git.sesse.net Git - nageru/blob - h264encode.cpp

   1 //#include "sysdeps.h"
   2 #include <stdio.h>
   3 #include <string.h>
   4 #include <stdlib.h>
   5 #include <getopt.h>
   6 #include <unistd.h>
   7 #include <sys/types.h>
   8 #include <sys/stat.h>
   9 #include <sys/time.h>
  10 #include <sys/mman.h>
  11 #include <fcntl.h>
  12 #include <assert.h>
  13 #include <pthread.h>
  14 #include <errno.h>
  15 #include <math.h>
  16 #include <va/va.h>
  17 #include <va/va_x11.h>
  18 #include <va/va_enc_h264.h>
  19 #include <va/va_drmcommon.h>
  20 #include <libdrm/drm_fourcc.h>
  21 #include <thread>
  22 #include <mutex>
  23 #include <queue>
  24 #include <condition_variable>
  25 #include "h264encode.h"
  26
  27 #define CHECK_VASTATUS(va_status, func)                                 \
  28     if (va_status != VA_STATUS_SUCCESS) {                               \
  29         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  30         exit(1);                                                        \
  31     }
  32
  33 //#include "loadsurface.h"
  34
  35 #define NAL_REF_IDC_NONE        0
  36 #define NAL_REF_IDC_LOW         1
  37 #define NAL_REF_IDC_MEDIUM      2
  38 #define NAL_REF_IDC_HIGH        3
  39
  40 #define NAL_NON_IDR             1
  41 #define NAL_IDR                 5
  42 #define NAL_SPS                 7
  43 #define NAL_PPS                 8
  44 #define NAL_SEI                 6
  45
  46 #define SLICE_TYPE_P            0
  47 #define SLICE_TYPE_B            1
  48 #define SLICE_TYPE_I            2
  49 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  50 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  51 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  52
  53
  54 #define ENTROPY_MODE_CAVLC      0
  55 #define ENTROPY_MODE_CABAC      1
  56
  57 #define PROFILE_IDC_BASELINE    66
  58 #define PROFILE_IDC_MAIN        77
  59 #define PROFILE_IDC_HIGH        100
  60
  61 #define BITSTREAM_ALLOCATE_STEPPING     4096
  62
  63 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  64 static  VADisplay va_dpy;
  65 static  VAProfile h264_profile = (VAProfile)~0;
  66 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  67 static  int config_attrib_num = 0, enc_packed_header_idx;
  68
  69 struct GLSurface {
  70         VASurfaceID src_surface, ref_surface;
  71         VABufferID coded_buf;
  72
  73         VAImage surface_image;
  74         GLuint y_tex, cbcr_tex;
  75         EGLImage y_egl_image, cbcr_egl_image;
  76 };
  77 GLSurface gl_surfaces[SURFACE_NUM];
  78
  79 static  VAConfigID config_id;
  80 static  VAContextID context_id;
  81 static  VAEncSequenceParameterBufferH264 seq_param;
  82 static  VAEncPictureParameterBufferH264 pic_param;
  83 static  VAEncSliceParameterBufferH264 slice_param;
  84 static  VAPictureH264 CurrentCurrPic;
  85 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  86
  87 static  unsigned int MaxFrameNum = (2<<16);
  88 static  unsigned int MaxPicOrderCntLsb = (2<<8);
  89 static  unsigned int Log2MaxFrameNum = 16;
  90 static  unsigned int Log2MaxPicOrderCntLsb = 8;
  91
  92 static  unsigned int num_ref_frames = 2;
  93 static  unsigned int numShortTerm = 0;
  94 static  int constraint_set_flag = 0;
  95 static  int h264_packedheader = 0; /* support pack header? */
  96 static  int h264_maxref = (1<<16|1);
  97 static  int h264_entropy_mode = 1; /* cabac */
  98
  99 static  char *coded_fn = NULL;
 100 static  FILE *coded_fp = NULL;
 101
 102 static  int frame_width = 176;
 103 static  int frame_height = 144;
 104 static  int frame_width_mbaligned;
 105 static  int frame_height_mbaligned;
 106 static  int frame_rate = 60;
 107 static  unsigned int frame_bitrate = 0;
 108 static  unsigned int frame_slices = 1;
 109 static  double frame_size = 0;
 110 static  int initial_qp = 15;
 111 //static  int initial_qp = 28;
 112 static  int minimal_qp = 0;
 113 static  int intra_period = 30;
 114 static  int intra_idr_period = 60;
 115 static  int ip_period = 1;
 116 static  int rc_mode = -1;
 117 static  int rc_default_modes[] = {
 118     VA_RC_VBR,
 119     VA_RC_CQP,
 120     VA_RC_VBR_CONSTRAINED,
 121     VA_RC_CBR,
 122     VA_RC_VCM,
 123     VA_RC_NONE,
 124 };
 125 static  unsigned long long current_frame_encoding = 0;
 126 static  unsigned long long current_frame_display = 0;
 127 static  unsigned long long current_IDR_display = 0;
 128 static  unsigned int current_frame_num = 0;
 129 static  int current_frame_type;
 130
 131 static  int misc_priv_type = 0;
 132 static  int misc_priv_value = 0;
 133
 134 /* thread to save coded data */
 135 #define SRC_SURFACE_FREE        0
 136 #define SRC_SURFACE_IN_ENCODING 1
 137
 138 struct __bitstream {
 139     unsigned int *buffer;
 140     int bit_offset;
 141     int max_size_in_dword;
 142 };
 143 typedef struct __bitstream bitstream;
 144
 145 using namespace std;
 146
 147 static unsigned int
 148 va_swap32(unsigned int val)
 149 {
 150     unsigned char *pval = (unsigned char *)&val;
 151
 152     return ((pval[0] << 24)     |
 153             (pval[1] << 16)     |
 154             (pval[2] << 8)      |
 155             (pval[3] << 0));
 156 }
 157
 158 static void
 159 bitstream_start(bitstream *bs)
 160 {
 161     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 162     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 163     bs->bit_offset = 0;
 164 }
 165
 166 static void
 167 bitstream_end(bitstream *bs)
 168 {
 169     int pos = (bs->bit_offset >> 5);
 170     int bit_offset = (bs->bit_offset & 0x1f);
 171     int bit_left = 32 - bit_offset;
 172
 173     if (bit_offset) {
 174         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 175     }
 176 }
 177
 178 static void
 179 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 180 {
 181     int pos = (bs->bit_offset >> 5);
 182     int bit_offset = (bs->bit_offset & 0x1f);
 183     int bit_left = 32 - bit_offset;
 184
 185     if (!size_in_bits)
 186         return;
 187
 188     bs->bit_offset += size_in_bits;
 189
 190     if (bit_left > size_in_bits) {
 191         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 192     } else {
 193         size_in_bits -= bit_left;
 194         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 195         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 196
 197         if (pos + 1 == bs->max_size_in_dword) {
 198             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 199             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 200         }
 201
 202         bs->buffer[pos + 1] = val;
 203     }
 204 }
 205
 206 static void
 207 bitstream_put_ue(bitstream *bs, unsigned int val)
 208 {
 209     int size_in_bits = 0;
 210     int tmp_val = ++val;
 211
 212     while (tmp_val) {
 213         tmp_val >>= 1;
 214         size_in_bits++;
 215     }
 216
 217     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 218     bitstream_put_ui(bs, val, size_in_bits);
 219 }
 220
 221 static void
 222 bitstream_put_se(bitstream *bs, int val)
 223 {
 224     unsigned int new_val;
 225
 226     if (val <= 0)
 227         new_val = -2 * val;
 228     else
 229         new_val = 2 * val - 1;
 230
 231     bitstream_put_ue(bs, new_val);
 232 }
 233
 234 static void
 235 bitstream_byte_aligning(bitstream *bs, int bit)
 236 {
 237     int bit_offset = (bs->bit_offset & 0x7);
 238     int bit_left = 8 - bit_offset;
 239     int new_val;
 240
 241     if (!bit_offset)
 242         return;
 243
 244     assert(bit == 0 || bit == 1);
 245
 246     if (bit)
 247         new_val = (1 << bit_left) - 1;
 248     else
 249         new_val = 0;
 250
 251     bitstream_put_ui(bs, new_val, bit_left);
 252 }
 253
 254 static void
 255 rbsp_trailing_bits(bitstream *bs)
 256 {
 257     bitstream_put_ui(bs, 1, 1);
 258     bitstream_byte_aligning(bs, 0);
 259 }
 260
 261 static void nal_start_code_prefix(bitstream *bs)
 262 {
 263     bitstream_put_ui(bs, 0x00000001, 32);
 264 }
 265
 266 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 267 {
 268     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 269     bitstream_put_ui(bs, nal_ref_idc, 2);
 270     bitstream_put_ui(bs, nal_unit_type, 5);
 271 }
 272
 273 static void sps_rbsp(bitstream *bs)
 274 {
 275     int profile_idc = PROFILE_IDC_BASELINE;
 276
 277     if (h264_profile  == VAProfileH264High)
 278         profile_idc = PROFILE_IDC_HIGH;
 279     else if (h264_profile  == VAProfileH264Main)
 280         profile_idc = PROFILE_IDC_MAIN;
 281
 282     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 283     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 284     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 285     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 286     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 287     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 288     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 289     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 290
 291     if ( profile_idc == PROFILE_IDC_HIGH) {
 292         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 293         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 294         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 295         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 296         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 297     }
 298
 299     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 300     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 301
 302     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 303         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 304     else {
 305         assert(0);
 306     }
 307
 308     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 309     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 310
 311     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 312     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 313     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 314
 315     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 316         assert(0);
 317     }
 318
 319     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 320     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 321
 322     if (seq_param.frame_cropping_flag) {
 323         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 324         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 325         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 326         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 327     }
 328
 329     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 330     if ( false ) {
 331         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 332     } else {
 333         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 334         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 335         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 336         bitstream_put_ui(bs, 0, 1); /* video_signal_type_present_flag */
 337         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 338         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 339         {
 340             bitstream_put_ui(bs, 1, 32);  // FPS
 341             bitstream_put_ui(bs, frame_rate * 2, 32);  // FPS
 342             bitstream_put_ui(bs, 1, 1);
 343         }
 344         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 345         {
 346             // hrd_parameters
 347             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 348             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 349             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 350
 351             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 352             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 353             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 354
 355             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 356             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 357             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 358             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 359         }
 360         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 361         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 362
 363         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 364         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 365     }
 366
 367     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 368 }
 369
 370
 371 static void pps_rbsp(bitstream *bs)
 372 {
 373     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 374     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 375
 376     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 377
 378     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 379
 380     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 381
 382     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 383     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 384
 385     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 386     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 387
 388     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 389     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 390     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 391
 392     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 393     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 394     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 395
 396     /* more_rbsp_data */
 397     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 398     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 399     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 400
 401     rbsp_trailing_bits(bs);
 402 }
 403
 404 static void slice_header(bitstream *bs)
 405 {
 406     int first_mb_in_slice = slice_param.macroblock_address;
 407
 408     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 409     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 410     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 411     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 412
 413     /* frame_mbs_only_flag == 1 */
 414     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 415         /* FIXME: */
 416         assert(0);
 417     }
 418
 419     if (pic_param.pic_fields.bits.idr_pic_flag)
 420         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 421
 422     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 423         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 424         /* pic_order_present_flag == 0 */
 425     } else {
 426         /* FIXME: */
 427         assert(0);
 428     }
 429
 430     /* redundant_pic_cnt_present_flag == 0 */
 431     /* slice type */
 432     if (IS_P_SLICE(slice_param.slice_type)) {
 433         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 434
 435         if (slice_param.num_ref_idx_active_override_flag)
 436             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 437
 438         /* ref_pic_list_reordering */
 439         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 440     } else if (IS_B_SLICE(slice_param.slice_type)) {
 441         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 442
 443         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 444
 445         if (slice_param.num_ref_idx_active_override_flag) {
 446             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 447             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 448         }
 449
 450         /* ref_pic_list_reordering */
 451         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 452         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 453     }
 454
 455     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 456          IS_P_SLICE(slice_param.slice_type)) ||
 457         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 458          IS_B_SLICE(slice_param.slice_type))) {
 459         /* FIXME: fill weight/offset table */
 460         assert(0);
 461     }
 462
 463     /* dec_ref_pic_marking */
 464     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 465         unsigned char no_output_of_prior_pics_flag = 0;
 466         unsigned char long_term_reference_flag = 0;
 467         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 468
 469         if (pic_param.pic_fields.bits.idr_pic_flag) {
 470             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 471             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 472         } else {
 473             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 474         }
 475     }
 476
 477     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 478         !IS_I_SLICE(slice_param.slice_type))
 479         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 480
 481     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 482
 483     /* ignore for SP/SI */
 484
 485     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 486         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 487
 488         if (slice_param.disable_deblocking_filter_idc != 1) {
 489             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 490             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 491         }
 492     }
 493
 494     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 495         bitstream_byte_aligning(bs, 1);
 496     }
 497 }
 498
 499 static int
 500 build_packed_pic_buffer(unsigned char **header_buffer)
 501 {
 502     bitstream bs;
 503
 504     bitstream_start(&bs);
 505     nal_start_code_prefix(&bs);
 506     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 507     pps_rbsp(&bs);
 508     bitstream_end(&bs);
 509
 510     *header_buffer = (unsigned char *)bs.buffer;
 511     return bs.bit_offset;
 512 }
 513
 514 static int
 515 build_packed_seq_buffer(unsigned char **header_buffer)
 516 {
 517     bitstream bs;
 518
 519     bitstream_start(&bs);
 520     nal_start_code_prefix(&bs);
 521     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 522     sps_rbsp(&bs);
 523     bitstream_end(&bs);
 524
 525     *header_buffer = (unsigned char *)bs.buffer;
 526     return bs.bit_offset;
 527 }
 528
 529 static int build_packed_slice_buffer(unsigned char **header_buffer)
 530 {
 531     bitstream bs;
 532     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 533     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 534
 535     bitstream_start(&bs);
 536     nal_start_code_prefix(&bs);
 537
 538     if (IS_I_SLICE(slice_param.slice_type)) {
 539         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 540     } else if (IS_P_SLICE(slice_param.slice_type)) {
 541         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 542     } else {
 543         assert(IS_B_SLICE(slice_param.slice_type));
 544         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 545     }
 546
 547     slice_header(&bs);
 548     bitstream_end(&bs);
 549
 550     *header_buffer = (unsigned char *)bs.buffer;
 551     return bs.bit_offset;
 552 }
 553
 554
 555 /*
 556   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 557   1) period between Frame #X and Frame #N = #X - #N
 558   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 559   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 560   4) intra_period and intra_idr_period take precedence over ip_period
 561   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 562      of I/IDR frames, see bellow examples
 563   -------------------------------------------------------------------
 564   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 565   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 566   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 567   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 568   1            1                ignored    IDR IDR IDR IDR...
 569   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 570   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 571   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 572                                               (PBB)(IBB)(PBB)(IBB)...
 573   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 574                                            IDRPPPPP IPPPPP IPPPPP...
 575   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 576                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 577                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 578                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 579                                            {IDR(PBB)(PBB)}                     (6/6/3)
 580                                            {IDR(PBB)(PBB)}.
 581 */
 582
 583 /*
 584  * Return displaying order with specified periods and encoding order
 585  * displaying_order: displaying order
 586  * frame_type: frame type
 587  */
 588 #define FRAME_P 0
 589 #define FRAME_B 1
 590 #define FRAME_I 2
 591 #define FRAME_IDR 7
 592 void encoding2display_order(
 593     unsigned long long encoding_order, int intra_period,
 594     int intra_idr_period, int ip_period,
 595     unsigned long long *displaying_order,
 596     int *frame_type)
 597 {
 598     int encoding_order_gop = 0;
 599
 600     if (intra_period == 1) { /* all are I/IDR frames */
 601         *displaying_order = encoding_order;
 602         if (intra_idr_period == 0)
 603             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 604         else
 605             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 606         return;
 607     }
 608
 609     if (intra_period == 0)
 610         intra_idr_period = 0;
 611
 612     /* new sequence like
 613      * IDR PPPPP IPPPPP
 614      * IDR (PBB)(PBB)(IBB)(PBB)
 615      */
 616     encoding_order_gop = (intra_idr_period == 0)? encoding_order:
 617         (encoding_order % (intra_idr_period + ((ip_period == 1)?0:1)));
 618
 619     if (encoding_order_gop == 0) { /* the first frame */
 620         *frame_type = FRAME_IDR;
 621         *displaying_order = encoding_order;
 622     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 623         *frame_type = FRAME_B;
 624         *displaying_order = encoding_order - 1;
 625     } else if ((intra_period != 0) && /* have I frames */
 626                (encoding_order_gop >= 2) &&
 627                ((ip_period == 1 && encoding_order_gop % intra_period == 0) || /* for IDR PPPPP IPPPP */
 628                 /* for IDR (PBB)(PBB)(IBB) */
 629                 (ip_period >= 2 && ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0))) {
 630         *frame_type = FRAME_I;
 631         *displaying_order = encoding_order + ip_period - 1;
 632     } else {
 633         *frame_type = FRAME_P;
 634         *displaying_order = encoding_order + ip_period - 1;
 635     }
 636 }
 637
 638
 639 static const char *rc_to_string(int rcmode)
 640 {
 641     switch (rc_mode) {
 642     case VA_RC_NONE:
 643         return "NONE";
 644     case VA_RC_CBR:
 645         return "CBR";
 646     case VA_RC_VBR:
 647         return "VBR";
 648     case VA_RC_VCM:
 649         return "VCM";
 650     case VA_RC_CQP:
 651         return "CQP";
 652     case VA_RC_VBR_CONSTRAINED:
 653         return "VBR_CONSTRAINED";
 654     default:
 655         return "Unknown";
 656     }
 657 }
 658
 659 #if 0
 660 static int process_cmdline(int argc, char *argv[])
 661 {
 662     char c;
 663     const struct option long_opts[] = {
 664         {"help", no_argument, NULL, 0 },
 665         {"bitrate", required_argument, NULL, 1 },
 666         {"minqp", required_argument, NULL, 2 },
 667         {"initialqp", required_argument, NULL, 3 },
 668         {"intra_period", required_argument, NULL, 4 },
 669         {"idr_period", required_argument, NULL, 5 },
 670         {"ip_period", required_argument, NULL, 6 },
 671         {"rcmode", required_argument, NULL, 7 },
 672         {"srcyuv", required_argument, NULL, 9 },
 673         {"recyuv", required_argument, NULL, 10 },
 674         {"fourcc", required_argument, NULL, 11 },
 675         {"syncmode", no_argument, NULL, 12 },
 676         {"enablePSNR", no_argument, NULL, 13 },
 677         {"prit", required_argument, NULL, 14 },
 678         {"priv", required_argument, NULL, 15 },
 679         {"framecount", required_argument, NULL, 16 },
 680         {"entropy", required_argument, NULL, 17 },
 681         {"profile", required_argument, NULL, 18 },
 682         {NULL, no_argument, NULL, 0 }};
 683     int long_index;
 684
 685     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 686         switch (c) {
 687         case 'w':
 688             frame_width = atoi(optarg);
 689             break;
 690         case 'h':
 691             frame_height = atoi(optarg);
 692             break;
 693         case 'n':
 694         case 'f':
 695             frame_rate = atoi(optarg);
 696             break;
 697         case 'o':
 698             coded_fn = strdup(optarg);
 699             break;
 700         case 0:
 701             print_help();
 702             exit(0);
 703         case 1:
 704             frame_bitrate = atoi(optarg);
 705             break;
 706         case 2:
 707             minimal_qp = atoi(optarg);
 708             break;
 709         case 3:
 710             initial_qp = atoi(optarg);
 711             break;
 712         case 4:
 713             intra_period = atoi(optarg);
 714             break;
 715         case 5:
 716             intra_idr_period = atoi(optarg);
 717             break;
 718         case 6:
 719             ip_period = atoi(optarg);
 720             break;
 721         case 7:
 722             rc_mode = string_to_rc(optarg);
 723             if (rc_mode < 0) {
 724                 print_help();
 725                 exit(1);
 726             }
 727             break;
 728         case 9:
 729             srcyuv_fn = strdup(optarg);
 730             break;
 731         case 11:
 732             srcyuv_fourcc = string_to_fourcc(optarg);
 733             if (srcyuv_fourcc <= 0) {
 734                 print_help();
 735                 exit(1);
 736             }
 737             break;
 738         case 13:
 739             calc_psnr = 1;
 740             break;
 741         case 14:
 742             misc_priv_type = strtol(optarg, NULL, 0);
 743             break;
 744         case 15:
 745             misc_priv_value = strtol(optarg, NULL, 0);
 746             break;
 747         case 17:
 748             h264_entropy_mode = atoi(optarg) ? 1: 0;
 749             break;
 750         case 18:
 751             if (strncmp(optarg, "BP", 2) == 0)
 752                 h264_profile = VAProfileH264Baseline;
 753             else if (strncmp(optarg, "MP", 2) == 0)
 754                 h264_profile = VAProfileH264Main;
 755             else if (strncmp(optarg, "HP", 2) == 0)
 756                 h264_profile = VAProfileH264High;
 757             else
 758                 h264_profile = (VAProfile)0;
 759             break;
 760         case ':':
 761         case '?':
 762             print_help();
 763             exit(0);
 764         }
 765     }
 766
 767     if (ip_period < 1) {
 768         printf(" ip_period must be greater than 0\n");
 769         exit(0);
 770     }
 771     if (intra_period != 1 && intra_period % ip_period != 0) {
 772         printf(" intra_period must be a multiplier of ip_period\n");
 773         exit(0);
 774     }
 775     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 776         printf(" intra_idr_period must be a multiplier of intra_period\n");
 777         exit(0);
 778     }
 779
 780     if (frame_bitrate == 0)
 781         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 782
 783     if (coded_fn == NULL) {
 784         struct stat buf;
 785         if (stat("/tmp", &buf) == 0)
 786             coded_fn = strdup("/tmp/test.264");
 787         else if (stat("/sdcard", &buf) == 0)
 788             coded_fn = strdup("/sdcard/test.264");
 789         else
 790             coded_fn = strdup("./test.264");
 791     }
 792
 793     /* store coded data into a file */
 794     coded_fp = fopen(coded_fn, "w+");
 795     if (coded_fp == NULL) {
 796         printf("Open file %s failed, exit\n", coded_fn);
 797         exit(1);
 798     }
 799
 800     frame_width_mbaligned = (frame_width + 15) & (~15);
 801     frame_height_mbaligned = (frame_height + 15) & (~15);
 802     if (frame_width != frame_width_mbaligned ||
 803         frame_height != frame_height_mbaligned) {
 804         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 805                frame_width, frame_height,
 806                frame_width_mbaligned, frame_height_mbaligned
 807                );
 808     }
 809
 810     return 0;
 811 }
 812 #endif
 813
 814 static Display *x11_display;
 815 static Window   x11_window;
 816
 817 VADisplay
 818 va_open_display(void)
 819 {
 820     x11_display = XOpenDisplay(NULL);
 821     if (!x11_display) {
 822         fprintf(stderr, "error: can't connect to X server!\n");
 823         return NULL;
 824     }
 825     return vaGetDisplay(x11_display);
 826 }
 827
 828 void
 829 va_close_display(VADisplay va_dpy)
 830 {
 831     if (!x11_display)
 832         return;
 833
 834     if (x11_window) {
 835         XUnmapWindow(x11_display, x11_window);
 836         XDestroyWindow(x11_display, x11_window);
 837         x11_window = None;
 838     }
 839     XCloseDisplay(x11_display);
 840     x11_display = NULL;
 841 }
 842
 843 static int init_va(void)
 844 {
 845     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 846     VAEntrypoint *entrypoints;
 847     int num_entrypoints, slice_entrypoint;
 848     int support_encode = 0;
 849     int major_ver, minor_ver;
 850     VAStatus va_status;
 851     unsigned int i;
 852
 853     va_dpy = va_open_display();
 854     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 855     CHECK_VASTATUS(va_status, "vaInitialize");
 856
 857     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 858     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 859     if (!entrypoints) {
 860         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 861         exit(1);
 862     }
 863
 864     /* use the highest profile */
 865     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 866         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 867             continue;
 868
 869         h264_profile = profile_list[i];
 870         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 871         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 872             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 873                 support_encode = 1;
 874                 break;
 875             }
 876         }
 877         if (support_encode == 1)
 878             break;
 879     }
 880
 881     if (support_encode == 0) {
 882         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 883         exit(1);
 884     } else {
 885         switch (h264_profile) {
 886             case VAProfileH264Baseline:
 887                 printf("Use profile VAProfileH264Baseline\n");
 888                 ip_period = 1;
 889                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 890                 h264_entropy_mode = 0;
 891                 break;
 892             case VAProfileH264ConstrainedBaseline:
 893                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 894                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 895                 ip_period = 1;
 896                 break;
 897
 898             case VAProfileH264Main:
 899                 printf("Use profile VAProfileH264Main\n");
 900                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 901                 break;
 902
 903             case VAProfileH264High:
 904                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 905                 printf("Use profile VAProfileH264High\n");
 906                 break;
 907             default:
 908                 printf("unknow profile. Set to Baseline");
 909                 h264_profile = VAProfileH264Baseline;
 910                 ip_period = 1;
 911                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 912                 break;
 913         }
 914     }
 915
 916     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 917
 918     /* find out the format for the render target, and rate control mode */
 919     for (i = 0; i < VAConfigAttribTypeMax; i++)
 920         attrib[i].type = (VAConfigAttribType)i;
 921
 922     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 923                                       &attrib[0], VAConfigAttribTypeMax);
 924     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 925     /* check the interested configattrib */
 926     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 927         printf("Not find desired YUV420 RT format\n");
 928         exit(1);
 929     } else {
 930         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 931         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 932         config_attrib_num++;
 933     }
 934
 935     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 936         int tmp = attrib[VAConfigAttribRateControl].value;
 937
 938         printf("Support rate control mode (0x%x):", tmp);
 939
 940         if (tmp & VA_RC_NONE)
 941             printf("NONE ");
 942         if (tmp & VA_RC_CBR)
 943             printf("CBR ");
 944         if (tmp & VA_RC_VBR)
 945             printf("VBR ");
 946         if (tmp & VA_RC_VCM)
 947             printf("VCM ");
 948         if (tmp & VA_RC_CQP)
 949             printf("CQP ");
 950         if (tmp & VA_RC_VBR_CONSTRAINED)
 951             printf("VBR_CONSTRAINED ");
 952
 953         printf("\n");
 954
 955         if (rc_mode == -1 || !(rc_mode & tmp))  {
 956             if (rc_mode != -1) {
 957                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
 958             }
 959
 960             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
 961                 if (rc_default_modes[i] & tmp) {
 962                     rc_mode = rc_default_modes[i];
 963                     break;
 964                 }
 965             }
 966
 967             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
 968         }
 969
 970         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 971         config_attrib[config_attrib_num].value = rc_mode;
 972         config_attrib_num++;
 973     }
 974
 975
 976     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 977         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 978
 979         printf("Support VAConfigAttribEncPackedHeaders\n");
 980
 981         h264_packedheader = 1;
 982         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 983         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 984
 985         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 986             printf("Support packed sequence headers\n");
 987             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 988         }
 989
 990         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 991             printf("Support packed picture headers\n");
 992             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 993         }
 994
 995         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 996             printf("Support packed slice headers\n");
 997             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 998         }
 999
1000         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1001             printf("Support packed misc headers\n");
1002             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1003         }
1004
1005         enc_packed_header_idx = config_attrib_num;
1006         config_attrib_num++;
1007     }
1008
1009     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1010         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1011
1012         printf("Support VAConfigAttribEncInterlaced\n");
1013
1014         if (tmp & VA_ENC_INTERLACED_FRAME)
1015             printf("support VA_ENC_INTERLACED_FRAME\n");
1016         if (tmp & VA_ENC_INTERLACED_FIELD)
1017             printf("Support VA_ENC_INTERLACED_FIELD\n");
1018         if (tmp & VA_ENC_INTERLACED_MBAFF)
1019             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1020         if (tmp & VA_ENC_INTERLACED_PAFF)
1021             printf("Support VA_ENC_INTERLACED_PAFF\n");
1022
1023         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1024         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1025         config_attrib_num++;
1026     }
1027
1028     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1029         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1030
1031         printf("Support %d RefPicList0 and %d RefPicList1\n",
1032                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1033     }
1034
1035     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1036         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1037
1038     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1039         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1040
1041         printf("Support VAConfigAttribEncSliceStructure\n");
1042
1043         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1044             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1045         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1046             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1047         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1048             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1049     }
1050     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1051         printf("Support VAConfigAttribEncMacroblockInfo\n");
1052     }
1053
1054     free(entrypoints);
1055     return 0;
1056 }
1057
1058 static int setup_encode()
1059 {
1060     VAStatus va_status;
1061     VASurfaceID *tmp_surfaceid;
1062     int codedbuf_size, i;
1063     static VASurfaceID src_surface[SURFACE_NUM];
1064     static VASurfaceID ref_surface[SURFACE_NUM];
1065
1066     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1067             &config_attrib[0], config_attrib_num, &config_id);
1068     CHECK_VASTATUS(va_status, "vaCreateConfig");
1069
1070     /* create source surfaces */
1071     va_status = vaCreateSurfaces(va_dpy,
1072                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1073                                  &src_surface[0], SURFACE_NUM,
1074                                  NULL, 0);
1075     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1076
1077     /* create reference surfaces */
1078     va_status = vaCreateSurfaces(va_dpy,
1079                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1080                                  &ref_surface[0], SURFACE_NUM,
1081                                  NULL, 0);
1082     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1083
1084     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1085     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1086     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1087
1088     /* Create a context for this encode pipe */
1089     va_status = vaCreateContext(va_dpy, config_id,
1090                                 frame_width_mbaligned, frame_height_mbaligned,
1091                                 VA_PROGRESSIVE,
1092                                 tmp_surfaceid, 2 * SURFACE_NUM,
1093                                 &context_id);
1094     CHECK_VASTATUS(va_status, "vaCreateContext");
1095     free(tmp_surfaceid);
1096
1097     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1098
1099     for (i = 0; i < SURFACE_NUM; i++) {
1100         /* create coded buffer once for all
1101          * other VA buffers which won't be used again after vaRenderPicture.
1102          * so APP can always vaCreateBuffer for every frame
1103          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1104          * so VA won't maintain the coded buffer
1105          */
1106         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1107                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1108         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1109     }
1110
1111     /* create OpenGL objects */
1112     //glGenFramebuffers(SURFACE_NUM, fbos);
1113
1114     for (i = 0; i < SURFACE_NUM; i++) {
1115         glGenTextures(1, &gl_surfaces[i].y_tex);
1116         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1117     }
1118
1119     for (i = 0; i < SURFACE_NUM; i++) {
1120         gl_surfaces[i].src_surface = src_surface[i];
1121         gl_surfaces[i].ref_surface = ref_surface[i];
1122     }
1123
1124     return 0;
1125 }
1126
1127
1128
1129 #define partition(ref, field, key, ascending)   \
1130     while (i <= j) {                            \
1131         if (ascending) {                        \
1132             while (ref[i].field < key)          \
1133                 i++;                            \
1134             while (ref[j].field > key)          \
1135                 j--;                            \
1136         } else {                                \
1137             while (ref[i].field > key)          \
1138                 i++;                            \
1139             while (ref[j].field < key)          \
1140                 j--;                            \
1141         }                                       \
1142         if (i <= j) {                           \
1143             tmp = ref[i];                       \
1144             ref[i] = ref[j];                    \
1145             ref[j] = tmp;                       \
1146             i++;                                \
1147             j--;                                \
1148         }                                       \
1149     }                                           \
1150
1151 static void sort_one(VAPictureH264 ref[], int left, int right,
1152                      int ascending, int frame_idx)
1153 {
1154     int i = left, j = right;
1155     unsigned int key;
1156     VAPictureH264 tmp;
1157
1158     if (frame_idx) {
1159         key = ref[(left + right) / 2].frame_idx;
1160         partition(ref, frame_idx, key, ascending);
1161     } else {
1162         key = ref[(left + right) / 2].TopFieldOrderCnt;
1163         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1164     }
1165
1166     /* recursion */
1167     if (left < j)
1168         sort_one(ref, left, j, ascending, frame_idx);
1169
1170     if (i < right)
1171         sort_one(ref, i, right, ascending, frame_idx);
1172 }
1173
1174 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1175                      int partition_ascending, int list0_ascending, int list1_ascending)
1176 {
1177     int i = left, j = right;
1178     VAPictureH264 tmp;
1179
1180     if (frame_idx) {
1181         partition(ref, frame_idx, key, partition_ascending);
1182     } else {
1183         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1184     }
1185
1186
1187     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1188     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1189 }
1190
1191 static int update_ReferenceFrames(void)
1192 {
1193     int i;
1194
1195     if (current_frame_type == FRAME_B)
1196         return 0;
1197
1198     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1199     numShortTerm++;
1200     if (numShortTerm > num_ref_frames)
1201         numShortTerm = num_ref_frames;
1202     for (i=numShortTerm-1; i>0; i--)
1203         ReferenceFrames[i] = ReferenceFrames[i-1];
1204     ReferenceFrames[0] = CurrentCurrPic;
1205
1206     if (current_frame_type != FRAME_B)
1207         current_frame_num++;
1208     if (current_frame_num > MaxFrameNum)
1209         current_frame_num = 0;
1210
1211     return 0;
1212 }
1213
1214
1215 static int update_RefPicList(void)
1216 {
1217     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1218
1219     if (current_frame_type == FRAME_P) {
1220         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1221         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1222     }
1223
1224     if (current_frame_type == FRAME_B) {
1225         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1226         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1227                  1, 0, 1);
1228
1229         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1230         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1231                  0, 1, 0);
1232     }
1233
1234     return 0;
1235 }
1236
1237
1238 static int render_sequence(void)
1239 {
1240     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1241     VAStatus va_status;
1242     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1243     VAEncMiscParameterRateControl *misc_rate_ctrl;
1244
1245     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1246     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1247     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1248     seq_param.bits_per_second = frame_bitrate;
1249
1250     seq_param.intra_period = intra_period;
1251     seq_param.intra_idr_period = intra_idr_period;
1252     seq_param.ip_period = ip_period;
1253
1254     seq_param.max_num_ref_frames = num_ref_frames;
1255     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1256     seq_param.time_scale = frame_rate * 2;
1257     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1258     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1259     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1260     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1261     seq_param.seq_fields.bits.chroma_format_idc = 1;
1262     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1263
1264     if (frame_width != frame_width_mbaligned ||
1265         frame_height != frame_height_mbaligned) {
1266         seq_param.frame_cropping_flag = 1;
1267         seq_param.frame_crop_left_offset = 0;
1268         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1269         seq_param.frame_crop_top_offset = 0;
1270         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1271     }
1272
1273     va_status = vaCreateBuffer(va_dpy, context_id,
1274                                VAEncSequenceParameterBufferType,
1275                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1276     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1277
1278     va_status = vaCreateBuffer(va_dpy, context_id,
1279                                VAEncMiscParameterBufferType,
1280                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1281                                1, NULL, &rc_param_buf);
1282     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1283
1284     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1285     misc_param->type = VAEncMiscParameterTypeRateControl;
1286     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1287     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1288     misc_rate_ctrl->bits_per_second = frame_bitrate;
1289     misc_rate_ctrl->target_percentage = 66;
1290     misc_rate_ctrl->window_size = 1000;
1291     misc_rate_ctrl->initial_qp = initial_qp;
1292     misc_rate_ctrl->min_qp = minimal_qp;
1293     misc_rate_ctrl->basic_unit_size = 0;
1294     vaUnmapBuffer(va_dpy, rc_param_buf);
1295
1296     render_id[0] = seq_param_buf;
1297     render_id[1] = rc_param_buf;
1298
1299     va_status = vaRenderPicture(va_dpy, context_id, &render_id[0], 2);
1300     CHECK_VASTATUS(va_status, "vaRenderPicture");;
1301
1302     if (misc_priv_type != 0) {
1303         va_status = vaCreateBuffer(va_dpy, context_id,
1304                                    VAEncMiscParameterBufferType,
1305                                    sizeof(VAEncMiscParameterBuffer),
1306                                    1, NULL, &misc_param_tmpbuf);
1307         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1308         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1309         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1310         misc_param_tmp->data[0] = misc_priv_value;
1311         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1312
1313         va_status = vaRenderPicture(va_dpy, context_id, &misc_param_tmpbuf, 1);
1314     }
1315
1316     return 0;
1317 }
1318
1319 static int calc_poc(int pic_order_cnt_lsb)
1320 {
1321     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1322     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1323     int PicOrderCntMsb, TopFieldOrderCnt;
1324
1325     if (current_frame_type == FRAME_IDR)
1326         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1327     else {
1328         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1329         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1330     }
1331
1332     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1333         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1334         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1335     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1336              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1337         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1338     else
1339         PicOrderCntMsb = prevPicOrderCntMsb;
1340
1341     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1342
1343     if (current_frame_type != FRAME_B) {
1344         PicOrderCntMsb_ref = PicOrderCntMsb;
1345         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1346     }
1347
1348     return TopFieldOrderCnt;
1349 }
1350
1351 static int render_picture(void)
1352 {
1353     VABufferID pic_param_buf;
1354     VAStatus va_status;
1355     int i = 0;
1356
1357     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1358     pic_param.CurrPic.frame_idx = current_frame_num;
1359     pic_param.CurrPic.flags = 0;
1360     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1361     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1362     CurrentCurrPic = pic_param.CurrPic;
1363
1364     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1365         update_RefPicList(); /* calc RefPicList */
1366         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1367         if (current_frame_type == FRAME_P) {
1368             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1369         } else if (current_frame_type == FRAME_B) {
1370             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1371             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1372         }
1373     } else {
1374         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1375         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1376             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1377             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1378         }
1379     }
1380
1381     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1382     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1383     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1384     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1385     pic_param.frame_num = current_frame_num;
1386     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1387     pic_param.last_picture = false;  // FIXME
1388     pic_param.pic_init_qp = initial_qp;
1389
1390     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1391                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1392     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1393
1394     va_status = vaRenderPicture(va_dpy, context_id, &pic_param_buf, 1);
1395     CHECK_VASTATUS(va_status, "vaRenderPicture");
1396
1397     return 0;
1398 }
1399
1400 static int render_packedsequence(void)
1401 {
1402     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1403     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1404     unsigned int length_in_bits;
1405     unsigned char *packedseq_buffer = NULL;
1406     VAStatus va_status;
1407
1408     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1409
1410     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1411
1412     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1413     packedheader_param_buffer.has_emulation_bytes = 0;
1414     va_status = vaCreateBuffer(va_dpy,
1415                                context_id,
1416                                VAEncPackedHeaderParameterBufferType,
1417                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1418                                &packedseq_para_bufid);
1419     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1420
1421     va_status = vaCreateBuffer(va_dpy,
1422                                context_id,
1423                                VAEncPackedHeaderDataBufferType,
1424                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1425                                &packedseq_data_bufid);
1426     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1427
1428     render_id[0] = packedseq_para_bufid;
1429     render_id[1] = packedseq_data_bufid;
1430     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1431     CHECK_VASTATUS(va_status, "vaRenderPicture");
1432
1433     free(packedseq_buffer);
1434
1435     return 0;
1436 }
1437
1438
1439 static int render_packedpicture(void)
1440 {
1441     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1442     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1443     unsigned int length_in_bits;
1444     unsigned char *packedpic_buffer = NULL;
1445     VAStatus va_status;
1446
1447     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1448     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1449     packedheader_param_buffer.bit_length = length_in_bits;
1450     packedheader_param_buffer.has_emulation_bytes = 0;
1451
1452     va_status = vaCreateBuffer(va_dpy,
1453                                context_id,
1454                                VAEncPackedHeaderParameterBufferType,
1455                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1456                                &packedpic_para_bufid);
1457     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1458
1459     va_status = vaCreateBuffer(va_dpy,
1460                                context_id,
1461                                VAEncPackedHeaderDataBufferType,
1462                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1463                                &packedpic_data_bufid);
1464     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1465
1466     render_id[0] = packedpic_para_bufid;
1467     render_id[1] = packedpic_data_bufid;
1468     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1469     CHECK_VASTATUS(va_status, "vaRenderPicture");
1470
1471     free(packedpic_buffer);
1472
1473     return 0;
1474 }
1475
1476 static void render_packedslice()
1477 {
1478     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1479     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1480     unsigned int length_in_bits;
1481     unsigned char *packedslice_buffer = NULL;
1482     VAStatus va_status;
1483
1484     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1485     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1486     packedheader_param_buffer.bit_length = length_in_bits;
1487     packedheader_param_buffer.has_emulation_bytes = 0;
1488
1489     va_status = vaCreateBuffer(va_dpy,
1490                                context_id,
1491                                VAEncPackedHeaderParameterBufferType,
1492                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1493                                &packedslice_para_bufid);
1494     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1495
1496     va_status = vaCreateBuffer(va_dpy,
1497                                context_id,
1498                                VAEncPackedHeaderDataBufferType,
1499                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1500                                &packedslice_data_bufid);
1501     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1502
1503     render_id[0] = packedslice_para_bufid;
1504     render_id[1] = packedslice_data_bufid;
1505     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1506     CHECK_VASTATUS(va_status, "vaRenderPicture");
1507
1508     free(packedslice_buffer);
1509 }
1510
1511 static int render_slice(void)
1512 {
1513     VABufferID slice_param_buf;
1514     VAStatus va_status;
1515     int i;
1516
1517     update_RefPicList();
1518
1519     /* one frame, one slice */
1520     slice_param.macroblock_address = 0;
1521     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1522     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1523     if (current_frame_type == FRAME_IDR) {
1524         if (current_frame_encoding != 0)
1525             ++slice_param.idr_pic_id;
1526     } else if (current_frame_type == FRAME_P) {
1527         int refpiclist0_max = h264_maxref & 0xffff;
1528         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1529
1530         for (i = refpiclist0_max; i < 32; i++) {
1531             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1532             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1533         }
1534     } else if (current_frame_type == FRAME_B) {
1535         int refpiclist0_max = h264_maxref & 0xffff;
1536         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1537
1538         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1539         for (i = refpiclist0_max; i < 32; i++) {
1540             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1541             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1542         }
1543
1544         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1545         for (i = refpiclist1_max; i < 32; i++) {
1546             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1547             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1548         }
1549     }
1550
1551     slice_param.slice_alpha_c0_offset_div2 = 0;
1552     slice_param.slice_beta_offset_div2 = 0;
1553     slice_param.direct_spatial_mv_pred_flag = 1;
1554     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1555
1556
1557     if (h264_packedheader &&
1558         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1559         render_packedslice();
1560
1561     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1562                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1563     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1564
1565     va_status = vaRenderPicture(va_dpy, context_id, &slice_param_buf, 1);
1566     CHECK_VASTATUS(va_status, "vaRenderPicture");
1567
1568     return 0;
1569 }
1570
1571
1572
1573 int H264Encoder::save_codeddata(unsigned long long display_order, unsigned long long encode_order, int frame_type)
1574 {
1575     VACodedBufferSegment *buf_list = NULL;
1576     VAStatus va_status;
1577     unsigned int coded_size = 0;
1578
1579     string data;
1580
1581     va_status = vaMapBuffer(va_dpy, gl_surfaces[display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1582     CHECK_VASTATUS(va_status, "vaMapBuffer");
1583     while (buf_list != NULL) {
1584         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1585         if (coded_fp != nullptr)
1586             coded_size += fwrite(buf_list->buf, 1, buf_list->size, coded_fp);
1587         buf_list = (VACodedBufferSegment *) buf_list->next;
1588
1589         frame_size += coded_size;
1590     }
1591     vaUnmapBuffer(va_dpy, gl_surfaces[display_order % SURFACE_NUM].coded_buf);
1592
1593     AVPacket pkt;
1594     memset(&pkt, 0, sizeof(pkt));
1595     pkt.buf = nullptr;
1596     pkt.pts = av_rescale_q(display_order, AVRational{1, frame_rate}, avstream->time_base);
1597     pkt.dts = av_rescale_q(encode_order, AVRational{1, frame_rate}, avstream->time_base);
1598     pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1599     pkt.size = data.size();
1600     pkt.stream_index = 0;
1601     if (frame_type == FRAME_IDR || frame_type == FRAME_I) {
1602         pkt.flags = AV_PKT_FLAG_KEY;
1603     } else {
1604         pkt.flags = 0;
1605     }
1606     pkt.duration = 1;
1607     av_interleaved_write_frame(avctx, &pkt);
1608
1609 #if 0
1610     printf("\r      "); /* return back to startpoint */
1611     switch (encode_order % 4) {
1612         case 0:
1613             printf("|");
1614             break;
1615         case 1:
1616             printf("/");
1617             break;
1618         case 2:
1619             printf("-");
1620             break;
1621         case 3:
1622             printf("\\");
1623             break;
1624     }
1625     printf("%08lld", encode_order);
1626     printf("(%06d bytes coded)", coded_size);
1627 #endif
1628
1629     return 0;
1630 }
1631
1632
1633 // this is weird. but it seems to put a new frame onto the queue
1634 void H264Encoder::storage_task_enqueue(unsigned long long display_order, unsigned long long encode_order, int frame_type)
1635 {
1636         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1637
1638         storage_task tmp;
1639         tmp.display_order = display_order;
1640         tmp.encode_order = encode_order;
1641         tmp.frame_type = frame_type;
1642         storage_task_queue.push(tmp);
1643         srcsurface_status[display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1644
1645         storage_task_queue_changed.notify_all();
1646 }
1647
1648 void H264Encoder::storage_task_thread()
1649 {
1650         for ( ;; ) {
1651                 storage_task current;
1652                 {
1653                         // wait until there's an encoded frame
1654                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1655                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1656                         if (storage_thread_should_quit) return;
1657                         current = storage_task_queue.front();
1658                         storage_task_queue.pop();
1659                 }
1660
1661                 VAStatus va_status;
1662
1663                 // waits for data, then saves it to disk.
1664                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1665                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1666                 save_codeddata(current.display_order, current.encode_order, current.frame_type);
1667
1668                 {
1669                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1670                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1671                         storage_task_queue_changed.notify_all();
1672                 }
1673         }
1674 }
1675
1676 static int release_encode()
1677 {
1678     int i;
1679
1680     for (i = 0; i < SURFACE_NUM; i++) {
1681         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1682         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1683         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1684     }
1685
1686     vaDestroyContext(va_dpy, context_id);
1687     vaDestroyConfig(va_dpy, config_id);
1688
1689     return 0;
1690 }
1691
1692 static int deinit_va()
1693 {
1694     vaTerminate(va_dpy);
1695
1696     va_close_display(va_dpy);
1697
1698     return 0;
1699 }
1700
1701
1702 static int print_input()
1703 {
1704     printf("\n\nINPUT:Try to encode H264...\n");
1705     if (rc_mode != -1)
1706         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1707     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1708     printf("INPUT: FrameRate    : %d\n", frame_rate);
1709     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1710     printf("INPUT: Slieces      : %d\n", frame_slices);
1711     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1712     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1713     printf("INPUT: IpPeriod     : %d\n", ip_period);
1714     printf("INPUT: Initial QP   : %d\n", initial_qp);
1715     printf("INPUT: Min QP       : %d\n", minimal_qp);
1716     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1717
1718     printf("\n\n"); /* return back to startpoint */
1719
1720     return 0;
1721 }
1722
1723
1724 //H264Encoder::H264Encoder(SDL_Window *window, SDL_GLContext context, int width, int height, const char *output_filename)
1725 H264Encoder::H264Encoder(QSurface *surface, int width, int height, const char *output_filename)
1726         : current_storage_frame(0), surface(surface)
1727         //: width(width), height(height), current_encoding_frame(0)
1728 {
1729         av_register_all();
1730         avctx = avformat_alloc_context();
1731         avctx->oformat = av_guess_format(NULL, output_filename, NULL);
1732         strcpy(avctx->filename, output_filename);
1733         if (avio_open2(&avctx->pb, output_filename, AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL) < 0) {
1734                 fprintf(stderr, "%s: avio_open2() failed\n", output_filename);
1735                 exit(1);
1736         }
1737         AVCodec *codec = avcodec_find_encoder(AV_CODEC_ID_H264);
1738         avstream = avformat_new_stream(avctx, codec);
1739         if (avstream == nullptr) {
1740                 fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
1741                 exit(1);
1742         }
1743         avstream->time_base = AVRational{1, frame_rate};
1744         avstream->codec->width = width;
1745         avstream->codec->height = height;
1746         avstream->codec->time_base = AVRational{1, frame_rate};
1747         avstream->codec->ticks_per_frame = 1;  // or 2?
1748
1749         if (avformat_write_header(avctx, NULL) < 0) {
1750                 fprintf(stderr, "%s: avformat_write_header() failed\n", output_filename);
1751                 exit(1);
1752         }
1753
1754         coded_fp = fopen("dump.h264", "wb");
1755         assert(coded_fp != NULL);
1756
1757         frame_width = width;
1758         frame_height = height;
1759         frame_width_mbaligned = (frame_width + 15) & (~15);
1760         frame_height_mbaligned = (frame_height + 15) & (~15);
1761         frame_bitrate = 15000000;  // / 60;
1762         current_frame_encoding = 0;
1763
1764         print_input();
1765
1766         init_va();
1767         setup_encode();
1768
1769         // No frames are ready yet.
1770         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1771
1772         memset(&seq_param, 0, sizeof(seq_param));
1773         memset(&pic_param, 0, sizeof(pic_param));
1774         memset(&slice_param, 0, sizeof(slice_param));
1775
1776         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1777
1778         copy_thread = std::thread([this]{
1779                 //SDL_GL_MakeCurrent(window, context);
1780                 QOpenGLContext *context = create_context();
1781                 eglBindAPI(EGL_OPENGL_API);
1782                 if (!make_current(context, this->surface)) {
1783                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1784                                 eglGetError());
1785                         exit(1);
1786                 }
1787                 copy_thread_func();
1788         });
1789 }
1790
1791 H264Encoder::~H264Encoder()
1792 {
1793         {
1794                 unique_lock<mutex> lock(storage_task_queue_mutex);
1795                 storage_thread_should_quit = true;
1796                 storage_task_queue_changed.notify_all();
1797         }
1798         {
1799                 unique_lock<mutex> lock(frame_queue_mutex);
1800                 copy_thread_should_quit = true;
1801                 frame_queue_nonempty.notify_one();
1802         }
1803         storage_thread.join();
1804         copy_thread.join();
1805
1806         release_encode();
1807         deinit_va();
1808
1809         av_write_trailer(avctx);
1810         avformat_free_context(avctx);
1811 }
1812
1813 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1814 {
1815         {
1816                 // Wait until this frame slot is done encoding.
1817                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1818                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1819                 if (storage_thread_should_quit) return false;
1820         }
1821
1822         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1823         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1824         *y_tex = surf->y_tex;
1825         *cbcr_tex = surf->cbcr_tex;
1826
1827         VASurfaceID surface = surf->src_surface;
1828         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1829         CHECK_VASTATUS(va_status, "vaDeriveImage");
1830
1831         VABufferInfo buf_info;
1832         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1833         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1834         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1835
1836         // Create Y image.
1837         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1838         EGLint y_attribs[] = {
1839                 EGL_WIDTH, frame_width,
1840                 EGL_HEIGHT, frame_height,
1841                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1842                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1843                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1844                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1845                 EGL_NONE
1846         };
1847
1848         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1849         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1850
1851         // Associate Y image to a texture.
1852         glBindTexture(GL_TEXTURE_2D, *y_tex);
1853         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1854
1855         // Create CbCr image.
1856         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1857         EGLint cbcr_attribs[] = {
1858                 EGL_WIDTH, frame_width,
1859                 EGL_HEIGHT, frame_height,
1860                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1861                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1862                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1863                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1864                 EGL_NONE
1865         };
1866
1867         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1868         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1869
1870         // Associate CbCr image to a texture.
1871         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1872         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1873
1874         return true;
1875 }
1876
1877 void H264Encoder::end_frame(GLsync fence, const std::vector<FrameAllocator::Frame> &input_frames_to_release)
1878 {
1879         {
1880                 unique_lock<mutex> lock(frame_queue_mutex);
1881                 pending_frames[current_storage_frame++] = PendingFrame{ fence, input_frames_to_release };
1882         }
1883         frame_queue_nonempty.notify_one();
1884 }
1885
1886 void H264Encoder::copy_thread_func()
1887 {
1888         for ( ;; ) {
1889                 PendingFrame frame;
1890                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1891                                        &current_frame_display, &current_frame_type);
1892                 if (current_frame_type == FRAME_IDR) {
1893                         numShortTerm = 0;
1894                         current_frame_num = 0;
1895                         current_IDR_display = current_frame_display;
1896                 }
1897
1898                 {
1899                         unique_lock<mutex> lock(frame_queue_mutex);
1900                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_frames.count(current_frame_display) != 0; });
1901                         if (copy_thread_should_quit) return;
1902                         frame = pending_frames[current_frame_display];
1903                         pending_frames.erase(current_frame_display);
1904                 }
1905
1906                 // Wait for the GPU to be done with the frame.
1907                 glClientWaitSync(frame.fence, 0, 0);
1908                 glDeleteSync(frame.fence);
1909
1910                 // Release back any input frames we needed to render this frame.
1911                 // (Actually, those that were needed one output frame ago.)
1912                 for (FrameAllocator::Frame input_frame : frame.input_frames_to_release) {
1913                         input_frame.owner->release_frame(input_frame);
1914                 }
1915
1916                 // Unmap the image.
1917                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
1918                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1919                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1920                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1921                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1922                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1923                 CHECK_VASTATUS(va_status, "vaDestroyImage");
1924
1925                 VASurfaceID surface = surf->src_surface;
1926
1927                 // Schedule the frame for encoding.
1928                 va_status = vaBeginPicture(va_dpy, context_id, surface);
1929                 CHECK_VASTATUS(va_status, "vaBeginPicture");
1930
1931                 if (current_frame_type == FRAME_IDR) {
1932                         render_sequence();
1933                         render_picture();
1934                         if (h264_packedheader) {
1935                                 render_packedsequence();
1936                                 render_packedpicture();
1937                         }
1938                 } else {
1939                         //render_sequence();
1940                         render_picture();
1941                 }
1942                 render_slice();
1943
1944                 va_status = vaEndPicture(va_dpy, context_id);
1945                 CHECK_VASTATUS(va_status, "vaEndPicture");
1946
1947                 // so now the data is done encoding (well, async job kicked off)...
1948                 // we send that to the storage thread
1949                 storage_task_enqueue(current_frame_display, current_frame_encoding, current_frame_type);
1950
1951                 update_ReferenceFrames();
1952                 ++current_frame_encoding;
1953         }
1954 }