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

   1 //#include "sysdeps.h"
   2 #include "h264encode.h"
   3
   4 #include <EGL/eglplatform.h>
   5 #include <X11/X.h>
   6 #include <X11/Xlib.h>
   7 #include <assert.h>
   8 #include <epoxy/egl.h>
   9 #include <libavcodec/avcodec.h>
  10 #include <libavformat/avio.h>
  11 #include <libavutil/mathematics.h>
  12 #include <libavutil/rational.h>
  13 #include <libdrm/drm_fourcc.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <string.h>
  18 #include <va/va.h>
  19 #include <va/va_drmcommon.h>
  20 #include <va/va_enc_h264.h>
  21 #include <va/va_x11.h>
  22 #include <condition_variable>
  23 #include <mutex>
  24 #include <queue>
  25 #include <string>
  26 #include <thread>
  27
  28 #include "context.h"
  29 #include "httpd.h"
  30 #include "timebase.h"
  31
  32 class QOpenGLContext;
  33 class QSurface;
  34
  35 #define CHECK_VASTATUS(va_status, func)                                 \
  36     if (va_status != VA_STATUS_SUCCESS) {                               \
  37         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  38         exit(1);                                                        \
  39     }
  40
  41 //#include "loadsurface.h"
  42
  43 #define NAL_REF_IDC_NONE        0
  44 #define NAL_REF_IDC_LOW         1
  45 #define NAL_REF_IDC_MEDIUM      2
  46 #define NAL_REF_IDC_HIGH        3
  47
  48 #define NAL_NON_IDR             1
  49 #define NAL_IDR                 5
  50 #define NAL_SPS                 7
  51 #define NAL_PPS                 8
  52 #define NAL_SEI                 6
  53
  54 #define SLICE_TYPE_P            0
  55 #define SLICE_TYPE_B            1
  56 #define SLICE_TYPE_I            2
  57 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  58 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  59 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  60
  61
  62 #define ENTROPY_MODE_CAVLC      0
  63 #define ENTROPY_MODE_CABAC      1
  64
  65 #define PROFILE_IDC_BASELINE    66
  66 #define PROFILE_IDC_MAIN        77
  67 #define PROFILE_IDC_HIGH        100
  68
  69 #define BITSTREAM_ALLOCATE_STEPPING     4096
  70
  71 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  72 static  VADisplay va_dpy;
  73 static  VAProfile h264_profile = (VAProfile)~0;
  74 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  75 static  int config_attrib_num = 0, enc_packed_header_idx;
  76
  77 struct GLSurface {
  78         VASurfaceID src_surface, ref_surface;
  79         VABufferID coded_buf;
  80
  81         VAImage surface_image;
  82         GLuint y_tex, cbcr_tex;
  83         EGLImage y_egl_image, cbcr_egl_image;
  84 };
  85 GLSurface gl_surfaces[SURFACE_NUM];
  86
  87 static  VAConfigID config_id;
  88 static  VAContextID context_id;
  89 static  VAEncSequenceParameterBufferH264 seq_param;
  90 static  VAEncPictureParameterBufferH264 pic_param;
  91 static  VAEncSliceParameterBufferH264 slice_param;
  92 static  VAPictureH264 CurrentCurrPic;
  93 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  94
  95 static  unsigned int MaxFrameNum = (2<<16);
  96 static  unsigned int MaxPicOrderCntLsb = (2<<8);
  97 static  unsigned int Log2MaxFrameNum = 16;
  98 static  unsigned int Log2MaxPicOrderCntLsb = 8;
  99
 100 static  unsigned int num_ref_frames = 2;
 101 static  unsigned int numShortTerm = 0;
 102 static  int constraint_set_flag = 0;
 103 static  int h264_packedheader = 0; /* support pack header? */
 104 static  int h264_maxref = (1<<16|1);
 105 static  int h264_entropy_mode = 1; /* cabac */
 106
 107 static  char *coded_fn = NULL;
 108
 109 static  int frame_width = 176;
 110 static  int frame_height = 144;
 111 static  int frame_width_mbaligned;
 112 static  int frame_height_mbaligned;
 113 static  int frame_rate = 60;
 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 = 15;
 118 //static  int initial_qp = 28;
 119 static  int minimal_qp = 0;
 120 static  int intra_period = 30;
 121 static  int intra_idr_period = 60;
 122 static  int ip_period = 3;
 123 static  int rc_mode = -1;
 124 static  int rc_default_modes[] = {
 125     VA_RC_VBR,
 126     VA_RC_CQP,
 127     VA_RC_VBR_CONSTRAINED,
 128     VA_RC_CBR,
 129     VA_RC_VCM,
 130     VA_RC_NONE,
 131 };
 132 static  unsigned long long current_frame_encoding = 0;
 133 static  unsigned long long current_frame_display = 0;
 134 static  unsigned long long current_IDR_display = 0;
 135 static  unsigned int current_frame_num = 0;
 136 static  int current_frame_type;
 137
 138 static  int misc_priv_type = 0;
 139 static  int misc_priv_value = 0;
 140
 141 /* thread to save coded data */
 142 #define SRC_SURFACE_FREE        0
 143 #define SRC_SURFACE_IN_ENCODING 1
 144
 145 struct __bitstream {
 146     unsigned int *buffer;
 147     int bit_offset;
 148     int max_size_in_dword;
 149 };
 150 typedef struct __bitstream bitstream;
 151
 152 using namespace std;
 153
 154 static unsigned int
 155 va_swap32(unsigned int val)
 156 {
 157     unsigned char *pval = (unsigned char *)&val;
 158
 159     return ((pval[0] << 24)     |
 160             (pval[1] << 16)     |
 161             (pval[2] << 8)      |
 162             (pval[3] << 0));
 163 }
 164
 165 static void
 166 bitstream_start(bitstream *bs)
 167 {
 168     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 169     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 170     bs->bit_offset = 0;
 171 }
 172
 173 static void
 174 bitstream_end(bitstream *bs)
 175 {
 176     int pos = (bs->bit_offset >> 5);
 177     int bit_offset = (bs->bit_offset & 0x1f);
 178     int bit_left = 32 - bit_offset;
 179
 180     if (bit_offset) {
 181         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 182     }
 183 }
 184
 185 static void
 186 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 187 {
 188     int pos = (bs->bit_offset >> 5);
 189     int bit_offset = (bs->bit_offset & 0x1f);
 190     int bit_left = 32 - bit_offset;
 191
 192     if (!size_in_bits)
 193         return;
 194
 195     bs->bit_offset += size_in_bits;
 196
 197     if (bit_left > size_in_bits) {
 198         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 199     } else {
 200         size_in_bits -= bit_left;
 201         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 202         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 203
 204         if (pos + 1 == bs->max_size_in_dword) {
 205             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 206             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 207         }
 208
 209         bs->buffer[pos + 1] = val;
 210     }
 211 }
 212
 213 static void
 214 bitstream_put_ue(bitstream *bs, unsigned int val)
 215 {
 216     int size_in_bits = 0;
 217     int tmp_val = ++val;
 218
 219     while (tmp_val) {
 220         tmp_val >>= 1;
 221         size_in_bits++;
 222     }
 223
 224     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 225     bitstream_put_ui(bs, val, size_in_bits);
 226 }
 227
 228 static void
 229 bitstream_put_se(bitstream *bs, int val)
 230 {
 231     unsigned int new_val;
 232
 233     if (val <= 0)
 234         new_val = -2 * val;
 235     else
 236         new_val = 2 * val - 1;
 237
 238     bitstream_put_ue(bs, new_val);
 239 }
 240
 241 static void
 242 bitstream_byte_aligning(bitstream *bs, int bit)
 243 {
 244     int bit_offset = (bs->bit_offset & 0x7);
 245     int bit_left = 8 - bit_offset;
 246     int new_val;
 247
 248     if (!bit_offset)
 249         return;
 250
 251     assert(bit == 0 || bit == 1);
 252
 253     if (bit)
 254         new_val = (1 << bit_left) - 1;
 255     else
 256         new_val = 0;
 257
 258     bitstream_put_ui(bs, new_val, bit_left);
 259 }
 260
 261 static void
 262 rbsp_trailing_bits(bitstream *bs)
 263 {
 264     bitstream_put_ui(bs, 1, 1);
 265     bitstream_byte_aligning(bs, 0);
 266 }
 267
 268 static void nal_start_code_prefix(bitstream *bs)
 269 {
 270     bitstream_put_ui(bs, 0x00000001, 32);
 271 }
 272
 273 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 274 {
 275     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 276     bitstream_put_ui(bs, nal_ref_idc, 2);
 277     bitstream_put_ui(bs, nal_unit_type, 5);
 278 }
 279
 280 static void sps_rbsp(bitstream *bs)
 281 {
 282     int profile_idc = PROFILE_IDC_BASELINE;
 283
 284     if (h264_profile  == VAProfileH264High)
 285         profile_idc = PROFILE_IDC_HIGH;
 286     else if (h264_profile  == VAProfileH264Main)
 287         profile_idc = PROFILE_IDC_MAIN;
 288
 289     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 290     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 291     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 292     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 293     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 294     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 295     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 296     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 297
 298     if ( profile_idc == PROFILE_IDC_HIGH) {
 299         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 300         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 301         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 302         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 303         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 304     }
 305
 306     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 307     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 308
 309     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 310         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 311     else {
 312         assert(0);
 313     }
 314
 315     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 316     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 317
 318     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 319     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 320     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 321
 322     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 323         assert(0);
 324     }
 325
 326     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 327     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 328
 329     if (seq_param.frame_cropping_flag) {
 330         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 331         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 332         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 333         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 334     }
 335
 336     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 337     if ( false ) {
 338         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 339     } else {
 340         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 341         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 342         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 343         bitstream_put_ui(bs, 0, 1); /* video_signal_type_present_flag */
 344         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 345         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 346         {
 347             bitstream_put_ui(bs, 1, 32);  // FPS
 348             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 349             bitstream_put_ui(bs, 1, 1);
 350         }
 351         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 352         {
 353             // hrd_parameters
 354             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 355             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 356             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 357
 358             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 359             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 360             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 361
 362             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 363             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 364             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 365             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 366         }
 367         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 368         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 369
 370         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 371         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 372     }
 373
 374     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 375 }
 376
 377
 378 static void pps_rbsp(bitstream *bs)
 379 {
 380     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 381     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 382
 383     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 384
 385     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 386
 387     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 388
 389     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 390     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 391
 392     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 393     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 394
 395     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 396     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 397     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 398
 399     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 400     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 401     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 402
 403     /* more_rbsp_data */
 404     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 405     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 406     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 407
 408     rbsp_trailing_bits(bs);
 409 }
 410
 411 static void slice_header(bitstream *bs)
 412 {
 413     int first_mb_in_slice = slice_param.macroblock_address;
 414
 415     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 416     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 417     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 418     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 419
 420     /* frame_mbs_only_flag == 1 */
 421     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 422         /* FIXME: */
 423         assert(0);
 424     }
 425
 426     if (pic_param.pic_fields.bits.idr_pic_flag)
 427         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 428
 429     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 430         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 431         /* pic_order_present_flag == 0 */
 432     } else {
 433         /* FIXME: */
 434         assert(0);
 435     }
 436
 437     /* redundant_pic_cnt_present_flag == 0 */
 438     /* slice type */
 439     if (IS_P_SLICE(slice_param.slice_type)) {
 440         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 441
 442         if (slice_param.num_ref_idx_active_override_flag)
 443             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 444
 445         /* ref_pic_list_reordering */
 446         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 447     } else if (IS_B_SLICE(slice_param.slice_type)) {
 448         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 449
 450         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 451
 452         if (slice_param.num_ref_idx_active_override_flag) {
 453             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 454             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 455         }
 456
 457         /* ref_pic_list_reordering */
 458         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 459         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 460     }
 461
 462     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 463          IS_P_SLICE(slice_param.slice_type)) ||
 464         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 465          IS_B_SLICE(slice_param.slice_type))) {
 466         /* FIXME: fill weight/offset table */
 467         assert(0);
 468     }
 469
 470     /* dec_ref_pic_marking */
 471     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 472         unsigned char no_output_of_prior_pics_flag = 0;
 473         unsigned char long_term_reference_flag = 0;
 474         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 475
 476         if (pic_param.pic_fields.bits.idr_pic_flag) {
 477             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 478             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 479         } else {
 480             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 481         }
 482     }
 483
 484     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 485         !IS_I_SLICE(slice_param.slice_type))
 486         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 487
 488     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 489
 490     /* ignore for SP/SI */
 491
 492     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 493         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 494
 495         if (slice_param.disable_deblocking_filter_idc != 1) {
 496             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 497             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 498         }
 499     }
 500
 501     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 502         bitstream_byte_aligning(bs, 1);
 503     }
 504 }
 505
 506 static int
 507 build_packed_pic_buffer(unsigned char **header_buffer)
 508 {
 509     bitstream bs;
 510
 511     bitstream_start(&bs);
 512     nal_start_code_prefix(&bs);
 513     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 514     pps_rbsp(&bs);
 515     bitstream_end(&bs);
 516
 517     *header_buffer = (unsigned char *)bs.buffer;
 518     return bs.bit_offset;
 519 }
 520
 521 static int
 522 build_packed_seq_buffer(unsigned char **header_buffer)
 523 {
 524     bitstream bs;
 525
 526     bitstream_start(&bs);
 527     nal_start_code_prefix(&bs);
 528     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 529     sps_rbsp(&bs);
 530     bitstream_end(&bs);
 531
 532     *header_buffer = (unsigned char *)bs.buffer;
 533     return bs.bit_offset;
 534 }
 535
 536 static int build_packed_slice_buffer(unsigned char **header_buffer)
 537 {
 538     bitstream bs;
 539     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 540     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 541
 542     bitstream_start(&bs);
 543     nal_start_code_prefix(&bs);
 544
 545     if (IS_I_SLICE(slice_param.slice_type)) {
 546         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 547     } else if (IS_P_SLICE(slice_param.slice_type)) {
 548         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 549     } else {
 550         assert(IS_B_SLICE(slice_param.slice_type));
 551         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 552     }
 553
 554     slice_header(&bs);
 555     bitstream_end(&bs);
 556
 557     *header_buffer = (unsigned char *)bs.buffer;
 558     return bs.bit_offset;
 559 }
 560
 561
 562 /*
 563   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 564   1) period between Frame #X and Frame #N = #X - #N
 565   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 566   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 567   4) intra_period and intra_idr_period take precedence over ip_period
 568   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 569      of I/IDR frames, see bellow examples
 570   -------------------------------------------------------------------
 571   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 572   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 573   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 574   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 575   1            1                ignored    IDR IDR IDR IDR...
 576   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 577   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 578   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 579                                               (PBB)(IBB)(PBB)(IBB)...
 580   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 581                                            IDRPPPPP IPPPPP IPPPPP...
 582   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 583                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 584                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 585                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 586                                            {IDR(PBB)(PBB)}                     (6/6/3)
 587                                            {IDR(PBB)(PBB)}.
 588 */
 589
 590 // General pts/dts strategy:
 591 //
 592 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 593 // bit tricky. We assume first of all that the frame rate never goes _above_
 594 // <frame_rate>, which gives us a frame period N. The decoder can always decode
 595 // in at least this speed, as long at dts <= pts (the frame is not attempted
 596 // presented before it is decoded). Furthermore, we never have longer chains of
 597 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 598 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 599 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 600 // frame, with an order O <= C.)
 601 //
 602 // Many strategies are possible, but we establish these rules:
 603 //
 604 //  - Tip frames have dts = pts - (C-O)*N.
 605 //  - Non-tip frames have dts = dts_last + N.
 606 //
 607 // An example, with C=2 and N=10 and the data flow showed with arrows:
 608 //
 609 //        I  B  P  B  B  P
 610 //   pts: 30 40 50 60 70 80
 611 //        ↓  ↓     ↓
 612 //   dts: 10 30 20 60 50←40
 613 //         |  |  ↑        ↑
 614 //         `--|--'        |
 615 //             `----------'
 616 //
 617 // To show that this works fine also with irregular spacings, let's say that
 618 // the third frame is delayed a bit (something earlier was dropped). Now the
 619 // situation looks like this:
 620 //
 621 //        I  B  P  B  B   P
 622 //   pts: 30 40 80 90 100 110
 623 //        ↓  ↓     ↓
 624 //   dts: 10 30 20 90 50←40
 625 //         |  |  ↑        ↑
 626 //         `--|--'        |
 627 //             `----------'
 628 //
 629 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 630 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 631 //
 632 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 633 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 634 // a tip frame and should be given a dts based on the previous one.
 635 #define FRAME_P 0
 636 #define FRAME_B 1
 637 #define FRAME_I 2
 638 #define FRAME_IDR 7
 639 void encoding2display_order(
 640     unsigned long long encoding_order, int intra_period,
 641     int intra_idr_period, int ip_period,
 642     unsigned long long *displaying_order,
 643     int *frame_type, int *pts_lag)
 644 {
 645     int encoding_order_gop = 0;
 646
 647     *pts_lag = 0;
 648
 649     if (intra_period == 1) { /* all are I/IDR frames */
 650         *displaying_order = encoding_order;
 651         if (intra_idr_period == 0)
 652             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 653         else
 654             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 655         return;
 656     }
 657
 658     if (intra_period == 0)
 659         intra_idr_period = 0;
 660
 661     if (ip_period == 1) {
 662         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 663         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 664         *displaying_order = encoding_order;
 665
 666         if (encoding_order_gop == 0) { /* the first frame */
 667             *frame_type = FRAME_IDR;
 668         } else if (intra_period != 0 && /* have I frames */
 669                    encoding_order_gop >= 2 &&
 670                    (encoding_order_gop % intra_period == 0)) {
 671             *frame_type = FRAME_I;
 672         } else {
 673             *frame_type = FRAME_P;
 674         }
 675         return;
 676     }
 677
 678     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 679     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 680     *pts_lag = -1;  // Most frames are not tip frames.
 681
 682     if (encoding_order_gop == 0) { /* the first frame */
 683         *frame_type = FRAME_IDR;
 684         *displaying_order = encoding_order;
 685         // IDR frames are a special case; I honestly can't find the logic behind
 686         // why this is the right thing, but it seems to line up nicely in practice :-)
 687         *pts_lag = TIMEBASE / frame_rate;
 688     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 689         *frame_type = FRAME_B;
 690         *displaying_order = encoding_order - 1;
 691         if ((encoding_order_gop % ip_period) == 0) {
 692             *pts_lag = 0;  // Last B-frame.
 693         }
 694     } else if (intra_period != 0 && /* have I frames */
 695                encoding_order_gop >= 2 &&
 696                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 697         *frame_type = FRAME_I;
 698         *displaying_order = encoding_order + ip_period - 1;
 699     } else {
 700         *frame_type = FRAME_P;
 701         *displaying_order = encoding_order + ip_period - 1;
 702     }
 703 }
 704
 705
 706 static const char *rc_to_string(int rcmode)
 707 {
 708     switch (rc_mode) {
 709     case VA_RC_NONE:
 710         return "NONE";
 711     case VA_RC_CBR:
 712         return "CBR";
 713     case VA_RC_VBR:
 714         return "VBR";
 715     case VA_RC_VCM:
 716         return "VCM";
 717     case VA_RC_CQP:
 718         return "CQP";
 719     case VA_RC_VBR_CONSTRAINED:
 720         return "VBR_CONSTRAINED";
 721     default:
 722         return "Unknown";
 723     }
 724 }
 725
 726 #if 0
 727 static int process_cmdline(int argc, char *argv[])
 728 {
 729     char c;
 730     const struct option long_opts[] = {
 731         {"help", no_argument, NULL, 0 },
 732         {"bitrate", required_argument, NULL, 1 },
 733         {"minqp", required_argument, NULL, 2 },
 734         {"initialqp", required_argument, NULL, 3 },
 735         {"intra_period", required_argument, NULL, 4 },
 736         {"idr_period", required_argument, NULL, 5 },
 737         {"ip_period", required_argument, NULL, 6 },
 738         {"rcmode", required_argument, NULL, 7 },
 739         {"srcyuv", required_argument, NULL, 9 },
 740         {"recyuv", required_argument, NULL, 10 },
 741         {"fourcc", required_argument, NULL, 11 },
 742         {"syncmode", no_argument, NULL, 12 },
 743         {"enablePSNR", no_argument, NULL, 13 },
 744         {"prit", required_argument, NULL, 14 },
 745         {"priv", required_argument, NULL, 15 },
 746         {"framecount", required_argument, NULL, 16 },
 747         {"entropy", required_argument, NULL, 17 },
 748         {"profile", required_argument, NULL, 18 },
 749         {NULL, no_argument, NULL, 0 }};
 750     int long_index;
 751
 752     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 753         switch (c) {
 754         case 'w':
 755             frame_width = atoi(optarg);
 756             break;
 757         case 'h':
 758             frame_height = atoi(optarg);
 759             break;
 760         case 'n':
 761         case 'f':
 762             frame_rate = atoi(optarg);
 763             break;
 764         case 'o':
 765             coded_fn = strdup(optarg);
 766             break;
 767         case 0:
 768             print_help();
 769             exit(0);
 770         case 1:
 771             frame_bitrate = atoi(optarg);
 772             break;
 773         case 2:
 774             minimal_qp = atoi(optarg);
 775             break;
 776         case 3:
 777             initial_qp = atoi(optarg);
 778             break;
 779         case 4:
 780             intra_period = atoi(optarg);
 781             break;
 782         case 5:
 783             intra_idr_period = atoi(optarg);
 784             break;
 785         case 6:
 786             ip_period = atoi(optarg);
 787             break;
 788         case 7:
 789             rc_mode = string_to_rc(optarg);
 790             if (rc_mode < 0) {
 791                 print_help();
 792                 exit(1);
 793             }
 794             break;
 795         case 9:
 796             srcyuv_fn = strdup(optarg);
 797             break;
 798         case 11:
 799             srcyuv_fourcc = string_to_fourcc(optarg);
 800             if (srcyuv_fourcc <= 0) {
 801                 print_help();
 802                 exit(1);
 803             }
 804             break;
 805         case 13:
 806             calc_psnr = 1;
 807             break;
 808         case 14:
 809             misc_priv_type = strtol(optarg, NULL, 0);
 810             break;
 811         case 15:
 812             misc_priv_value = strtol(optarg, NULL, 0);
 813             break;
 814         case 17:
 815             h264_entropy_mode = atoi(optarg) ? 1: 0;
 816             break;
 817         case 18:
 818             if (strncmp(optarg, "BP", 2) == 0)
 819                 h264_profile = VAProfileH264Baseline;
 820             else if (strncmp(optarg, "MP", 2) == 0)
 821                 h264_profile = VAProfileH264Main;
 822             else if (strncmp(optarg, "HP", 2) == 0)
 823                 h264_profile = VAProfileH264High;
 824             else
 825                 h264_profile = (VAProfile)0;
 826             break;
 827         case ':':
 828         case '?':
 829             print_help();
 830             exit(0);
 831         }
 832     }
 833
 834     if (ip_period < 1) {
 835         printf(" ip_period must be greater than 0\n");
 836         exit(0);
 837     }
 838     if (intra_period != 1 && intra_period % ip_period != 0) {
 839         printf(" intra_period must be a multiplier of ip_period\n");
 840         exit(0);
 841     }
 842     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 843         printf(" intra_idr_period must be a multiplier of intra_period\n");
 844         exit(0);
 845     }
 846
 847     if (frame_bitrate == 0)
 848         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 849
 850     if (coded_fn == NULL) {
 851         struct stat buf;
 852         if (stat("/tmp", &buf) == 0)
 853             coded_fn = strdup("/tmp/test.264");
 854         else if (stat("/sdcard", &buf) == 0)
 855             coded_fn = strdup("/sdcard/test.264");
 856         else
 857             coded_fn = strdup("./test.264");
 858     }
 859
 860
 861     frame_width_mbaligned = (frame_width + 15) & (~15);
 862     frame_height_mbaligned = (frame_height + 15) & (~15);
 863     if (frame_width != frame_width_mbaligned ||
 864         frame_height != frame_height_mbaligned) {
 865         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 866                frame_width, frame_height,
 867                frame_width_mbaligned, frame_height_mbaligned
 868                );
 869     }
 870
 871     return 0;
 872 }
 873 #endif
 874
 875 static Display *x11_display;
 876 static Window   x11_window;
 877
 878 VADisplay
 879 va_open_display(void)
 880 {
 881     x11_display = XOpenDisplay(NULL);
 882     if (!x11_display) {
 883         fprintf(stderr, "error: can't connect to X server!\n");
 884         return NULL;
 885     }
 886     return vaGetDisplay(x11_display);
 887 }
 888
 889 void
 890 va_close_display(VADisplay va_dpy)
 891 {
 892     if (!x11_display)
 893         return;
 894
 895     if (x11_window) {
 896         XUnmapWindow(x11_display, x11_window);
 897         XDestroyWindow(x11_display, x11_window);
 898         x11_window = None;
 899     }
 900     XCloseDisplay(x11_display);
 901     x11_display = NULL;
 902 }
 903
 904 static int init_va(void)
 905 {
 906     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 907     VAEntrypoint *entrypoints;
 908     int num_entrypoints, slice_entrypoint;
 909     int support_encode = 0;
 910     int major_ver, minor_ver;
 911     VAStatus va_status;
 912     unsigned int i;
 913
 914     va_dpy = va_open_display();
 915     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 916     CHECK_VASTATUS(va_status, "vaInitialize");
 917
 918     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 919     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 920     if (!entrypoints) {
 921         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 922         exit(1);
 923     }
 924
 925     /* use the highest profile */
 926     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 927         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 928             continue;
 929
 930         h264_profile = profile_list[i];
 931         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 932         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 933             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 934                 support_encode = 1;
 935                 break;
 936             }
 937         }
 938         if (support_encode == 1)
 939             break;
 940     }
 941
 942     if (support_encode == 0) {
 943         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 944         exit(1);
 945     } else {
 946         switch (h264_profile) {
 947             case VAProfileH264Baseline:
 948                 printf("Use profile VAProfileH264Baseline\n");
 949                 ip_period = 1;
 950                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 951                 h264_entropy_mode = 0;
 952                 break;
 953             case VAProfileH264ConstrainedBaseline:
 954                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 955                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 956                 ip_period = 1;
 957                 break;
 958
 959             case VAProfileH264Main:
 960                 printf("Use profile VAProfileH264Main\n");
 961                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 962                 break;
 963
 964             case VAProfileH264High:
 965                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 966                 printf("Use profile VAProfileH264High\n");
 967                 break;
 968             default:
 969                 printf("unknow profile. Set to Baseline");
 970                 h264_profile = VAProfileH264Baseline;
 971                 ip_period = 1;
 972                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 973                 break;
 974         }
 975     }
 976
 977     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 978
 979     /* find out the format for the render target, and rate control mode */
 980     for (i = 0; i < VAConfigAttribTypeMax; i++)
 981         attrib[i].type = (VAConfigAttribType)i;
 982
 983     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 984                                       &attrib[0], VAConfigAttribTypeMax);
 985     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 986     /* check the interested configattrib */
 987     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 988         printf("Not find desired YUV420 RT format\n");
 989         exit(1);
 990     } else {
 991         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 992         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 993         config_attrib_num++;
 994     }
 995
 996     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 997         int tmp = attrib[VAConfigAttribRateControl].value;
 998
 999         printf("Support rate control mode (0x%x):", tmp);
1000
1001         if (tmp & VA_RC_NONE)
1002             printf("NONE ");
1003         if (tmp & VA_RC_CBR)
1004             printf("CBR ");
1005         if (tmp & VA_RC_VBR)
1006             printf("VBR ");
1007         if (tmp & VA_RC_VCM)
1008             printf("VCM ");
1009         if (tmp & VA_RC_CQP)
1010             printf("CQP ");
1011         if (tmp & VA_RC_VBR_CONSTRAINED)
1012             printf("VBR_CONSTRAINED ");
1013
1014         printf("\n");
1015
1016         if (rc_mode == -1 || !(rc_mode & tmp))  {
1017             if (rc_mode != -1) {
1018                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1019             }
1020
1021             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1022                 if (rc_default_modes[i] & tmp) {
1023                     rc_mode = rc_default_modes[i];
1024                     break;
1025                 }
1026             }
1027
1028             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
1029         }
1030
1031         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1032         config_attrib[config_attrib_num].value = rc_mode;
1033         config_attrib_num++;
1034     }
1035
1036
1037     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1038         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1039
1040         printf("Support VAConfigAttribEncPackedHeaders\n");
1041
1042         h264_packedheader = 1;
1043         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1044         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1045
1046         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1047             printf("Support packed sequence headers\n");
1048             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1049         }
1050
1051         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1052             printf("Support packed picture headers\n");
1053             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1054         }
1055
1056         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1057             printf("Support packed slice headers\n");
1058             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1059         }
1060
1061         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1062             printf("Support packed misc headers\n");
1063             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1064         }
1065
1066         enc_packed_header_idx = config_attrib_num;
1067         config_attrib_num++;
1068     }
1069
1070     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1071         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1072
1073         printf("Support VAConfigAttribEncInterlaced\n");
1074
1075         if (tmp & VA_ENC_INTERLACED_FRAME)
1076             printf("support VA_ENC_INTERLACED_FRAME\n");
1077         if (tmp & VA_ENC_INTERLACED_FIELD)
1078             printf("Support VA_ENC_INTERLACED_FIELD\n");
1079         if (tmp & VA_ENC_INTERLACED_MBAFF)
1080             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1081         if (tmp & VA_ENC_INTERLACED_PAFF)
1082             printf("Support VA_ENC_INTERLACED_PAFF\n");
1083
1084         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1085         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1086         config_attrib_num++;
1087     }
1088
1089     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1090         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1091
1092         printf("Support %d RefPicList0 and %d RefPicList1\n",
1093                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1094     }
1095
1096     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1097         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1098
1099     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1100         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1101
1102         printf("Support VAConfigAttribEncSliceStructure\n");
1103
1104         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1105             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1106         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1107             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1108         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1109             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1110     }
1111     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1112         printf("Support VAConfigAttribEncMacroblockInfo\n");
1113     }
1114
1115     free(entrypoints);
1116     return 0;
1117 }
1118
1119 static int setup_encode()
1120 {
1121     VAStatus va_status;
1122     VASurfaceID *tmp_surfaceid;
1123     int codedbuf_size, i;
1124     static VASurfaceID src_surface[SURFACE_NUM];
1125     static VASurfaceID ref_surface[SURFACE_NUM];
1126
1127     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1128             &config_attrib[0], config_attrib_num, &config_id);
1129     CHECK_VASTATUS(va_status, "vaCreateConfig");
1130
1131     /* create source surfaces */
1132     va_status = vaCreateSurfaces(va_dpy,
1133                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1134                                  &src_surface[0], SURFACE_NUM,
1135                                  NULL, 0);
1136     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1137
1138     /* create reference surfaces */
1139     va_status = vaCreateSurfaces(va_dpy,
1140                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1141                                  &ref_surface[0], SURFACE_NUM,
1142                                  NULL, 0);
1143     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1144
1145     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1146     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1147     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1148
1149     /* Create a context for this encode pipe */
1150     va_status = vaCreateContext(va_dpy, config_id,
1151                                 frame_width_mbaligned, frame_height_mbaligned,
1152                                 VA_PROGRESSIVE,
1153                                 tmp_surfaceid, 2 * SURFACE_NUM,
1154                                 &context_id);
1155     CHECK_VASTATUS(va_status, "vaCreateContext");
1156     free(tmp_surfaceid);
1157
1158     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1159
1160     for (i = 0; i < SURFACE_NUM; i++) {
1161         /* create coded buffer once for all
1162          * other VA buffers which won't be used again after vaRenderPicture.
1163          * so APP can always vaCreateBuffer for every frame
1164          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1165          * so VA won't maintain the coded buffer
1166          */
1167         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1168                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1169         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1170     }
1171
1172     /* create OpenGL objects */
1173     //glGenFramebuffers(SURFACE_NUM, fbos);
1174
1175     for (i = 0; i < SURFACE_NUM; i++) {
1176         glGenTextures(1, &gl_surfaces[i].y_tex);
1177         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1178     }
1179
1180     for (i = 0; i < SURFACE_NUM; i++) {
1181         gl_surfaces[i].src_surface = src_surface[i];
1182         gl_surfaces[i].ref_surface = ref_surface[i];
1183     }
1184
1185     return 0;
1186 }
1187
1188
1189
1190 #define partition(ref, field, key, ascending)   \
1191     while (i <= j) {                            \
1192         if (ascending) {                        \
1193             while (ref[i].field < key)          \
1194                 i++;                            \
1195             while (ref[j].field > key)          \
1196                 j--;                            \
1197         } else {                                \
1198             while (ref[i].field > key)          \
1199                 i++;                            \
1200             while (ref[j].field < key)          \
1201                 j--;                            \
1202         }                                       \
1203         if (i <= j) {                           \
1204             tmp = ref[i];                       \
1205             ref[i] = ref[j];                    \
1206             ref[j] = tmp;                       \
1207             i++;                                \
1208             j--;                                \
1209         }                                       \
1210     }                                           \
1211
1212 static void sort_one(VAPictureH264 ref[], int left, int right,
1213                      int ascending, int frame_idx)
1214 {
1215     int i = left, j = right;
1216     unsigned int key;
1217     VAPictureH264 tmp;
1218
1219     if (frame_idx) {
1220         key = ref[(left + right) / 2].frame_idx;
1221         partition(ref, frame_idx, key, ascending);
1222     } else {
1223         key = ref[(left + right) / 2].TopFieldOrderCnt;
1224         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1225     }
1226
1227     /* recursion */
1228     if (left < j)
1229         sort_one(ref, left, j, ascending, frame_idx);
1230
1231     if (i < right)
1232         sort_one(ref, i, right, ascending, frame_idx);
1233 }
1234
1235 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1236                      int partition_ascending, int list0_ascending, int list1_ascending)
1237 {
1238     int i = left, j = right;
1239     VAPictureH264 tmp;
1240
1241     if (frame_idx) {
1242         partition(ref, frame_idx, key, partition_ascending);
1243     } else {
1244         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1245     }
1246
1247
1248     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1249     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1250 }
1251
1252 static int update_ReferenceFrames(void)
1253 {
1254     int i;
1255
1256     if (current_frame_type == FRAME_B)
1257         return 0;
1258
1259     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1260     numShortTerm++;
1261     if (numShortTerm > num_ref_frames)
1262         numShortTerm = num_ref_frames;
1263     for (i=numShortTerm-1; i>0; i--)
1264         ReferenceFrames[i] = ReferenceFrames[i-1];
1265     ReferenceFrames[0] = CurrentCurrPic;
1266
1267     if (current_frame_type != FRAME_B)
1268         current_frame_num++;
1269     if (current_frame_num > MaxFrameNum)
1270         current_frame_num = 0;
1271
1272     return 0;
1273 }
1274
1275
1276 static int update_RefPicList(void)
1277 {
1278     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1279
1280     if (current_frame_type == FRAME_P) {
1281         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1282         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1283     }
1284
1285     if (current_frame_type == FRAME_B) {
1286         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1287         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1288                  1, 0, 1);
1289
1290         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1291         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1292                  0, 1, 0);
1293     }
1294
1295     return 0;
1296 }
1297
1298
1299 static int render_sequence(void)
1300 {
1301     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1302     VAStatus va_status;
1303     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1304     VAEncMiscParameterRateControl *misc_rate_ctrl;
1305
1306     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1307     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1308     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1309     seq_param.bits_per_second = frame_bitrate;
1310
1311     seq_param.intra_period = intra_period;
1312     seq_param.intra_idr_period = intra_idr_period;
1313     seq_param.ip_period = ip_period;
1314
1315     seq_param.max_num_ref_frames = num_ref_frames;
1316     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1317     seq_param.time_scale = TIMEBASE * 2;
1318     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1319     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1320     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1321     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1322     seq_param.seq_fields.bits.chroma_format_idc = 1;
1323     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1324
1325     if (frame_width != frame_width_mbaligned ||
1326         frame_height != frame_height_mbaligned) {
1327         seq_param.frame_cropping_flag = 1;
1328         seq_param.frame_crop_left_offset = 0;
1329         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1330         seq_param.frame_crop_top_offset = 0;
1331         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1332     }
1333
1334     va_status = vaCreateBuffer(va_dpy, context_id,
1335                                VAEncSequenceParameterBufferType,
1336                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1337     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1338
1339     va_status = vaCreateBuffer(va_dpy, context_id,
1340                                VAEncMiscParameterBufferType,
1341                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1342                                1, NULL, &rc_param_buf);
1343     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1344
1345     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1346     misc_param->type = VAEncMiscParameterTypeRateControl;
1347     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1348     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1349     misc_rate_ctrl->bits_per_second = frame_bitrate;
1350     misc_rate_ctrl->target_percentage = 66;
1351     misc_rate_ctrl->window_size = 1000;
1352     misc_rate_ctrl->initial_qp = initial_qp;
1353     misc_rate_ctrl->min_qp = minimal_qp;
1354     misc_rate_ctrl->basic_unit_size = 0;
1355     vaUnmapBuffer(va_dpy, rc_param_buf);
1356
1357     render_id[0] = seq_param_buf;
1358     render_id[1] = rc_param_buf;
1359
1360     va_status = vaRenderPicture(va_dpy, context_id, &render_id[0], 2);
1361     CHECK_VASTATUS(va_status, "vaRenderPicture");;
1362
1363     if (misc_priv_type != 0) {
1364         va_status = vaCreateBuffer(va_dpy, context_id,
1365                                    VAEncMiscParameterBufferType,
1366                                    sizeof(VAEncMiscParameterBuffer),
1367                                    1, NULL, &misc_param_tmpbuf);
1368         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1369         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1370         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1371         misc_param_tmp->data[0] = misc_priv_value;
1372         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1373
1374         va_status = vaRenderPicture(va_dpy, context_id, &misc_param_tmpbuf, 1);
1375     }
1376
1377     return 0;
1378 }
1379
1380 static int calc_poc(int pic_order_cnt_lsb)
1381 {
1382     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1383     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1384     int PicOrderCntMsb, TopFieldOrderCnt;
1385
1386     if (current_frame_type == FRAME_IDR)
1387         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1388     else {
1389         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1390         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1391     }
1392
1393     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1394         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1395         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1396     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1397              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1398         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1399     else
1400         PicOrderCntMsb = prevPicOrderCntMsb;
1401
1402     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1403
1404     if (current_frame_type != FRAME_B) {
1405         PicOrderCntMsb_ref = PicOrderCntMsb;
1406         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1407     }
1408
1409     return TopFieldOrderCnt;
1410 }
1411
1412 static int render_picture(void)
1413 {
1414     VABufferID pic_param_buf;
1415     VAStatus va_status;
1416     int i = 0;
1417
1418     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1419     pic_param.CurrPic.frame_idx = current_frame_num;
1420     pic_param.CurrPic.flags = 0;
1421     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1422     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1423     CurrentCurrPic = pic_param.CurrPic;
1424
1425     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1426         update_RefPicList(); /* calc RefPicList */
1427         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1428         if (current_frame_type == FRAME_P) {
1429             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1430         } else if (current_frame_type == FRAME_B) {
1431             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1432             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1433         }
1434     } else {
1435         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1436         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1437             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1438             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1439         }
1440     }
1441
1442     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1443     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1444     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1445     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1446     pic_param.frame_num = current_frame_num;
1447     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1448     pic_param.last_picture = false;  // FIXME
1449     pic_param.pic_init_qp = initial_qp;
1450
1451     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1452                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1453     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1454
1455     va_status = vaRenderPicture(va_dpy, context_id, &pic_param_buf, 1);
1456     CHECK_VASTATUS(va_status, "vaRenderPicture");
1457
1458     return 0;
1459 }
1460
1461 static int render_packedsequence(void)
1462 {
1463     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1464     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1465     unsigned int length_in_bits;
1466     unsigned char *packedseq_buffer = NULL;
1467     VAStatus va_status;
1468
1469     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1470
1471     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1472
1473     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1474     packedheader_param_buffer.has_emulation_bytes = 0;
1475     va_status = vaCreateBuffer(va_dpy,
1476                                context_id,
1477                                VAEncPackedHeaderParameterBufferType,
1478                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1479                                &packedseq_para_bufid);
1480     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1481
1482     va_status = vaCreateBuffer(va_dpy,
1483                                context_id,
1484                                VAEncPackedHeaderDataBufferType,
1485                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1486                                &packedseq_data_bufid);
1487     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1488
1489     render_id[0] = packedseq_para_bufid;
1490     render_id[1] = packedseq_data_bufid;
1491     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1492     CHECK_VASTATUS(va_status, "vaRenderPicture");
1493
1494     free(packedseq_buffer);
1495
1496     return 0;
1497 }
1498
1499
1500 static int render_packedpicture(void)
1501 {
1502     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1503     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1504     unsigned int length_in_bits;
1505     unsigned char *packedpic_buffer = NULL;
1506     VAStatus va_status;
1507
1508     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1509     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1510     packedheader_param_buffer.bit_length = length_in_bits;
1511     packedheader_param_buffer.has_emulation_bytes = 0;
1512
1513     va_status = vaCreateBuffer(va_dpy,
1514                                context_id,
1515                                VAEncPackedHeaderParameterBufferType,
1516                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1517                                &packedpic_para_bufid);
1518     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1519
1520     va_status = vaCreateBuffer(va_dpy,
1521                                context_id,
1522                                VAEncPackedHeaderDataBufferType,
1523                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1524                                &packedpic_data_bufid);
1525     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1526
1527     render_id[0] = packedpic_para_bufid;
1528     render_id[1] = packedpic_data_bufid;
1529     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1530     CHECK_VASTATUS(va_status, "vaRenderPicture");
1531
1532     free(packedpic_buffer);
1533
1534     return 0;
1535 }
1536
1537 static void render_packedslice()
1538 {
1539     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1540     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1541     unsigned int length_in_bits;
1542     unsigned char *packedslice_buffer = NULL;
1543     VAStatus va_status;
1544
1545     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1546     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1547     packedheader_param_buffer.bit_length = length_in_bits;
1548     packedheader_param_buffer.has_emulation_bytes = 0;
1549
1550     va_status = vaCreateBuffer(va_dpy,
1551                                context_id,
1552                                VAEncPackedHeaderParameterBufferType,
1553                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1554                                &packedslice_para_bufid);
1555     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1556
1557     va_status = vaCreateBuffer(va_dpy,
1558                                context_id,
1559                                VAEncPackedHeaderDataBufferType,
1560                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1561                                &packedslice_data_bufid);
1562     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1563
1564     render_id[0] = packedslice_para_bufid;
1565     render_id[1] = packedslice_data_bufid;
1566     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1567     CHECK_VASTATUS(va_status, "vaRenderPicture");
1568
1569     free(packedslice_buffer);
1570 }
1571
1572 static int render_slice(void)
1573 {
1574     VABufferID slice_param_buf;
1575     VAStatus va_status;
1576     int i;
1577
1578     update_RefPicList();
1579
1580     /* one frame, one slice */
1581     slice_param.macroblock_address = 0;
1582     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1583     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1584     if (current_frame_type == FRAME_IDR) {
1585         if (current_frame_encoding != 0)
1586             ++slice_param.idr_pic_id;
1587     } else if (current_frame_type == FRAME_P) {
1588         int refpiclist0_max = h264_maxref & 0xffff;
1589         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1590
1591         for (i = refpiclist0_max; i < 32; i++) {
1592             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1593             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1594         }
1595     } else if (current_frame_type == FRAME_B) {
1596         int refpiclist0_max = h264_maxref & 0xffff;
1597         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1598
1599         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1600         for (i = refpiclist0_max; i < 32; i++) {
1601             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1602             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1603         }
1604
1605         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1606         for (i = refpiclist1_max; i < 32; i++) {
1607             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1608             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1609         }
1610     }
1611
1612     slice_param.slice_alpha_c0_offset_div2 = 0;
1613     slice_param.slice_beta_offset_div2 = 0;
1614     slice_param.direct_spatial_mv_pred_flag = 1;
1615     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1616
1617
1618     if (h264_packedheader &&
1619         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1620         render_packedslice();
1621
1622     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1623                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1624     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1625
1626     va_status = vaRenderPicture(va_dpy, context_id, &slice_param_buf, 1);
1627     CHECK_VASTATUS(va_status, "vaRenderPicture");
1628
1629     return 0;
1630 }
1631
1632
1633
1634 int H264Encoder::save_codeddata(storage_task task)
1635 {
1636     VACodedBufferSegment *buf_list = NULL;
1637     VAStatus va_status;
1638     unsigned int coded_size = 0;
1639
1640     string data;
1641
1642     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / frame_rate);  // So we never get negative dts.
1643
1644     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1645     CHECK_VASTATUS(va_status, "vaMapBuffer");
1646     while (buf_list != NULL) {
1647         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1648         buf_list = (VACodedBufferSegment *) buf_list->next;
1649
1650         frame_size += coded_size;
1651     }
1652     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1653
1654     {
1655         // Add video.
1656         AVPacket pkt;
1657         memset(&pkt, 0, sizeof(pkt));
1658         pkt.buf = nullptr;
1659         pkt.pts = av_rescale_q(task.pts + global_delay, AVRational{1, TIMEBASE}, avstream_video->time_base);
1660         pkt.dts = av_rescale_q(task.dts + global_delay, AVRational{1, TIMEBASE}, avstream_video->time_base);
1661         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1662         pkt.size = data.size();
1663         pkt.stream_index = 0;
1664         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1665             pkt.flags = AV_PKT_FLAG_KEY;
1666         } else {
1667             pkt.flags = 0;
1668         }
1669         //pkt.duration = 1;
1670         httpd->add_packet(pkt);
1671         av_interleaved_write_frame(avctx, &pkt);
1672     }
1673     // Encode and add all audio frames up to and including the pts of this video frame.
1674     // (They can never be queued to us after the video frame they belong to, only before.)
1675     for ( ;; ) {
1676         int64_t audio_pts;
1677         std::vector<float> audio;
1678         {
1679              unique_lock<mutex> lock(frame_queue_mutex);
1680              if (pending_audio_frames.empty()) break;
1681              auto it = pending_audio_frames.begin();
1682              if (it->first > task.pts) break;
1683              audio_pts = it->first;
1684              audio = move(it->second);
1685              pending_audio_frames.erase(it);
1686         }
1687         AVFrame *frame = avcodec_alloc_frame();
1688         frame->nb_samples = audio.size() / 2;
1689         frame->format = AV_SAMPLE_FMT_FLT;
1690         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1691
1692         unique_ptr<float[]> planar_samples(new float[audio.size()]);
1693         avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), audio.size() * sizeof(float), 0);
1694         for (int i = 0; i < frame->nb_samples; ++i) {
1695             planar_samples[i] = audio[i * 2 + 0];
1696             planar_samples[i + frame->nb_samples] = audio[i * 2 + 1];
1697         }
1698
1699         AVPacket pkt;
1700         av_init_packet(&pkt);
1701         pkt.data = nullptr;
1702         pkt.size = 0;
1703         int got_output;
1704         avcodec_encode_audio2(avstream_audio->codec, &pkt, frame, &got_output);
1705         if (got_output) {
1706             pkt.pts = av_rescale_q(audio_pts + global_delay, AVRational{1, TIMEBASE}, avstream_audio->time_base);
1707             pkt.dts = pkt.pts;
1708             pkt.stream_index = 1;
1709             httpd->add_packet(pkt);
1710             av_interleaved_write_frame(avctx, &pkt);
1711         }
1712         // TODO: Delayed frames.
1713         avcodec_free_frame(&frame);
1714     }
1715
1716 #if 0
1717     printf("\r      "); /* return back to startpoint */
1718     switch (encode_order % 4) {
1719         case 0:
1720             printf("|");
1721             break;
1722         case 1:
1723             printf("/");
1724             break;
1725         case 2:
1726             printf("-");
1727             break;
1728         case 3:
1729             printf("\\");
1730             break;
1731     }
1732     printf("%08lld", encode_order);
1733     printf("(%06d bytes coded)", coded_size);
1734 #endif
1735
1736     return 0;
1737 }
1738
1739
1740 // this is weird. but it seems to put a new frame onto the queue
1741 void H264Encoder::storage_task_enqueue(storage_task task)
1742 {
1743         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1744         storage_task_queue.push(move(task));
1745         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1746         storage_task_queue_changed.notify_all();
1747 }
1748
1749 void H264Encoder::storage_task_thread()
1750 {
1751         for ( ;; ) {
1752                 storage_task current;
1753                 {
1754                         // wait until there's an encoded frame
1755                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1756                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1757                         if (storage_thread_should_quit) return;
1758                         current = move(storage_task_queue.front());
1759                         storage_task_queue.pop();
1760                 }
1761
1762                 VAStatus va_status;
1763
1764                 // waits for data, then saves it to disk.
1765                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1766                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1767                 save_codeddata(move(current));
1768
1769                 {
1770                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1771                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1772                         storage_task_queue_changed.notify_all();
1773                 }
1774         }
1775 }
1776
1777 static int release_encode()
1778 {
1779     int i;
1780
1781     for (i = 0; i < SURFACE_NUM; i++) {
1782         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1783         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1784         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1785     }
1786
1787     vaDestroyContext(va_dpy, context_id);
1788     vaDestroyConfig(va_dpy, config_id);
1789
1790     return 0;
1791 }
1792
1793 static int deinit_va()
1794 {
1795     vaTerminate(va_dpy);
1796
1797     va_close_display(va_dpy);
1798
1799     return 0;
1800 }
1801
1802
1803 static int print_input()
1804 {
1805     printf("\n\nINPUT:Try to encode H264...\n");
1806     if (rc_mode != -1)
1807         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1808     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1809     printf("INPUT: FrameRate    : %d\n", frame_rate);
1810     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1811     printf("INPUT: Slieces      : %d\n", frame_slices);
1812     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1813     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1814     printf("INPUT: IpPeriod     : %d\n", ip_period);
1815     printf("INPUT: Initial QP   : %d\n", initial_qp);
1816     printf("INPUT: Min QP       : %d\n", minimal_qp);
1817     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1818
1819     printf("\n\n"); /* return back to startpoint */
1820
1821     return 0;
1822 }
1823
1824
1825 //H264Encoder::H264Encoder(SDL_Window *window, SDL_GLContext context, int width, int height, const char *output_filename)
1826 H264Encoder::H264Encoder(QSurface *surface, int width, int height, const char *output_filename, HTTPD *httpd)
1827         : current_storage_frame(0), surface(surface), httpd(httpd)
1828         //: width(width), height(height), current_encoding_frame(0)
1829 {
1830         avctx = avformat_alloc_context();
1831         avctx->oformat = av_guess_format(NULL, output_filename, NULL);
1832         strcpy(avctx->filename, output_filename);
1833         if (avio_open2(&avctx->pb, output_filename, AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL) < 0) {
1834                 fprintf(stderr, "%s: avio_open2() failed\n", output_filename);
1835                 exit(1);
1836         }
1837         AVCodec *codec_video = avcodec_find_encoder(AV_CODEC_ID_H264);
1838         avstream_video = avformat_new_stream(avctx, codec_video);
1839         if (avstream_video == nullptr) {
1840                 fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
1841                 exit(1);
1842         }
1843         avstream_video->time_base = AVRational{1, TIMEBASE};
1844         avstream_video->codec->width = width;
1845         avstream_video->codec->height = height;
1846         avstream_video->codec->time_base = AVRational{1, TIMEBASE};
1847         avstream_video->codec->ticks_per_frame = 1;  // or 2?
1848
1849         AVCodec *codec_audio = avcodec_find_encoder(AV_CODEC_ID_MP3);
1850         avstream_audio = avformat_new_stream(avctx, codec_audio);
1851         if (avstream_audio == nullptr) {
1852                 fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
1853                 exit(1);
1854         }
1855         avstream_audio->time_base = AVRational{1, TIMEBASE};
1856         avstream_audio->codec->bit_rate = 256000;
1857         avstream_audio->codec->sample_rate = 48000;
1858         avstream_audio->codec->sample_fmt = AV_SAMPLE_FMT_FLTP;
1859         avstream_audio->codec->channels = 2;
1860         avstream_audio->codec->channel_layout = AV_CH_LAYOUT_STEREO;
1861         avstream_audio->codec->time_base = AVRational{1, TIMEBASE};
1862
1863         /* open it */
1864         if (avcodec_open2(avstream_audio->codec, codec_audio, NULL) < 0) {
1865                 fprintf(stderr, "Could not open codec\n");
1866                 exit(1);
1867         }
1868
1869         if (avformat_write_header(avctx, NULL) < 0) {
1870                 fprintf(stderr, "%s: avformat_write_header() failed\n", output_filename);
1871                 exit(1);
1872         }
1873
1874         frame_width = width;
1875         frame_height = height;
1876         frame_width_mbaligned = (frame_width + 15) & (~15);
1877         frame_height_mbaligned = (frame_height + 15) & (~15);
1878         frame_bitrate = 15000000;  // / 60;
1879         current_frame_encoding = 0;
1880
1881         print_input();
1882
1883         init_va();
1884         setup_encode();
1885
1886         // No frames are ready yet.
1887         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1888
1889         memset(&seq_param, 0, sizeof(seq_param));
1890         memset(&pic_param, 0, sizeof(pic_param));
1891         memset(&slice_param, 0, sizeof(slice_param));
1892
1893         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1894
1895         copy_thread = std::thread([this]{
1896                 //SDL_GL_MakeCurrent(window, context);
1897                 QOpenGLContext *context = create_context();
1898                 eglBindAPI(EGL_OPENGL_API);
1899                 if (!make_current(context, this->surface)) {
1900                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1901                                 eglGetError());
1902                         exit(1);
1903                 }
1904                 copy_thread_func();
1905         });
1906 }
1907
1908 H264Encoder::~H264Encoder()
1909 {
1910         {
1911                 unique_lock<mutex> lock(storage_task_queue_mutex);
1912                 storage_thread_should_quit = true;
1913                 storage_task_queue_changed.notify_all();
1914         }
1915         {
1916                 unique_lock<mutex> lock(frame_queue_mutex);
1917                 copy_thread_should_quit = true;
1918                 frame_queue_nonempty.notify_one();
1919         }
1920         storage_thread.join();
1921         copy_thread.join();
1922
1923         release_encode();
1924         deinit_va();
1925
1926         av_write_trailer(avctx);
1927         avformat_free_context(avctx);
1928 }
1929
1930 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1931 {
1932         {
1933                 // Wait until this frame slot is done encoding.
1934                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1935                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1936                 if (storage_thread_should_quit) return false;
1937         }
1938
1939         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1940         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1941         *y_tex = surf->y_tex;
1942         *cbcr_tex = surf->cbcr_tex;
1943
1944         VASurfaceID surface = surf->src_surface;
1945         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1946         CHECK_VASTATUS(va_status, "vaDeriveImage");
1947
1948         VABufferInfo buf_info;
1949         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1950         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1951         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1952
1953         // Create Y image.
1954         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1955         EGLint y_attribs[] = {
1956                 EGL_WIDTH, frame_width,
1957                 EGL_HEIGHT, frame_height,
1958                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1959                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1960                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1961                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1962                 EGL_NONE
1963         };
1964
1965         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1966         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1967
1968         // Associate Y image to a texture.
1969         glBindTexture(GL_TEXTURE_2D, *y_tex);
1970         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1971
1972         // Create CbCr image.
1973         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1974         EGLint cbcr_attribs[] = {
1975                 EGL_WIDTH, frame_width,
1976                 EGL_HEIGHT, frame_height,
1977                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1978                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1979                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1980                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1981                 EGL_NONE
1982         };
1983
1984         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1985         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1986
1987         // Associate CbCr image to a texture.
1988         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1989         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1990
1991         return true;
1992 }
1993
1994 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1995 {
1996         {
1997                 unique_lock<mutex> lock(frame_queue_mutex);
1998                 pending_audio_frames[pts] = move(audio);
1999         }
2000         frame_queue_nonempty.notify_one();
2001 }
2002
2003
2004 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
2005 {
2006         {
2007                 unique_lock<mutex> lock(frame_queue_mutex);
2008                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
2009                 ++current_storage_frame;
2010         }
2011         frame_queue_nonempty.notify_one();
2012 }
2013
2014 void H264Encoder::copy_thread_func()
2015 {
2016         int64_t last_dts = -1;
2017         for ( ;; ) {
2018                 PendingFrame frame;
2019                 int pts_lag;
2020                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
2021                                        &current_frame_display, &current_frame_type, &pts_lag);
2022                 if (current_frame_type == FRAME_IDR) {
2023                         numShortTerm = 0;
2024                         current_frame_num = 0;
2025                         current_IDR_display = current_frame_display;
2026                 }
2027
2028                 {
2029                         unique_lock<mutex> lock(frame_queue_mutex);
2030                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
2031                         if (copy_thread_should_quit) return;
2032                         frame = move(pending_video_frames[current_frame_display]);
2033                         pending_video_frames.erase(current_frame_display);
2034                 }
2035
2036                 // Wait for the GPU to be done with the frame.
2037                 glClientWaitSync(frame.fence.get(), 0, 0);
2038
2039                 // Release back any input frames we needed to render this frame.
2040                 frame.input_frames.clear();
2041
2042                 // Unmap the image.
2043                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
2044                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2045                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2046                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2047                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2048                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2049                 CHECK_VASTATUS(va_status, "vaDestroyImage");
2050
2051                 VASurfaceID surface = surf->src_surface;
2052
2053                 // Schedule the frame for encoding.
2054                 va_status = vaBeginPicture(va_dpy, context_id, surface);
2055                 CHECK_VASTATUS(va_status, "vaBeginPicture");
2056
2057                 if (current_frame_type == FRAME_IDR) {
2058                         render_sequence();
2059                         render_picture();
2060                         if (h264_packedheader) {
2061                                 render_packedsequence();
2062                                 render_packedpicture();
2063                         }
2064                 } else {
2065                         //render_sequence();
2066                         render_picture();
2067                 }
2068                 render_slice();
2069
2070                 va_status = vaEndPicture(va_dpy, context_id);
2071                 CHECK_VASTATUS(va_status, "vaEndPicture");
2072
2073                 // Determine the pts and dts of this frame.
2074                 int64_t pts = frame.pts;
2075                 int64_t dts;
2076                 if (pts_lag == -1) {
2077                         assert(last_dts != -1);
2078                         dts = last_dts + (TIMEBASE / frame_rate);
2079                 } else {
2080                         dts = pts - pts_lag;
2081                 }
2082                 last_dts = dts;
2083
2084                 // so now the data is done encoding (well, async job kicked off)...
2085                 // we send that to the storage thread
2086                 storage_task tmp;
2087                 tmp.display_order = current_frame_display;
2088                 tmp.frame_type = current_frame_type;
2089                 tmp.pts = pts;
2090                 tmp.dts = dts;
2091                 storage_task_enqueue(move(tmp));
2092
2093                 update_ReferenceFrames();
2094                 ++current_frame_encoding;
2095         }
2096 }