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 RefPicList0 and %d RefPicList1\n",
1029                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1030     }
1031
1032     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1033         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1034
1035     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1036         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1037
1038         printf("Support VAConfigAttribEncSliceStructure\n");
1039
1040         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1041             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1042         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1043             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1044         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1045             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1046     }
1047     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1048         printf("Support VAConfigAttribEncMacroblockInfo\n");
1049     }
1050
1051     return 0;
1052 }
1053
1054 static int setup_encode()
1055 {
1056     VAStatus va_status;
1057     VASurfaceID *tmp_surfaceid;
1058     int codedbuf_size, i;
1059
1060     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1061             &config_attrib[0], config_attrib_num, &config_id);
1062     CHECK_VASTATUS(va_status, "vaCreateConfig");
1063
1064     /* create source surfaces */
1065     va_status = vaCreateSurfaces(va_dpy,
1066                                  VA_RT_FORMAT_YUV420, frame_width, frame_height,
1067                                  &src_surface[0], SURFACE_NUM,
1068                                  NULL, 0);
1069     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1070
1071     /* create reference surfaces */
1072     va_status = vaCreateSurfaces(
1073         va_dpy,
1074         VA_RT_FORMAT_YUV420, frame_width, frame_height,
1075         &ref_surface[0], SURFACE_NUM,
1076         NULL, 0
1077         );
1078     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1079
1080     tmp_surfaceid = calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1081     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1082     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1083
1084     /* Create a context for this encode pipe */
1085     va_status = vaCreateContext(va_dpy, config_id,
1086                                 frame_width, frame_height,
1087                                 VA_PROGRESSIVE,
1088                                 tmp_surfaceid, 2 * SURFACE_NUM,
1089                                 &context_id);
1090     CHECK_VASTATUS(va_status, "vaCreateContext");
1091     free(tmp_surfaceid);
1092
1093     codedbuf_size = (frame_width * frame_height * 400) / (16*16);
1094
1095     for (i = 0; i < SURFACE_NUM; i++) {
1096         /* create coded buffer once for all
1097          * other VA buffers which won't be used again after vaRenderPicture.
1098          * so APP can always vaCreateBuffer for every frame
1099          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1100          * so VA won't maintain the coded buffer
1101          */
1102         va_status = vaCreateBuffer(va_dpy,context_id,VAEncCodedBufferType,
1103                 codedbuf_size, 1, NULL, &coded_buf[i]);
1104         CHECK_VASTATUS(va_status,"vaCreateBuffer");
1105     }
1106
1107     return 0;
1108 }
1109
1110
1111
1112 #define partition(ref, field, key, ascending)   \
1113     while (i <= j) {                            \
1114         if (ascending) {                        \
1115             while (ref[i].field < key)          \
1116                 i++;                            \
1117             while (ref[j].field > key)          \
1118                 j--;                            \
1119         } else {                                \
1120             while (ref[i].field > key)          \
1121                 i++;                            \
1122             while (ref[j].field < key)          \
1123                 j--;                            \
1124         }                                       \
1125         if (i <= j) {                           \
1126             tmp = ref[i];                       \
1127             ref[i] = ref[j];                    \
1128             ref[j] = tmp;                       \
1129             i++;                                \
1130             j--;                                \
1131         }                                       \
1132     }                                           \
1133
1134 static void sort_one(VAPictureH264 ref[], int left, int right,
1135                      int ascending, int frame_idx)
1136 {
1137     int i = left, j = right;
1138     unsigned int key;
1139     VAPictureH264 tmp;
1140
1141     if (frame_idx) {
1142         key = ref[(left + right) / 2].frame_idx;
1143         partition(ref, frame_idx, key, ascending);
1144     } else {
1145         key = ref[(left + right) / 2].TopFieldOrderCnt;
1146         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1147     }
1148
1149     /* recursion */
1150     if (left < j)
1151         sort_one(ref, left, j, ascending, frame_idx);
1152
1153     if (i < right)
1154         sort_one(ref, i, right, ascending, frame_idx);
1155 }
1156
1157 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1158                      int partition_ascending, int list0_ascending, int list1_ascending)
1159 {
1160     int i = left, j = right;
1161     VAPictureH264 tmp;
1162
1163     if (frame_idx) {
1164         partition(ref, frame_idx, key, partition_ascending);
1165     } else {
1166         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1167     }
1168
1169
1170     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1171     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1172 }
1173
1174 static int update_ReferenceFrames(void)
1175 {
1176     int i;
1177
1178     if (current_frame_type == FRAME_B)
1179         return 0;
1180
1181     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1182     numShortTerm++;
1183     if (numShortTerm > num_ref_frames)
1184         numShortTerm = num_ref_frames;
1185     for (i=numShortTerm-1; i>0; i--)
1186         ReferenceFrames[i] = ReferenceFrames[i-1];
1187     ReferenceFrames[0] = CurrentCurrPic;
1188
1189     if (current_frame_type != FRAME_B)
1190         current_frame_num++;
1191     if (current_frame_num > MaxFrameNum)
1192         current_frame_num = 0;
1193
1194     return 0;
1195 }
1196
1197
1198 static int update_RefPicList(void)
1199 {
1200     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1201
1202     if (current_frame_type == FRAME_P) {
1203         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1204         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1205     }
1206
1207     if (current_frame_type == FRAME_B) {
1208         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1209         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1210                  1, 0, 1);
1211
1212         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1213         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1214                  0, 1, 0);
1215     }
1216
1217     return 0;
1218 }
1219
1220
1221 static int render_sequence(void)
1222 {
1223     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1224     VAStatus va_status;
1225     VAEncMiscParameterBuffer *misc_param;
1226     VAEncMiscParameterRateControl *misc_rate_ctrl;
1227
1228     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1229     seq_param.picture_width_in_mbs = frame_width / 16;
1230     seq_param.picture_height_in_mbs = frame_height / 16;
1231     seq_param.bits_per_second = frame_bitrate;
1232
1233     seq_param.intra_period = intra_period;
1234     seq_param.intra_idr_period = intra_idr_period;
1235     seq_param.ip_period = ip_period;
1236
1237     seq_param.max_num_ref_frames = num_ref_frames;
1238     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1239     seq_param.time_scale = 900;
1240     seq_param.num_units_in_tick = 15; /* Tc = num_units_in_tick / time_sacle */
1241     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1242     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1243     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1244
1245     va_status = vaCreateBuffer(va_dpy, context_id,
1246                                VAEncSequenceParameterBufferType,
1247                                sizeof(seq_param),1,&seq_param,&seq_param_buf);
1248     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1249
1250     va_status = vaCreateBuffer(va_dpy, context_id,
1251                                VAEncMiscParameterBufferType,
1252                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1253                                1,NULL,&rc_param_buf);
1254     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1255
1256     vaMapBuffer(va_dpy, rc_param_buf,(void **)&misc_param);
1257     misc_param->type = VAEncMiscParameterTypeRateControl;
1258     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1259     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1260     misc_rate_ctrl->bits_per_second = frame_bitrate;
1261     misc_rate_ctrl->initial_qp = initial_qp;
1262     misc_rate_ctrl->min_qp = minimal_qp;
1263     misc_rate_ctrl->basic_unit_size = 0;
1264     vaUnmapBuffer(va_dpy, rc_param_buf);
1265
1266     render_id[0] = seq_param_buf;
1267     render_id[1] = rc_param_buf;
1268
1269     va_status = vaRenderPicture(va_dpy,context_id, &render_id[0], 2);
1270     CHECK_VASTATUS(va_status,"vaRenderPicture");;
1271
1272     return 0;
1273 }
1274
1275 static int calc_poc(int pic_order_cnt_lsb)
1276 {
1277     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1278     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1279     int PicOrderCntMsb, TopFieldOrderCnt;
1280
1281     if (current_frame_type == FRAME_IDR)
1282         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1283     else {
1284         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1285         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1286     }
1287
1288     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1289         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1290         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1291     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1292              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1293         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1294     else
1295         PicOrderCntMsb = prevPicOrderCntMsb;
1296
1297     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1298
1299     if (current_frame_type != FRAME_B) {
1300         PicOrderCntMsb_ref = PicOrderCntMsb;
1301         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1302     }
1303
1304     return TopFieldOrderCnt;
1305 }
1306
1307 static int render_picture(void)
1308 {
1309     VABufferID pic_param_buf;
1310     VAStatus va_status;
1311     int i = 0;
1312
1313     pic_param.CurrPic.picture_id = ref_surface[current_slot];
1314     pic_param.CurrPic.frame_idx = current_frame_num;
1315     pic_param.CurrPic.flags = 0;
1316     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1317     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1318     CurrentCurrPic = pic_param.CurrPic;
1319
1320     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1321         update_RefPicList(); /* calc RefPicList */
1322         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1323         if (current_frame_type == FRAME_P) {
1324             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1325         } else if (current_frame_type == FRAME_B) {
1326             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1327             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1328         }
1329     } else {
1330         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1331         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1332             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1333             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1334         }
1335     }
1336
1337     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1338     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1339     pic_param.pic_fields.bits.entropy_coding_mode_flag = 1;
1340     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1341     pic_param.frame_num = current_frame_num;
1342     pic_param.coded_buf = coded_buf[current_slot];
1343     pic_param.last_picture = (current_frame_encoding == frame_count);
1344     pic_param.pic_init_qp = initial_qp;
1345
1346     va_status = vaCreateBuffer(va_dpy, context_id,VAEncPictureParameterBufferType,
1347                                sizeof(pic_param),1,&pic_param, &pic_param_buf);
1348     CHECK_VASTATUS(va_status,"vaCreateBuffer");;
1349
1350     va_status = vaRenderPicture(va_dpy,context_id, &pic_param_buf, 1);
1351     CHECK_VASTATUS(va_status,"vaRenderPicture");
1352
1353     return 0;
1354 }
1355
1356 static int render_packedsequence(void)
1357 {
1358     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1359     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1360     unsigned int length_in_bits;
1361     unsigned char *packedseq_buffer = NULL;
1362     VAStatus va_status;
1363
1364     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1365
1366     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1367
1368     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1369     packedheader_param_buffer.has_emulation_bytes = 0;
1370     va_status = vaCreateBuffer(va_dpy,
1371                                context_id,
1372                                VAEncPackedHeaderParameterBufferType,
1373                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1374                                &packedseq_para_bufid);
1375     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1376
1377     va_status = vaCreateBuffer(va_dpy,
1378                                context_id,
1379                                VAEncPackedHeaderDataBufferType,
1380                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1381                                &packedseq_data_bufid);
1382     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1383
1384     render_id[0] = packedseq_para_bufid;
1385     render_id[1] = packedseq_data_bufid;
1386     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1387     CHECK_VASTATUS(va_status,"vaRenderPicture");
1388
1389     free(packedseq_buffer);
1390
1391     return 0;
1392 }
1393
1394
1395 static int render_packedpicture(void)
1396 {
1397     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1398     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1399     unsigned int length_in_bits;
1400     unsigned char *packedpic_buffer = NULL;
1401     VAStatus va_status;
1402
1403     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1404     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1405     packedheader_param_buffer.bit_length = length_in_bits;
1406     packedheader_param_buffer.has_emulation_bytes = 0;
1407
1408     va_status = vaCreateBuffer(va_dpy,
1409                                context_id,
1410                                VAEncPackedHeaderParameterBufferType,
1411                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1412                                &packedpic_para_bufid);
1413     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1414
1415     va_status = vaCreateBuffer(va_dpy,
1416                                context_id,
1417                                VAEncPackedHeaderDataBufferType,
1418                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1419                                &packedpic_data_bufid);
1420     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1421
1422     render_id[0] = packedpic_para_bufid;
1423     render_id[1] = packedpic_data_bufid;
1424     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1425     CHECK_VASTATUS(va_status,"vaRenderPicture");
1426
1427     free(packedpic_buffer);
1428
1429     return 0;
1430 }
1431
1432 static void render_packedsei(void)
1433 {
1434     VAEncPackedHeaderParameterBuffer packed_header_param_buffer;
1435     VABufferID packed_sei_header_param_buf_id, packed_sei_buf_id, render_id[2];
1436     unsigned int length_in_bits /*offset_in_bytes*/;
1437     unsigned char *packed_sei_buffer = NULL;
1438     VAStatus va_status;
1439     int init_cpb_size, target_bit_rate, i_initial_cpb_removal_delay_length, i_initial_cpb_removal_delay;
1440     int i_cpb_removal_delay, i_dpb_output_delay_length, i_cpb_removal_delay_length;
1441
1442     /* it comes for the bps defined in SPS */
1443     target_bit_rate = frame_bitrate;
1444     init_cpb_size = (target_bit_rate * 8) >> 10;
1445     i_initial_cpb_removal_delay = init_cpb_size * 0.5 * 1024 / target_bit_rate * 90000;
1446
1447     i_cpb_removal_delay = 2;
1448     i_initial_cpb_removal_delay_length = 24;
1449     i_cpb_removal_delay_length = 24;
1450     i_dpb_output_delay_length = 24;
1451
1452
1453     length_in_bits = build_packed_sei_buffer_timing(
1454         i_initial_cpb_removal_delay_length,
1455         i_initial_cpb_removal_delay,
1456         0,
1457         i_cpb_removal_delay_length,
1458         i_cpb_removal_delay * current_frame_encoding,
1459         i_dpb_output_delay_length,
1460         0,
1461         &packed_sei_buffer);
1462
1463     //offset_in_bytes = 0;
1464     packed_header_param_buffer.type = VAEncPackedHeaderH264_SEI;
1465     packed_header_param_buffer.bit_length = length_in_bits;
1466     packed_header_param_buffer.has_emulation_bytes = 0;
1467
1468     va_status = vaCreateBuffer(va_dpy,
1469                                context_id,
1470                                VAEncPackedHeaderParameterBufferType,
1471                                sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
1472                                &packed_sei_header_param_buf_id);
1473     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1474
1475     va_status = vaCreateBuffer(va_dpy,
1476                                context_id,
1477                                VAEncPackedHeaderDataBufferType,
1478                                (length_in_bits + 7) / 8, 1, packed_sei_buffer,
1479                                &packed_sei_buf_id);
1480     CHECK_VASTATUS(va_status,"vaCreateBuffer");
1481
1482
1483     render_id[0] = packed_sei_header_param_buf_id;
1484     render_id[1] = packed_sei_buf_id;
1485     va_status = vaRenderPicture(va_dpy,context_id, render_id, 2);
1486     CHECK_VASTATUS(va_status,"vaRenderPicture");
1487
1488
1489     free(packed_sei_buffer);
1490
1491     return;
1492 }
1493
1494
1495 static int render_hrd(void)
1496 {
1497     VABufferID misc_parameter_hrd_buf_id;
1498     VAStatus va_status;
1499     VAEncMiscParameterBuffer *misc_param;
1500     VAEncMiscParameterHRD *misc_hrd_param;
1501
1502     va_status = vaCreateBuffer(va_dpy, context_id,
1503                    VAEncMiscParameterBufferType,
1504                    sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterHRD),
1505                    1,
1506                    NULL,
1507                    &misc_parameter_hrd_buf_id);
1508     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1509
1510     vaMapBuffer(va_dpy,
1511                 misc_parameter_hrd_buf_id,
1512                 (void **)&misc_param);
1513     misc_param->type = VAEncMiscParameterTypeHRD;
1514     misc_hrd_param = (VAEncMiscParameterHRD *)misc_param->data;
1515
1516     if (frame_bitrate > 0) {
1517         misc_hrd_param->initial_buffer_fullness = frame_bitrate * 1024 * 4;
1518         misc_hrd_param->buffer_size = frame_bitrate * 1024 * 8;
1519     } else {
1520         misc_hrd_param->initial_buffer_fullness = 0;
1521         misc_hrd_param->buffer_size = 0;
1522     }
1523     vaUnmapBuffer(va_dpy, misc_parameter_hrd_buf_id);
1524
1525     va_status = vaRenderPicture(va_dpy,context_id, &misc_parameter_hrd_buf_id, 1);
1526     CHECK_VASTATUS(va_status,"vaRenderPicture");;
1527
1528     return 0;
1529 }
1530
1531 static int render_slice(void)
1532 {
1533     VABufferID slice_param_buf;
1534     VAStatus va_status;
1535     int i;
1536
1537     update_RefPicList();
1538
1539     /* one frame, one slice */
1540     slice_param.macroblock_address = 0;
1541     slice_param.num_macroblocks = frame_width*frame_height/(16*16); /* Measured by MB */
1542     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1543     if (current_frame_type == FRAME_IDR) {
1544         ++slice_param.idr_pic_id;
1545     } else if (current_frame_type == FRAME_P) {
1546         int refpiclist0_max = h264_maxref & 0xffff;
1547         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1548
1549         for (i = refpiclist0_max; i < 32; i++) {
1550             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1551             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1552         }
1553     } else if (current_frame_type == FRAME_B) {
1554         int refpiclist0_max = h264_maxref & 0xffff;
1555         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1556
1557         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1558         for (i = refpiclist0_max; i < 32; i++) {
1559             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1560             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1561         }
1562
1563         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1564         for (i = refpiclist1_max; i < 32; i++) {
1565             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1566             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1567         }
1568     }
1569
1570     slice_param.slice_alpha_c0_offset_div2 = 2;
1571     slice_param.slice_beta_offset_div2 = 2;
1572     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1573
1574     va_status = vaCreateBuffer(va_dpy,context_id,VAEncSliceParameterBufferType,
1575                                sizeof(slice_param),1,&slice_param,&slice_param_buf);
1576     CHECK_VASTATUS(va_status,"vaCreateBuffer");;
1577
1578     va_status = vaRenderPicture(va_dpy,context_id, &slice_param_buf, 1);
1579     CHECK_VASTATUS(va_status,"vaRenderPicture");
1580
1581     return 0;
1582 }
1583
1584
1585 static int upload_source_YUV_once_for_all()
1586 {
1587     int box_width=8;
1588     int row_shift=0;
1589     int i;
1590
1591     for (i = 0; i < SURFACE_NUM; i++) {
1592         printf("\rLoading data into surface %d.....", i);
1593         upload_surface(va_dpy, src_surface[i], box_width, row_shift, 0);
1594
1595         row_shift++;
1596         if (row_shift==(2*box_width)) row_shift= 0;
1597     }
1598     printf("Complete surface loading\n");
1599
1600     return 0;
1601 }
1602
1603 #define check_ret(ret)                                  \
1604 if (ret != 1) {                                         \
1605     printf("fread doesn't return enough data\n");       \
1606     exit(1);                                            \
1607 }
1608 static int load_surface(VASurfaceID surface_id, unsigned long long display_order)
1609 {
1610     unsigned char *src_Y = NULL, *src_U = NULL, *src_V = NULL;
1611     int ret = 0;
1612
1613     if (srcyuv_fp == NULL)
1614         return 0;
1615
1616     /* rewind the file pointer if encoding more than srcyuv_frames */
1617     display_order = display_order % srcyuv_frames;
1618     fseek(srcyuv_fp, display_order * frame_width * frame_height * 1.5, SEEK_SET);
1619
1620
1621     if (srcyuv_fourcc == VA_FOURCC_NV12) {
1622         int uv_size = 2 * (frame_width/2) * (frame_height/2);
1623
1624         src_Y = malloc(2 * uv_size);
1625         src_U = malloc(uv_size);
1626
1627         ret = fread(src_Y, frame_width * frame_height, 1, srcyuv_fp);
1628         check_ret(ret);
1629         ret = fread(src_U, uv_size, 1, srcyuv_fp);
1630         check_ret(ret);
1631     } else if (srcyuv_fourcc == VA_FOURCC_IYUV ||
1632         srcyuv_fourcc == VA_FOURCC_YV12) {
1633         int uv_size = (frame_width/2) * (frame_height/2);
1634
1635         src_Y = malloc(4 * uv_size);
1636         src_U = malloc(uv_size);
1637         src_V = malloc(uv_size);
1638
1639         ret = fread(src_Y, frame_width * frame_height, 1, srcyuv_fp);
1640         check_ret(ret);
1641         if (srcyuv_fourcc == VA_FOURCC_IYUV) {
1642             ret = fread(src_U, uv_size, 1, srcyuv_fp);
1643             check_ret(ret);
1644             ret = fread(src_V, uv_size, 1, srcyuv_fp);
1645             check_ret(ret);
1646         } else { /* YV12 */
1647             ret = fread(src_V, uv_size, 1, srcyuv_fp);
1648             check_ret(ret);
1649             ret = fread(src_U, uv_size, 1, srcyuv_fp);
1650             check_ret(ret);
1651         }
1652     } else {
1653         printf("Unsupported source YUV format\n");
1654         exit(1);
1655     }
1656
1657     upload_surface_yuv(va_dpy, surface_id,
1658                        srcyuv_fourcc, frame_width, frame_height,
1659                        src_Y, src_U, src_V);
1660     if (src_Y)
1661         free(src_Y);
1662     if (src_U)
1663         free(src_U);
1664     if (src_V)
1665         free(src_V);
1666
1667     return 0;
1668 }
1669
1670
1671 static int save_recyuv(VASurfaceID surface_id,
1672                        unsigned long long display_order,
1673                        unsigned long long encode_order)
1674 {
1675     unsigned char *dst_Y = NULL, *dst_U = NULL, *dst_V = NULL;
1676
1677     if (recyuv_fp == NULL)
1678         return 0;
1679
1680     if (srcyuv_fourcc == VA_FOURCC_NV12) {
1681         int uv_size = 2 * (frame_width/2) * (frame_height/2);
1682         dst_Y = malloc(2*uv_size);
1683         dst_U = malloc(uv_size);
1684     } else if (srcyuv_fourcc == VA_FOURCC_IYUV ||
1685                srcyuv_fourcc == VA_FOURCC_YV12) {
1686         int uv_size = (frame_width/2) * (frame_height/2);
1687         dst_Y = malloc(4*uv_size);
1688         dst_U = malloc(uv_size);
1689         dst_V = malloc(uv_size);
1690     } else {
1691         printf("Unsupported source YUV format\n");
1692         exit(1);
1693     }
1694
1695     download_surface_yuv(va_dpy, surface_id,
1696                          srcyuv_fourcc, frame_width, frame_height,
1697                          dst_Y, dst_U, dst_V);
1698     fseek(recyuv_fp, display_order * frame_width * frame_height * 1.5, SEEK_SET);
1699
1700     if (srcyuv_fourcc == VA_FOURCC_NV12) {
1701         int uv_size = 2 * (frame_width/2) * (frame_height/2);
1702         fwrite(dst_Y, uv_size * 2, 1, recyuv_fp);
1703         fwrite(dst_U, uv_size, 1, recyuv_fp);
1704     } else if (srcyuv_fourcc == VA_FOURCC_IYUV ||
1705                srcyuv_fourcc == VA_FOURCC_YV12) {
1706         int uv_size = (frame_width/2) * (frame_height/2);
1707         fwrite(dst_Y, uv_size * 4, 1, recyuv_fp);
1708
1709         if (srcyuv_fourcc == VA_FOURCC_IYUV) {
1710             fwrite(dst_U, uv_size, 1, recyuv_fp);
1711             fwrite(dst_V, uv_size, 1, recyuv_fp);
1712         } else {
1713             fwrite(dst_V, uv_size, 1, recyuv_fp);
1714             fwrite(dst_U, uv_size, 1, recyuv_fp);
1715         }
1716     } else {
1717         printf("Unsupported YUV format\n");
1718         exit(1);
1719     }
1720
1721     if (dst_Y)
1722         free(dst_Y);
1723     if (dst_U)
1724         free(dst_U);
1725     if (dst_V)
1726         free(dst_V);
1727
1728     return 0;
1729 }
1730
1731
1732 static int save_codeddata(unsigned long long display_order, unsigned long long encode_order)
1733 {
1734     VACodedBufferSegment *buf_list = NULL;
1735     VAStatus va_status;
1736     unsigned int coded_size = 0;
1737
1738     va_status = vaMapBuffer(va_dpy,coded_buf[display_order % SURFACE_NUM],(void **)(&buf_list));
1739     CHECK_VASTATUS(va_status,"vaMapBuffer");
1740     while (buf_list != NULL) {
1741         coded_size += fwrite(buf_list->buf, 1, buf_list->size, coded_fp);
1742         buf_list = (VACodedBufferSegment *) buf_list->next;
1743
1744         frame_size += coded_size;
1745     }
1746     vaUnmapBuffer(va_dpy,coded_buf[display_order % SURFACE_NUM]);
1747
1748     printf("\r      "); /* return back to startpoint */
1749     switch (encode_order % 4) {
1750         case 0:
1751             printf("|");
1752             break;
1753         case 1:
1754             printf("/");
1755             break;
1756         case 2:
1757             printf("-");
1758             break;
1759         case 3:
1760             printf("\\");
1761             break;
1762     }
1763     printf("%08lld", encode_order);
1764     /*
1765     if (current_frame_encoding % intra_count == 0)
1766         printf("(I)");
1767     else
1768         printf("(P)");
1769     */
1770     printf("(%06d bytes coded)",coded_size);
1771     /* skipped frame ? */
1772     printf("                                    ");
1773
1774     return 0;
1775 }
1776
1777
1778 static struct storage_task_t * storage_task_dequeue(void)
1779 {
1780     struct storage_task_t *header;
1781
1782     pthread_mutex_lock(&encode_mutex);
1783
1784     header = storage_task_header;
1785     if (storage_task_header != NULL) {
1786         if (storage_task_tail == storage_task_header)
1787             storage_task_tail = NULL;
1788         storage_task_header = header->next;
1789     }
1790
1791     pthread_mutex_unlock(&encode_mutex);
1792
1793     return header;
1794 }
1795
1796 static int storage_task_queue(unsigned long long display_order, unsigned long long encode_order)
1797 {
1798     struct storage_task_t *tmp;
1799
1800     tmp = calloc(1, sizeof(struct storage_task_t));
1801     tmp->display_order = display_order;
1802     tmp->encode_order = encode_order;
1803
1804     pthread_mutex_lock(&encode_mutex);
1805
1806     if (storage_task_header == NULL) {
1807         storage_task_header = tmp;
1808         storage_task_tail = tmp;
1809     } else {
1810         storage_task_tail->next = tmp;
1811         storage_task_tail = tmp;
1812     }
1813
1814     srcsurface_status[display_order % SURFACE_NUM] = SRC_SURFACE_IN_STORAGE;
1815     pthread_cond_signal(&encode_cond);
1816
1817     pthread_mutex_unlock(&encode_mutex);
1818
1819     return 0;
1820 }
1821
1822 static void storage_task(unsigned long long display_order, unsigned long long encode_order)
1823 {
1824     unsigned int tmp;
1825     VAStatus va_status;
1826
1827     tmp = GetTickCount();
1828     va_status = vaSyncSurface(va_dpy, src_surface[display_order % SURFACE_NUM]);
1829     CHECK_VASTATUS(va_status,"vaSyncSurface");
1830     SyncPictureTicks += GetTickCount() - tmp;
1831     tmp = GetTickCount();
1832     save_codeddata(display_order, encode_order);
1833     SavePictureTicks += GetTickCount() - tmp;
1834
1835     save_recyuv(ref_surface[display_order % SURFACE_NUM], display_order, encode_order);
1836
1837     /* reload a new frame data */
1838     tmp = GetTickCount();
1839     if (srcyuv_fp != NULL)
1840         load_surface(src_surface[display_order % SURFACE_NUM], display_order + SURFACE_NUM);
1841     UploadPictureTicks += GetTickCount() - tmp;
1842
1843     pthread_mutex_lock(&encode_mutex);
1844     srcsurface_status[display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1845     pthread_mutex_unlock(&encode_mutex);
1846 }
1847
1848
1849 static void * storage_task_thread(void *t)
1850 {
1851     while (1) {
1852         struct storage_task_t *current;
1853
1854         current = storage_task_dequeue();
1855         if (current == NULL) {
1856             pthread_mutex_lock(&encode_mutex);
1857             pthread_cond_wait(&encode_cond, &encode_mutex);
1858             pthread_mutex_unlock(&encode_mutex);
1859             continue;
1860         }
1861
1862         storage_task(current->display_order, current->encode_order);
1863
1864         free(current);
1865
1866         /* all frames are saved, exit the thread */
1867         if (++frame_coded >= frame_count)
1868             break;
1869     }
1870
1871     return 0;
1872 }
1873
1874
1875 static int encode_frames(void)
1876 {
1877     unsigned int i, tmp;
1878     VAStatus va_status;
1879     //VASurfaceStatus surface_status;
1880
1881     /* upload RAW YUV data into all surfaces */
1882     tmp = GetTickCount();
1883     if (srcyuv_fp != NULL) {
1884         for (i = 0; i < SURFACE_NUM; i++)
1885             load_surface(src_surface[i], i);
1886     } else
1887         upload_source_YUV_once_for_all();
1888     UploadPictureTicks += GetTickCount() - tmp;
1889
1890     /* ready for encoding */
1891     memset(srcsurface_status, SRC_SURFACE_IN_ENCODING, sizeof(srcsurface_status));
1892
1893     memset(&seq_param, 0, sizeof(seq_param));
1894     memset(&pic_param, 0, sizeof(pic_param));
1895     memset(&slice_param, 0, sizeof(slice_param));
1896
1897     if (encode_syncmode == 0)
1898         pthread_create(&encode_thread, NULL, storage_task_thread, NULL);
1899
1900     for (current_frame_encoding = 0; current_frame_encoding < frame_count; current_frame_encoding++) {
1901         encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1902                                &current_frame_display, &current_frame_type);
1903         if (current_frame_type == FRAME_IDR) {
1904             numShortTerm = 0;
1905             current_frame_num = 0;
1906             current_IDR_display = current_frame_display;
1907         }
1908
1909         /* check if the source frame is ready */
1910         while (srcsurface_status[current_slot] != SRC_SURFACE_IN_ENCODING) {
1911             usleep(1);
1912         }
1913
1914         tmp = GetTickCount();
1915         va_status = vaBeginPicture(va_dpy, context_id, src_surface[current_slot]);
1916         CHECK_VASTATUS(va_status,"vaBeginPicture");
1917         BeginPictureTicks += GetTickCount() - tmp;
1918
1919         tmp = GetTickCount();
1920         if (current_frame_type == FRAME_IDR) {
1921             render_sequence();
1922             render_picture();
1923             if (h264_packedheader) {
1924                 render_packedsequence();
1925                 render_packedpicture();
1926             }
1927             //if (rc_mode == VA_RC_CBR)
1928             //    render_packedsei();
1929             //render_hrd();
1930         } else {
1931             //render_sequence();
1932             render_picture();
1933             //if (rc_mode == VA_RC_CBR)
1934             //    render_packedsei();
1935             //render_hrd();
1936         }
1937         render_slice();
1938         RenderPictureTicks += GetTickCount() - tmp;
1939
1940         tmp = GetTickCount();
1941         va_status = vaEndPicture(va_dpy,context_id);
1942         CHECK_VASTATUS(va_status,"vaEndPicture");;
1943         EndPictureTicks += GetTickCount() - tmp;
1944
1945         if (encode_syncmode)
1946             storage_task(current_frame_display, current_frame_encoding);
1947         else /* queue the storage task queue */
1948             storage_task_queue(current_frame_display, current_frame_encoding);
1949
1950         /* how to process skipped frames
1951            surface_status = (VASurfaceStatus) 0;
1952            va_status = vaQuerySurfaceStatus(va_dpy, src_surface[i%SURFACE_NUM],&surface_status);
1953            frame_skipped = (surface_status & VASurfaceSkipped);
1954         */
1955         update_ReferenceFrames();
1956     }
1957
1958     if (encode_syncmode == 0) {
1959         int ret;
1960         pthread_join(encode_thread, (void **)&ret);
1961     }
1962
1963     return 0;
1964 }
1965
1966
1967 static int release_encode()
1968 {
1969     int i;
1970
1971     vaDestroySurfaces(va_dpy,&src_surface[0],SURFACE_NUM);
1972     vaDestroySurfaces(va_dpy,&ref_surface[0],SURFACE_NUM);
1973
1974     for (i = 0; i < SURFACE_NUM; i++)
1975         vaDestroyBuffer(va_dpy,coded_buf[i]);
1976
1977     vaDestroyContext(va_dpy,context_id);
1978     vaDestroyConfig(va_dpy,config_id);
1979
1980     return 0;
1981 }
1982
1983 static int deinit_va()
1984 {
1985     vaTerminate(va_dpy);
1986
1987     va_close_display(va_dpy);
1988
1989     return 0;
1990 }
1991
1992
1993 static int print_input()
1994 {
1995     printf("\n\nINPUT:Try to encode H264...\n");
1996     printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1997     printf("INPUT: Resolution   : %dx%d, %d frames\n",
1998            frame_width, frame_height, frame_count);
1999     printf("INPUT: FrameRate    : %d\n", frame_rate);
2000     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
2001     printf("INPUT: Slieces      : %d\n", frame_slices);
2002     printf("INPUT: IntraPeriod  : %d\n", intra_period);
2003     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
2004     printf("INPUT: IpPeriod     : %d\n", ip_period);
2005     printf("INPUT: Initial QP   : %d\n", initial_qp);
2006     printf("INPUT: Min QP       : %d\n", minimal_qp);
2007     printf("INPUT: Source YUV   : %s", srcyuv_fp?"FILE":"AUTO generated");
2008     if (srcyuv_fp)
2009         printf(":%s (fourcc %s)\n", srcyuv_fn, fourcc_to_string(srcyuv_fourcc));
2010     else
2011         printf("\n");
2012     printf("INPUT: Coded Clip   : %s\n", coded_fn);
2013     if (recyuv_fp == NULL)
2014         printf("INPUT: Rec   Clip   : %s\n", "Not save reconstructed frame");
2015     else
2016         printf("INPUT: Rec   Clip   : Save reconstructed frame into %s (fourcc %s)\n", recyuv_fn,
2017                fourcc_to_string(srcyuv_fourcc));
2018
2019     printf("\n\n"); /* return back to startpoint */
2020
2021     return 0;
2022 }
2023
2024 static int calc_PSNR(double *psnr)
2025 {
2026     unsigned long srcyuv_size, recyuv_size, min_size;
2027     char *srcyuv_ptr, *recyuv_ptr;
2028     unsigned long i, sse=0;
2029     double ssemean;
2030
2031     fseek(srcyuv_fp, 0L, SEEK_END);
2032     srcyuv_size = ftell(srcyuv_fp);
2033     fseek(recyuv_fp, 0L, SEEK_END);
2034     recyuv_size = ftell(recyuv_fp);
2035
2036     fseek(srcyuv_fp, 0L, SEEK_SET);
2037     fseek(recyuv_fp, 0L, SEEK_SET);
2038
2039     min_size = MIN(srcyuv_size,recyuv_size);
2040     srcyuv_ptr = mmap (0, min_size, PROT_READ, MAP_SHARED, fileno(srcyuv_fp), 0);
2041     recyuv_ptr = mmap (0, min_size, PROT_READ, MAP_SHARED, fileno(recyuv_fp), 0);
2042     if ((srcyuv_ptr == MAP_FAILED) || (recyuv_ptr == MAP_FAILED)) {
2043         printf("Failed to mmap YUV files\n");
2044         return 1;
2045     }
2046
2047     for (i=0; i<min_size; i++) {
2048         char tmp = srcyuv_ptr[i] - recyuv_ptr[i];
2049         sse += tmp * tmp;
2050     }
2051     ssemean = (double)sse/(double)min_size;
2052     *psnr = 20.0*log10(255) - 10.0*log10(ssemean);
2053
2054     munmap(srcyuv_ptr, min_size);
2055     munmap(recyuv_ptr, min_size);
2056
2057     return 0;
2058 }
2059
2060 static int print_performance(unsigned int PictureCount)
2061 {
2062     unsigned int psnr_ret = 1, others = 0;
2063     double psnr = 0, total_size = frame_width * frame_height * 1.5 * frame_count;
2064
2065     if (calc_psnr && srcyuv_fp && recyuv_fp)
2066         psnr_ret = calc_PSNR(&psnr);
2067
2068     others = TotalTicks - UploadPictureTicks - BeginPictureTicks
2069         - RenderPictureTicks - EndPictureTicks - SyncPictureTicks - SavePictureTicks;
2070
2071     printf("\n\n");
2072
2073     printf("PERFORMANCE:   Frame Rate           : %.2f fps (%d frames, %d ms (%.2f ms per frame))\n",
2074            (double) 1000*PictureCount / TotalTicks, PictureCount,
2075            TotalTicks, ((double)  TotalTicks) / (double) PictureCount);
2076     printf("PERFORMANCE:   Compression ratio    : %d:1\n", (unsigned int)(total_size / frame_size));
2077     if (psnr_ret == 0)
2078         printf("PERFORMANCE:   PSNR                 : %.2f (%lld frames calculated)\n",
2079                psnr, MIN(frame_count, srcyuv_frames));
2080
2081     printf("PERFORMANCE:     UploadPicture      : %d ms (%.2f, %.2f%% percent)\n",
2082            (int) UploadPictureTicks, ((double)  UploadPictureTicks) / (double) PictureCount,
2083            UploadPictureTicks/(double) TotalTicks/0.01);
2084     printf("PERFORMANCE:     vaBeginPicture     : %d ms (%.2f, %.2f%% percent)\n",
2085            (int) BeginPictureTicks, ((double)  BeginPictureTicks) / (double) PictureCount,
2086            BeginPictureTicks/(double) TotalTicks/0.01);
2087     printf("PERFORMANCE:     vaRenderHeader     : %d ms (%.2f, %.2f%% percent)\n",
2088            (int) RenderPictureTicks, ((double)  RenderPictureTicks) / (double) PictureCount,
2089            RenderPictureTicks/(double) TotalTicks/0.01);
2090     printf("PERFORMANCE:     vaEndPicture       : %d ms (%.2f, %.2f%% percent)\n",
2091            (int) EndPictureTicks, ((double)  EndPictureTicks) / (double) PictureCount,
2092            EndPictureTicks/(double) TotalTicks/0.01);
2093     printf("PERFORMANCE:     vaSyncSurface      : %d ms (%.2f, %.2f%% percent)\n",
2094            (int) SyncPictureTicks, ((double) SyncPictureTicks) / (double) PictureCount,
2095            SyncPictureTicks/(double) TotalTicks/0.01);
2096     printf("PERFORMANCE:     SavePicture        : %d ms (%.2f, %.2f%% percent)\n",
2097            (int) SavePictureTicks, ((double)  SavePictureTicks) / (double) PictureCount,
2098            SavePictureTicks/(double) TotalTicks/0.01);
2099     printf("PERFORMANCE:     Others             : %d ms (%.2f, %.2f%% percent)\n",
2100            (int) others, ((double) others) / (double) PictureCount,
2101            others/(double) TotalTicks/0.01);
2102
2103     if (encode_syncmode == 0)
2104         printf("(Multithread enabled, the timing is only for reference)\n");
2105
2106     return 0;
2107 }
2108
2109
2110 int main(int argc,char **argv)
2111 {
2112     unsigned int start;
2113
2114     process_cmdline(argc, argv);
2115
2116     print_input();
2117
2118     start = GetTickCount();
2119
2120     init_va();
2121     setup_encode();
2122
2123     encode_frames();
2124
2125     release_encode();
2126     deinit_va();
2127
2128     TotalTicks += GetTickCount() - start;
2129     print_performance(frame_count);
2130
2131     return 0;
2132 }