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 <sys/mman.h>
  34 #include <fcntl.h>
  35 #include <assert.h>
  36 #include <pthread.h>
  37 #include <errno.h>
  38 #include <math.h>
  39 #include <va/va.h>
  40 #include <va/va_enc_h264.h>
  41 #include "va_display.h"
  42
  43 #define CHECK_VASTATUS(va_status,func)                                  \
  44     if (va_status != VA_STATUS_SUCCESS) {                               \
  45         fprintf(stderr,"%s:%s (%d) failed,exit\n", __func__, func, __LINE__); \
  46         exit(1);                                                        \
  47     }
  48
  49 #include "../loadsurface.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 ENTROPY_MODE_CAVLC      0
  67 #define ENTROPY_MODE_CABAC      1
  68
  69 #define PROFILE_IDC_BASELINE    66
  70 #define PROFILE_IDC_MAIN        77
  71 #define PROFILE_IDC_HIGH        100
  72
  73 #define BITSTREAM_ALLOCATE_STEPPING     4096
  74
  75 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  76 #define SURFACE_NUM 16 /* 16 surfaces for reference */
  77 static  VADisplay va_dpy;
  78 static  VAProfile h264_profile;
  79 static  VAConfigAttrib attrib[VAConfigAttribTypeMax];
  80 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  81 static  int config_attrib_num = 0;
  82 static  VASurfaceID src_surface[SURFACE_NUM];
  83 static  VABufferID  coded_buf[SURFACE_NUM];
  84 static  VASurfaceID ref_surface[SURFACE_NUM];
  85 static  VAConfigID config_id;
  86 static  VAContextID context_id;
  87 static  VAEncSequenceParameterBufferH264 seq_param;
  88 static  VAEncPictureParameterBufferH264 pic_param;
  89 static  VAEncSliceParameterBufferH264 slice_param;
  90 static  VAPictureH264 CurrentCurrPic;
  91 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  92
  93 static  unsigned int MaxFrameNum = (2<<16);
  94 static  unsigned int MaxPicOrderCntLsb = (2<<8);
  95 static  unsigned int Log2MaxFrameNum = 16;
  96 static  unsigned int Log2MaxPicOrderCntLsb = 8;
  97
  98 static  unsigned int num_ref_frames = 2;
  99 static  unsigned int numShortTerm = 0;
 100 static  int constraint_set_flag = 0;
 101 static  int h264_packedheader = 0; /* support pack header? */
 102 static  int h264_maxref = (1<<16|1);
 103 static  char *coded_fn = NULL, *srcyuv_fn = NULL, *recyuv_fn = NULL;
 104 static  FILE *coded_fp = NULL, *srcyuv_fp = NULL, *recyuv_fp = NULL;
 105 static  unsigned long long srcyuv_frames = 0;
 106 static  int srcyuv_fourcc = VA_FOURCC_NV12;
 107 static  int calc_psnr = 0;
 108
 109 static  int frame_width = 176;
 110 static  int frame_height = 144;
 111 static  int frame_rate = 30;
 112 static  unsigned int frame_count = 60;
 113 static  unsigned int frame_coded = 0;
 114 static  unsigned int frame_bitrate = 0;
 115 static  unsigned int frame_slices = 1;
 116 static  double frame_size = 0;
 117 static  int initial_qp = 28;
 118 static  int minimal_qp = 0;
 119 static  int intra_period = 30;
 120 static  int intra_idr_period = 60;
 121 static  int ip_period = 1;
 122 static  int rc_mode = VA_RC_VBR;
 123 static  unsigned long long current_frame_encoding = 0;
 124 static  unsigned long long current_frame_display = 0;
 125 static  unsigned long long current_IDR_display = 0;
 126 static  unsigned int current_frame_num = 0;
 127 static  int current_frame_type;
 128 #define current_slot (current_frame_display % SURFACE_NUM)
 129
 130 #define MIN(a, b) ((a)>(b)?(b):(a))
 131 #define MAX(a, b) ((a)>(b)?(a):(b))
 132
 133 /* thread to save coded data/upload source YUV */
 134 struct storage_task_t {
 135     void *next;
 136     unsigned long long display_order;
 137     unsigned long long encode_order;
 138 };
 139 static  struct storage_task_t *storage_task_header = NULL, *storage_task_tail = NULL;
 140 #define SRC_SURFACE_IN_ENCODING 0
 141 #define SRC_SURFACE_IN_STORAGE  1
 142 static  int srcsurface_status[SURFACE_NUM];
 143 static  int encode_syncmode = 0;
 144 static  pthread_mutex_t encode_mutex = PTHREAD_MUTEX_INITIALIZER;
 145 static  pthread_cond_t  encode_cond = PTHREAD_COND_INITIALIZER;
 146 static  pthread_t encode_thread;
 147
 148 /* for performance profiling */
 149 static unsigned int UploadPictureTicks=0;
 150 static unsigned int BeginPictureTicks=0;
 151 static unsigned int RenderPictureTicks=0;
 152 static unsigned int EndPictureTicks=0;
 153 static unsigned int SyncPictureTicks=0;
 154 static unsigned int SavePictureTicks=0;
 155 static unsigned int TotalTicks=0;
 156
 157 struct __bitstream {
 158     unsigned int *buffer;
 159     int bit_offset;
 160     int max_size_in_dword;
 161 };
 162 typedef struct __bitstream bitstream;
 163
 164
 165 static unsigned int
 166 va_swap32(unsigned int val)
 167 {
 168     unsigned char *pval = (unsigned char *)&val;
 169
 170     return ((pval[0] << 24)     |
 171             (pval[1] << 16)     |
 172             (pval[2] << 8)      |
 173             (pval[3] << 0));
 174 }
 175
 176 static void
 177 bitstream_start(bitstream *bs)
 178 {
 179     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 180     bs->buffer = calloc(bs->max_size_in_dword * sizeof(int), 1);
 181     bs->bit_offset = 0;
 182 }
 183
 184 static void
 185 bitstream_end(bitstream *bs)
 186 {
 187     int pos = (bs->bit_offset >> 5);
 188     int bit_offset = (bs->bit_offset & 0x1f);
 189     int bit_left = 32 - bit_offset;
 190
 191     if (bit_offset) {
 192         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 193     }
 194 }
 195
 196 static void
 197 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 198 {
 199     int pos = (bs->bit_offset >> 5);
 200     int bit_offset = (bs->bit_offset & 0x1f);
 201     int bit_left = 32 - bit_offset;
 202
 203     if (!size_in_bits)
 204         return;
 205
 206     bs->bit_offset += size_in_bits;
 207
 208     if (bit_left > size_in_bits) {
 209         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 210     } else {
 211         size_in_bits -= bit_left;
 212         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 213         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 214
 215         if (pos + 1 == bs->max_size_in_dword) {
 216             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 217             bs->buffer = realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 218         }
 219
 220         bs->buffer[pos + 1] = val;
 221     }
 222 }
 223
 224 static void
 225 bitstream_put_ue(bitstream *bs, unsigned int val)
 226 {
 227     int size_in_bits = 0;
 228     int tmp_val = ++val;
 229
 230     while (tmp_val) {
 231         tmp_val >>= 1;
 232         size_in_bits++;
 233     }
 234
 235     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 236     bitstream_put_ui(bs, val, size_in_bits);
 237 }
 238
 239 static void
 240 bitstream_put_se(bitstream *bs, int val)
 241 {
 242     unsigned int new_val;
 243
 244     if (val <= 0)
 245         new_val = -2 * val;
 246     else
 247         new_val = 2 * val - 1;
 248
 249     bitstream_put_ue(bs, new_val);
 250 }
 251
 252 static void
 253 bitstream_byte_aligning(bitstream *bs, int bit)
 254 {
 255     int bit_offset = (bs->bit_offset & 0x7);
 256     int bit_left = 8 - bit_offset;
 257     int new_val;
 258
 259     if (!bit_offset)
 260         return;
 261
 262     assert(bit == 0 || bit == 1);
 263
 264     if (bit)
 265         new_val = (1 << bit_left) - 1;
 266     else
 267         new_val = 0;
 268
 269     bitstream_put_ui(bs, new_val, bit_left);
 270 }
 271
 272 static void
 273 rbsp_trailing_bits(bitstream *bs)
 274 {
 275     bitstream_put_ui(bs, 1, 1);
 276     bitstream_byte_aligning(bs, 0);
 277 }
 278
 279 static void nal_start_code_prefix(bitstream *bs)
 280 {
 281     bitstream_put_ui(bs, 0x00000001, 32);
 282 }
 283
 284 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 285 {
 286     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 287     bitstream_put_ui(bs, nal_ref_idc, 2);
 288     bitstream_put_ui(bs, nal_unit_type, 5);
 289 }
 290
 291 static void sps_rbsp(bitstream *bs)
 292 {
 293     int profile_idc = PROFILE_IDC_BASELINE;
 294
 295     if (h264_profile  == VAProfileH264High)
 296         profile_idc = PROFILE_IDC_HIGH;
 297     else if (h264_profile  == VAProfileH264Main)
 298         profile_idc = PROFILE_IDC_MAIN;
 299
 300     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 301     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 302     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 303     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 304     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 305     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 306     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 307     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 308
 309     if ( profile_idc == PROFILE_IDC_HIGH) {
 310         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 311         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 312         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 313         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 314         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 315     }
 316
 317     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 318     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 319
 320     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 321         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 322     else {
 323         assert(0);
 324     }
 325
 326     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 327     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 328
 329     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 330     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 331     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 332
 333     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 334         assert(0);
 335     }
 336
 337     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 338     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 339
 340     if (seq_param.frame_cropping_flag) {
 341         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 342         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 343         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 344         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 345     }
 346
 347     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 348     if ( 1 ) {
 349         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 350     } else {
 351         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 352         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 353         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 354         bitstream_put_ui(bs, 0, 1); /* video_signal_type_present_flag */
 355         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 356         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 357         {
 358             bitstream_put_ui(bs, 15, 32);
 359             bitstream_put_ui(bs, 900, 32);
 360             bitstream_put_ui(bs, 1, 1);
 361         }
 362         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 363         {
 364             // hrd_parameters
 365             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 366             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 367             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 368
 369             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 370             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 371             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 372
 373             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 374             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 375             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 376             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 377         }
 378         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 379         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 380
 381         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 382         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 383     }
 384
 385     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 386 }
 387
 388
 389 static void pps_rbsp(bitstream *bs)
 390 {
 391     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 392     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 393
 394     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 395
 396     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 397
 398     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 399
 400     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 401     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 402
 403     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 404     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 405
 406     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 407     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 408     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 409
 410     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 411     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 412     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 413
 414     /* more_rbsp_data */
 415     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 416     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 417     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 418
 419     rbsp_trailing_bits(bs);
 420 }
 421
 422
 423 static int
 424 build_packed_pic_buffer(unsigned char **header_buffer)
 425 {
 426     bitstream bs;
 427
 428     bitstream_start(&bs);
 429     nal_start_code_prefix(&bs);
 430     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 431     pps_rbsp(&bs);
 432     bitstream_end(&bs);
 433
 434     *header_buffer = (unsigned char *)bs.buffer;
 435     return bs.bit_offset;
 436 }
 437
 438 static int
 439 build_packed_seq_buffer(unsigned char **header_buffer)
 440 {
 441     bitstream bs;
 442
 443     bitstream_start(&bs);
 444     nal_start_code_prefix(&bs);
 445     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 446     sps_rbsp(&bs);
 447     bitstream_end(&bs);
 448
 449     *header_buffer = (unsigned char *)bs.buffer;
 450     return bs.bit_offset;
 451 }
 452
 453 static int
 454 build_packed_sei_buffer_timing(unsigned int init_cpb_removal_length,
 455                                 unsigned int init_cpb_removal_delay,
 456                                 unsigned int init_cpb_removal_delay_offset,
 457                                 unsigned int cpb_removal_length,
 458                                 unsigned int cpb_removal_delay,
 459                                 unsigned int dpb_output_length,
 460                                 unsigned int dpb_output_delay,
 461                                 unsigned char **sei_buffer)
 462 {
 463     unsigned char *byte_buf;
 464     int bp_byte_size, i, pic_byte_size;
 465
 466     bitstream nal_bs;
 467     bitstream sei_bp_bs, sei_pic_bs;
 468
 469     bitstream_start(&sei_bp_bs);
 470     bitstream_put_ue(&sei_bp_bs, 0);       /*seq_parameter_set_id*/
 471     bitstream_put_ui(&sei_bp_bs, init_cpb_removal_delay, cpb_removal_length);
 472     bitstream_put_ui(&sei_bp_bs, init_cpb_removal_delay_offset, cpb_removal_length);
 473     if ( sei_bp_bs.bit_offset & 0x7) {
 474         bitstream_put_ui(&sei_bp_bs, 1, 1);
 475     }
 476     bitstream_end(&sei_bp_bs);
 477     bp_byte_size = (sei_bp_bs.bit_offset + 7) / 8;
 478
 479     bitstream_start(&sei_pic_bs);
 480     bitstream_put_ui(&sei_pic_bs, cpb_removal_delay, cpb_removal_length);
 481     bitstream_put_ui(&sei_pic_bs, dpb_output_delay, dpb_output_length);
 482     if ( sei_pic_bs.bit_offset & 0x7) {
 483         bitstream_put_ui(&sei_pic_bs, 1, 1);
 484     }
 485     bitstream_end(&sei_pic_bs);
 486     pic_byte_size = (sei_pic_bs.bit_offset + 7) / 8;
 487
 488     bitstream_start(&nal_bs);
 489     nal_start_code_prefix(&nal_bs);
 490     nal_header(&nal_bs, NAL_REF_IDC_NONE, NAL_SEI);
 491
 492         /* Write the SEI buffer period data */
 493     bitstream_put_ui(&nal_bs, 0, 8);
 494     bitstream_put_ui(&nal_bs, bp_byte_size, 8);
 495
 496     byte_buf = (unsigned char *)sei_bp_bs.buffer;
 497     for(i = 0; i < bp_byte_size; i++) {
 498         bitstream_put_ui(&nal_bs, byte_buf[i], 8);
 499     }
 500     free(byte_buf);
 501         /* write the SEI timing data */
 502     bitstream_put_ui(&nal_bs, 0x01, 8);
 503     bitstream_put_ui(&nal_bs, pic_byte_size, 8);
 504
 505     byte_buf = (unsigned char *)sei_pic_bs.buffer;
 506     for(i = 0; i < pic_byte_size; i++) {
 507         bitstream_put_ui(&nal_bs, byte_buf[i], 8);
 508     }
 509     free(byte_buf);
 510
 511     rbsp_trailing_bits(&nal_bs);
 512     bitstream_end(&nal_bs);
 513
 514     *sei_buffer = (unsigned char *)nal_bs.buffer;
 515
 516     return nal_bs.bit_offset;
 517 }
 518
 519
 520
 521 /*
 522  * Helper function for profiling purposes
 523  */
 524 static unsigned int GetTickCount()
 525 {
 526     struct timeval tv;
 527     if (gettimeofday(&tv, NULL))
 528         return 0;
 529     return tv.tv_usec/1000+tv.tv_sec*1000;
 530 }
 531
 532 /*
 533   Assume frame sequence is: Frame#0,#1,#2,...,#M,...,#X,... (encoding order)
 534   1) period between Frame #X and Frame #N = #X - #N
 535   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 536   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 537   4) intra_period and intra_idr_period take precedence over ip_period
 538   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 539      of I/IDR frames, see bellow examples
 540   -------------------------------------------------------------------
 541   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 542   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 543   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 544   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 545   1            1                ignored    IDR IDR IDR IDR...
 546   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 547   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 548   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 549                                               (PBB)(IBB)(PBB)(IBB)...
 550   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 551                                            IDRPPPPP IPPPPP IPPPPP...
 552   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 553                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 554                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 555                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 556                                            {IDR(PBB)(PBB)}                     (6/6/3)
 557                                            {IDR(PBB)(PBB)}.
 558 */
 559
 560 /*
 561  * Return displaying order with specified periods and encoding order
 562  * displaying_order: displaying order
 563  * frame_type: frame type
 564  */
 565 #define FRAME_P 0
 566 #define FRAME_B 1
 567 #define FRAME_I 2
 568 #define FRAME_IDR 7
 569 void encoding2display_order(
 570     unsigned long long encoding_order,int intra_period,
 571     int intra_idr_period,int ip_period,
 572     unsigned long long *displaying_order,
 573     int *frame_type)
 574 {
 575     int encoding_order_gop = 0;
 576
 577     if (intra_period == 1) { /* all are I/IDR frames */
 578         *displaying_order = encoding_order;
 579         if (intra_idr_period == 0)
 580             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 581         else
 582             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 583         return;
 584     }
 585
 586     if (intra_period == 0)
 587         intra_idr_period = 0;
 588
 589     /* new sequence like
 590      * IDR PPPPP IPPPPP
 591      * IDR (PBB)(PBB)(IBB)(PBB)
 592      */
 593     encoding_order_gop = (intra_idr_period == 0)? encoding_order:
 594         (encoding_order % (intra_idr_period + ((ip_period == 1)?0:1)));
 595
 596     if (encoding_order_gop == 0) { /* the first frame */
 597         *frame_type = FRAME_IDR;
 598         *displaying_order = encoding_order;
 599     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 600         *frame_type = FRAME_B;
 601         *displaying_order = encoding_order - 1;
 602     } else if ((intra_period != 0) && /* have I frames */
 603                (encoding_order_gop >= 2) &&
 604                ((ip_period == 1 && encoding_order_gop % intra_period == 0) || /* for IDR PPPPP IPPPP */
 605                 /* for IDR (PBB)(PBB)(IBB) */
 606                 (ip_period >= 2 && ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0))) {
 607         *frame_type = FRAME_I;
 608         *displaying_order = encoding_order + ip_period - 1;
 609     } else {
 610         *frame_type = FRAME_P;
 611         *displaying_order = encoding_order + ip_period - 1;
 612     }
 613 }
 614
 615
 616 static char *fourcc_to_string(int fourcc)
 617 {
 618     switch (fourcc) {
 619     case VA_FOURCC_NV12:
 620         return "NV12";
 621     case VA_FOURCC_IYUV:
 622         return "IYUV";
 623     case VA_FOURCC_YV12:
 624         return "YV12";
 625     case VA_FOURCC_UYVY:
 626         return "UYVY";
 627     default:
 628         return "Unknown";
 629     }
 630 }
 631
 632 static int string_to_fourcc(char *str)
 633 {
 634     int fourcc;
 635
 636     if (!strncmp(str, "NV12", 4))
 637         fourcc = VA_FOURCC_NV12;
 638     else if (!strncmp(str, "IYUV", 4))
 639         fourcc = VA_FOURCC_IYUV;
 640     else if (!strncmp(str, "YV12", 4))
 641         fourcc = VA_FOURCC_YV12;
 642     else if (!strncmp(str, "UYVY", 4))
 643         fourcc = VA_FOURCC_UYVY;
 644     else {
 645         printf("Unknow FOURCC\n");
 646         fourcc = -1;
 647     }
 648     return fourcc;
 649 }
 650
 651
 652 static char *rc_to_string(int rcmode)
 653 {
 654     switch (rc_mode) {
 655     case VA_RC_NONE:
 656         return "NONE";
 657     case VA_RC_CBR:
 658         return "CBR";
 659     case VA_RC_VBR:
 660         return "VBR";
 661     case VA_RC_VCM:
 662         return "VCM";
 663     case VA_RC_CQP:
 664         return "CQP";
 665     case VA_RC_VBR_CONSTRAINED:
 666         return "VBR_CONSTRAINED";
 667     default:
 668         return "Unknown";
 669     }
 670 }
 671
 672 static int string_to_rc(char *str)
 673 {
 674     int rc_mode;
 675
 676     if (!strncmp(str, "NONE", 4))
 677         rc_mode = VA_RC_NONE;
 678     else if (!strncmp(str, "CBR", 3))
 679         rc_mode = VA_RC_CBR;
 680     else if (!strncmp(str, "VBR", 3))
 681         rc_mode = VA_RC_VBR;
 682     else if (!strncmp(str, "VCM", 3))
 683         rc_mode = VA_RC_VCM;
 684     else if (!strncmp(str, "CQP", 3))
 685         rc_mode = VA_RC_CQP;
 686     else if (!strncmp(str, "VBR_CONSTRAINED", 15))
 687         rc_mode = VA_RC_VBR_CONSTRAINED;
 688     else {
 689         printf("Unknown RC mode\n");
 690         rc_mode = -1;
 691     }
 692     return rc_mode;
 693 }
 694
 695
 696 static int print_help(void)
 697 {
 698     printf("./h264encode <options>\n");
 699     printf("   -w <width> -h <height>\n");
 700     printf("   -n <frame number>\n");
 701     printf("   -o <coded file>\n");
 702     printf("   -f <frame rate>\n");
 703     printf("   --intra_period <number>\n");
 704     printf("   --idr_period <number>\n");
 705     printf("   --ip_period <number>\n");
 706     printf("   --bitrate <bitrate>\n");
 707     printf("   --initialqp <number>\n");
 708     printf("   --minqp <number>\n");
 709     printf("   --rcmode <NONE|CBR|VBR|VCM|CQP|VBR_CONTRAINED>\n");
 710     printf("   --syncmode: sequentially upload source, encoding, save result, no multi-thread\n");
 711     printf("   --srcyuv <filename> load YUV from a file\n");
 712     printf("   --fourcc <NV12|IYUV|YV12> source YUV fourcc\n");
 713     printf("   --recyuv <filename> save reconstructed YUV into a file\n");
 714     printf("   --enablePSNR calculate PSNR of recyuv vs. srcyuv\n");
 715
 716     return 0;
 717 }
 718
 719 static int process_cmdline(int argc, char *argv[])
 720 {
 721     char c;
 722     const struct option long_opts[] = {
 723         {"help", no_argument, NULL, 0 },
 724         {"bitrate", required_argument, NULL, 1 },
 725         {"minqp", required_argument, NULL, 2 },
 726         {"initialqp", required_argument, NULL, 3 },
 727         {"intra_period", required_argument, NULL, 4 },
 728         {"idr_period", required_argument, NULL, 5 },
 729         {"ip_period", required_argument, NULL, 6 },
 730         {"rcmode", required_argument, NULL, 7 },
 731         {"srcyuv", required_argument, NULL, 9 },
 732         {"recyuv", required_argument, NULL, 10 },
 733         {"fourcc", required_argument, NULL, 11 },
 734         {"syncmode", no_argument, NULL, 12 },
 735         {"enablePSNR", no_argument, NULL, 13 },
 736         {NULL, no_argument, NULL, 0 }};
 737     int long_index;
 738
 739     while ((c =getopt_long_only(argc,argv,"w:h:n:f:o:?",long_opts,&long_index)) != EOF) {
 740         switch (c) {
 741         case 'w':
 742             frame_width = atoi(optarg);
 743             break;
 744         case 'h':
 745             frame_height = atoi(optarg);
 746             break;
 747         case 'n':
 748             frame_count = atoi(optarg);
 749             break;
 750         case 'f':
 751             frame_rate = atoi(optarg);
 752             break;
 753         case 'o':
 754             coded_fn = strdup(optarg);
 755             break;
 756         case 0:
 757             print_help();
 758             exit(0);
 759         case 1:
 760             frame_bitrate = atoi(optarg);
 761             break;
 762         case 2:
 763             minimal_qp = atoi(optarg);
 764             break;
 765         case 3:
 766             initial_qp = atoi(optarg);
 767             break;
 768         case 4:
 769             intra_period = atoi(optarg);
 770             break;
 771         case 5:
 772             intra_idr_period = atoi(optarg);
 773             break;
 774         case 6:
 775             ip_period = atoi(optarg);
 776             break;
 777         case 7:
 778             rc_mode = string_to_rc(optarg);
 779             if (rc_mode < 0) {
 780                 print_help();
 781                 exit(1);
 782             }
 783             break;
 784         case 9:
 785             srcyuv_fn = strdup(optarg);
 786             break;
 787         case 10:
 788             recyuv_fn = strdup(optarg);
 789             break;
 790         case 11:
 791             srcyuv_fourcc = string_to_fourcc(optarg);
 792             if (srcyuv_fourcc <= 0) {
 793                 print_help();
 794                 exit(1);
 795             }
 796             break;
 797         case 12:
 798             encode_syncmode = 1;
 799             break;
 800         case 13:
 801             calc_psnr = 1;
 802             break;
 803         case ':':
 804         case '?':
 805             print_help();
 806             exit(0);
 807         }
 808     }
 809
 810     if (ip_period < 1) {
 811         printf(" ip_period must be greater than 0\n");
 812         exit(0);
 813     }
 814     if (intra_period != 1 && intra_period % ip_period != 0) {
 815         printf(" intra_period must be a multiplier of ip_period\n");
 816         exit(0);
 817     }
 818     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 819         printf(" intra_idr_period must be a multiplier of intra_period\n");
 820         exit(0);
 821     }
 822
 823     if (frame_bitrate == 0)
 824         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 825
 826     /* open source file */
 827     if (srcyuv_fn) {
 828         srcyuv_fp = fopen(srcyuv_fn,"r");
 829
 830         if (srcyuv_fp == NULL)
 831             printf("Open source YUV file %s failed, use auto-generated YUV data\n", srcyuv_fn);
 832         else {
 833             fseek(srcyuv_fp, 0L, SEEK_END);
 834             srcyuv_frames = ftell(srcyuv_fp) / (frame_width * frame_height * 1.5);
 835             printf("Source YUV file %s with %llu frames\n", srcyuv_fn, srcyuv_frames);
 836         }
 837     }
 838
 839     /* open source file */
 840     if (recyuv_fn) {
 841         recyuv_fp = fopen(recyuv_fn,"w+");
 842
 843         if (recyuv_fp == NULL)
 844             printf("Open reconstructed YUV file %s failed\n", recyuv_fn);
 845     }
 846
 847     if (coded_fn == NULL) {
 848         struct stat buf;
 849         if (stat("/tmp", &buf) == 0)
 850             coded_fn = strdup("/tmp/test.264");
 851         else if (stat("/sdcard", &buf) == 0)
 852             coded_fn = strdup("/sdcard/test.264");
 853         else
 854             coded_fn = strdup("./test.264");
 855     }
 856
 857     /* store coded data into a file */
 858     coded_fp = fopen(coded_fn,"w+");
 859     if (coded_fp == NULL) {
 860         printf("Open file %s failed, exit\n", coded_fn);
 861         exit(1);
 862     }
 863
 864
 865     return 0;
 866 }
 867
 868 static int init_va(void)
 869 {
 870     VAProfile profile_list[]={VAProfileH264High,VAProfileH264Main,VAProfileH264Baseline,VAProfileH264ConstrainedBaseline};
 871     VAEntrypoint entrypoints[VAEntrypointMax]={0};
 872     int num_entrypoints,slice_entrypoint;
 873     int support_encode = 0;
 874     int major_ver, minor_ver;
 875     VAStatus va_status;
 876     unsigned int i;
 877
 878     va_dpy = va_open_display();
 879     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 880     CHECK_VASTATUS(va_status, "vaInitialize");
 881
 882     /* use the highest profile */
 883     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 884         h264_profile = profile_list[i];
 885         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 886         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 887             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 888                 support_encode = 1;
 889                 break;
 890             }
 891         }
 892         if (support_encode == 1)
 893             break;
 894     }
 895
 896     if (support_encode == 0) {
 897         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 898         exit(1);
 899     } else {
 900         switch (h264_profile) {
 901             case VAProfileH264Baseline:
 902                 printf("Use profile VAProfileH264Baseline\n");
 903                 ip_period = 1;
 904                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 905                 break;
 906             case VAProfileH264ConstrainedBaseline:
 907                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 908                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 909                 ip_period = 1;
 910                 break;
 911
 912             case VAProfileH264Main:
 913                 printf("Use profile VAProfileH264Main\n");
 914                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 915                 break;
 916
 917             case VAProfileH264High:
 918                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 919                 printf("Use profile VAProfileH264High\n");
 920                 break;
 921             default:
 922                 printf("unknow profile. Set to Baseline");
 923                 h264_profile = VAProfileH264Baseline;
 924                 ip_period = 1;
 925                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 926                 break;
 927         }
 928     }
 929
 930     /* find out the format for the render target, and rate control mode */
 931     for (i = 0; i < VAConfigAttribTypeMax; i++)
 932         attrib[i].type = i;
 933
 934     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 935                                       &attrib[0], VAConfigAttribTypeMax);
 936     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 937     /* check the interested configattrib */
 938     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 939         printf("Not find desired YUV420 RT format\n");
 940         exit(1);
 941     } else {
 942         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 943         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 944         config_attrib_num++;
 945     }
 946
 947     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 948         int tmp = attrib[VAConfigAttribRateControl].value;
 949
 950         printf("Support rate control mode (0x%x):", tmp);
 951
 952         if (tmp & VA_RC_NONE)
 953             printf("NONE ");
 954         if (tmp & VA_RC_CBR)
 955             printf("CBR ");
 956         if (tmp & VA_RC_VBR)
 957             printf("VBR ");
 958         if (tmp & VA_RC_VCM)
 959             printf("VCM ");
 960         if (tmp & VA_RC_CQP)
 961             printf("CQP ");
 962         if (tmp & VA_RC_VBR_CONSTRAINED)
 963             printf("VBR_CONSTRAINED ");
 964
 965         printf("\n");
 966
 967         /* need to check if support rc_mode */
 968         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 969         config_attrib[config_attrib_num].value = rc_mode;
 970         config_attrib_num++;
 971     }
 972
 973
 974     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 975         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 976
 977         printf("Support VAConfigAttribEncPackedHeaders\n");
 978
 979         h264_packedheader = 1;
 980         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 981         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 982
 983         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 984             printf("Support packed sequence headers\n");
 985             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 986         }
 987
 988         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 989             printf("Support packed picture headers\n");
 990             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 991         }
 992
 993         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 994             printf("Support packed slice headers\n");
 995             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 996         }
 997
 998         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
 999             printf("Support packed misc headers\n");
1000             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1001         }
1002
1003         config_attrib_num++;
1004     }
1005
1006     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1007         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1008
1009         printf("Support VAConfigAttribEncInterlaced\n");
1010
1011         if (tmp & VA_ENC_INTERLACED_FRAME)
1012             printf("support VA_ENC_INTERLACED_FRAME\n");
1013         if (tmp & VA_ENC_INTERLACED_FIELD)
1014             printf("Support VA_ENC_INTERLACED_FIELD\n");
1015         if (tmp & VA_ENC_INTERLACED_MBAFF)
1016             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1017         if (tmp & VA_ENC_INTERLACED_PAFF)
1018             printf("Support VA_ENC_INTERLACED_PAFF\n");
1019
1020         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1021         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1022         config_attrib_num++;
1023     }
1024
1025     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1026         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1027
1028         printf("Support %d reference frames\n", h264_maxref);
1029     }
1030
1031     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1032         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1033
1034     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1035         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1036
1037         printf("Support VAConfigAttribEncSliceStructure\n");
1038
1039         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1040             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1041         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1042             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1043         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1044             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1045     }
1046     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1047         printf("Support VAConfigAttribEncMacroblockInfo\n");
1048     }
1049
1050     return 0;
1051 }
1052
1053 static int setup_encode()
1054 {
1055     VAStatus va_status;
1056     VASurfaceID *tmp_surfaceid;
1057     int codedbuf_size, i;
1058
1059     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1060             &config_attrib[0], config_attrib_num, &config_id);
1061     CHECK_VASTATUS(va_status, "vaCreateConfig");
1062
1063     /* create source surfaces */
1064     va_status = vaCreateSurfaces(va_dpy,
1065                                  VA_RT_FORMAT_YUV420, frame_width, frame_height,
1066                                  &src_surface[0], SURFACE_NUM,
1067                                  NULL, 0);
1068     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1069
1070     /* create reference surfaces */
1071     va_status = vaCreateSurfaces(
1072         va_dpy,
1073         VA_RT_FORMAT_YUV420, frame_width, frame_height,
1074         &ref_surface[0], SURFACE_NUM,
1075         NULL, 0
1076         );
1077     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1078
1079     tmp_surfaceid = calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1080     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1081     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1082
1083     /* Create a context for this encode pipe */
1084     va_status = vaCreateContext(va_dpy, config_id,
1085                                 frame_width, frame_height,
1086                                 VA_PROGRESSIVE,
1087                                 tmp_surfaceid, 2 * SURFACE_NUM,
1088                                 &context_id);
1089     CHECK_VASTATUS(va_status, "vaCreateContext");
1090     free(tmp_surfaceid);
1091
1092     codedbuf_size = (frame_width * frame_height * 400) / (16*16);
1093
1094     for (i = 0; i < SURFACE_NUM; i++) {
1095         /* create coded buffer once for all
1096          * other VA buffers which won't be used again after vaRenderPicture.
1097          * so APP can always vaCreateBuffer for every frame
1098          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1099          * so VA won't maintain the coded buffer
1100          */
1101         va_status = vaCreateBuffer(va_dpy,context_id,VAEncCodedBufferType,
1102                 codedbuf_size, 1, NULL, &coded_buf[i]);
1103         CHECK_VASTATUS(va_status,"vaCreateBuffer");
1104     }
1105
1106     return 0;
1107 }
1108
1109
1110
1111 #define partition(ref, field, key, ascending)   \
1112     while (i <= j) {                            \
1113         if (ascending) {                        \
1114             while (ref[i].field < key)          \
1115                 i++;                            \
1116             while (ref[j].field > key)          \
1117                 j--;                            \
1118         } else {                                \
1119             while (ref[i].field > key)          \
1120                 i++;                            \
1121             while (ref[j].field < key)          \
1122                 j--;                            \
1123         }                                       \
1124         if (i <= j) {                           \
1125             tmp = ref[i];                       \
1126             ref[i] = ref[j];                    \
1127             ref[j] = tmp;                       \
1128             i++;                                \
1129             j--;                                \
1130         }                                       \
1131     }                                           \
1132
1133 static void sort_one(VAPictureH264 ref[], int left, int right,
1134                      int ascending, int frame_idx)
1135 {
1136     int i = left, j = right;
1137     unsigned int key;
1138     VAPictureH264 tmp;
1139
1140     if (frame_idx) {
1141         key = ref[(left + right) / 2].frame_idx;
1142         partition(ref, frame_idx, key, ascending);
1143     } else {
1144         key = ref[(left + right) / 2].TopFieldOrderCnt;
1145         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1146     }
1147
1148     /* recursion */
1149     if (left < j)
1150         sort_one(ref, left, j, ascending, frame_idx);
1151
1152     if (i < right)
1153         sort_one(ref, i, right, ascending, frame_idx);
1154 }
1155
1156 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1157                      int partition_ascending, int list0_ascending, int list1_ascending)
1158 {
1159     int i = left, j = right;
1160     VAPictureH264 tmp;
1161
1162     if (frame_idx) {
1163         partition(ref, frame_idx, key, partition_ascending);
1164     } else {
1165         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1166     }
1167
1168
1169     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1170     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1171 }
1172
1173 static int update_ReferenceFrames(void)
1174 {
1175     int i;
1176
1177     if (current_frame_type == FRAME_B)
1178         return 0;
1179
1180     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1181     numShortTerm++;
1182     if (numShortTerm > num_ref_frames)
1183         numShortTerm = num_ref_frames;
1184     for (i=numShortTerm-1; i>0; i--)
1185         ReferenceFrames[i] = ReferenceFrames[i-1];
1186     ReferenceFrames[0] = CurrentCurrPic;
1187
1188     if (current_frame_type != FRAME_B)
1189         current_frame_num++;
1190     if (current_frame_num > MaxFrameNum)
1191         current_frame_num = 0;
1192
1193     return 0;
1194 }
1195
1196
1197 static int update_RefPicList(void)
1198 {
1199     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1200
1201     if (current_frame_type == FRAME_P) {
1202         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1203         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1204     }
1205
1206     if (current_frame_type == FRAME_B) {
1207         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1208         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1209                  1, 0, 1);
1210
1211         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1212         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1213                  0, 1, 0);
1214     }
1215
1216     return 0;
1217 }
1218
1219
1220 static int render_sequence(void)
1221 {
1222     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1223     VAStatus va_status;
1224     VAEncMiscParameterBuffer *misc_param;
1225     VAEncMiscParameterRateControl *misc_rate_ctrl;
1226
1227     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1228     seq_param.picture_width_in_mbs = frame_width / 16;
1229     seq_param.picture_height_in_mbs = frame_height / 16;
1230     seq_param.bits_per_second = frame_bitrate;
1231
1232     seq_param.intra_period = intra_period;
1233     seq_param.intra_idr_period = intra_idr_period;
1234     seq_param.ip_period = ip_period;
1235
1236     seq_param.max_num_ref_frames = num_ref_frames;
1237     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1238     seq_param.time_scale = 900;
1239     seq_param.num_units_in_tick = 15; /* Tc = num_units_in_tick / time_sacle */
1240     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1241     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1242     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1243
1244     va_status = vaCreateBuffer(va_dpy, context_id,
1245                                VAEncSequenceParameterBufferType,
1246                                sizeof(seq_param),1,&seq_param,&seq_param_buf);
1247     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1248
1249     va_status = vaCreateBuffer(va_dpy, context_id,
1250                                VAEncMiscParameterBufferType,
1251                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1252                                1,NULL,&rc_param_buf);
1253     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1254
1255     vaMapBuffer(va_dpy, rc_param_buf,(void **)&misc_param);
1256     misc_param->type = VAEncMiscParameterTypeRateControl;
1257     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1258     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1259     misc_rate_ctrl->bits_per_second = frame_bitrate;
1260     misc_rate_ctrl->initial_qp = initial_qp;
1261     misc_rate_ctrl->min_qp = minimal_qp;
1262     misc_rate_ctrl->basic_unit_size = 0;
1263     vaUnmapBuffer(va_dpy, rc_param_buf);
1264
1265     render_id[0] = seq_param_buf;
1266     render_id[1] = rc_param_buf;
1267
1268     va_status = vaRenderPicture(va_dpy,context_id, &render_id[0], 2);
1269     CHECK_VASTATUS(va_status,"vaRenderPicture");;
1270
1271     return 0;
1272 }
1273
1274 static int calc_poc(int pic_order_cnt_lsb)
1275 {
1276     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1277     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1278     int PicOrderCntMsb, TopFieldOrderCnt;
1279
1280     if (current_frame_type == FRAME_IDR)
1281         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1282     else {
1283         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1284         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1285     }
1286
1287     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1288         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1289         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1290     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1291              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1292         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1293     else
1294         PicOrderCntMsb = prevPicOrderCntMsb;
1295
1296     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1297
1298     if (current_frame_type != FRAME_B) {
1299         PicOrderCntMsb_ref = PicOrderCntMsb;
1300         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1301     }
1302
1303     return TopFieldOrderCnt;
1304 }
1305
1306 static int render_picture(void)
1307 {
1308     VABufferID pic_param_buf;
1309     VAStatus va_status;
1310     int i = 0;
1311
1312     pic_param.CurrPic.picture_id = ref_surface[current_slot];
1313     pic_param.CurrPic.frame_idx = current_frame_num;
1314     pic_param.CurrPic.flags = 0;
1315     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1316     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1317     CurrentCurrPic = pic_param.CurrPic;
1318
1319     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1320         update_RefPicList(); /* calc RefPicList */
1321         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1322         if (current_frame_type == FRAME_P) {
1323             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1324         } else if (current_frame_type == FRAME_B) {
1325             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1326             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1327         }
1328     } else {
1329         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1330         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1331             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1332             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1333         }
1334     }
1335
1336     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1337     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1338     pic_param.pic_fields.bits.entropy_coding_mode_flag = 1;
1339     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1340     pic_param.frame_num = current_frame_num;
1341     pic_param.coded_buf = coded_buf[current_slot];
1342     pic_param.last_picture = (current_frame_encoding == frame_count);
1343     pic_param.pic_init_qp = initial_qp;
1344
1345     va_status = vaCreateBuffer(va_dpy, context_id,VAEncPictureParameterBufferType,
1346                                sizeof(pic_param),1,&pic_param, &pic_param_buf);
1347     CHECK_VASTATUS(va_status,"vaCreateBuffer");;
1348
1349     va_status = vaRenderPicture(va_dpy,context_id, &pic_param_buf, 1);
1350     CHECK_VASTATUS(va_status,"vaRenderPicture");
1351
1352     return 0;
1353 }
1354
1355 static int render_packedsequence(void)
1356 {
1357     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1358     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1359     unsigned int length_in_bits;
1360     unsigned char *packedseq_buffer = NULL;
1361     VAStatus va_status;
1362
1363     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1364
1365     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1366
1367     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1368     packedheader_param_buffer.has_emulation_bytes = 0;
1369     va_status = vaCreateBuffer(va_dpy,
1370                                context_id,
1371                                VAEncPackedHeaderParameterBufferType,
1372                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1373                                &packedseq_para_bufid);
1374     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1375
1376     va_status = vaCreateBuffer(va_dpy,
1377                                context_id,
1378                                VAEncPackedHeaderDataBufferType,
1379                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1380                                &packedseq_data_bufid);
1381     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1382
1383     render_id[0] = packedseq_para_bufid;
1384     render_id[1] = packedseq_data_bufid;
1385     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1386     CHECK_VASTATUS(va_status,"vaRenderPicture");
1387
1388     free(packedseq_buffer);
1389
1390     return 0;
1391 }
1392
1393
1394 static int render_packedpicture(void)
1395 {
1396     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1397     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1398     unsigned int length_in_bits;
1399     unsigned char *packedpic_buffer = NULL;
1400     VAStatus va_status;
1401
1402     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1403     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1404     packedheader_param_buffer.bit_length = length_in_bits;
1405     packedheader_param_buffer.has_emulation_bytes = 0;
1406
1407     va_status = vaCreateBuffer(va_dpy,
1408                                context_id,
1409                                VAEncPackedHeaderParameterBufferType,
1410                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1411                                &packedpic_para_bufid);
1412     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1413
1414     va_status = vaCreateBuffer(va_dpy,
1415                                context_id,
1416                                VAEncPackedHeaderDataBufferType,
1417                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1418                                &packedpic_data_bufid);
1419     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1420
1421     render_id[0] = packedpic_para_bufid;
1422     render_id[1] = packedpic_data_bufid;
1423     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1424     CHECK_VASTATUS(va_status,"vaRenderPicture");
1425
1426     free(packedpic_buffer);
1427
1428     return 0;
1429 }
1430
1431 static void render_packedsei(void)
1432 {
1433     VAEncPackedHeaderParameterBuffer packed_header_param_buffer;
1434     VABufferID packed_sei_header_param_buf_id, packed_sei_buf_id, render_id[2];
1435     unsigned int length_in_bits /*offset_in_bytes*/;
1436     unsigned char *packed_sei_buffer = NULL;
1437     VAStatus va_status;
1438     int init_cpb_size, target_bit_rate, i_initial_cpb_removal_delay_length, i_initial_cpb_removal_delay;
1439     int i_cpb_removal_delay, i_dpb_output_delay_length, i_cpb_removal_delay_length;
1440
1441     /* it comes for the bps defined in SPS */
1442     target_bit_rate = frame_bitrate;
1443     init_cpb_size = (target_bit_rate * 8) >> 10;
1444     i_initial_cpb_removal_delay = init_cpb_size * 0.5 * 1024 / target_bit_rate * 90000;
1445
1446     i_cpb_removal_delay = 2;
1447     i_initial_cpb_removal_delay_length = 24;
1448     i_cpb_removal_delay_length = 24;
1449     i_dpb_output_delay_length = 24;
1450
1451
1452     length_in_bits = build_packed_sei_buffer_timing(
1453         i_initial_cpb_removal_delay_length,
1454         i_initial_cpb_removal_delay,
1455         0,
1456         i_cpb_removal_delay_length,
1457         i_cpb_removal_delay * current_frame_encoding,
1458         i_dpb_output_delay_length,
1459         0,
1460         &packed_sei_buffer);
1461
1462     //offset_in_bytes = 0;
1463     packed_header_param_buffer.type = VAEncPackedHeaderH264_SEI;
1464     packed_header_param_buffer.bit_length = length_in_bits;
1465     packed_header_param_buffer.has_emulation_bytes = 0;
1466
1467     va_status = vaCreateBuffer(va_dpy,
1468                                context_id,
1469                                VAEncPackedHeaderParameterBufferType,
1470                                sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
1471                                &packed_sei_header_param_buf_id);
1472     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1473
1474     va_status = vaCreateBuffer(va_dpy,
1475                                context_id,
1476                                VAEncPackedHeaderDataBufferType,
1477                                (length_in_bits + 7) / 8, 1, packed_sei_buffer,
1478                                &packed_sei_buf_id);
1479     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1480
1481
1482     render_id[0] = packed_sei_header_param_buf_id;
1483     render_id[1] = packed_sei_buf_id;
1484     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1485     CHECK_VASTATUS(va_status,"vaRenderPicture");
1486
1487
1488     free(packed_sei_buffer);
1489
1490     return;
1491 }
1492
1493
1494 static int render_hrd(void)
1495 {
1496     VABufferID misc_parameter_hrd_buf_id;
1497     VAStatus va_status;
1498     VAEncMiscParameterBuffer *misc_param;
1499     VAEncMiscParameterHRD *misc_hrd_param;
1500
1501     va_status = vaCreateBuffer(va_dpy, context_id,
1502                    VAEncMiscParameterBufferType,
1503                    sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterHRD),
1504                    1,
1505                    NULL,
1506                    &misc_parameter_hrd_buf_id);
1507     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1508
1509     vaMapBuffer(va_dpy,
1510                 misc_parameter_hrd_buf_id,
1511                 (void **)&misc_param);
1512     misc_param->type = VAEncMiscParameterTypeHRD;
1513     misc_hrd_param = (VAEncMiscParameterHRD *)misc_param->data;
1514
1515     if (frame_bitrate > 0) {
1516         misc_hrd_param->initial_buffer_fullness = frame_bitrate * 1024 * 4;
1517         misc_hrd_param->buffer_size = frame_bitrate * 1024 * 8;
1518     } else {
1519         misc_hrd_param->initial_buffer_fullness = 0;
1520         misc_hrd_param->buffer_size = 0;
1521     }
1522     vaUnmapBuffer(va_dpy, misc_parameter_hrd_buf_id);
1523
1524     va_status = vaRenderPicture(va_dpy,context_id, &misc_parameter_hrd_buf_id, 1);
1525     CHECK_VASTATUS(va_status,"vaRenderPicture");;
1526
1527     return 0;
1528 }
1529
1530 static int render_slice(void)
1531 {
1532     VABufferID slice_param_buf;
1533     VAStatus va_status;
1534     int i;
1535
1536     update_RefPicList();
1537
1538     /* one frame, one slice */
1539     slice_param.macroblock_address = 0;
1540     slice_param.num_macroblocks = frame_width*frame_height/(16*16); /* Measured by MB */
1541     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1542     if (current_frame_type == FRAME_IDR) {
1543         ++slice_param.idr_pic_id;
1544     } else if (current_frame_type == FRAME_P) {
1545         int refpiclist0_max = h264_maxref & 0xffff;
1546         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1547
1548         for (i = refpiclist0_max; i < 32; i++) {
1549             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1550             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1551         }
1552     } else if (current_frame_type == FRAME_B) {
1553         int refpiclist0_max = h264_maxref & 0xffff;
1554         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1555
1556         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1557         for (i = refpiclist0_max; i < 32; i++) {
1558             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1559             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1560         }
1561
1562         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1563         for (i = refpiclist1_max; i < 32; i++) {
1564             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1565             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1566         }
1567     }
1568
1569     slice_param.slice_alpha_c0_offset_div2 = 2;
1570     slice_param.slice_beta_offset_div2 = 2;
1571     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1572
1573     va_status = vaCreateBuffer(va_dpy,context_id,VAEncSliceParameterBufferType,
1574                                sizeof(slice_param),1,&slice_param,&slice_param_buf);
1575     CHECK_VASTATUS(va_status,"vaCreateBuffer");;
1576
1577     va_status = vaRenderPicture(va_dpy,context_id, &slice_param_buf, 1);
1578     CHECK_VASTATUS(va_status,"vaRenderPicture");
1579
1580     return 0;
1581 }
1582
1583
1584 static int upload_source_YUV_once_for_all()
1585 {
1586     int box_width=8;
1587     int row_shift=0;
1588     int i;
1589
1590     for (i = 0; i < SURFACE_NUM; i++) {
1591         printf("\rLoading data into surface %d.....", i);
1592         upload_surface(va_dpy, src_surface[i], box_width, row_shift, 0);
1593
1594         row_shift++;
1595         if (row_shift==(2*box_width)) row_shift= 0;
1596     }
1597     printf("Complete surface loading\n");
1598
1599     return 0;
1600 }
1601
1602 #define check_ret(ret)                                  \
1603 if (ret != 1) {                                         \
1604     printf("fread doesn't return enough data\n");       \
1605     exit(1);                                            \
1606 }
1607 static int load_surface(VASurfaceID surface_id, unsigned long long display_order)
1608 {
1609     unsigned char *src_Y = NULL, *src_U = NULL, *src_V = NULL;
1610     int ret = 0;
1611
1612     if (srcyuv_fp == NULL)
1613         return 0;
1614
1615     /* rewind the file pointer if encoding more than srcyuv_frames */
1616     display_order = display_order % srcyuv_frames;
1617     fseek(srcyuv_fp, display_order * frame_width * frame_height * 1.5, SEEK_SET);
1618
1619
1620     if (srcyuv_fourcc == VA_FOURCC_NV12) {
1621         int uv_size = 2 * (frame_width/2) * (frame_height/2);
1622
1623         src_Y = malloc(2 * uv_size);
1624         src_U = malloc(uv_size);
1625
1626         ret = fread(src_Y, frame_width * frame_height, 1, srcyuv_fp);
1627         check_ret(ret);
1628         ret = fread(src_U, uv_size, 1, srcyuv_fp);
1629         check_ret(ret);
1630     } else if (srcyuv_fourcc == VA_FOURCC_IYUV ||
1631         srcyuv_fourcc == VA_FOURCC_YV12) {
1632         int uv_size = (frame_width/2) * (frame_height/2);
1633
1634         src_Y = malloc(4 * uv_size);
1635         src_U = malloc(uv_size);
1636         src_V = malloc(uv_size);
1637
1638         ret = fread(src_Y, frame_width * frame_height, 1, srcyuv_fp);
1639         check_ret(ret);
1640         if (srcyuv_fourcc == VA_FOURCC_IYUV) {
1641             ret = fread(src_U, uv_size, 1, srcyuv_fp);
1642             check_ret(ret);
1643             ret = fread(src_V, uv_size, 1, srcyuv_fp);
1644             check_ret(ret);
1645         } else { /* YV12 */
1646             ret = fread(src_V, uv_size, 1, srcyuv_fp);
1647             check_ret(ret);
1648             ret = fread(src_U, uv_size, 1, srcyuv_fp);
1649             check_ret(ret);
1650         }
1651     } else {
1652         printf("Unsupported source YUV format\n");
1653         exit(1);
1654     }
1655
1656     upload_surface_yuv(va_dpy, surface_id,
1657                        srcyuv_fourcc, frame_width, frame_height,
1658                        src_Y, src_U, src_V);
1659     if (src_Y)
1660         free(src_Y);
1661     if (src_U)
1662         free(src_U);
1663     if (src_V)
1664         free(src_V);
1665
1666     return 0;
1667 }
1668
1669
1670 static int save_recyuv(VASurfaceID surface_id,
1671                        unsigned long long display_order,
1672                        unsigned long long encode_order)
1673 {
1674     unsigned char *dst_Y = NULL, *dst_U = NULL, *dst_V = NULL;
1675
1676     if (recyuv_fp == NULL)
1677         return 0;
1678
1679     if (srcyuv_fourcc == VA_FOURCC_NV12) {
1680         int uv_size = 2 * (frame_width/2) * (frame_height/2);
1681         dst_Y = malloc(2*uv_size);
1682         dst_U = malloc(uv_size);
1683     } else if (srcyuv_fourcc == VA_FOURCC_IYUV ||
1684                srcyuv_fourcc == VA_FOURCC_YV12) {
1685         int uv_size = (frame_width/2) * (frame_height/2);
1686         dst_Y = malloc(4*uv_size);
1687         dst_U = malloc(uv_size);
1688         dst_V = malloc(uv_size);
1689     } else {
1690         printf("Unsupported source YUV format\n");
1691         exit(1);
1692     }
1693
1694     download_surface_yuv(va_dpy, surface_id,
1695                          srcyuv_fourcc, frame_width, frame_height,
1696                          dst_Y, dst_U, dst_V);
1697     fseek(recyuv_fp, display_order * frame_width * frame_height * 1.5, SEEK_SET);
1698
1699     if (srcyuv_fourcc == VA_FOURCC_NV12) {
1700         int uv_size = 2 * (frame_width/2) * (frame_height/2);
1701         fwrite(dst_Y, uv_size * 2, 1, recyuv_fp);
1702         fwrite(dst_U, uv_size, 1, recyuv_fp);
1703     } else if (srcyuv_fourcc == VA_FOURCC_IYUV ||
1704                srcyuv_fourcc == VA_FOURCC_YV12) {
1705         int uv_size = (frame_width/2) * (frame_height/2);
1706         fwrite(dst_Y, uv_size * 4, 1, recyuv_fp);
1707
1708         if (srcyuv_fourcc == VA_FOURCC_IYUV) {
1709             fwrite(dst_U, uv_size, 1, recyuv_fp);
1710             fwrite(dst_V, uv_size, 1, recyuv_fp);
1711         } else {
1712             fwrite(dst_V, uv_size, 1, recyuv_fp);
1713             fwrite(dst_U, uv_size, 1, recyuv_fp);
1714         }
1715     } else {
1716         printf("Unsupported YUV format\n");
1717         exit(1);
1718     }
1719
1720     if (dst_Y)
1721         free(dst_Y);
1722     if (dst_U)
1723         free(dst_U);
1724     if (dst_V)
1725         free(dst_V);
1726
1727     return 0;
1728 }
1729
1730
1731 static int save_codeddata(unsigned long long display_order, unsigned long long encode_order)
1732 {
1733     VACodedBufferSegment *buf_list = NULL;
1734     VAStatus va_status;
1735     unsigned int coded_size = 0;
1736
1737     va_status = vaMapBuffer(va_dpy,coded_buf[display_order % SURFACE_NUM],(void **)(&buf_list));
1738     CHECK_VASTATUS(va_status,"vaMapBuffer");
1739     while (buf_list != NULL) {
1740         coded_size += fwrite(buf_list->buf, 1, buf_list->size, coded_fp);
1741         buf_list = (VACodedBufferSegment *) buf_list->next;
1742
1743         frame_size += coded_size;
1744     }
1745     vaUnmapBuffer(va_dpy,coded_buf[display_order % SURFACE_NUM]);
1746
1747     printf("\r      "); /* return back to startpoint */
1748     switch (encode_order % 4) {
1749         case 0:
1750             printf("|");
1751             break;
1752         case 1:
1753             printf("/");
1754             break;
1755         case 2:
1756             printf("-");
1757             break;
1758         case 3:
1759             printf("\\");
1760             break;
1761     }
1762     printf("%08lld", encode_order);
1763     /*
1764     if (current_frame_encoding % intra_count == 0)
1765         printf("(I)");
1766     else
1767         printf("(P)");
1768     */
1769     printf("(%06d bytes coded)",coded_size);
1770     /* skipped frame ? */
1771     printf("                                    ");
1772
1773     return 0;
1774 }
1775
1776
1777 static struct storage_task_t * storage_task_dequeue(void)
1778 {
1779     struct storage_task_t *header;
1780
1781     pthread_mutex_lock(&encode_mutex);
1782
1783     header = storage_task_header;
1784     if (storage_task_header != NULL) {
1785         if (storage_task_tail == storage_task_header)
1786             storage_task_tail = NULL;
1787         storage_task_header = header->next;
1788     }
1789
1790     pthread_mutex_unlock(&encode_mutex);
1791
1792     return header;
1793 }
1794
1795 static int storage_task_queue(unsigned long long display_order, unsigned long long encode_order)
1796 {
1797     struct storage_task_t *tmp;
1798
1799     tmp = calloc(1, sizeof(struct storage_task_t));
1800     tmp->display_order = display_order;
1801     tmp->encode_order = encode_order;
1802
1803     pthread_mutex_lock(&encode_mutex);
1804
1805     if (storage_task_header == NULL) {
1806         storage_task_header = tmp;
1807         storage_task_tail = tmp;
1808     } else {
1809         storage_task_tail->next = tmp;
1810         storage_task_tail = tmp;
1811     }
1812
1813     srcsurface_status[display_order % SURFACE_NUM] = SRC_SURFACE_IN_STORAGE;
1814     pthread_cond_signal(&encode_cond);
1815
1816     pthread_mutex_unlock(&encode_mutex);
1817
1818     return 0;
1819 }
1820
1821 static void storage_task(unsigned long long display_order, unsigned long long encode_order)
1822 {
1823     unsigned int tmp;
1824     VAStatus va_status;
1825
1826     tmp = GetTickCount();
1827     va_status = vaSyncSurface(va_dpy, src_surface[display_order % SURFACE_NUM]);
1828     CHECK_VASTATUS(va_status,"vaSyncSurface");
1829     SyncPictureTicks += GetTickCount() - tmp;
1830     tmp = GetTickCount();
1831     save_codeddata(display_order, encode_order);
1832     SavePictureTicks += GetTickCount() - tmp;
1833
1834     save_recyuv(ref_surface[display_order % SURFACE_NUM], display_order, encode_order);
1835
1836     /* reload a new frame data */
1837     tmp = GetTickCount();
1838     if (srcyuv_fp != NULL)
1839         load_surface(src_surface[display_order % SURFACE_NUM], display_order + SURFACE_NUM);
1840     UploadPictureTicks += GetTickCount() - tmp;
1841
1842     pthread_mutex_lock(&encode_mutex);
1843     srcsurface_status[display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1844     pthread_mutex_unlock(&encode_mutex);
1845 }
1846
1847
1848 static void * storage_task_thread(void *t)
1849 {
1850     while (1) {
1851         struct storage_task_t *current;
1852
1853         current = storage_task_dequeue();
1854         if (current == NULL) {
1855             pthread_mutex_lock(&encode_mutex);
1856             pthread_cond_wait(&encode_cond, &encode_mutex);
1857             pthread_mutex_unlock(&encode_mutex);
1858             continue;
1859         }
1860
1861         storage_task(current->display_order, current->encode_order);
1862
1863         free(current);
1864
1865         /* all frames are saved, exit the thread */
1866         if (++frame_coded >= frame_count)
1867             break;
1868     }
1869
1870     return 0;
1871 }
1872
1873
1874 static int encode_frames(void)
1875 {
1876     unsigned int i, tmp;
1877     VAStatus va_status;
1878     //VASurfaceStatus surface_status;
1879
1880     /* upload RAW YUV data into all surfaces */
1881     tmp = GetTickCount();
1882     if (srcyuv_fp != NULL) {
1883         for (i = 0; i < SURFACE_NUM; i++)
1884             load_surface(src_surface[i], i);
1885     } else
1886         upload_source_YUV_once_for_all();
1887     UploadPictureTicks += GetTickCount() - tmp;
1888
1889     /* ready for encoding */
1890     memset(srcsurface_status, SRC_SURFACE_IN_ENCODING, sizeof(srcsurface_status));
1891
1892     memset(&seq_param, 0, sizeof(seq_param));
1893     memset(&pic_param, 0, sizeof(pic_param));
1894     memset(&slice_param, 0, sizeof(slice_param));
1895
1896     if (encode_syncmode == 0)
1897         pthread_create(&encode_thread, NULL, storage_task_thread, NULL);
1898
1899     for (current_frame_encoding = 0; current_frame_encoding < frame_count; current_frame_encoding++) {
1900         encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1901                                &current_frame_display, &current_frame_type);
1902         if (current_frame_type == FRAME_IDR) {
1903             numShortTerm = 0;
1904             current_frame_num = 0;
1905             current_IDR_display = current_frame_display;
1906         }
1907
1908         /* check if the source frame is ready */
1909         while (srcsurface_status[current_slot] != SRC_SURFACE_IN_ENCODING) {
1910             usleep(1);
1911         }
1912
1913         tmp = GetTickCount();
1914         va_status = vaBeginPicture(va_dpy, context_id, src_surface[current_slot]);
1915         CHECK_VASTATUS(va_status,"vaBeginPicture");
1916         BeginPictureTicks += GetTickCount() - tmp;
1917
1918         tmp = GetTickCount();
1919         if (current_frame_type == FRAME_IDR) {
1920             render_sequence();
1921             render_picture();
1922             if (h264_packedheader) {
1923                 render_packedsequence();
1924                 render_packedpicture();
1925             }
1926             //if (rc_mode == VA_RC_CBR)
1927             //    render_packedsei();
1928             //render_hrd();
1929         } else {
1930             //render_sequence();
1931             render_picture();
1932             //if (rc_mode == VA_RC_CBR)
1933             //    render_packedsei();
1934             //render_hrd();
1935         }
1936         render_slice();
1937         RenderPictureTicks += GetTickCount() - tmp;
1938
1939         tmp = GetTickCount();
1940         va_status = vaEndPicture(va_dpy,context_id);
1941         CHECK_VASTATUS(va_status,"vaEndPicture");;
1942         EndPictureTicks += GetTickCount() - tmp;
1943
1944         if (encode_syncmode)
1945             storage_task(current_frame_display, current_frame_encoding);
1946         else /* queue the storage task queue */
1947             storage_task_queue(current_frame_display, current_frame_encoding);
1948
1949         /* how to process skipped frames
1950            surface_status = (VASurfaceStatus) 0;
1951            va_status = vaQuerySurfaceStatus(va_dpy, src_surface[i%SURFACE_NUM],&surface_status);
1952            frame_skipped = (surface_status & VASurfaceSkipped);
1953         */
1954         update_ReferenceFrames();
1955     }
1956
1957     if (encode_syncmode == 0) {
1958         int ret;
1959         pthread_join(encode_thread, (void **)&ret);
1960     }
1961
1962     return 0;
1963 }
1964
1965
1966 static int release_encode()
1967 {
1968     int i;
1969
1970     vaDestroySurfaces(va_dpy,&src_surface[0],SURFACE_NUM);
1971     vaDestroySurfaces(va_dpy,&ref_surface[0],SURFACE_NUM);
1972
1973     for (i = 0; i < SURFACE_NUM; i++)
1974         vaDestroyBuffer(va_dpy,coded_buf[i]);
1975
1976     vaDestroyContext(va_dpy,context_id);
1977     vaDestroyConfig(va_dpy,config_id);
1978
1979     return 0;
1980 }
1981
1982 static int deinit_va()
1983 {
1984     vaTerminate(va_dpy);
1985
1986     va_close_display(va_dpy);
1987
1988     return 0;
1989 }
1990
1991
1992 static int print_input()
1993 {
1994     printf("\n\nINPUT:Try to encode H264...\n");
1995     printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1996     printf("INPUT: Resolution   : %dx%d, %d frames\n",
1997            frame_width, frame_height, frame_count);
1998     printf("INPUT: FrameRate    : %d\n", frame_rate);
1999     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
2000     printf("INPUT: Slieces      : %d\n", frame_slices);
2001     printf("INPUT: IntraPeriod  : %d\n", intra_period);
2002     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
2003     printf("INPUT: IpPeriod     : %d\n", ip_period);
2004     printf("INPUT: Initial QP   : %d\n", initial_qp);
2005     printf("INPUT: Min QP       : %d\n", minimal_qp);
2006     printf("INPUT: Source YUV   : %s", srcyuv_fp?"FILE":"AUTO generated");
2007     if (srcyuv_fp)
2008         printf(":%s (fourcc %s)\n", srcyuv_fn, fourcc_to_string(srcyuv_fourcc));
2009     else
2010         printf("\n");
2011     printf("INPUT: Coded Clip   : %s\n", coded_fn);
2012     if (recyuv_fp == NULL)
2013         printf("INPUT: Rec   Clip   : %s\n", "Not save reconstructed frame");
2014     else
2015         printf("INPUT: Rec   Clip   : Save reconstructed frame into %s (fourcc %s)\n", recyuv_fn,
2016                fourcc_to_string(srcyuv_fourcc));
2017
2018     printf("\n\n"); /* return back to startpoint */
2019
2020     return 0;
2021 }
2022
2023 static int calc_PSNR(double *psnr)
2024 {
2025     unsigned long srcyuv_size, recyuv_size, min_size;
2026     char *srcyuv_ptr, *recyuv_ptr;
2027     unsigned long i, sse=0;
2028     double ssemean;
2029
2030     fseek(srcyuv_fp, 0L, SEEK_END);
2031     srcyuv_size = ftell(srcyuv_fp);
2032     fseek(recyuv_fp, 0L, SEEK_END);
2033     recyuv_size = ftell(recyuv_fp);
2034
2035     fseek(srcyuv_fp, 0L, SEEK_SET);
2036     fseek(recyuv_fp, 0L, SEEK_SET);
2037
2038     min_size = MIN(srcyuv_size,recyuv_size);
2039     srcyuv_ptr = mmap (0, min_size, PROT_READ, MAP_SHARED, fileno(srcyuv_fp), 0);
2040     recyuv_ptr = mmap (0, min_size, PROT_READ, MAP_SHARED, fileno(recyuv_fp), 0);
2041     if ((srcyuv_ptr == MAP_FAILED) || (recyuv_ptr == MAP_FAILED)) {
2042         printf("Failed to mmap YUV files\n");
2043         return 1;
2044     }
2045
2046     for (i=0; i<min_size; i++) {
2047         char tmp = srcyuv_ptr[i] - recyuv_ptr[i];
2048         sse += tmp * tmp;
2049     }
2050     ssemean = (double)sse/(double)min_size;
2051     *psnr = 20.0*log10(255) - 10.0*log10(ssemean);
2052
2053     munmap(srcyuv_ptr, min_size);
2054     munmap(recyuv_ptr, min_size);
2055
2056     return 0;
2057 }
2058
2059 static int print_performance(unsigned int PictureCount)
2060 {
2061     unsigned int psnr_ret = 1, others = 0;
2062     double psnr = 0, total_size = frame_width * frame_height * 1.5 * frame_count;
2063
2064     if (calc_psnr && srcyuv_fp && recyuv_fp)
2065         psnr_ret = calc_PSNR(&psnr);
2066
2067     others = TotalTicks - UploadPictureTicks - BeginPictureTicks
2068         - RenderPictureTicks - EndPictureTicks - SyncPictureTicks - SavePictureTicks;
2069
2070     printf("\n\n");
2071
2072     printf("PERFORMANCE:   Frame Rate           : %.2f fps (%d frames, %d ms (%.2f ms per frame))\n",
2073            (double) 1000*PictureCount / TotalTicks, PictureCount,
2074            TotalTicks, ((double)  TotalTicks) / (double) PictureCount);
2075     printf("PERFORMANCE:   Compression ratio    : %d:1\n", (unsigned int)(total_size / frame_size));
2076     if (psnr_ret == 0)
2077         printf("PERFORMANCE:   PSNR                 : %.2f (%lld frames calculated)\n",
2078                psnr, MIN(frame_count, srcyuv_frames));
2079
2080     printf("PERFORMANCE:     UploadPicture      : %d ms (%.2f, %.2f%% percent)\n",
2081            (int) UploadPictureTicks, ((double)  UploadPictureTicks) / (double) PictureCount,
2082            UploadPictureTicks/(double) TotalTicks/0.01);
2083     printf("PERFORMANCE:     vaBeginPicture     : %d ms (%.2f, %.2f%% percent)\n",
2084            (int) BeginPictureTicks, ((double)  BeginPictureTicks) / (double) PictureCount,
2085            BeginPictureTicks/(double) TotalTicks/0.01);
2086     printf("PERFORMANCE:     vaRenderHeader     : %d ms (%.2f, %.2f%% percent)\n",
2087            (int) RenderPictureTicks, ((double)  RenderPictureTicks) / (double) PictureCount,
2088            RenderPictureTicks/(double) TotalTicks/0.01);
2089     printf("PERFORMANCE:     vaEndPicture       : %d ms (%.2f, %.2f%% percent)\n",
2090            (int) EndPictureTicks, ((double)  EndPictureTicks) / (double) PictureCount,
2091            EndPictureTicks/(double) TotalTicks/0.01);
2092     printf("PERFORMANCE:     vaSyncSurface      : %d ms (%.2f, %.2f%% percent)\n",
2093            (int) SyncPictureTicks, ((double) SyncPictureTicks) / (double) PictureCount,
2094            SyncPictureTicks/(double) TotalTicks/0.01);
2095     printf("PERFORMANCE:     SavePicture        : %d ms (%.2f, %.2f%% percent)\n",
2096            (int) SavePictureTicks, ((double)  SavePictureTicks) / (double) PictureCount,
2097            SavePictureTicks/(double) TotalTicks/0.01);
2098     printf("PERFORMANCE:     Others             : %d ms (%.2f, %.2f%% percent)\n",
2099            (int) others, ((double) others) / (double) PictureCount,
2100            others/(double) TotalTicks/0.01);
2101
2102     if (encode_syncmode == 0)
2103         printf("(Multithread enabled, the timing is only for reference)\n");
2104
2105     return 0;
2106 }
2107
2108
2109 int main(int argc,char **argv)
2110 {
2111     unsigned int start;
2112
2113     process_cmdline(argc, argv);
2114
2115     print_input();
2116
2117     start = GetTickCount();
2118
2119     init_va();
2120     setup_encode();
2121
2122     encode_frames();
2123
2124     release_encode();
2125     deinit_va();
2126
2127     TotalTicks += GetTickCount() - start;
2128     print_performance(frame_count);
2129
2130     return 0;
2131 }