test/encode/h264encode.c

   1 /*
   2  * Copyright (c) 2007-2013 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 #include "sysdeps.h"
  25 #include <stdio.h>
  26 #include <string.h>
  27 #include <stdlib.h>
  28 #include <getopt.h>
  29 #include <unistd.h>
  30 #include <sys/types.h>
  31 #include <sys/stat.h>
  32 #include <sys/time.h>
  33 #include <fcntl.h>
  34 #include <assert.h>
  35 #include <pthread.h>
  36 #include <va/va.h>
  37 #include <va/va_enc_h264.h>
  38 #include "va_display.h"
  39
  40 #define CHECK_VASTATUS(va_status,func)                                  \
  41     if (va_status != VA_STATUS_SUCCESS) {                               \
  42         fprintf(stderr,"%s:%s (%d) failed,exit\n", __func__, func, __LINE__); \
  43         exit(1);                                                        \
  44     }
  45
  46 #include "../loadsurface.h"
  47
  48 #define NAL_REF_IDC_NONE        0
  49 #define NAL_REF_IDC_LOW         1
  50 #define NAL_REF_IDC_MEDIUM      2
  51 #define NAL_REF_IDC_HIGH        3
  52
  53 #define NAL_NON_IDR             1
  54 #define NAL_IDR                 5
  55 #define NAL_SPS                 7
  56 #define NAL_PPS                 8
  57 #define NAL_SEI                 6
  58
  59 #define SLICE_TYPE_P            0
  60 #define SLICE_TYPE_B            1
  61 #define SLICE_TYPE_I            2
  62
  63 #define ENTROPY_MODE_CAVLC      0
  64 #define ENTROPY_MODE_CABAC      1
  65
  66 #define PROFILE_IDC_BASELINE    66
  67 #define PROFILE_IDC_MAIN        77
  68 #define PROFILE_IDC_HIGH        100
  69
  70 #define BITSTREAM_ALLOCATE_STEPPING     4096
  71
  72
  73 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  74 #define SURFACE_NUM 16 /* 16 surfaces for reference */
  75 static  VADisplay va_dpy;
  76 static  VAProfile h264_profile;
  77 static  VAConfigAttrib attrib[VAConfigAttribTypeMax];
  78 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  79 static  int config_attrib_num = 0;
  80 static  VASurfaceID src_surface[SURFACE_NUM];
  81 static  VABufferID  coded_buf[SURFACE_NUM];
  82 static  VASurfaceID ref_surface[SURFACE_NUM];
  83 static  VAConfigID config_id;
  84 static  VAContextID context_id;
  85 static  VAEncSequenceParameterBufferH264 seq_param;
  86 static  VAEncPictureParameterBufferH264 pic_param;
  87 static  VAEncSliceParameterBufferH264 slice_param;
  88 static  VAPictureH264 CurrentCurrPic;
  89 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  90
  91 static  unsigned int MaxFrameNum = (2<<16);
  92 static  unsigned int MaxPicOrderCntLsb = (2<<8);
  93 static  unsigned int Log2MaxFrameNum = 16;
  94 static  unsigned int Log2MaxPicOrderCntLsb = 8;
  95
  96 static  unsigned int num_ref_frames = 2;
  97 static  unsigned int numShortTerm = 0;
  98 static  int constraint_set_flag = 0;
  99 static  int h264_packedheader = 0; /* support pack header? */
 100 static  int h264_maxref = (1<<16|1);
 101 static  char *coded_fn = NULL, *srcyuv_fn = NULL, *recyuv_fn = NULL;
 102 static  FILE *coded_fp = NULL, *srcyuv_fp = NULL, *recyuv_fp = NULL;
 103 static  unsigned long long srcyuv_frames = 0;
 104 static  int srcyuv_fourcc = VA_FOURCC_NV12;
 105
 106 static  int frame_width = 176;
 107 static  int frame_height = 144;
 108 static  int frame_rate = 30;
 109 static  unsigned int frame_count = 60;
 110 static  unsigned int frame_coded = 0;
 111 static  unsigned int frame_bitrate = 0;
 112 static  unsigned int frame_slices = 1;
 113 static  int initial_qp = 28;
 114 static  int minimal_qp = 0;
 115 static  int intra_period = 30;
 116 static  int intra_idr_period = 60;
 117 static  int ip_period = 1;
 118 static  int rc_mode = VA_RC_VBR;
 119 static  unsigned long long current_frame_encoding = 0;
 120 static  unsigned long long current_frame_display = 0;
 121 static  int current_frame_num = 0;
 122 static  int current_frame_type;
 123 #define current_slot (current_frame_display % SURFACE_NUM)
 124
 125 /* thread to save coded data/upload source YUV */
 126 struct storage_task_t {
 127     void *next;
 128     unsigned long long display_order;
 129     unsigned long long encode_order;
 130 };
 131 static  struct storage_task_t *storage_task_header = NULL, *storage_task_tail = NULL;
 132 #define SRC_SURFACE_IN_ENCODING 0
 133 #define SRC_SURFACE_IN_STORAGE  1
 134 static  int srcsurface_status[SURFACE_NUM];
 135 static  int encode_syncmode = 0;
 136 static  pthread_mutex_t encode_mutex = PTHREAD_MUTEX_INITIALIZER;
 137 static  pthread_cond_t  encode_cond = PTHREAD_COND_INITIALIZER;
 138 static  pthread_t encode_thread;
 139
 140 /* for performance profiling */
 141 static unsigned int UploadPictureTicks=0;
 142 static unsigned int BeginPictureTicks=0;
 143 static unsigned int RenderPictureTicks=0;
 144 static unsigned int EndPictureTicks=0;
 145 static unsigned int SyncPictureTicks=0;
 146 static unsigned int SavePictureTicks=0;
 147 static unsigned int TotalTicks=0;
 148
 149 struct __bitstream {
 150     unsigned int *buffer;
 151     int bit_offset;
 152     int max_size_in_dword;
 153 };
 154 typedef struct __bitstream bitstream;
 155
 156
 157 static unsigned int
 158 va_swap32(unsigned int val)
 159 {
 160     unsigned char *pval = (unsigned char *)&val;
 161
 162     return ((pval[0] << 24)     |
 163             (pval[1] << 16)     |
 164             (pval[2] << 8)      |
 165             (pval[3] << 0));
 166 }
 167
 168 static void
 169 bitstream_start(bitstream *bs)
 170 {
 171     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 172     bs->buffer = calloc(bs->max_size_in_dword * sizeof(int), 1);
 173     bs->bit_offset = 0;
 174 }
 175
 176 static void
 177 bitstream_end(bitstream *bs)
 178 {
 179     int pos = (bs->bit_offset >> 5);
 180     int bit_offset = (bs->bit_offset & 0x1f);
 181     int bit_left = 32 - bit_offset;
 182
 183     if (bit_offset) {
 184         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 185     }
 186 }
 187
 188 static void
 189 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 190 {
 191     int pos = (bs->bit_offset >> 5);
 192     int bit_offset = (bs->bit_offset & 0x1f);
 193     int bit_left = 32 - bit_offset;
 194
 195     if (!size_in_bits)
 196         return;
 197
 198     bs->bit_offset += size_in_bits;
 199
 200     if (bit_left > size_in_bits) {
 201         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 202     } else {
 203         size_in_bits -= bit_left;
 204         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 205         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 206
 207         if (pos + 1 == bs->max_size_in_dword) {
 208             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 209             bs->buffer = realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 210         }
 211
 212         bs->buffer[pos + 1] = val;
 213     }
 214 }
 215
 216 static void
 217 bitstream_put_ue(bitstream *bs, unsigned int val)
 218 {
 219     int size_in_bits = 0;
 220     int tmp_val = ++val;
 221
 222     while (tmp_val) {
 223         tmp_val >>= 1;
 224         size_in_bits++;
 225     }
 226
 227     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 228     bitstream_put_ui(bs, val, size_in_bits);
 229 }
 230
 231 static void
 232 bitstream_put_se(bitstream *bs, int val)
 233 {
 234     unsigned int new_val;
 235
 236     if (val <= 0)
 237         new_val = -2 * val;
 238     else
 239         new_val = 2 * val - 1;
 240
 241     bitstream_put_ue(bs, new_val);
 242 }
 243
 244 static void
 245 bitstream_byte_aligning(bitstream *bs, int bit)
 246 {
 247     int bit_offset = (bs->bit_offset & 0x7);
 248     int bit_left = 8 - bit_offset;
 249     int new_val;
 250
 251     if (!bit_offset)
 252         return;
 253
 254     assert(bit == 0 || bit == 1);
 255
 256     if (bit)
 257         new_val = (1 << bit_left) - 1;
 258     else
 259         new_val = 0;
 260
 261     bitstream_put_ui(bs, new_val, bit_left);
 262 }
 263
 264 static void
 265 rbsp_trailing_bits(bitstream *bs)
 266 {
 267     bitstream_put_ui(bs, 1, 1);
 268     bitstream_byte_aligning(bs, 0);
 269 }
 270
 271 static void nal_start_code_prefix(bitstream *bs)
 272 {
 273     bitstream_put_ui(bs, 0x00000001, 32);
 274 }
 275
 276 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 277 {
 278     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 279     bitstream_put_ui(bs, nal_ref_idc, 2);
 280     bitstream_put_ui(bs, nal_unit_type, 5);
 281 }
 282
 283 static void sps_rbsp(bitstream *bs)
 284 {
 285     int profile_idc = PROFILE_IDC_BASELINE;
 286
 287     if (h264_profile  == VAProfileH264High)
 288         profile_idc = PROFILE_IDC_HIGH;
 289     else if (h264_profile  == VAProfileH264Main)
 290         profile_idc = PROFILE_IDC_MAIN;
 291
 292     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 293     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 294     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 295     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 296     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 297     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 298     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 299     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 300
 301     if ( profile_idc == PROFILE_IDC_HIGH) {
 302         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 303         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 304         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 305         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 306         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 307     }
 308
 309     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 310     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 311
 312     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 313         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 314     else {
 315         assert(0);
 316     }
 317
 318     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 319     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 320
 321     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 322     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 323     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 324
 325     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 326         assert(0);
 327     }
 328
 329     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 330     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 331
 332     if (seq_param.frame_cropping_flag) {
 333         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 334         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 335         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 336         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 337     }
 338
 339     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 340     if ( 1 ) {
 341         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 342     } else {
 343         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 344         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 345         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 346         bitstream_put_ui(bs, 0, 1); /* video_signal_type_present_flag */
 347         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 348         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 349         {
 350             bitstream_put_ui(bs, 15, 32);
 351             bitstream_put_ui(bs, 900, 32);
 352             bitstream_put_ui(bs, 1, 1);
 353         }
 354         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 355         {
 356             // hrd_parameters
 357             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 358             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 359             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 360
 361             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 362             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 363             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 364
 365             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 366             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 367             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 368             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 369         }
 370         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 371         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 372
 373         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 374         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 375     }
 376
 377     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 378 }
 379
 380
 381 static void pps_rbsp(bitstream *bs)
 382 {
 383     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 384     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 385
 386     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 387
 388     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 389
 390     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 391
 392     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 393     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 394
 395     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 396     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 397
 398     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 399     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 400     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 401
 402     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 403     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 404     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 405
 406     /* more_rbsp_data */
 407     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 408     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 409     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 410
 411     rbsp_trailing_bits(bs);
 412 }
 413
 414
 415 static int
 416 build_packed_pic_buffer(unsigned char **header_buffer)
 417 {
 418     bitstream bs;
 419
 420     bitstream_start(&bs);
 421     nal_start_code_prefix(&bs);
 422     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 423     pps_rbsp(&bs);
 424     bitstream_end(&bs);
 425
 426     *header_buffer = (unsigned char *)bs.buffer;
 427     return bs.bit_offset;
 428 }
 429
 430 static int
 431 build_packed_seq_buffer(unsigned char **header_buffer)
 432 {
 433     bitstream bs;
 434
 435     bitstream_start(&bs);
 436     nal_start_code_prefix(&bs);
 437     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 438     sps_rbsp(&bs);
 439     bitstream_end(&bs);
 440
 441     *header_buffer = (unsigned char *)bs.buffer;
 442     return bs.bit_offset;
 443 }
 444
 445 static int
 446 build_packed_sei_buffer_timing(unsigned int init_cpb_removal_length,
 447                                 unsigned int init_cpb_removal_delay,
 448                                 unsigned int init_cpb_removal_delay_offset,
 449                                 unsigned int cpb_removal_length,
 450                                 unsigned int cpb_removal_delay,
 451                                 unsigned int dpb_output_length,
 452                                 unsigned int dpb_output_delay,
 453                                 unsigned char **sei_buffer)
 454 {
 455     unsigned char *byte_buf;
 456     int bp_byte_size, i, pic_byte_size;
 457
 458     bitstream nal_bs;
 459     bitstream sei_bp_bs, sei_pic_bs;
 460
 461     bitstream_start(&sei_bp_bs);
 462     bitstream_put_ue(&sei_bp_bs, 0);       /*seq_parameter_set_id*/
 463     bitstream_put_ui(&sei_bp_bs, init_cpb_removal_delay, cpb_removal_length);
 464     bitstream_put_ui(&sei_bp_bs, init_cpb_removal_delay_offset, cpb_removal_length);
 465     if ( sei_bp_bs.bit_offset & 0x7) {
 466         bitstream_put_ui(&sei_bp_bs, 1, 1);
 467     }
 468     bitstream_end(&sei_bp_bs);
 469     bp_byte_size = (sei_bp_bs.bit_offset + 7) / 8;
 470
 471     bitstream_start(&sei_pic_bs);
 472     bitstream_put_ui(&sei_pic_bs, cpb_removal_delay, cpb_removal_length);
 473     bitstream_put_ui(&sei_pic_bs, dpb_output_delay, dpb_output_length);
 474     if ( sei_pic_bs.bit_offset & 0x7) {
 475         bitstream_put_ui(&sei_pic_bs, 1, 1);
 476     }
 477     bitstream_end(&sei_pic_bs);
 478     pic_byte_size = (sei_pic_bs.bit_offset + 7) / 8;
 479
 480     bitstream_start(&nal_bs);
 481     nal_start_code_prefix(&nal_bs);
 482     nal_header(&nal_bs, NAL_REF_IDC_NONE, NAL_SEI);
 483
 484         /* Write the SEI buffer period data */
 485     bitstream_put_ui(&nal_bs, 0, 8);
 486     bitstream_put_ui(&nal_bs, bp_byte_size, 8);
 487
 488     byte_buf = (unsigned char *)sei_bp_bs.buffer;
 489     for(i = 0; i < bp_byte_size; i++) {
 490         bitstream_put_ui(&nal_bs, byte_buf[i], 8);
 491     }
 492     free(byte_buf);
 493         /* write the SEI timing data */
 494     bitstream_put_ui(&nal_bs, 0x01, 8);
 495     bitstream_put_ui(&nal_bs, pic_byte_size, 8);
 496
 497     byte_buf = (unsigned char *)sei_pic_bs.buffer;
 498     for(i = 0; i < pic_byte_size; i++) {
 499         bitstream_put_ui(&nal_bs, byte_buf[i], 8);
 500     }
 501     free(byte_buf);
 502
 503     rbsp_trailing_bits(&nal_bs);
 504     bitstream_end(&nal_bs);
 505
 506     *sei_buffer = (unsigned char *)nal_bs.buffer;
 507
 508     return nal_bs.bit_offset;
 509 }
 510
 511
 512
 513 /*
 514  * Helper function for profiling purposes
 515  */
 516 static unsigned int GetTickCount()
 517 {
 518     struct timeval tv;
 519     if (gettimeofday(&tv, NULL))
 520         return 0;
 521     return tv.tv_usec/1000+tv.tv_sec*1000;
 522 }
 523
 524 /*
 525   Assume frame sequence is: Frame#0,#1,#2,...,#M,...,#X,... (encoding order)
 526   1) period between Frame #X and Frame #N = #X - #N
 527   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 528   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 529   4) intra_period and intra_idr_period take precedence over ip_period
 530   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 531      of I/IDR frames, see bellow examples
 532   -------------------------------------------------------------------
 533   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 534   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 535   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 536   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 537   1            1                ignored    IDR IDR IDR IDR...
 538   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 539   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 540   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 541                                               (PBB)(IBB)(PBB)(IBB)...
 542   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 543                                            IDRPPPPP IPPPPP IPPPPP...
 544   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 545                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 546                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 547                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 548                                            {IDR(PBB)(PBB)}                     (6/6/3)
 549                                            {IDR(PBB)(PBB)}.
 550 */
 551
 552 /*
 553  * Return displaying order with specified periods and encoding order
 554  * displaying_order: displaying order
 555  * frame_type: frame type
 556  */
 557 #define FRAME_P 0
 558 #define FRAME_B 1
 559 #define FRAME_I 2
 560 #define FRAME_IDR 7
 561 void encoding2display_order(
 562     unsigned long long encoding_order,int intra_period,
 563     int intra_idr_period,int ip_period,
 564     unsigned long long *displaying_order,
 565     int *frame_type)
 566 {
 567     int encoding_order_gop = 0;
 568
 569     if (intra_period == 1) { /* all are I/IDR frames */
 570         *displaying_order = encoding_order;
 571         if (intra_idr_period == 0)
 572             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 573         else
 574             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 575         return;
 576     }
 577
 578     if (intra_period == 0)
 579         intra_idr_period = 0;
 580
 581     /* new sequence like
 582      * IDR PPPPP IPPPPP
 583      * IDR (PBB)(PBB)(IBB)(PBB)
 584      */
 585     encoding_order_gop = (intra_idr_period == 0)? encoding_order:
 586         (encoding_order % (intra_idr_period + ((ip_period == 1)?0:1)));
 587
 588     if (encoding_order_gop == 0) { /* the first frame */
 589         *frame_type = FRAME_IDR;
 590         *displaying_order = encoding_order;
 591     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 592         *frame_type = FRAME_B;
 593         *displaying_order = encoding_order - 1;
 594     } else if ((intra_period != 0) && /* have I frames */
 595                (encoding_order_gop >= 2) &&
 596                ((ip_period == 1 && encoding_order_gop % intra_period == 0) || /* for IDR PPPPP IPPPP */
 597                 /* for IDR (PBB)(PBB)(IBB) */
 598                 (ip_period >= 2 && ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0))) {
 599         *frame_type = FRAME_I;
 600         *displaying_order = encoding_order + ip_period - 1;
 601     } else {
 602         *frame_type = FRAME_P;
 603         *displaying_order = encoding_order + ip_period - 1;
 604     }
 605 }
 606
 607
 608 static char *fourcc_to_string(int fourcc)
 609 {
 610     switch (fourcc) {
 611     case VA_FOURCC_NV12:
 612         return "NV12";
 613     case VA_FOURCC_IYUV:
 614         return "IYUV";
 615     case VA_FOURCC_YV12:
 616         return "YV12";
 617     case VA_FOURCC_UYVY:
 618         return "UYVY";
 619     default:
 620         return "Unknown";
 621     }
 622 }
 623
 624 static int string_to_fourcc(char *str)
 625 {
 626     int fourcc;
 627
 628     if (!strncmp(str, "NV12", 4))
 629         fourcc = VA_FOURCC_NV12;
 630     else if (!strncmp(str, "IYUV", 4))
 631         fourcc = VA_FOURCC_IYUV;
 632     else if (!strncmp(str, "YV12", 4))
 633         fourcc = VA_FOURCC_YV12;
 634     else if (!strncmp(str, "UYVY", 4))
 635         fourcc = VA_FOURCC_UYVY;
 636     else {
 637         printf("Unknow FOURCC\n");
 638         fourcc = -1;
 639     }
 640     return fourcc;
 641 }
 642
 643
 644 static char *rc_to_string(int rcmode)
 645 {
 646     switch (rc_mode) {
 647     case VA_RC_NONE:
 648         return "NONE";
 649     case VA_RC_CBR:
 650         return "CBR";
 651     case VA_RC_VBR:
 652         return "VBR";
 653     case VA_RC_VCM:
 654         return "VCM";
 655     case VA_RC_CQP:
 656         return "CQP";
 657     case VA_RC_VBR_CONSTRAINED:
 658         return "VBR_CONSTRAINED";
 659     default:
 660         return "Unknown";
 661     }
 662 }
 663
 664 static int string_to_rc(char *str)
 665 {
 666     int rc_mode;
 667
 668     if (!strncmp(str, "NONE", 4))
 669         rc_mode = VA_RC_NONE;
 670     else if (!strncmp(str, "CBR", 3))
 671         rc_mode = VA_RC_CBR;
 672     else if (!strncmp(str, "VBR", 3))
 673         rc_mode = VA_RC_VBR;
 674     else if (!strncmp(str, "VCM", 3))
 675         rc_mode = VA_RC_VCM;
 676     else if (!strncmp(str, "CQP", 3))
 677         rc_mode = VA_RC_CQP;
 678     else if (!strncmp(str, "VBR_CONSTRAINED", 15))
 679         rc_mode = VA_RC_VBR_CONSTRAINED;
 680     else {
 681         printf("Unknown RC mode\n");
 682         rc_mode = -1;
 683     }
 684     return rc_mode;
 685 }
 686
 687
 688 static int print_help(void)
 689 {
 690     printf("./h264encode <options>\n");
 691     printf("   -w <width> -h <height>\n");
 692     printf("   -n <frame number>\n");
 693     printf("   -o <coded file>\n");
 694     printf("   -f <frame rate>\n");
 695     printf("   --intra_period <number>\n");
 696     printf("   --idr_period <number>\n");
 697     printf("   --ip_period <number>\n");
 698     printf("   --bitrate <bitrate>\n");
 699     printf("   --initialqp <number>\n");
 700     printf("   --minqp <number>\n");
 701     printf("   --rcmode <NONE|CBR|VBR|VCM|CQP|VBR_CONTRAINED>\n");
 702     printf("   --syncmode: sequentially upload source, encoding, save result, no multi-thread\n");
 703     printf("   --srcyuv <filename> load YUV from a file\n");
 704     printf("   --fourcc <NV12|IYUV|I420|YV12> source YUV fourcc\n");
 705
 706     return 0;
 707 }
 708
 709 static int process_cmdline(int argc, char *argv[])
 710 {
 711     char c;
 712     const struct option long_opts[] = {
 713         {"bitrate", required_argument, NULL, 1 },
 714         {"minqp", required_argument, NULL, 2 },
 715         {"initialqp", required_argument, NULL, 3 },
 716         {"intra_period", required_argument, NULL, 4 },
 717         {"idr_period", required_argument, NULL, 5 },
 718         {"ip_period", required_argument, NULL, 6 },
 719         {"rcmode", required_argument, NULL, 7 },
 720         {"srcyuv", required_argument, NULL, 9 },
 721         {"fourcc", required_argument, NULL, 10 },
 722         {"syncmode", no_argument, NULL, 11 },
 723         {NULL, no_argument, NULL, 0 }};
 724     int long_index;
 725
 726     while ((c =getopt_long_only(argc,argv,"w:h:n:f:o:?",long_opts,&long_index)) != EOF) {
 727         switch (c) {
 728         case 'w':
 729             frame_width = atoi(optarg);
 730             break;
 731         case 'h':
 732             frame_height = atoi(optarg);
 733             break;
 734         case 'n':
 735             frame_count = atoi(optarg);
 736             break;
 737         case 'f':
 738             frame_rate = atoi(optarg);
 739             break;
 740         case 'o':
 741             coded_fn = strdup(optarg);
 742             break;
 743         case 1:
 744             frame_bitrate = atoi(optarg);
 745             break;
 746         case 2:
 747             minimal_qp = atoi(optarg);
 748             break;
 749         case 3:
 750             initial_qp = atoi(optarg);
 751             break;
 752         case 4:
 753             intra_period = atoi(optarg);
 754             break;
 755         case 5:
 756             intra_idr_period = atoi(optarg);
 757             break;
 758         case 6:
 759             ip_period = atoi(optarg);
 760             break;
 761         case 7:
 762             rc_mode = string_to_rc(optarg);
 763             if (rc_mode < 0) {
 764                 print_help();
 765                 exit(1);
 766             }
 767             break;
 768         case 9:
 769             srcyuv_fn = strdup(optarg);
 770             break;
 771         case 10:
 772             srcyuv_fourcc = string_to_fourcc(optarg);
 773             if (srcyuv_fourcc <= 0) {
 774                 print_help();
 775                 exit(1);
 776             }
 777             break;
 778         case 11:
 779             encode_syncmode = 1;
 780             break;
 781         case ':':
 782         case '?':
 783             print_help();
 784             exit(0);
 785         }
 786     }
 787
 788     if (ip_period < 1) {
 789         printf(" ip_period must be greater than 0\n");
 790         exit(0);
 791     }
 792     if (intra_period != 1 && intra_period % ip_period != 0) {
 793         printf(" intra_period must be a multiplier of ip_period\n");
 794         exit(0);
 795     }
 796     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 797         printf(" intra_idr_period must be a multiplier of intra_period\n");
 798         exit(0);
 799     }
 800
 801     if (frame_bitrate == 0)
 802         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 803
 804     /* open source file */
 805     if (srcyuv_fn) {
 806         srcyuv_fp = fopen(srcyuv_fn,"r");
 807
 808         if (srcyuv_fp == NULL)
 809             printf("Open source YUV file %s failed, use auto-generated YUV data\n", srcyuv_fn);
 810         else {
 811             fseek(srcyuv_fp, 0L, SEEK_END);
 812             srcyuv_frames = ftell(srcyuv_fp) / (frame_width * frame_height * 1.5);
 813             printf("Source YUV file %s with %llu frames\n", srcyuv_fn, srcyuv_frames);
 814         }
 815     }
 816
 817     if (coded_fn == NULL) {
 818         struct stat buf;
 819         if (stat("/tmp", &buf) == 0)
 820             coded_fn = strdup("/tmp/test.264");
 821         else if (stat("/sdcard", &buf) == 0)
 822             coded_fn = strdup("/sdcard/test.264");
 823         else
 824             coded_fn = strdup("./test.264");
 825     }
 826
 827     /* store coded data into a file */
 828     coded_fp = fopen(coded_fn,"w+");
 829     if (coded_fp == NULL) {
 830         printf("Open file %s failed, exit\n", coded_fn);
 831         exit(1);
 832     }
 833
 834
 835     return 0;
 836 }
 837
 838 static int init_va(void)
 839 {
 840     VAProfile profile_list[]={VAProfileH264High,VAProfileH264Main,VAProfileH264Baseline,VAProfileH264ConstrainedBaseline};
 841     VAEntrypoint entrypoints[VAEntrypointMax]={0};
 842     int num_entrypoints,slice_entrypoint;
 843     int support_encode = 0;
 844     int major_ver, minor_ver;
 845     VAStatus va_status;
 846     unsigned int i;
 847
 848     va_dpy = va_open_display();
 849     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 850     CHECK_VASTATUS(va_status, "vaInitialize");
 851
 852     /* use the highest profile */
 853     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 854         h264_profile = profile_list[i];
 855         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 856         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 857             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 858                 support_encode = 1;
 859                 break;
 860             }
 861         }
 862         if (support_encode == 1)
 863             break;
 864     }
 865
 866     if (support_encode == 0) {
 867         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 868         exit(1);
 869     } else {
 870         switch (h264_profile) {
 871             case VAProfileH264Baseline:
 872                 printf("Use profile VAProfileH264Baseline\n");
 873                 ip_period = 1;
 874                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 875                 break;
 876             case VAProfileH264ConstrainedBaseline:
 877                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 878                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 879                 ip_period = 1;
 880                 break;
 881
 882             case VAProfileH264Main:
 883                 printf("Use profile VAProfileH264Main\n");
 884                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 885                 break;
 886
 887             case VAProfileH264High:
 888                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 889                 printf("Use profile VAProfileH264High\n");
 890                 break;
 891             default:
 892                 printf("unknow profile. Set to Baseline");
 893                 h264_profile = VAProfileH264Baseline;
 894                 ip_period = 1;
 895                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 896                 break;
 897         }
 898     }
 899
 900     /* find out the format for the render target, and rate control mode */
 901     for (i = 0; i < VAConfigAttribTypeMax; i++)
 902         attrib[i].type = i;
 903
 904     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 905                                       &attrib[0], VAConfigAttribTypeMax);
 906     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 907     /* check the interested configattrib */
 908     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 909         printf("Not find desired YUV420 RT format\n");
 910         exit(1);
 911     } else {
 912         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 913         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 914         config_attrib_num++;
 915     }
 916
 917     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 918         int tmp = attrib[VAConfigAttribRateControl].value;
 919
 920         printf("Supported rate control mode (0x%x):", tmp);
 921
 922         if (tmp & VA_RC_NONE)
 923             printf("NONE ");
 924         if (tmp & VA_RC_CBR)
 925             printf("CBR ");
 926         if (tmp & VA_RC_VBR)
 927             printf("VBR ");
 928         if (tmp & VA_RC_VCM)
 929             printf("VCM ");
 930         if (tmp & VA_RC_CQP)
 931             printf("CQP ");
 932         if (tmp & VA_RC_VBR_CONSTRAINED)
 933             printf("VBR_CONSTRAINED ");
 934
 935         printf("\n");
 936
 937         /* need to check if support rc_mode */
 938         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 939         config_attrib[config_attrib_num].value = rc_mode;
 940         config_attrib_num++;
 941     }
 942
 943
 944     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 945         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 946
 947         printf("Support VAConfigAttribEncPackedHeaders\n");
 948
 949         h264_packedheader = 1;
 950         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 951         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 952
 953         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 954             printf("Support packed sequence headers\n");
 955             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 956         }
 957
 958         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 959             printf("Support packed picture headers\n");
 960             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 961         }
 962
 963         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 964             printf("Support packed slice headers\n");
 965             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 966         }
 967
 968         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
 969             printf("Support packed misc headers\n");
 970             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
 971         }
 972
 973         config_attrib_num++;
 974     }
 975
 976     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
 977         int tmp = attrib[VAConfigAttribEncInterlaced].value;
 978
 979         printf("Support VAConfigAttribEncInterlaced\n");
 980
 981         if (tmp & VA_ENC_INTERLACED_FRAME)
 982             printf("support VA_ENC_INTERLACED_FRAME\n");
 983         if (tmp & VA_ENC_INTERLACED_FIELD)
 984             printf("Support VA_ENC_INTERLACED_FIELD\n");
 985         if (tmp & VA_ENC_INTERLACED_MBAFF)
 986             printf("Support VA_ENC_INTERLACED_MBAFF\n");
 987         if (tmp & VA_ENC_INTERLACED_PAFF)
 988             printf("Support VA_ENC_INTERLACED_PAFF\n");
 989
 990         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
 991         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 992         config_attrib_num++;
 993     }
 994
 995     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
 996         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
 997
 998         printf("Support %d reference frames\n", h264_maxref);
 999     }
1000
1001     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1002         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1003
1004     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1005         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1006
1007         printf("Support VAConfigAttribEncSliceStructure\n");
1008
1009         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1010             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1011         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1012             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1013         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1014             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1015     }
1016     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1017         printf("Support VAConfigAttribEncMacroblockInfo\n");
1018     }
1019
1020     return 0;
1021 }
1022
1023 static int setup_encode()
1024 {
1025     VAStatus va_status;
1026     VASurfaceID *tmp_surfaceid;
1027     int codedbuf_size, i;
1028
1029     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1030             &config_attrib[0], config_attrib_num, &config_id);
1031     CHECK_VASTATUS(va_status, "vaCreateConfig");
1032
1033     /* create source surfaces */
1034     va_status = vaCreateSurfaces(va_dpy,
1035                                  VA_RT_FORMAT_YUV420, frame_width, frame_height,
1036                                  &src_surface[0], SURFACE_NUM,
1037                                  NULL, 0);
1038     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1039
1040     /* create reference surfaces */
1041     va_status = vaCreateSurfaces(
1042         va_dpy,
1043         VA_RT_FORMAT_YUV420, frame_width, frame_height,
1044         &ref_surface[0], SURFACE_NUM,
1045         NULL, 0
1046         );
1047     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1048
1049     tmp_surfaceid = calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1050     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1051     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1052
1053     /* Create a context for this encode pipe */
1054     va_status = vaCreateContext(va_dpy, config_id,
1055                                 frame_width, frame_height,
1056                                 VA_PROGRESSIVE,
1057                                 tmp_surfaceid, 2 * SURFACE_NUM,
1058                                 &context_id);
1059     CHECK_VASTATUS(va_status, "vaCreateContext");
1060     free(tmp_surfaceid);
1061
1062     codedbuf_size = (frame_width * frame_height * 400) / (16*16);
1063
1064     for (i = 0; i < SURFACE_NUM; i++) {
1065         /* create coded buffer once for all
1066          * other VA buffers which won't be used again after vaRenderPicture.
1067          * so APP can always vaCreateBuffer for every frame
1068          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1069          * so VA won't maintain the coded buffer
1070          */
1071         va_status = vaCreateBuffer(va_dpy,context_id,VAEncCodedBufferType,
1072                 codedbuf_size, 1, NULL, &coded_buf[i]);
1073         CHECK_VASTATUS(va_status,"vaCreateBuffer");
1074     }
1075
1076     return 0;
1077 }
1078
1079
1080
1081 #define partition(ref, field, key, ascending)   \
1082     while (i <= j) {                            \
1083         if (ascending) {                        \
1084             while (ref[i].field < key)          \
1085                 i++;                            \
1086             while (ref[j].field > key)          \
1087                 j--;                            \
1088         } else {                                \
1089             while (ref[i].field > key)          \
1090                 i++;                            \
1091             while (ref[j].field < key)          \
1092                 j--;                            \
1093         }                                       \
1094         if (i <= j) {                           \
1095             tmp = ref[i];                       \
1096             ref[i] = ref[j];                    \
1097             ref[j] = tmp;                       \
1098             i++;                                \
1099             j--;                                \
1100         }                                       \
1101     }                                           \
1102
1103 static void sort_one(VAPictureH264 ref[], int left, int right,
1104                      int ascending, int frame_idx)
1105 {
1106     int i = left, j = right;
1107     unsigned int key;
1108     VAPictureH264 tmp;
1109
1110     if (frame_idx) {
1111         key = ref[(left + right) / 2].frame_idx;
1112         partition(ref, frame_idx, key, ascending);
1113     } else {
1114         key = ref[(left + right) / 2].TopFieldOrderCnt;
1115         partition(ref, TopFieldOrderCnt, key, ascending);
1116     }
1117
1118     /* recursion */
1119     if (left < j)
1120         sort_one(ref, left, j, ascending, frame_idx);
1121
1122     if (i < right)
1123         sort_one(ref, i, right, ascending, frame_idx);
1124 }
1125
1126 static void sort_two(VAPictureH264 ref[], int left, int right, int key, int frame_idx,
1127                        int divide_ascending, int list0_ascending, int list1_ascending)
1128 {
1129     int i = left, j = right;
1130     VAPictureH264 tmp;
1131
1132     if (frame_idx) {
1133         partition(ref, frame_idx, key, divide_ascending);
1134     } else {
1135         partition(ref, TopFieldOrderCnt, key, divide_ascending);
1136     }
1137
1138
1139     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1140     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1141 }
1142
1143 static int update_ReferenceFrames(void)
1144 {
1145     int i;
1146
1147     if (current_frame_type == FRAME_B)
1148         return 0;
1149
1150     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1151     numShortTerm++;
1152     if (numShortTerm > num_ref_frames)
1153         numShortTerm = num_ref_frames;
1154     for (i=numShortTerm-1; i>0; i--)
1155         ReferenceFrames[i] = ReferenceFrames[i-1];
1156     ReferenceFrames[0] = CurrentCurrPic;
1157
1158     if (current_frame_type != FRAME_B)
1159         current_frame_num++;
1160     if (current_frame_num > MaxFrameNum)
1161         current_frame_num = 0;
1162
1163     return 0;
1164 }
1165
1166 static int update_RefPicList(void)
1167 {
1168     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1169
1170     if (current_frame_type == FRAME_IDR) {
1171         numShortTerm = 0;
1172         current_frame_num = 0;
1173         return 0;
1174     }
1175
1176     if (current_frame_type == FRAME_P) {
1177         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1178         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1179     }
1180
1181     if (current_frame_type == FRAME_B) {
1182         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1183         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1184                  1, 0, 1);
1185
1186         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1187         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1188                  0, 1, 0);
1189     }
1190
1191     return 0;
1192 }
1193
1194
1195 static int render_sequence(void)
1196 {
1197     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1198     VAStatus va_status;
1199     VAEncMiscParameterBuffer *misc_param;
1200     VAEncMiscParameterRateControl *misc_rate_ctrl;
1201
1202     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1203     seq_param.picture_width_in_mbs = frame_width / 16;
1204     seq_param.picture_height_in_mbs = frame_height / 16;
1205     seq_param.bits_per_second = frame_bitrate;
1206
1207     seq_param.intra_period = intra_period;
1208     seq_param.intra_idr_period = intra_idr_period;
1209     seq_param.ip_period = ip_period;
1210
1211     seq_param.max_num_ref_frames = num_ref_frames;
1212     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1213     seq_param.time_scale = 900;
1214     seq_param.num_units_in_tick = 15; /* Tc = num_units_in_tick / time_sacle */
1215     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1216     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1217     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1218
1219     va_status = vaCreateBuffer(va_dpy, context_id,
1220                                VAEncSequenceParameterBufferType,
1221                                sizeof(seq_param),1,&seq_param,&seq_param_buf);
1222     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1223
1224     va_status = vaCreateBuffer(va_dpy, context_id,
1225                                VAEncMiscParameterBufferType,
1226                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1227                                1,NULL,&rc_param_buf);
1228     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1229
1230     vaMapBuffer(va_dpy, rc_param_buf,(void **)&misc_param);
1231     misc_param->type = VAEncMiscParameterTypeRateControl;
1232     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1233     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1234     misc_rate_ctrl->bits_per_second = frame_bitrate;
1235     misc_rate_ctrl->initial_qp = initial_qp;
1236     misc_rate_ctrl->min_qp = minimal_qp;
1237     misc_rate_ctrl->basic_unit_size = 0;
1238     vaUnmapBuffer(va_dpy, rc_param_buf);
1239
1240     render_id[0] = seq_param_buf;
1241     render_id[1] = rc_param_buf;
1242
1243     va_status = vaRenderPicture(va_dpy,context_id, &render_id[0], 2);
1244     CHECK_VASTATUS(va_status,"vaRenderPicture");;
1245
1246     return 0;
1247 }
1248
1249 static unsigned int calc_poc(unsigned int pic_order_cnt_lsb)
1250 {
1251     static unsigned int prevPicOrderCntMsb = 0, prevPicOrderCntLsb = 0;
1252     static unsigned int PicOrderCntMsb = 0;
1253     unsigned int TopFieldOrderCnt;
1254
1255     if (current_frame_type == FRAME_IDR)
1256         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1257
1258     prevPicOrderCntMsb = PicOrderCntMsb;
1259     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1260         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (MaxPicOrderCntLsb / 2)))
1261         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1262     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1263              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (MaxPicOrderCntLsb / 2)))
1264         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1265     else
1266         PicOrderCntMsb = prevPicOrderCntMsb;
1267
1268     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1269
1270     return TopFieldOrderCnt;
1271 }
1272
1273 static int render_picture(void)
1274 {
1275     VABufferID pic_param_buf;
1276     VAStatus va_status;
1277     int i = 0;
1278
1279     pic_param.CurrPic.picture_id = ref_surface[current_slot];
1280     pic_param.CurrPic.frame_idx = current_frame_num;
1281     pic_param.CurrPic.flags = 0;
1282     pic_param.CurrPic.TopFieldOrderCnt = calc_poc(current_frame_display % MaxPicOrderCntLsb);
1283     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1284     CurrentCurrPic = pic_param.CurrPic;
1285
1286     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1287         update_RefPicList(); /* calc RefPicList */
1288         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1289         if (current_frame_type == FRAME_P) {
1290             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1291         } else if (current_frame_type == FRAME_B) {
1292             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1293             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1294         }
1295     } else {
1296         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1297         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1298             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1299             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1300         }
1301     }
1302
1303     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1304     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1305     pic_param.pic_fields.bits.entropy_coding_mode_flag = 1;
1306     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1307     pic_param.frame_num = current_frame_num;
1308     pic_param.coded_buf = coded_buf[current_slot];
1309     pic_param.last_picture = (current_frame_encoding == frame_count);
1310     pic_param.pic_init_qp = initial_qp;
1311
1312     va_status = vaCreateBuffer(va_dpy, context_id,VAEncPictureParameterBufferType,
1313                                sizeof(pic_param),1,&pic_param, &pic_param_buf);
1314     CHECK_VASTATUS(va_status,"vaCreateBuffer");;
1315
1316     va_status = vaRenderPicture(va_dpy,context_id, &pic_param_buf, 1);
1317     CHECK_VASTATUS(va_status,"vaRenderPicture");
1318
1319     return 0;
1320 }
1321
1322 static int render_packedsequence(void)
1323 {
1324     VAEncPackedHeaderParameterBuffer packedheader_param_buffer={0};
1325     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1326     unsigned int length_in_bits;
1327     unsigned char *packedseq_buffer = NULL;
1328     VAStatus va_status;
1329
1330     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1331
1332     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1333
1334     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1335     packedheader_param_buffer.has_emulation_bytes = 0;
1336     va_status = vaCreateBuffer(va_dpy,
1337                                context_id,
1338                                VAEncPackedHeaderParameterBufferType,
1339                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1340                                &packedseq_para_bufid);
1341     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1342
1343     va_status = vaCreateBuffer(va_dpy,
1344                                context_id,
1345                                VAEncPackedHeaderDataBufferType,
1346                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1347                                &packedseq_data_bufid);
1348     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1349
1350     render_id[0] = packedseq_para_bufid;
1351     render_id[1] = packedseq_data_bufid;
1352     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1353     CHECK_VASTATUS(va_status,"vaRenderPicture");
1354
1355     free(packedseq_buffer);
1356
1357     return 0;
1358 }
1359
1360
1361 static int render_packedpicture(void)
1362 {
1363     VAEncPackedHeaderParameterBuffer packedheader_param_buffer={0};
1364     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1365     unsigned int length_in_bits;
1366     unsigned char *packedpic_buffer = NULL;
1367     VAStatus va_status;
1368
1369     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1370     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1371     packedheader_param_buffer.bit_length = length_in_bits;
1372     packedheader_param_buffer.has_emulation_bytes = 0;
1373
1374     va_status = vaCreateBuffer(va_dpy,
1375                                context_id,
1376                                VAEncPackedHeaderParameterBufferType,
1377                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1378                                &packedpic_para_bufid);
1379     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1380
1381     va_status = vaCreateBuffer(va_dpy,
1382                                context_id,
1383                                VAEncPackedHeaderDataBufferType,
1384                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1385                                &packedpic_data_bufid);
1386     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1387
1388     render_id[0] = packedpic_para_bufid;
1389     render_id[1] = packedpic_data_bufid;
1390     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1391     CHECK_VASTATUS(va_status,"vaRenderPicture");
1392
1393     free(packedpic_buffer);
1394
1395     return 0;
1396 }
1397
1398 static void render_packedsei(void)
1399 {
1400     VAEncPackedHeaderParameterBuffer packed_header_param_buffer;
1401     VABufferID packed_sei_header_param_buf_id, packed_sei_buf_id, render_id[2];
1402     unsigned int length_in_bits, offset_in_bytes;
1403     unsigned char *packed_sei_buffer = NULL;
1404     VAStatus va_status;
1405     int init_cpb_size, target_bit_rate, i_initial_cpb_removal_delay_length, i_initial_cpb_removal_delay;
1406     int i_cpb_removal_delay, i_dpb_output_delay_length, i_cpb_removal_delay_length;
1407
1408     /* it comes for the bps defined in SPS */
1409     target_bit_rate = frame_bitrate;
1410     init_cpb_size = (target_bit_rate * 8) >> 10;
1411     i_initial_cpb_removal_delay = init_cpb_size * 0.5 * 1024 / target_bit_rate * 90000;
1412
1413     i_cpb_removal_delay = 2;
1414     i_initial_cpb_removal_delay_length = 24;
1415     i_cpb_removal_delay_length = 24;
1416     i_dpb_output_delay_length = 24;
1417
1418
1419     length_in_bits = build_packed_sei_buffer_timing(
1420         i_initial_cpb_removal_delay_length,
1421         i_initial_cpb_removal_delay,
1422         0,
1423         i_cpb_removal_delay_length,
1424         i_cpb_removal_delay * current_frame_encoding,
1425         i_dpb_output_delay_length,
1426         0,
1427         &packed_sei_buffer);
1428
1429     offset_in_bytes = 0;
1430     packed_header_param_buffer.type = VAEncPackedHeaderH264_SEI;
1431     packed_header_param_buffer.bit_length = length_in_bits;
1432     packed_header_param_buffer.has_emulation_bytes = 0;
1433
1434     va_status = vaCreateBuffer(va_dpy,
1435                                context_id,
1436                                VAEncPackedHeaderParameterBufferType,
1437                                sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
1438                                &packed_sei_header_param_buf_id);
1439     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1440
1441     va_status = vaCreateBuffer(va_dpy,
1442                                context_id,
1443                                VAEncPackedHeaderDataBufferType,
1444                                (length_in_bits + 7) / 8, 1, packed_sei_buffer,
1445                                &packed_sei_buf_id);
1446     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1447
1448
1449     render_id[0] = packed_sei_header_param_buf_id;
1450     render_id[1] = packed_sei_buf_id;
1451     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1452     CHECK_VASTATUS(va_status,"vaRenderPicture");
1453
1454
1455     free(packed_sei_buffer);
1456
1457     return;
1458 }
1459
1460
1461 static int render_hrd(void)
1462 {
1463     VABufferID misc_parameter_hrd_buf_id;
1464     VAStatus va_status;
1465     VAEncMiscParameterBuffer *misc_param;
1466     VAEncMiscParameterHRD *misc_hrd_param;
1467
1468     va_status = vaCreateBuffer(va_dpy, context_id,
1469                    VAEncMiscParameterBufferType,
1470                    sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterHRD),
1471                    1,
1472                    NULL,
1473                    &misc_parameter_hrd_buf_id);
1474     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1475
1476     vaMapBuffer(va_dpy,
1477                 misc_parameter_hrd_buf_id,
1478                 (void **)&misc_param);
1479     misc_param->type = VAEncMiscParameterTypeHRD;
1480     misc_hrd_param = (VAEncMiscParameterHRD *)misc_param->data;
1481
1482     if (frame_bitrate > 0) {
1483         misc_hrd_param->initial_buffer_fullness = frame_bitrate * 1024 * 4;
1484         misc_hrd_param->buffer_size = frame_bitrate * 1024 * 8;
1485     } else {
1486         misc_hrd_param->initial_buffer_fullness = 0;
1487         misc_hrd_param->buffer_size = 0;
1488     }
1489     vaUnmapBuffer(va_dpy, misc_parameter_hrd_buf_id);
1490
1491     va_status = vaRenderPicture(va_dpy,context_id, &misc_parameter_hrd_buf_id, 1);
1492     CHECK_VASTATUS(va_status,"vaRenderPicture");;
1493
1494     return 0;
1495 }
1496
1497 static int render_slice(void)
1498 {
1499     VABufferID slice_param_buf;
1500     VAStatus va_status;
1501     int i;
1502
1503     update_RefPicList();
1504
1505     /* one frame, one slice */
1506     slice_param.macroblock_address = 0;
1507     slice_param.num_macroblocks = frame_width*frame_height/(16*16); /* Measured by MB */
1508     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1509     if (current_frame_type == FRAME_I || current_frame_type == FRAME_IDR) {
1510         ;
1511     } else if (current_frame_type == FRAME_P) {
1512         int refpiclist0_max = h264_maxref & 0xffff;
1513         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1514
1515         for (i = refpiclist0_max; i < 32; i++) {
1516             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1517             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1518         }
1519     } else if (current_frame_type == FRAME_B) {
1520         int refpiclist0_max = h264_maxref & 0xffff;
1521         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1522
1523         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1524         for (i = refpiclist0_max; i < 32; i++) {
1525             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1526             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1527         }
1528
1529         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1530         for (i = refpiclist1_max; i < 32; i++) {
1531             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1532             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1533         }
1534     }
1535
1536     slice_param.slice_alpha_c0_offset_div2 = 2;
1537     slice_param.slice_beta_offset_div2 = 2;
1538     slice_param.pic_order_cnt_lsb = current_frame_display % MaxPicOrderCntLsb;
1539
1540     va_status = vaCreateBuffer(va_dpy,context_id,VAEncSliceParameterBufferType,
1541                                sizeof(slice_param),1,&slice_param,&slice_param_buf);
1542     CHECK_VASTATUS(va_status,"vaCreateBuffer");;
1543
1544     va_status = vaRenderPicture(va_dpy,context_id, &slice_param_buf, 1);
1545     CHECK_VASTATUS(va_status,"vaRenderPicture");
1546
1547     return 0;
1548 }
1549
1550
1551 static int upload_source_YUV_once_for_all()
1552 {
1553     int box_width=8;
1554     int row_shift=0;
1555     int i;
1556
1557     for (i = 0; i < SURFACE_NUM; i++) {
1558         printf("\rLoading data into surface %d.....", i);
1559         upload_surface(va_dpy, src_surface[i], box_width, row_shift, 0);
1560
1561         row_shift++;
1562         if (row_shift==(2*box_width)) row_shift= 0;
1563     }
1564     printf("Completed surface loading\n");
1565
1566     return 0;
1567 }
1568
1569
1570 static int load_surface(VASurfaceID surface_id, unsigned long long display_order)
1571 {
1572     VAImage surface_image;
1573     unsigned char *surface_p=NULL, *Y_start=NULL, *U_start=NULL,*V_start=NULL, *uv_ptr;
1574     int Y_pitch=0, U_pitch=0, row, V_pitch, uv_size;
1575     VAStatus va_status;
1576
1577     if (srcyuv_fp == NULL)
1578         return 0;
1579
1580     /* rewind the file pointer if encoding more than srcyuv_frames */
1581     display_order = display_order % srcyuv_frames;
1582
1583     fseek(srcyuv_fp, display_order * frame_width * frame_height * 1.5, SEEK_SET);
1584
1585     va_status = vaDeriveImage(va_dpy,surface_id, &surface_image);
1586     CHECK_VASTATUS(va_status,"vaDeriveImage");
1587
1588     vaMapBuffer(va_dpy,surface_image.buf,(void **)&surface_p);
1589     assert(VA_STATUS_SUCCESS == va_status);
1590
1591     Y_start = surface_p;
1592     Y_pitch = surface_image.pitches[0];
1593     switch (surface_image.format.fourcc) {
1594     case VA_FOURCC_NV12:
1595         U_start = (unsigned char *)surface_p + surface_image.offsets[1];
1596         V_start = U_start + 1;
1597         U_pitch = surface_image.pitches[1];
1598         V_pitch = surface_image.pitches[1];
1599         break;
1600     case VA_FOURCC_IYUV:
1601         U_start = (unsigned char *)surface_p + surface_image.offsets[1];
1602         V_start = (unsigned char *)surface_p + surface_image.offsets[2];
1603         U_pitch = surface_image.pitches[1];
1604         V_pitch = surface_image.pitches[2];
1605         break;
1606     case VA_FOURCC_YV12:
1607         U_start = (unsigned char *)surface_p + surface_image.offsets[2];
1608         V_start = (unsigned char *)surface_p + surface_image.offsets[1];
1609         U_pitch = surface_image.pitches[2];
1610         V_pitch = surface_image.pitches[1];
1611         break;
1612     case VA_FOURCC_YUY2:
1613         U_start = surface_p + 1;
1614         V_start = surface_p + 3;
1615         U_pitch = surface_image.pitches[0];
1616         V_pitch = surface_image.pitches[0];
1617         break;
1618     default:
1619         assert(0);
1620     }
1621
1622     /* copy Y plane */
1623     for (row=0;row<surface_image.height;row++) {
1624         unsigned char *Y_row = Y_start + row * Y_pitch;
1625         (void)fread(Y_row, 1, surface_image.width, srcyuv_fp);
1626     }
1627
1628     /* copy UV data, reset file pointer,
1629      * surface_image.height may not be equal to source YUV height/frame_height
1630      */
1631     fseek(srcyuv_fp,
1632           display_order * frame_width * frame_height * 1.5 + frame_width * frame_height,
1633           SEEK_SET);
1634
1635     uv_size = 2 * (frame_width/2) * (frame_height/2);
1636     uv_ptr = malloc(uv_size);
1637     fread(uv_ptr, uv_size, 1, srcyuv_fp);
1638
1639     for (row =0; row < surface_image.height/2; row++) {
1640         unsigned char *U_row = U_start + row * U_pitch;
1641         unsigned char *u_ptr, *v_ptr;
1642         int j;
1643         switch (surface_image.format.fourcc) {
1644         case VA_FOURCC_NV12:
1645             if (srcyuv_fourcc == VA_FOURCC_NV12) {
1646                 memcpy(U_row, uv_ptr + row * frame_width, frame_width);
1647                 break;
1648             } else if (srcyuv_fourcc == VA_FOURCC_IYUV) {
1649                 u_ptr = uv_ptr + row * (frame_width/2);
1650                 v_ptr = uv_ptr + (frame_width/2) * (frame_height/2) + row * (frame_width/2);
1651             } else if (srcyuv_fourcc == VA_FOURCC_YV12) {
1652                 v_ptr = uv_ptr + row * (frame_height/2);
1653                 u_ptr = uv_ptr + (frame_width/2) * (frame_height/2) + row * (frame_width/2);
1654             }
1655             for(j = 0; j < frame_width/2; j++) {
1656                 U_row[2*j] = u_ptr[j];
1657                 U_row[2*j+1] = v_ptr[j];
1658             }
1659             break;
1660         case VA_FOURCC_IYUV:
1661         case VA_FOURCC_YV12:
1662         case VA_FOURCC_YUY2:
1663         default:
1664             printf("unsupported fourcc in load_surface\n");
1665             assert(0);
1666         }
1667     }
1668     free(uv_ptr);
1669
1670     vaUnmapBuffer(va_dpy,surface_image.buf);
1671
1672     vaDestroyImage(va_dpy,surface_image.image_id);
1673
1674     return 0;
1675 }
1676
1677
1678 static int save_codeddata(unsigned long long display_order, unsigned long long encode_order)
1679 {
1680     VACodedBufferSegment *buf_list = NULL;
1681     VAStatus va_status;
1682     unsigned int coded_size = 0;
1683
1684     va_status = vaMapBuffer(va_dpy,coded_buf[display_order % SURFACE_NUM],(void **)(&buf_list));
1685     CHECK_VASTATUS(va_status,"vaMapBuffer");
1686     while (buf_list != NULL) {
1687         coded_size += fwrite(buf_list->buf, 1, buf_list->size, coded_fp);
1688         buf_list = (VACodedBufferSegment *) buf_list->next;
1689     }
1690     vaUnmapBuffer(va_dpy,coded_buf[display_order % SURFACE_NUM]);
1691
1692     printf("\r      "); /* return back to startpoint */
1693     switch (encode_order % 4) {
1694         case 0:
1695             printf("|");
1696             break;
1697         case 1:
1698             printf("/");
1699             break;
1700         case 2:
1701             printf("-");
1702             break;
1703         case 3:
1704             printf("\\");
1705             break;
1706     }
1707     printf("%08lld", encode_order);
1708     /*
1709     if (current_frame_encoding % intra_count == 0)
1710         printf("(I)");
1711     else
1712         printf("(P)");
1713     */
1714     printf("(%06d bytes coded)",coded_size);
1715     /* skipped frame ? */
1716     printf("                                    ");
1717
1718     return 0;
1719 }
1720
1721
1722 static struct storage_task_t * storage_task_dequque(void)
1723 {
1724     struct storage_task_t *header;
1725
1726     pthread_mutex_lock(&encode_mutex);
1727
1728     header = storage_task_header;
1729     if (storage_task_header != NULL) {
1730         if (storage_task_tail == storage_task_header)
1731             storage_task_tail = NULL;
1732         storage_task_header = header->next;
1733     }
1734
1735     pthread_mutex_unlock(&encode_mutex);
1736
1737     return header;
1738 }
1739
1740 static int storage_task_queue(unsigned long long display_order, unsigned long long encode_order)
1741 {
1742     struct storage_task_t *tmp;
1743
1744     tmp = calloc(1, sizeof(struct storage_task_t));
1745     tmp->display_order = display_order;
1746     tmp->encode_order = encode_order;
1747
1748     pthread_mutex_lock(&encode_mutex);
1749
1750     if (storage_task_header == NULL) {
1751         storage_task_header = tmp;
1752         storage_task_tail = tmp;
1753     } else {
1754         storage_task_tail->next = tmp;
1755         storage_task_tail = tmp;
1756     }
1757
1758     srcsurface_status[display_order % SURFACE_NUM] = SRC_SURFACE_IN_STORAGE;
1759     pthread_cond_signal(&encode_cond);
1760
1761     pthread_mutex_unlock(&encode_mutex);
1762
1763     return 0;
1764 }
1765
1766 static void storage_task(unsigned long long display_order, unsigned long encode_order)
1767 {
1768     unsigned int tmp;
1769     VAStatus va_status;
1770
1771     tmp = GetTickCount();
1772     va_status = vaSyncSurface(va_dpy, src_surface[display_order % SURFACE_NUM]);
1773     CHECK_VASTATUS(va_status,"vaSyncSurface");
1774     SyncPictureTicks += GetTickCount() - tmp;
1775     tmp = GetTickCount();
1776     save_codeddata(display_order, encode_order);
1777     SavePictureTicks += GetTickCount() - tmp;
1778     /* tbd: save reconstructed frame */
1779
1780     /* reload a new frame data */
1781     tmp = GetTickCount();
1782     if (srcyuv_fp != NULL)
1783         load_surface(src_surface[display_order % SURFACE_NUM], display_order + SURFACE_NUM);
1784     UploadPictureTicks += GetTickCount() - tmp;
1785
1786     pthread_mutex_lock(&encode_mutex);
1787     srcsurface_status[display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1788     pthread_mutex_unlock(&encode_mutex);
1789 }
1790
1791
1792 static void * storage_task_thread(void *t)
1793 {
1794     while (1) {
1795         struct storage_task_t *current;
1796
1797         current = storage_task_dequque();
1798         if (current == NULL) {
1799             pthread_mutex_lock(&encode_mutex);
1800             pthread_cond_wait(&encode_cond, &encode_mutex);
1801             pthread_mutex_unlock(&encode_mutex);
1802             continue;
1803         }
1804
1805         storage_task(current->display_order, current->encode_order);
1806
1807         free(current);
1808
1809         /* all frames are saved, exit the thread */
1810         if (++frame_coded >= frame_count)
1811             break;
1812     }
1813
1814     return 0;
1815 }
1816
1817
1818 static int encode_frames(void)
1819 {
1820     unsigned int i, tmp;
1821     VAStatus va_status;
1822     //VASurfaceStatus surface_status;
1823
1824     /* upload RAW YUV data into all surfaces */
1825     tmp = GetTickCount();
1826     if (srcyuv_fp != NULL) {
1827         for (i = 0; i < SURFACE_NUM; i++)
1828             load_surface(src_surface[i], i);
1829     } else
1830         upload_source_YUV_once_for_all();
1831     UploadPictureTicks += GetTickCount() - tmp;
1832
1833     /* ready for encoding */
1834     memset(srcsurface_status, SRC_SURFACE_IN_ENCODING, sizeof(srcsurface_status));
1835
1836     memset(&seq_param, 0, sizeof(seq_param));
1837     memset(&pic_param, 0, sizeof(pic_param));
1838     memset(&slice_param, 0, sizeof(slice_param));
1839
1840     if (encode_syncmode == 0)
1841         pthread_create(&encode_thread, NULL, storage_task_thread, NULL);
1842
1843     for (current_frame_encoding = 0; current_frame_encoding < frame_count; current_frame_encoding++) {
1844         encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1845                                &current_frame_display, &current_frame_type);
1846
1847         /* check if the source frame is ready */
1848         while (srcsurface_status[current_slot] != SRC_SURFACE_IN_ENCODING);
1849
1850         tmp = GetTickCount();
1851         va_status = vaBeginPicture(va_dpy, context_id, src_surface[current_slot]);
1852         CHECK_VASTATUS(va_status,"vaBeginPicture");
1853         BeginPictureTicks += GetTickCount() - tmp;
1854
1855         tmp = GetTickCount();
1856         if (current_frame_encoding  == 0) {
1857             render_sequence();
1858             render_picture();
1859             if (h264_packedheader) {
1860                 render_packedsequence();
1861                 render_packedpicture();
1862             }
1863             //if (rc_mode == VA_RC_CBR)
1864             //    render_packedsei();
1865             //render_hrd();
1866         } else {
1867             //render_sequence();
1868             render_picture();
1869             //if (rc_mode == VA_RC_CBR)
1870             //    render_packedsei();
1871             //render_hrd();
1872         }
1873         render_slice();
1874         RenderPictureTicks += GetTickCount() - tmp;
1875
1876         tmp = GetTickCount();
1877         va_status = vaEndPicture(va_dpy,context_id);
1878         CHECK_VASTATUS(va_status,"vaEndPicture");;
1879         EndPictureTicks += GetTickCount() - tmp;
1880
1881         if (encode_syncmode)
1882             storage_task(current_frame_display, current_frame_encoding);
1883         else /* queue the storage task queue */
1884             storage_task_queue(current_frame_display, current_frame_encoding);
1885
1886         /* how to process skipped frames
1887            surface_status = (VASurfaceStatus) 0;
1888            va_status = vaQuerySurfaceStatus(va_dpy, src_surface[i%SURFACE_NUM],&surface_status);
1889            frame_skipped = (surface_status & VASurfaceSkipped);
1890         */
1891         update_ReferenceFrames();
1892     }
1893
1894     if (encode_syncmode == 0) {
1895         int ret;
1896         pthread_join(encode_thread, (void **)&ret);
1897     }
1898
1899     return 0;
1900 }
1901
1902
1903 static int release_encode()
1904 {
1905     int i;
1906
1907     vaDestroySurfaces(va_dpy,&src_surface[0],SURFACE_NUM);
1908     vaDestroySurfaces(va_dpy,&ref_surface[0],SURFACE_NUM);
1909
1910     for (i = 0; i < SURFACE_NUM; i++)
1911         vaDestroyBuffer(va_dpy,coded_buf[i]);
1912
1913     vaDestroyContext(va_dpy,context_id);
1914     vaDestroyConfig(va_dpy,config_id);
1915
1916     return 0;
1917 }
1918
1919 static int deinit_va()
1920 {
1921     vaTerminate(va_dpy);
1922
1923     va_close_display(va_dpy);
1924
1925     return 0;
1926 }
1927
1928
1929 static int print_input()
1930 {
1931     printf("\n\nINPUT:Try to encode H264...\n");
1932     printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1933     printf("INPUT: Resolution   : %dx%d, %d frames\n",
1934            frame_width, frame_height, frame_count);
1935     printf("INPUT: FrameRate    : %d\n", frame_rate);
1936     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1937     printf("INPUT: Slieces      : %d\n", frame_slices);
1938     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1939     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1940     printf("INPUT: IpPeriod     : %d\n", ip_period);
1941     printf("INPUT: Initial QP   : %d\n", initial_qp);
1942     printf("INPUT: Min QP       : %d\n", minimal_qp);
1943     printf("INPUT: Source YUV   : %s", srcyuv_fp?"FILE":"AUTO generated");
1944     if (srcyuv_fp)
1945         printf(":%s (fourcc %s)\n", srcyuv_fn, fourcc_to_string(srcyuv_fourcc));
1946     else
1947         printf("\n");
1948     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1949     if (recyuv_fp == NULL)
1950         printf("INPUT: Rec   Clip   : %s\n", "Not save reconstructed frame");
1951     else
1952         printf("INPUT: Rec   Clip   : Save reconstructed frame into %s (fourcc %s)\n", recyuv_fn,
1953                fourcc_to_string(srcyuv_fourcc));
1954
1955     printf("\n\n"); /* return back to startpoint */
1956
1957     return 0;
1958 }
1959
1960
1961 static int print_performance(unsigned int PictureCount)
1962 {
1963     unsigned int others = 0;
1964
1965     others = TotalTicks - UploadPictureTicks - BeginPictureTicks
1966         - RenderPictureTicks - EndPictureTicks - SyncPictureTicks - SavePictureTicks;
1967
1968     printf("\n\n");
1969
1970     printf("PERFORMANCE:   Frame Rate           : %.2f fps (%d frames, %d ms (%.2f ms per frame))\n",
1971            (double) 1000*PictureCount / TotalTicks, PictureCount,
1972            TotalTicks, ((double)  TotalTicks) / (double) PictureCount);
1973
1974     printf("PERFORMANCE:     UploadPicture      : %d ms (%.2f, %.2f%% percent)\n",
1975            (int) UploadPictureTicks, ((double)  UploadPictureTicks) / (double) PictureCount,
1976            UploadPictureTicks/(double) TotalTicks/0.01);
1977     printf("PERFORMANCE:     vaBeginPicture     : %d ms (%.2f, %.2f%% percent)\n",
1978            (int) BeginPictureTicks, ((double)  BeginPictureTicks) / (double) PictureCount,
1979            BeginPictureTicks/(double) TotalTicks/0.01);
1980     printf("PERFORMANCE:     vaRenderHeader     : %d ms (%.2f, %.2f%% percent)\n",
1981            (int) RenderPictureTicks, ((double)  RenderPictureTicks) / (double) PictureCount,
1982            RenderPictureTicks/(double) TotalTicks/0.01);
1983     printf("PERFORMANCE:     vaEndPicture       : %d ms (%.2f, %.2f%% percent)\n",
1984            (int) EndPictureTicks, ((double)  EndPictureTicks) / (double) PictureCount,
1985            EndPictureTicks/(double) TotalTicks/0.01);
1986     printf("PERFORMANCE:     vaSyncSurface      : %d ms (%.2f, %.2f%% percent)\n",
1987            (int) SyncPictureTicks, ((double) SyncPictureTicks) / (double) PictureCount,
1988            SyncPictureTicks/(double) TotalTicks/0.01);
1989     printf("PERFORMANCE:     SavePicture        : %d ms (%.2f, %.2f%% percent)\n",
1990            (int) SavePictureTicks, ((double)  SavePictureTicks) / (double) PictureCount,
1991            SavePictureTicks/(double) TotalTicks/0.01);
1992     printf("PERFORMANCE:     Others             : %d ms (%.2f, %.2f%% percent)\n",
1993            (int) others, ((double) others) / (double) PictureCount,
1994            others/(double) TotalTicks/0.01);
1995
1996     if (encode_syncmode == 0)
1997         printf("(Multithread enabled, the profiling is only for reference)\n");
1998
1999     return 0;
2000 }
2001
2002
2003 int main(int argc,char **argv)
2004 {
2005     unsigned int start;
2006
2007     process_cmdline(argc, argv);
2008
2009     print_input();
2010
2011     start = GetTickCount();
2012
2013     init_va();
2014     setup_encode();
2015
2016     encode_frames();
2017
2018     release_encode();
2019     deinit_va();
2020
2021     TotalTicks += GetTickCount() - start;
2022     print_performance(frame_count);
2023
2024     return 0;
2025 }