));
1492     CHECK_VASTATUS(va_status,"vaMapBuffer");
1493     coded_mem = coded_buffer_segment->buf;
1494
1495     slice_data_length = get_coded_bitsteam_length(coded_mem, codedbuf_size);
1496
1497     for (i = 0; i < slice_data_length; i++) {
1498         bitstream_put_ui(bs, *coded_mem, 8);
1499         coded_mem++;
1500     }
1501
1502     vaUnmapBuffer(va_dpy, avcenc_context.codedbuf_buf_id);
1503 }
1504
1505 static void
1506 build_nal_slice(FILE *avc_fp, int frame_num, int display_frame, int slice_type, int is_idr)
1507 {
1508     bitstream bs;
1509
1510     bitstream_start(&bs);
1511     slice_data(&bs);
1512     bitstream_end(&bs, avc_fp);
1513 }
1514
1515 #endif
1516
1517 static int
1518 store_coded_buffer(FILE *avc_fp, int slice_type)
1519 {
1520     VACodedBufferSegment *coded_buffer_segment;
1521     unsigned char *coded_mem;
1522     int slice_data_length;
1523     VAStatus va_status;
1524     VASurfaceStatus surface_status;
1525     size_t w_items;
1526
1527     va_status = vaSyncSurface(va_dpy, surface_ids[avcenc_context.current_input_surface]);
1528     CHECK_VASTATUS(va_status,"vaSyncSurface");
1529
1530     surface_status = 0;
1531     va_status = vaQuerySurfaceStatus(va_dpy, surface_ids[avcenc_context.current_input_surface], &surface_status);
1532     CHECK_VASTATUS(va_status,"vaQuerySurfaceStatus");
1533
1534     va_status = vaMapBuffer(va_dpy, avcenc_context.codedbuf_buf_id, (void **)(&coded_buffer_segment));
1535     CHECK_VASTATUS(va_status,"vaMapBuffer");
1536     coded_mem = coded_buffer_segment->buf;
1537
1538     if (coded_buffer_segment->status & VA_CODED_BUF_STATUS_SLICE_OVERFLOW_MASK) {
1539         if (slice_type == SLICE_TYPE_I)
1540             avcenc_context.codedbuf_i_size *= 2;
1541         else
1542             avcenc_context.codedbuf_pb_size *= 2;
1543
1544         vaUnmapBuffer(va_dpy, avcenc_context.codedbuf_buf_id);
1545         return -1;
1546     }
1547
1548     slice_data_length = coded_buffer_segment->size;
1549
1550     do {
1551         w_items = fwrite(coded_mem, slice_data_length, 1, avc_fp);
1552     } while (w_items != 1);
1553
1554     vaUnmapBuffer(va_dpy, avcenc_context.codedbuf_buf_id);
1555
1556     return 0;
1557 }
1558
1559 /*
1560  * It is from the h264encode.c but it simplifies something.
1561  * For example: When one frame is encoded as I-frame under the scenario with
1562  * P-B frames, it will be regarded as IDR frame(key-frame) and then new GOP is
1563  * started. If the video clip is encoded as all I-frames, the first frame
1564  * is regarded as IDR and the remaining is regarded as I-frame.
1565  *
1566  */
1567
1568 static void encoding2display_order(
1569     unsigned long long encoding_order,int gop_size,
1570     int ip_period,
1571     unsigned long long *displaying_order,
1572     int *frame_type)
1573 {
1574     int encoding_order_gop = 0;
1575
1576     /* When ip_period is 0, all are I/IDR frames */
1577     if (ip_period == 0) { /* all are I/IDR frames */
1578         if (encoding_order == 0)
1579             *frame_type = FRAME_IDR;
1580         else
1581             *frame_type = SLICE_TYPE_I;
1582
1583         *displaying_order = encoding_order;
1584         return;
1585     }
1586
1587     /* new sequence like
1588      * IDR PPPPP IDRPPPPP
1589      * IDR (PBB)(PBB)(PBB)(PBB) IDR (PBB)(PBB)(PBB)(PBB)
1590      */
1591     encoding_order_gop = encoding_order % gop_size;
1592
1593     if (encoding_order_gop == 0) { /* the first frame */
1594         *frame_type = FRAME_IDR;
1595         *displaying_order = encoding_order;
1596     } else {
1597         int gop_delta;
1598
1599         gop_delta = 1;
1600
1601         if ((ip_period != 1) && ((gop_size - 1) % ip_period)) {
1602             int ipb_size;
1603             ipb_size = (gop_size - 1) / ip_period * ip_period + 1;
1604             if (encoding_order_gop >= ipb_size) {
1605                 gop_delta = ipb_size;
1606                 ip_period = gop_size - ipb_size;
1607             }
1608         }
1609
1610         if (((encoding_order_gop - gop_delta) % ip_period) == 0) { /* P frames */
1611             *frame_type = SLICE_TYPE_P;
1612             *displaying_order = encoding_order + ip_period - 1;
1613         } else {
1614             *frame_type = SLICE_TYPE_B;
1615             *displaying_order = encoding_order - 1;
1616         }
1617     }
1618 }
1619
1620
1621 static void
1622 encode_picture(FILE *yuv_fp, FILE *avc_fp,
1623                int frame_num, int display_num,
1624                int is_idr,
1625                int slice_type, int next_is_bpic,
1626                int next_display_num)
1627 {
1628     VAStatus va_status;
1629     int ret = 0, codedbuf_size;
1630
1631     begin_picture(yuv_fp, frame_num, display_num, slice_type, is_idr);
1632
1633     //if (next_display_num < frame_number) {
1634     if (1) {
1635         int index;
1636
1637         /* prepare for next frame */
1638         if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_0)
1639             index = SID_INPUT_PICTURE_1;
1640         else
1641             index = SID_INPUT_PICTURE_0;
1642         if ( next_display_num >= frame_number )
1643             next_display_num = frame_number - 1;
1644         fseeko(yuv_fp, (off_t)frame_size * next_display_num, SEEK_SET);
1645
1646         avcenc_context.upload_thread_param.yuv_fp = yuv_fp;
1647         avcenc_context.upload_thread_param.surface_id = surface_ids[index];
1648
1649         avcenc_context.upload_thread_value = pthread_create(&avcenc_context.upload_thread_id,
1650                                                             NULL,
1651                                                             upload_thread_function,
1652                                                             (void*)&avcenc_context.upload_thread_param);
1653     }
1654
1655     do {
1656         avcenc_destroy_buffers(&avcenc_context.codedbuf_buf_id, 1);
1657         avcenc_destroy_buffers(&avcenc_context.pic_param_buf_id, 1);
1658
1659
1660         if (SLICE_TYPE_I == slice_type) {
1661             codedbuf_size = avcenc_context.codedbuf_i_size;
1662         } else {
1663             codedbuf_size = avcenc_context.codedbuf_pb_size;
1664         }
1665
1666         /* coded buffer */
1667         va_status = vaCreateBuffer(va_dpy,
1668                                    avcenc_context.context_id,
1669                                    VAEncCodedBufferType,
1670                                    codedbuf_size, 1, NULL,
1671                                    &avcenc_context.codedbuf_buf_id);
1672         CHECK_VASTATUS(va_status,"vaCreateBuffer");
1673
1674         /* Update the RefPicList */
1675         update_RefPicList();
1676
1677         /* picture parameter set */
1678         avcenc_update_picture_parameter(slice_type, is_idr);
1679
1680         /* slice parameter */
1681         avcenc_update_slice_parameter(slice_type);
1682
1683         if (avcenc_context.rate_control_method == VA_RC_CBR)
1684                 avcenc_update_sei_param(is_idr);
1685
1686         avcenc_render_picture();
1687
1688         ret = store_coded_buffer(avc_fp, slice_type);
1689     } while (ret);
1690
1691     end_picture(slice_type, next_is_bpic);
1692 }
1693
1694 static void show_help()
1695 {
1696     printf("Usage: avnenc <width> <height> <input_yuvfile> <output_avcfile> [--qp=qpvalue|--fb=framebitrate] [--mode=0(I frames only)/1(I and P frames)/2(I, P and B frames)] [--low-power]\n");
1697 }
1698
1699 static void avcenc_context_seq_param_init(VAEncSequenceParameterBufferH264 *seq_param,
1700                                           int width, int height)
1701
1702 {
1703     int width_in_mbs = (width + 15) / 16;
1704     int height_in_mbs = (height + 15) / 16;
1705     int frame_cropping_flag = 0;
1706     int frame_crop_bottom_offset = 0;
1707
1708     seq_param->seq_parameter_set_id = 0;
1709     seq_param->level_idc = 41;
1710     seq_param->intra_period = intra_period;
1711     seq_param->intra_idr_period = seq_param->intra_period;
1712     seq_param->ip_period = ip_period;
1713     seq_param->max_num_ref_frames = 4;
1714     seq_param->picture_width_in_mbs = width_in_mbs;
1715     seq_param->picture_height_in_mbs = height_in_mbs;
1716     seq_param->seq_fields.bits.frame_mbs_only_flag = 1;
1717     seq_param->seq_fields.bits.chroma_format_idc = 1;
1718
1719
1720     if (frame_bit_rate > 0)
1721         seq_param->bits_per_second = 1000 * frame_bit_rate; /* use kbps as input */
1722     else
1723         seq_param->bits_per_second = 0;
1724
1725     seq_param->time_scale = frame_rate * 2;
1726     seq_param->num_units_in_tick = 1;                   /* Tc = num_units_in_tick / time_sacle */
1727
1728     if (height_in_mbs * 16 - height) {
1729         frame_cropping_flag = 1;
1730         frame_crop_bottom_offset =
1731             (height_in_mbs * 16 - height) / (2 * (!seq_param->seq_fields.bits.frame_mbs_only_flag + 1));
1732     }
1733
1734     seq_param->frame_cropping_flag = frame_cropping_flag;
1735     seq_param->frame_crop_left_offset = 0;
1736     seq_param->frame_crop_right_offset = 0;
1737     seq_param->frame_crop_top_offset = 0;
1738     seq_param->frame_crop_bottom_offset = frame_crop_bottom_offset;
1739
1740     seq_param->seq_fields.bits.pic_order_cnt_type = 0;
1741     seq_param->seq_fields.bits.direct_8x8_inference_flag = 0;
1742
1743     seq_param->seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;
1744     seq_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1745
1746     if (frame_bit_rate > 0)
1747         seq_param->vui_parameters_present_flag = 1;     //HRD info located in vui
1748     else
1749         seq_param->vui_parameters_present_flag = 0;
1750 }
1751
1752 static void avcenc_context_pic_param_init(VAEncPictureParameterBufferH264 *pic_param)
1753 {
1754     pic_param->seq_parameter_set_id = 0;
1755     pic_param->pic_parameter_set_id = 0;
1756
1757     pic_param->last_picture = 0;
1758     pic_param->frame_num = 0;
1759
1760     pic_param->pic_init_qp = (qp_value >= 0 ?  qp_value : 26);
1761     pic_param->num_ref_idx_l0_active_minus1 = 0;
1762     pic_param->num_ref_idx_l1_active_minus1 = 0;
1763
1764     pic_param->pic_fields.bits.idr_pic_flag = 0;
1765     pic_param->pic_fields.bits.reference_pic_flag = 0;
1766     pic_param->pic_fields.bits.entropy_coding_mode_flag = ENTROPY_MODE_CABAC;
1767     pic_param->pic_fields.bits.weighted_pred_flag = 0;
1768     pic_param->pic_fields.bits.weighted_bipred_idc = 0;
1769
1770     if (avcenc_context.constraint_set_flag & 0x7)
1771         pic_param->pic_fields.bits.transform_8x8_mode_flag = 0;
1772     else
1773         pic_param->pic_fields.bits.transform_8x8_mode_flag = 1;
1774
1775     pic_param->pic_fields.bits.deblocking_filter_control_present_flag = 1;
1776
1777     memset(pic_param->ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1778 }
1779
1780 static void avcenc_context_sei_init()
1781 {
1782         /* it comes for the bps defined in SPS */
1783         avcenc_context.i_initial_cpb_removal_delay = 2 * 90000;
1784         avcenc_context.i_initial_cpb_removal_delay_offset = 2 * 90000;
1785
1786         avcenc_context.i_cpb_removal_delay = 2;
1787         avcenc_context.i_initial_cpb_removal_delay_length = 24;
1788         avcenc_context.i_cpb_removal_delay_length = 24;
1789         avcenc_context.i_dpb_output_delay_length = 24;
1790         avcenc_context.time_offset_length = 24;
1791
1792         avcenc_context.prev_idr_cpb_removal = avcenc_context.i_initial_cpb_removal_delay / 90000;
1793         avcenc_context.current_idr_cpb_removal = avcenc_context.prev_idr_cpb_removal;
1794         avcenc_context.current_cpb_removal = 0;
1795         avcenc_context.idr_frame_num = 0;
1796 }
1797
1798 static void avcenc_context_init(int width, int height)
1799 {
1800     int i;
1801     memset(&avcenc_context, 0, sizeof(avcenc_context));
1802     avcenc_context.profile = VAProfileH264Main;
1803
1804     memset(&use_slot, 0, sizeof(use_slot));
1805     switch (avcenc_context.profile) {
1806     case VAProfileH264Baseline:
1807         avcenc_context.constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1808         break;
1809
1810     case VAProfileH264Main:
1811         avcenc_context.constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1812         break;
1813
1814     case VAProfileH264High:
1815         avcenc_context.constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1816         break;
1817
1818     default:
1819         break;
1820     }
1821
1822     avcenc_context.seq_param_buf_id = VA_INVALID_ID;
1823     avcenc_context.pic_param_buf_id = VA_INVALID_ID;
1824     avcenc_context.packed_seq_header_param_buf_id = VA_INVALID_ID;
1825     avcenc_context.packed_seq_buf_id = VA_INVALID_ID;
1826     avcenc_context.packed_pic_header_param_buf_id = VA_INVALID_ID;
1827     avcenc_context.packed_pic_buf_id = VA_INVALID_ID;
1828     avcenc_context.codedbuf_buf_id = VA_INVALID_ID;
1829     avcenc_context.misc_parameter_hrd_buf_id = VA_INVALID_ID;
1830     avcenc_context.codedbuf_i_size = width * height;
1831     avcenc_context.codedbuf_pb_size = width * height;
1832     avcenc_context.current_input_surface = SID_INPUT_PICTURE_0;
1833     avcenc_context.upload_thread_value = -1;
1834     avcenc_context.packed_sei_header_param_buf_id = VA_INVALID_ID;
1835     avcenc_context.packed_sei_buf_id = VA_INVALID_ID;
1836
1837     if (qp_value == -1)
1838         avcenc_context.rate_control_method = VA_RC_CBR;
1839     else if (qp_value == -2)
1840         avcenc_context.rate_control_method = VA_RC_VBR;
1841     else {
1842         assert(qp_value >= 0 && qp_value <= 51);
1843         avcenc_context.rate_control_method = VA_RC_CQP;
1844     }
1845
1846     for (i = 0; i < MAX_SLICES; i++) {
1847         avcenc_context.slice_param_buf_id[i] = VA_INVALID_ID;
1848     }
1849
1850     avcenc_context_seq_param_init(&avcenc_context.seq_param, width, height);
1851     avcenc_context_pic_param_init(&avcenc_context.pic_param);
1852     if (avcenc_context.rate_control_method == VA_RC_CBR)
1853         avcenc_context_sei_init();
1854 }
1855
1856 int main(int argc, char *argv[])
1857 {
1858     int f;
1859     FILE *yuv_fp;
1860     FILE *avc_fp;
1861     off_t file_size;
1862     int mode_value;
1863     struct timeval tpstart,tpend;
1864     float  timeuse;
1865
1866     va_init_display_args(&argc, argv);
1867
1868     if(argc < 5) {
1869         show_help();
1870         return -1;
1871     }
1872
1873     picture_width = atoi(argv[1]);
1874     picture_height = atoi(argv[2]);
1875     picture_width_in_mbs = (picture_width + 15) / 16;
1876     picture_height_in_mbs = (picture_height + 15) / 16;
1877
1878     if (argc > 5) {
1879         char o;
1880
1881         optind = 5;
1882
1883         while ((o = getopt_long_only(argc, argv, "", longopts, NULL)) != -1) {
1884             switch (o) {
1885             case 1:     // qp
1886                 frame_bit_rate = -1;
1887                 qp_value = atoi(optarg);
1888
1889                 if (qp_value > 51)
1890                     qp_value = 51;
1891
1892                 if (qp_value < 0)
1893                     qp_value = 0;
1894
1895                 break;
1896
1897             case 2:     // fb
1898                 qp_value = -1;
1899                 frame_bit_rate = atoi(optarg);
1900
1901                 if (frame_bit_rate <= 0) {
1902                     show_help();
1903
1904                     return -1;
1905                 }
1906
1907                 break;
1908
1909             case 3:     // mode
1910                 mode_value = atoi(optarg);
1911
1912                 if (mode_value == 0)
1913                     ip_period = 0;
1914                 else if (mode_value == 1)
1915                     ip_period = 1;
1916                 else if (mode_value == 2)
1917                     /* Hack mechanism before adding the parameter of B-frame number */
1918                     ip_period = 2;
1919                 else {
1920                     printf("mode_value = %d\n", mode_value);
1921                     show_help();
1922                     return -1;
1923                 }
1924
1925                 break;
1926
1927             case 4:     // low-power mode
1928                 select_entrypoint = VAEntrypointEncSliceLP;
1929                 break;
1930
1931             default:
1932                 show_help();
1933                 return -1;
1934             }
1935         }
1936     } else
1937         qp_value = 28;                          //default const QP mode
1938
1939     yuv_fp = fopen(argv[3],"rb");
1940     if ( yuv_fp == NULL){
1941         printf("Can't open input YUV file\n");
1942         return -1;
1943     }
1944     fseeko(yuv_fp, (off_t)0, SEEK_END);
1945     file_size = ftello(yuv_fp);
1946     frame_size = picture_width * picture_height +  ((picture_width * picture_height) >> 1) ;
1947
1948     if ( (file_size < frame_size) || (file_size % frame_size) ) {
1949         fclose(yuv_fp);
1950         printf("The YUV file's size is not correct\n");
1951         return -1;
1952     }
1953     frame_number = file_size / frame_size;
1954     fseeko(yuv_fp, (off_t)0, SEEK_SET);
1955
1956     avc_fp = fopen(argv[4], "wb");
1957     if ( avc_fp == NULL) {
1958         fclose(yuv_fp);
1959         printf("Can't open output avc file\n");
1960         return -1;
1961     }
1962     gettimeofday(&tpstart,NULL);
1963     avcenc_context_init(picture_width, picture_height);
1964     create_encode_pipe();
1965     alloc_encode_resource(yuv_fp);
1966
1967     enc_frame_number = 0;
1968     for ( f = 0; f < frame_number; f++) {               //picture level loop
1969         unsigned long long next_frame_display;
1970         int next_frame_type;
1971
1972         enc_frame_number = f;
1973
1974         encoding2display_order(enc_frame_number, intra_period, ip_period,
1975                                &current_frame_display, &current_frame_type);
1976
1977         encoding2display_order(enc_frame_number + 1, intra_period, ip_period,
1978                                &next_frame_display, &next_frame_type);
1979
1980         if (current_frame_type == FRAME_IDR) {
1981             numShortTerm = 0;
1982             current_frame_num = 0;
1983             memset(&use_slot, 0, sizeof(use_slot));
1984             current_IDR_display = current_frame_display;
1985             if (avcenc_context.rate_control_method == VA_RC_CBR) {
1986                 unsigned long long frame_interval;
1987
1988                 frame_interval = enc_frame_number - avcenc_context.idr_frame_num;
1989
1990                 /* Based on the H264 spec the removal time of the IDR access
1991                  * unit is derived as the following:
1992                  * the removal time of previous IDR unit + Tc * cpb_removal_delay(n)
1993                  */
1994                 avcenc_context.current_cpb_removal = avcenc_context.prev_idr_cpb_removal +
1995                                 frame_interval * 2;
1996                 avcenc_context.idr_frame_num = enc_frame_number;
1997                 avcenc_context.current_idr_cpb_removal = avcenc_context.current_cpb_removal;
1998                 if (ip_period)
1999                     avcenc_context.current_dpb_removal_delta = (ip_period + 1) * 2;
2000                 else
2001                     avcenc_context.current_dpb_removal_delta = 2;
2002             }
2003         } else {
2004             if (avcenc_context.rate_control_method == VA_RC_CBR) {
2005                 unsigned long long frame_interval;
2006
2007                 frame_interval = enc_frame_number - avcenc_context.idr_frame_num;
2008
2009                 /* Based on the H264 spec the removal time of the non-IDR access
2010                  * unit is derived as the following:
2011                  * the removal time of current IDR unit + Tc * cpb_removal_delay(n)
2012                  */
2013                 avcenc_context.current_cpb_removal = avcenc_context.current_idr_cpb_removal +
2014                                 frame_interval * 2;
2015                 if (current_frame_type == SLICE_TYPE_I ||
2016                     current_frame_type == SLICE_TYPE_P) {
2017                     if (ip_period)
2018                         avcenc_context.current_dpb_removal_delta = (ip_period + 1) * 2;
2019                     else
2020                         avcenc_context.current_dpb_removal_delta = 2;
2021                 } else
2022                    avcenc_context.current_dpb_removal_delta = 2;
2023             }
2024         }
2025
2026         /* use the simple mechanism to calc the POC */
2027         current_poc = (current_frame_display - current_IDR_display) * 2;
2028
2029         encode_picture(yuv_fp, avc_fp, frame_number, current_frame_display,
2030                       (current_frame_type == FRAME_IDR) ? 1 : 0,
2031                       (current_frame_type == FRAME_IDR) ? SLICE_TYPE_I : current_frame_type,
2032                       (next_frame_type == SLICE_TYPE_B) ? 1 : 0,
2033                 next_frame_display);
2034         if ((current_frame_type == FRAME_IDR) &&
2035             (avcenc_context.rate_control_method == VA_RC_CBR)) {
2036            /* after one IDR frame is written, it needs to update the
2037             * prev_idr_cpb_removal for next IDR
2038             */
2039            avcenc_context.prev_idr_cpb_removal = avcenc_context.current_idr_cpb_removal;
2040         }
2041         printf("\r %d/%d ...", f, frame_number);
2042         fflush(stdout);
2043     }
2044
2045     gettimeofday(&tpend,NULL);
2046     timeuse=1000000*(tpend.tv_sec-tpstart.tv_sec)+ tpend.tv_usec-tpstart.tv_usec;
2047     timeuse/=1000000;
2048     printf("\ndone!\n");
2049     printf("encode %d frames in %f secondes, FPS is %.1f\n",frame_number, timeuse, frame_number/timeuse);
2050     release_encode_resource();
2051     destory_encode_pipe();
2052
2053     fclose(yuv_fp);
2054     fclose(avc_fp);
2055
2056     return 0;
2057 }