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 <libavutil/channel_layout.h>
  11 #include <libavutil/frame.h>
  12 #include <libavutil/rational.h>
  13 #include <libavutil/samplefmt.h>
  14 #include <libdrm/drm_fourcc.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 <cstdint>
  24 #include <memory>
  25 #include <mutex>
  26 #include <queue>
  27 #include <string>
  28 #include <thread>
  29 #include <utility>
  30
  31 #include "context.h"
  32 #include "defs.h"
  33 #include "httpd.h"
  34 #include "timebase.h"
  35
  36 using namespace std;
  37
  38 class QOpenGLContext;
  39 class QSurface;
  40
  41 #define CHECK_VASTATUS(va_status, func)                                 \
  42     if (va_status != VA_STATUS_SUCCESS) {                               \
  43         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  44         exit(1);                                                        \
  45     }
  46
  47 //#include "loadsurface.h"
  48
  49 #define NAL_REF_IDC_NONE        0
  50 #define NAL_REF_IDC_LOW         1
  51 #define NAL_REF_IDC_MEDIUM      2
  52 #define NAL_REF_IDC_HIGH        3
  53
  54 #define NAL_NON_IDR             1
  55 #define NAL_IDR                 5
  56 #define NAL_SPS                 7
  57 #define NAL_PPS                 8
  58 #define NAL_SEI                 6
  59
  60 #define SLICE_TYPE_P            0
  61 #define SLICE_TYPE_B            1
  62 #define SLICE_TYPE_I            2
  63 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  64 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  65 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  66
  67
  68 #define ENTROPY_MODE_CAVLC      0
  69 #define ENTROPY_MODE_CABAC      1
  70
  71 #define PROFILE_IDC_BASELINE    66
  72 #define PROFILE_IDC_MAIN        77
  73 #define PROFILE_IDC_HIGH        100
  74
  75 #define BITSTREAM_ALLOCATE_STEPPING     4096
  76
  77 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  78 static  VADisplay va_dpy;
  79 static  VAProfile h264_profile = (VAProfile)~0;
  80 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  81 static  int config_attrib_num = 0, enc_packed_header_idx;
  82
  83 struct GLSurface {
  84         VASurfaceID src_surface, ref_surface;
  85         VABufferID coded_buf;
  86
  87         VAImage surface_image;
  88         GLuint y_tex, cbcr_tex;
  89         EGLImage y_egl_image, cbcr_egl_image;
  90 };
  91 GLSurface gl_surfaces[SURFACE_NUM];
  92
  93 static  VAConfigID config_id;
  94 static  VAContextID context_id;
  95 static  VAEncSequenceParameterBufferH264 seq_param;
  96 static  VAEncPictureParameterBufferH264 pic_param;
  97 static  VAEncSliceParameterBufferH264 slice_param;
  98 static  VAPictureH264 CurrentCurrPic;
  99 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
 100
 101 static  unsigned int MaxFrameNum = (2<<16);
 102 static  unsigned int MaxPicOrderCntLsb = (2<<8);
 103 static  unsigned int Log2MaxFrameNum = 16;
 104 static  unsigned int Log2MaxPicOrderCntLsb = 8;
 105
 106 static  unsigned int num_ref_frames = 2;
 107 static  unsigned int numShortTerm = 0;
 108 static  int constraint_set_flag = 0;
 109 static  int h264_packedheader = 0; /* support pack header? */
 110 static  int h264_maxref = (1<<16|1);
 111 static  int h264_entropy_mode = 1; /* cabac */
 112
 113 static  int frame_width = 176;
 114 static  int frame_height = 144;
 115 static  int frame_width_mbaligned;
 116 static  int frame_height_mbaligned;
 117 static  unsigned int frame_bitrate = 0;
 118 static  double frame_size = 0;
 119 static  int initial_qp = 15;
 120 //static  int initial_qp = 28;
 121 static  int minimal_qp = 0;
 122 static  int intra_period = 30;
 123 static  int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 124 static  int ip_period = 3;
 125 static  int rc_mode = -1;
 126 static  int rc_default_modes[] = {
 127     VA_RC_VBR,
 128     VA_RC_CQP,
 129     VA_RC_VBR_CONSTRAINED,
 130     VA_RC_CBR,
 131     VA_RC_VCM,
 132     VA_RC_NONE,
 133 };
 134 static  unsigned int current_frame_num = 0;
 135
 136 static  int misc_priv_type = 0;
 137 static  int misc_priv_value = 0;
 138
 139 /* thread to save coded data */
 140 #define SRC_SURFACE_FREE        0
 141 #define SRC_SURFACE_IN_ENCODING 1
 142
 143 struct __bitstream {
 144     unsigned int *buffer;
 145     int bit_offset;
 146     int max_size_in_dword;
 147 };
 148 typedef struct __bitstream bitstream;
 149
 150 using namespace std;
 151
 152 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 153 // but if we don't delete it here, we get leaks. The GStreamer implementation
 154 // does the same.
 155 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 156 {
 157     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 158     CHECK_VASTATUS(va_status, "vaRenderPicture");
 159
 160     for (int i = 0; i < num_buffers; ++i) {
 161         va_status = vaDestroyBuffer(va_dpy, buffers[i]);
 162         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 163     }
 164 }
 165
 166 static unsigned int
 167 va_swap32(unsigned int val)
 168 {
 169     unsigned char *pval = (unsigned char *)&val;
 170
 171     return ((pval[0] << 24)     |
 172             (pval[1] << 16)     |
 173             (pval[2] << 8)      |
 174             (pval[3] << 0));
 175 }
 176
 177 static void
 178 bitstream_start(bitstream *bs)
 179 {
 180     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 181     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 182     bs->bit_offset = 0;
 183 }
 184
 185 static void
 186 bitstream_end(bitstream *bs)
 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 (bit_offset) {
 193         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 194     }
 195 }
 196
 197 static void
 198 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 199 {
 200     int pos = (bs->bit_offset >> 5);
 201     int bit_offset = (bs->bit_offset & 0x1f);
 202     int bit_left = 32 - bit_offset;
 203
 204     if (!size_in_bits)
 205         return;
 206
 207     bs->bit_offset += size_in_bits;
 208
 209     if (bit_left > size_in_bits) {
 210         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 211     } else {
 212         size_in_bits -= bit_left;
 213         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 214         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 215
 216         if (pos + 1 == bs->max_size_in_dword) {
 217             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 218             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 219         }
 220
 221         bs->buffer[pos + 1] = val;
 222     }
 223 }
 224
 225 static void
 226 bitstream_put_ue(bitstream *bs, unsigned int val)
 227 {
 228     int size_in_bits = 0;
 229     int tmp_val = ++val;
 230
 231     while (tmp_val) {
 232         tmp_val >>= 1;
 233         size_in_bits++;
 234     }
 235
 236     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 237     bitstream_put_ui(bs, val, size_in_bits);
 238 }
 239
 240 static void
 241 bitstream_put_se(bitstream *bs, int val)
 242 {
 243     unsigned int new_val;
 244
 245     if (val <= 0)
 246         new_val = -2 * val;
 247     else
 248         new_val = 2 * val - 1;
 249
 250     bitstream_put_ue(bs, new_val);
 251 }
 252
 253 static void
 254 bitstream_byte_aligning(bitstream *bs, int bit)
 255 {
 256     int bit_offset = (bs->bit_offset & 0x7);
 257     int bit_left = 8 - bit_offset;
 258     int new_val;
 259
 260     if (!bit_offset)
 261         return;
 262
 263     assert(bit == 0 || bit == 1);
 264
 265     if (bit)
 266         new_val = (1 << bit_left) - 1;
 267     else
 268         new_val = 0;
 269
 270     bitstream_put_ui(bs, new_val, bit_left);
 271 }
 272
 273 static void
 274 rbsp_trailing_bits(bitstream *bs)
 275 {
 276     bitstream_put_ui(bs, 1, 1);
 277     bitstream_byte_aligning(bs, 0);
 278 }
 279
 280 static void nal_start_code_prefix(bitstream *bs)
 281 {
 282     bitstream_put_ui(bs, 0x00000001, 32);
 283 }
 284
 285 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 286 {
 287     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 288     bitstream_put_ui(bs, nal_ref_idc, 2);
 289     bitstream_put_ui(bs, nal_unit_type, 5);
 290 }
 291
 292 static void sps_rbsp(bitstream *bs)
 293 {
 294     int profile_idc = PROFILE_IDC_BASELINE;
 295
 296     if (h264_profile  == VAProfileH264High)
 297         profile_idc = PROFILE_IDC_HIGH;
 298     else if (h264_profile  == VAProfileH264Main)
 299         profile_idc = PROFILE_IDC_MAIN;
 300
 301     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 302     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 303     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 304     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 305     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 306     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 307     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 308     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 309
 310     if ( profile_idc == PROFILE_IDC_HIGH) {
 311         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 312         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 313         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 314         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 315         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 316     }
 317
 318     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 319     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 320
 321     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 322         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 323     else {
 324         assert(0);
 325     }
 326
 327     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 328     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 329
 330     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 331     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 332     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 333
 334     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 335         assert(0);
 336     }
 337
 338     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 339     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 340
 341     if (seq_param.frame_cropping_flag) {
 342         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 343         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 344         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 345         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 346     }
 347
 348     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 349     if ( false ) {
 350         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 351     } else {
 352         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 353         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 354         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 355         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 356         {
 357             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 358             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 359             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 360             {
 361                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 362                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 363                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 364             }
 365         }
 366         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 367         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 368         {
 369             bitstream_put_ui(bs, 1, 32);  // FPS
 370             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 371             bitstream_put_ui(bs, 1, 1);
 372         }
 373         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 374         {
 375             // hrd_parameters
 376             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 377             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 378             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 379
 380             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 381             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 382             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 383
 384             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 385             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 386             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 387             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 388         }
 389         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 390         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 391
 392         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 393         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 394     }
 395
 396     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 397 }
 398
 399
 400 static void pps_rbsp(bitstream *bs)
 401 {
 402     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 403     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 404
 405     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 406
 407     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 408
 409     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 410
 411     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 412     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 413
 414     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 415     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 416
 417     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 418     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 419     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 420
 421     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 422     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 423     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 424
 425     /* more_rbsp_data */
 426     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 427     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 428     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 429
 430     rbsp_trailing_bits(bs);
 431 }
 432
 433 static void slice_header(bitstream *bs)
 434 {
 435     int first_mb_in_slice = slice_param.macroblock_address;
 436
 437     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 438     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 439     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 440     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 441
 442     /* frame_mbs_only_flag == 1 */
 443     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 444         /* FIXME: */
 445         assert(0);
 446     }
 447
 448     if (pic_param.pic_fields.bits.idr_pic_flag)
 449         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 450
 451     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 452         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 453         /* pic_order_present_flag == 0 */
 454     } else {
 455         /* FIXME: */
 456         assert(0);
 457     }
 458
 459     /* redundant_pic_cnt_present_flag == 0 */
 460     /* slice type */
 461     if (IS_P_SLICE(slice_param.slice_type)) {
 462         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 463
 464         if (slice_param.num_ref_idx_active_override_flag)
 465             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 466
 467         /* ref_pic_list_reordering */
 468         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 469     } else if (IS_B_SLICE(slice_param.slice_type)) {
 470         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 471
 472         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 473
 474         if (slice_param.num_ref_idx_active_override_flag) {
 475             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 476             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 477         }
 478
 479         /* ref_pic_list_reordering */
 480         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 481         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 482     }
 483
 484     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 485          IS_P_SLICE(slice_param.slice_type)) ||
 486         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 487          IS_B_SLICE(slice_param.slice_type))) {
 488         /* FIXME: fill weight/offset table */
 489         assert(0);
 490     }
 491
 492     /* dec_ref_pic_marking */
 493     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 494         unsigned char no_output_of_prior_pics_flag = 0;
 495         unsigned char long_term_reference_flag = 0;
 496         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 497
 498         if (pic_param.pic_fields.bits.idr_pic_flag) {
 499             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 500             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 501         } else {
 502             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 503         }
 504     }
 505
 506     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 507         !IS_I_SLICE(slice_param.slice_type))
 508         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 509
 510     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 511
 512     /* ignore for SP/SI */
 513
 514     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 515         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 516
 517         if (slice_param.disable_deblocking_filter_idc != 1) {
 518             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 519             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 520         }
 521     }
 522
 523     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 524         bitstream_byte_aligning(bs, 1);
 525     }
 526 }
 527
 528 static int
 529 build_packed_pic_buffer(unsigned char **header_buffer)
 530 {
 531     bitstream bs;
 532
 533     bitstream_start(&bs);
 534     nal_start_code_prefix(&bs);
 535     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 536     pps_rbsp(&bs);
 537     bitstream_end(&bs);
 538
 539     *header_buffer = (unsigned char *)bs.buffer;
 540     return bs.bit_offset;
 541 }
 542
 543 static int
 544 build_packed_seq_buffer(unsigned char **header_buffer)
 545 {
 546     bitstream bs;
 547
 548     bitstream_start(&bs);
 549     nal_start_code_prefix(&bs);
 550     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 551     sps_rbsp(&bs);
 552     bitstream_end(&bs);
 553
 554     *header_buffer = (unsigned char *)bs.buffer;
 555     return bs.bit_offset;
 556 }
 557
 558 static int build_packed_slice_buffer(unsigned char **header_buffer)
 559 {
 560     bitstream bs;
 561     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 562     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 563
 564     bitstream_start(&bs);
 565     nal_start_code_prefix(&bs);
 566
 567     if (IS_I_SLICE(slice_param.slice_type)) {
 568         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 569     } else if (IS_P_SLICE(slice_param.slice_type)) {
 570         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 571     } else {
 572         assert(IS_B_SLICE(slice_param.slice_type));
 573         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 574     }
 575
 576     slice_header(&bs);
 577     bitstream_end(&bs);
 578
 579     *header_buffer = (unsigned char *)bs.buffer;
 580     return bs.bit_offset;
 581 }
 582
 583
 584 /*
 585   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 586   1) period between Frame #X and Frame #N = #X - #N
 587   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 588   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 589   4) intra_period and intra_idr_period take precedence over ip_period
 590   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 591      of I/IDR frames, see bellow examples
 592   -------------------------------------------------------------------
 593   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 594   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 595   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 596   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 597   1            1                ignored    IDR IDR IDR IDR...
 598   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 599   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 600   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 601                                               (PBB)(IBB)(PBB)(IBB)...
 602   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 603                                            IDRPPPPP IPPPPP IPPPPP...
 604   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 605                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 606                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 607                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 608                                            {IDR(PBB)(PBB)}                     (6/6/3)
 609                                            {IDR(PBB)(PBB)}.
 610 */
 611
 612 // General pts/dts strategy:
 613 //
 614 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 615 // bit tricky. We assume first of all that the frame rate never goes _above_
 616 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 617 // in at least this speed, as long at dts <= pts (the frame is not attempted
 618 // presented before it is decoded). Furthermore, we never have longer chains of
 619 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 620 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 621 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 622 // frame, with an order O <= C.)
 623 //
 624 // Many strategies are possible, but we establish these rules:
 625 //
 626 //  - Tip frames have dts = pts - (C-O)*N.
 627 //  - Non-tip frames have dts = dts_last + N.
 628 //
 629 // An example, with C=2 and N=10 and the data flow showed with arrows:
 630 //
 631 //        I  B  P  B  B  P
 632 //   pts: 30 40 50 60 70 80
 633 //        ↓  ↓     ↓
 634 //   dts: 10 30 20 60 50←40
 635 //         |  |  ↑        ↑
 636 //         `--|--'        |
 637 //             `----------'
 638 //
 639 // To show that this works fine also with irregular spacings, let's say that
 640 // the third frame is delayed a bit (something earlier was dropped). Now the
 641 // situation looks like this:
 642 //
 643 //        I  B  P  B  B   P
 644 //   pts: 30 40 80 90 100 110
 645 //        ↓  ↓     ↓
 646 //   dts: 10 30 20 90 50←40
 647 //         |  |  ↑        ↑
 648 //         `--|--'        |
 649 //             `----------'
 650 //
 651 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 652 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 653 //
 654 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 655 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 656 // a tip frame and should be given a dts based on the previous one.
 657 #define FRAME_P 0
 658 #define FRAME_B 1
 659 #define FRAME_I 2
 660 #define FRAME_IDR 7
 661 void encoding2display_order(
 662     int encoding_order, int intra_period,
 663     int intra_idr_period, int ip_period,
 664     int *displaying_order,
 665     int *frame_type, int *pts_lag)
 666 {
 667     int encoding_order_gop = 0;
 668
 669     *pts_lag = 0;
 670
 671     if (intra_period == 1) { /* all are I/IDR frames */
 672         *displaying_order = encoding_order;
 673         if (intra_idr_period == 0)
 674             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 675         else
 676             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 677         return;
 678     }
 679
 680     if (intra_period == 0)
 681         intra_idr_period = 0;
 682
 683     if (ip_period == 1) {
 684         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 685         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 686         *displaying_order = encoding_order;
 687
 688         if (encoding_order_gop == 0) { /* the first frame */
 689             *frame_type = FRAME_IDR;
 690         } else if (intra_period != 0 && /* have I frames */
 691                    encoding_order_gop >= 2 &&
 692                    (encoding_order_gop % intra_period == 0)) {
 693             *frame_type = FRAME_I;
 694         } else {
 695             *frame_type = FRAME_P;
 696         }
 697         return;
 698     }
 699
 700     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 701     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 702     *pts_lag = -1;  // Most frames are not tip frames.
 703
 704     if (encoding_order_gop == 0) { /* the first frame */
 705         *frame_type = FRAME_IDR;
 706         *displaying_order = encoding_order;
 707         // IDR frames are a special case; I honestly can't find the logic behind
 708         // why this is the right thing, but it seems to line up nicely in practice :-)
 709         *pts_lag = TIMEBASE / MAX_FPS;
 710     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 711         *frame_type = FRAME_B;
 712         *displaying_order = encoding_order - 1;
 713         if ((encoding_order_gop % ip_period) == 0) {
 714             *pts_lag = 0;  // Last B-frame.
 715         }
 716     } else if (intra_period != 0 && /* have I frames */
 717                encoding_order_gop >= 2 &&
 718                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 719         *frame_type = FRAME_I;
 720         *displaying_order = encoding_order + ip_period - 1;
 721     } else {
 722         *frame_type = FRAME_P;
 723         *displaying_order = encoding_order + ip_period - 1;
 724     }
 725 }
 726
 727
 728 static const char *rc_to_string(int rcmode)
 729 {
 730     switch (rc_mode) {
 731     case VA_RC_NONE:
 732         return "NONE";
 733     case VA_RC_CBR:
 734         return "CBR";
 735     case VA_RC_VBR:
 736         return "VBR";
 737     case VA_RC_VCM:
 738         return "VCM";
 739     case VA_RC_CQP:
 740         return "CQP";
 741     case VA_RC_VBR_CONSTRAINED:
 742         return "VBR_CONSTRAINED";
 743     default:
 744         return "Unknown";
 745     }
 746 }
 747
 748 #if 0
 749 static int process_cmdline(int argc, char *argv[])
 750 {
 751     char c;
 752     const struct option long_opts[] = {
 753         {"help", no_argument, NULL, 0 },
 754         {"bitrate", required_argument, NULL, 1 },
 755         {"minqp", required_argument, NULL, 2 },
 756         {"initialqp", required_argument, NULL, 3 },
 757         {"intra_period", required_argument, NULL, 4 },
 758         {"idr_period", required_argument, NULL, 5 },
 759         {"ip_period", required_argument, NULL, 6 },
 760         {"rcmode", required_argument, NULL, 7 },
 761         {"srcyuv", required_argument, NULL, 9 },
 762         {"recyuv", required_argument, NULL, 10 },
 763         {"fourcc", required_argument, NULL, 11 },
 764         {"syncmode", no_argument, NULL, 12 },
 765         {"enablePSNR", no_argument, NULL, 13 },
 766         {"prit", required_argument, NULL, 14 },
 767         {"priv", required_argument, NULL, 15 },
 768         {"framecount", required_argument, NULL, 16 },
 769         {"entropy", required_argument, NULL, 17 },
 770         {"profile", required_argument, NULL, 18 },
 771         {NULL, no_argument, NULL, 0 }};
 772     int long_index;
 773
 774     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 775         switch (c) {
 776         case 'w':
 777             frame_width = atoi(optarg);
 778             break;
 779         case 'h':
 780             frame_height = atoi(optarg);
 781             break;
 782         case 'n':
 783         case 'f':
 784             frame_rate = atoi(optarg);
 785             break;
 786         case 'o':
 787             coded_fn = strdup(optarg);
 788             break;
 789         case 0:
 790             print_help();
 791             exit(0);
 792         case 1:
 793             frame_bitrate = atoi(optarg);
 794             break;
 795         case 2:
 796             minimal_qp = atoi(optarg);
 797             break;
 798         case 3:
 799             initial_qp = atoi(optarg);
 800             break;
 801         case 4:
 802             intra_period = atoi(optarg);
 803             break;
 804         case 5:
 805             intra_idr_period = atoi(optarg);
 806             break;
 807         case 6:
 808             ip_period = atoi(optarg);
 809             break;
 810         case 7:
 811             rc_mode = string_to_rc(optarg);
 812             if (rc_mode < 0) {
 813                 print_help();
 814                 exit(1);
 815             }
 816             break;
 817         case 9:
 818             srcyuv_fn = strdup(optarg);
 819             break;
 820         case 11:
 821             srcyuv_fourcc = string_to_fourcc(optarg);
 822             if (srcyuv_fourcc <= 0) {
 823                 print_help();
 824                 exit(1);
 825             }
 826             break;
 827         case 13:
 828             calc_psnr = 1;
 829             break;
 830         case 14:
 831             misc_priv_type = strtol(optarg, NULL, 0);
 832             break;
 833         case 15:
 834             misc_priv_value = strtol(optarg, NULL, 0);
 835             break;
 836         case 17:
 837             h264_entropy_mode = atoi(optarg) ? 1: 0;
 838             break;
 839         case 18:
 840             if (strncmp(optarg, "BP", 2) == 0)
 841                 h264_profile = VAProfileH264Baseline;
 842             else if (strncmp(optarg, "MP", 2) == 0)
 843                 h264_profile = VAProfileH264Main;
 844             else if (strncmp(optarg, "HP", 2) == 0)
 845                 h264_profile = VAProfileH264High;
 846             else
 847                 h264_profile = (VAProfile)0;
 848             break;
 849         case ':':
 850         case '?':
 851             print_help();
 852             exit(0);
 853         }
 854     }
 855
 856     if (ip_period < 1) {
 857         printf(" ip_period must be greater than 0\n");
 858         exit(0);
 859     }
 860     if (intra_period != 1 && intra_period % ip_period != 0) {
 861         printf(" intra_period must be a multiplier of ip_period\n");
 862         exit(0);
 863     }
 864     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 865         printf(" intra_idr_period must be a multiplier of intra_period\n");
 866         exit(0);
 867     }
 868
 869     if (frame_bitrate == 0)
 870         frame_bitrate = frame_width * frame_height * 12 * MAX_FPS / 50;
 871
 872     if (coded_fn == NULL) {
 873         struct stat buf;
 874         if (stat("/tmp", &buf) == 0)
 875             coded_fn = strdup("/tmp/test.264");
 876         else if (stat("/sdcard", &buf) == 0)
 877             coded_fn = strdup("/sdcard/test.264");
 878         else
 879             coded_fn = strdup("./test.264");
 880     }
 881
 882
 883     frame_width_mbaligned = (frame_width + 15) & (~15);
 884     frame_height_mbaligned = (frame_height + 15) & (~15);
 885     if (frame_width != frame_width_mbaligned ||
 886         frame_height != frame_height_mbaligned) {
 887         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 888                frame_width, frame_height,
 889                frame_width_mbaligned, frame_height_mbaligned
 890                );
 891     }
 892
 893     return 0;
 894 }
 895 #endif
 896
 897 static Display *x11_display;
 898 static Window   x11_window;
 899
 900 VADisplay
 901 va_open_display(void)
 902 {
 903     x11_display = XOpenDisplay(NULL);
 904     if (!x11_display) {
 905         fprintf(stderr, "error: can't connect to X server!\n");
 906         return NULL;
 907     }
 908     return vaGetDisplay(x11_display);
 909 }
 910
 911 void
 912 va_close_display(VADisplay va_dpy)
 913 {
 914     if (!x11_display)
 915         return;
 916
 917     if (x11_window) {
 918         XUnmapWindow(x11_display, x11_window);
 919         XDestroyWindow(x11_display, x11_window);
 920         x11_window = None;
 921     }
 922     XCloseDisplay(x11_display);
 923     x11_display = NULL;
 924 }
 925
 926 static int init_va(void)
 927 {
 928     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 929     VAEntrypoint *entrypoints;
 930     int num_entrypoints, slice_entrypoint;
 931     int support_encode = 0;
 932     int major_ver, minor_ver;
 933     VAStatus va_status;
 934     unsigned int i;
 935
 936     va_dpy = va_open_display();
 937     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 938     CHECK_VASTATUS(va_status, "vaInitialize");
 939
 940     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 941     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 942     if (!entrypoints) {
 943         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 944         exit(1);
 945     }
 946
 947     /* use the highest profile */
 948     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 949         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 950             continue;
 951
 952         h264_profile = profile_list[i];
 953         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 954         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 955             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 956                 support_encode = 1;
 957                 break;
 958             }
 959         }
 960         if (support_encode == 1)
 961             break;
 962     }
 963
 964     if (support_encode == 0) {
 965         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 966         exit(1);
 967     } else {
 968         switch (h264_profile) {
 969             case VAProfileH264Baseline:
 970                 ip_period = 1;
 971                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 972                 h264_entropy_mode = 0;
 973                 break;
 974             case VAProfileH264ConstrainedBaseline:
 975                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 976                 ip_period = 1;
 977                 break;
 978
 979             case VAProfileH264Main:
 980                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 981                 break;
 982
 983             case VAProfileH264High:
 984                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 985                 break;
 986             default:
 987                 h264_profile = VAProfileH264Baseline;
 988                 ip_period = 1;
 989                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 990                 break;
 991         }
 992     }
 993
 994     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 995
 996     /* find out the format for the render target, and rate control mode */
 997     for (i = 0; i < VAConfigAttribTypeMax; i++)
 998         attrib[i].type = (VAConfigAttribType)i;
 999
1000     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1001                                       &attrib[0], VAConfigAttribTypeMax);
1002     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1003     /* check the interested configattrib */
1004     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1005         printf("Not find desired YUV420 RT format\n");
1006         exit(1);
1007     } else {
1008         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1009         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1010         config_attrib_num++;
1011     }
1012
1013     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1014         int tmp = attrib[VAConfigAttribRateControl].value;
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
1029         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1030         config_attrib[config_attrib_num].value = rc_mode;
1031         config_attrib_num++;
1032     }
1033
1034
1035     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1036         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1037
1038         h264_packedheader = 1;
1039         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1040         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1041
1042         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1043             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1044         }
1045
1046         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1047             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1048         }
1049
1050         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1051             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1052         }
1053
1054         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1055             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1056         }
1057
1058         enc_packed_header_idx = config_attrib_num;
1059         config_attrib_num++;
1060     }
1061
1062     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1063         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1064         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1065         config_attrib_num++;
1066     }
1067
1068     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1069         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1070     }
1071
1072     free(entrypoints);
1073     return 0;
1074 }
1075
1076 static int setup_encode()
1077 {
1078     VAStatus va_status;
1079     VASurfaceID *tmp_surfaceid;
1080     int codedbuf_size, i;
1081     static VASurfaceID src_surface[SURFACE_NUM];
1082     static VASurfaceID ref_surface[SURFACE_NUM];
1083
1084     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1085             &config_attrib[0], config_attrib_num, &config_id);
1086     CHECK_VASTATUS(va_status, "vaCreateConfig");
1087
1088     /* create source surfaces */
1089     va_status = vaCreateSurfaces(va_dpy,
1090                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1091                                  &src_surface[0], SURFACE_NUM,
1092                                  NULL, 0);
1093     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1094
1095     /* create reference surfaces */
1096     va_status = vaCreateSurfaces(va_dpy,
1097                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1098                                  &ref_surface[0], SURFACE_NUM,
1099                                  NULL, 0);
1100     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1101
1102     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1103     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1104     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1105
1106     /* Create a context for this encode pipe */
1107     va_status = vaCreateContext(va_dpy, config_id,
1108                                 frame_width_mbaligned, frame_height_mbaligned,
1109                                 VA_PROGRESSIVE,
1110                                 tmp_surfaceid, 2 * SURFACE_NUM,
1111                                 &context_id);
1112     CHECK_VASTATUS(va_status, "vaCreateContext");
1113     free(tmp_surfaceid);
1114
1115     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1116
1117     for (i = 0; i < SURFACE_NUM; i++) {
1118         /* create coded buffer once for all
1119          * other VA buffers which won't be used again after vaRenderPicture.
1120          * so APP can always vaCreateBuffer for every frame
1121          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1122          * so VA won't maintain the coded buffer
1123          */
1124         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1125                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1126         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1127     }
1128
1129     /* create OpenGL objects */
1130     //glGenFramebuffers(SURFACE_NUM, fbos);
1131
1132     for (i = 0; i < SURFACE_NUM; i++) {
1133         glGenTextures(1, &gl_surfaces[i].y_tex);
1134         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1135     }
1136
1137     for (i = 0; i < SURFACE_NUM; i++) {
1138         gl_surfaces[i].src_surface = src_surface[i];
1139         gl_surfaces[i].ref_surface = ref_surface[i];
1140     }
1141
1142     return 0;
1143 }
1144
1145
1146
1147 #define partition(ref, field, key, ascending)   \
1148     while (i <= j) {                            \
1149         if (ascending) {                        \
1150             while (ref[i].field < key)          \
1151                 i++;                            \
1152             while (ref[j].field > key)          \
1153                 j--;                            \
1154         } else {                                \
1155             while (ref[i].field > key)          \
1156                 i++;                            \
1157             while (ref[j].field < key)          \
1158                 j--;                            \
1159         }                                       \
1160         if (i <= j) {                           \
1161             tmp = ref[i];                       \
1162             ref[i] = ref[j];                    \
1163             ref[j] = tmp;                       \
1164             i++;                                \
1165             j--;                                \
1166         }                                       \
1167     }                                           \
1168
1169 static void sort_one(VAPictureH264 ref[], int left, int right,
1170                      int ascending, int frame_idx)
1171 {
1172     int i = left, j = right;
1173     unsigned int key;
1174     VAPictureH264 tmp;
1175
1176     if (frame_idx) {
1177         key = ref[(left + right) / 2].frame_idx;
1178         partition(ref, frame_idx, key, ascending);
1179     } else {
1180         key = ref[(left + right) / 2].TopFieldOrderCnt;
1181         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1182     }
1183
1184     /* recursion */
1185     if (left < j)
1186         sort_one(ref, left, j, ascending, frame_idx);
1187
1188     if (i < right)
1189         sort_one(ref, i, right, ascending, frame_idx);
1190 }
1191
1192 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1193                      int partition_ascending, int list0_ascending, int list1_ascending)
1194 {
1195     int i = left, j = right;
1196     VAPictureH264 tmp;
1197
1198     if (frame_idx) {
1199         partition(ref, frame_idx, key, partition_ascending);
1200     } else {
1201         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1202     }
1203
1204
1205     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1206     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1207 }
1208
1209 static void update_ReferenceFrames(int frame_type)
1210 {
1211     int i;
1212
1213     if (frame_type == FRAME_B)
1214         return;
1215
1216     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1217     numShortTerm++;
1218     if (numShortTerm > num_ref_frames)
1219         numShortTerm = num_ref_frames;
1220     for (i=numShortTerm-1; i>0; i--)
1221         ReferenceFrames[i] = ReferenceFrames[i-1];
1222     ReferenceFrames[0] = CurrentCurrPic;
1223
1224     current_frame_num++;
1225     if (current_frame_num > MaxFrameNum)
1226         current_frame_num = 0;
1227 }
1228
1229
1230 static int update_RefPicList(int frame_type)
1231 {
1232     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1233
1234     if (frame_type == FRAME_P) {
1235         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1236         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1237     }
1238
1239     if (frame_type == FRAME_B) {
1240         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1241         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1242                  1, 0, 1);
1243
1244         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1245         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1246                  0, 1, 0);
1247     }
1248
1249     return 0;
1250 }
1251
1252
1253 static int render_sequence(void)
1254 {
1255     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1256     VAStatus va_status;
1257     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1258     VAEncMiscParameterRateControl *misc_rate_ctrl;
1259
1260     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1261     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1262     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1263     seq_param.bits_per_second = frame_bitrate;
1264
1265     seq_param.intra_period = intra_period;
1266     seq_param.intra_idr_period = intra_idr_period;
1267     seq_param.ip_period = ip_period;
1268
1269     seq_param.max_num_ref_frames = num_ref_frames;
1270     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1271     seq_param.time_scale = TIMEBASE * 2;
1272     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1273     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1274     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1275     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1276     seq_param.seq_fields.bits.chroma_format_idc = 1;
1277     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1278
1279     if (frame_width != frame_width_mbaligned ||
1280         frame_height != frame_height_mbaligned) {
1281         seq_param.frame_cropping_flag = 1;
1282         seq_param.frame_crop_left_offset = 0;
1283         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1284         seq_param.frame_crop_top_offset = 0;
1285         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1286     }
1287
1288     va_status = vaCreateBuffer(va_dpy, context_id,
1289                                VAEncSequenceParameterBufferType,
1290                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1291     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1292
1293     va_status = vaCreateBuffer(va_dpy, context_id,
1294                                VAEncMiscParameterBufferType,
1295                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1296                                1, NULL, &rc_param_buf);
1297     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1298
1299     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1300     misc_param->type = VAEncMiscParameterTypeRateControl;
1301     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1302     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1303     misc_rate_ctrl->bits_per_second = frame_bitrate;
1304     misc_rate_ctrl->target_percentage = 66;
1305     misc_rate_ctrl->window_size = 1000;
1306     misc_rate_ctrl->initial_qp = initial_qp;
1307     misc_rate_ctrl->min_qp = minimal_qp;
1308     misc_rate_ctrl->basic_unit_size = 0;
1309     vaUnmapBuffer(va_dpy, rc_param_buf);
1310
1311     render_id[0] = seq_param_buf;
1312     render_id[1] = rc_param_buf;
1313
1314     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1315
1316     if (misc_priv_type != 0) {
1317         va_status = vaCreateBuffer(va_dpy, context_id,
1318                                    VAEncMiscParameterBufferType,
1319                                    sizeof(VAEncMiscParameterBuffer),
1320                                    1, NULL, &misc_param_tmpbuf);
1321         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1322         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1323         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1324         misc_param_tmp->data[0] = misc_priv_value;
1325         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1326
1327         render_picture_and_delete(va_dpy, context_id, &misc_param_tmpbuf, 1);
1328     }
1329
1330     return 0;
1331 }
1332
1333 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1334 {
1335     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1336     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1337     int PicOrderCntMsb, TopFieldOrderCnt;
1338
1339     if (frame_type == FRAME_IDR)
1340         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1341     else {
1342         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1343         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1344     }
1345
1346     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1347         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1348         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1349     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1350              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1351         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1352     else
1353         PicOrderCntMsb = prevPicOrderCntMsb;
1354
1355     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1356
1357     if (frame_type != FRAME_B) {
1358         PicOrderCntMsb_ref = PicOrderCntMsb;
1359         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1360     }
1361
1362     return TopFieldOrderCnt;
1363 }
1364
1365 static int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1366 {
1367     VABufferID pic_param_buf;
1368     VAStatus va_status;
1369     int i = 0;
1370
1371     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1372     pic_param.CurrPic.frame_idx = current_frame_num;
1373     pic_param.CurrPic.flags = 0;
1374     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1375     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1376     CurrentCurrPic = pic_param.CurrPic;
1377
1378     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1379     for (i = numShortTerm; i < SURFACE_NUM; i++) {
1380         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1381         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1382     }
1383
1384     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1385     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1386     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1387     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1388     pic_param.frame_num = current_frame_num;
1389     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1390     pic_param.last_picture = false;  // FIXME
1391     pic_param.pic_init_qp = initial_qp;
1392
1393     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1394                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1395     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1396
1397     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1398
1399     return 0;
1400 }
1401
1402 static int render_packedsequence(void)
1403 {
1404     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1405     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1406     unsigned int length_in_bits;
1407     unsigned char *packedseq_buffer = NULL;
1408     VAStatus va_status;
1409
1410     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1411
1412     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1413
1414     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1415     packedheader_param_buffer.has_emulation_bytes = 0;
1416     va_status = vaCreateBuffer(va_dpy,
1417                                context_id,
1418                                VAEncPackedHeaderParameterBufferType,
1419                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1420                                &packedseq_para_bufid);
1421     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1422
1423     va_status = vaCreateBuffer(va_dpy,
1424                                context_id,
1425                                VAEncPackedHeaderDataBufferType,
1426                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1427                                &packedseq_data_bufid);
1428     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1429
1430     render_id[0] = packedseq_para_bufid;
1431     render_id[1] = packedseq_data_bufid;
1432     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1433
1434     free(packedseq_buffer);
1435
1436     return 0;
1437 }
1438
1439
1440 static int render_packedpicture(void)
1441 {
1442     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1443     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1444     unsigned int length_in_bits;
1445     unsigned char *packedpic_buffer = NULL;
1446     VAStatus va_status;
1447
1448     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1449     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1450     packedheader_param_buffer.bit_length = length_in_bits;
1451     packedheader_param_buffer.has_emulation_bytes = 0;
1452
1453     va_status = vaCreateBuffer(va_dpy,
1454                                context_id,
1455                                VAEncPackedHeaderParameterBufferType,
1456                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1457                                &packedpic_para_bufid);
1458     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1459
1460     va_status = vaCreateBuffer(va_dpy,
1461                                context_id,
1462                                VAEncPackedHeaderDataBufferType,
1463                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1464                                &packedpic_data_bufid);
1465     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1466
1467     render_id[0] = packedpic_para_bufid;
1468     render_id[1] = packedpic_data_bufid;
1469     render_picture_and_delete(va_dpy, context_id, render_id, 2);
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     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1506
1507     free(packedslice_buffer);
1508 }
1509
1510 static int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1511 {
1512     VABufferID slice_param_buf;
1513     VAStatus va_status;
1514     int i;
1515
1516     update_RefPicList(frame_type);
1517
1518     /* one frame, one slice */
1519     slice_param.macroblock_address = 0;
1520     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1521     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1522     if (frame_type == FRAME_IDR) {
1523         if (encoding_frame_num != 0)
1524             ++slice_param.idr_pic_id;
1525     } else if (frame_type == FRAME_P) {
1526         int refpiclist0_max = h264_maxref & 0xffff;
1527         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1528
1529         for (i = refpiclist0_max; i < 32; i++) {
1530             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1531             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1532         }
1533     } else if (frame_type == FRAME_B) {
1534         int refpiclist0_max = h264_maxref & 0xffff;
1535         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1536
1537         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1538         for (i = refpiclist0_max; i < 32; i++) {
1539             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1540             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1541         }
1542
1543         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1544         for (i = refpiclist1_max; i < 32; i++) {
1545             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1546             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1547         }
1548     }
1549
1550     slice_param.slice_alpha_c0_offset_div2 = 0;
1551     slice_param.slice_beta_offset_div2 = 0;
1552     slice_param.direct_spatial_mv_pred_flag = 1;
1553     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1554
1555
1556     if (h264_packedheader &&
1557         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1558         render_packedslice();
1559
1560     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1561                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1562     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1563
1564     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1565
1566     return 0;
1567 }
1568
1569
1570
1571 void H264Encoder::save_codeddata(storage_task task)
1572 {
1573     VACodedBufferSegment *buf_list = NULL;
1574     VAStatus va_status;
1575     unsigned int coded_size = 0;
1576
1577     string data;
1578
1579     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1580
1581     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.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         buf_list = (VACodedBufferSegment *) buf_list->next;
1586
1587         frame_size += coded_size;
1588     }
1589     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1590
1591     {
1592         // Add video.
1593         AVPacket pkt;
1594         memset(&pkt, 0, sizeof(pkt));
1595         pkt.buf = nullptr;
1596         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1597         pkt.size = data.size();
1598         pkt.stream_index = 0;
1599         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1600             pkt.flags = AV_PKT_FLAG_KEY;
1601         } else {
1602             pkt.flags = 0;
1603         }
1604         //pkt.duration = 1;
1605         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1606     }
1607     // Encode and add all audio frames up to and including the pts of this video frame.
1608     for ( ;; ) {
1609         int64_t audio_pts;
1610         vector<float> audio;
1611         {
1612              unique_lock<mutex> lock(frame_queue_mutex);
1613              frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1614              if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1615              auto it = pending_audio_frames.begin();
1616              if (it->first > task.pts) break;
1617              audio_pts = it->first;
1618              audio = move(it->second);
1619              pending_audio_frames.erase(it);
1620         }
1621
1622         AVFrame *frame = avcodec_alloc_frame();
1623         frame->nb_samples = audio.size() / 2;
1624         frame->format = AV_SAMPLE_FMT_S32;
1625         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1626
1627         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1628         int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1629         if (ret < 0) {
1630             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1631             exit(1);
1632         }
1633         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1634             if (audio[i] >= 1.0f) {
1635                 int_samples[i] = 2147483647;
1636             } else if (audio[i] <= -1.0f) {
1637                 int_samples[i] = -2147483647;
1638             } else {
1639                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1640             }
1641         }
1642
1643         AVPacket pkt;
1644         av_init_packet(&pkt);
1645         pkt.data = nullptr;
1646         pkt.size = 0;
1647         int got_output;
1648         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1649         if (got_output) {
1650             pkt.stream_index = 1;
1651             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1652         }
1653         // TODO: Delayed frames.
1654         avcodec_free_frame(&frame);
1655         av_free_packet(&pkt);
1656         if (audio_pts == task.pts) break;
1657     }
1658
1659 #if 0
1660     printf("\r      "); /* return back to startpoint */
1661     switch (encode_order % 4) {
1662         case 0:
1663             printf("|");
1664             break;
1665         case 1:
1666             printf("/");
1667             break;
1668         case 2:
1669             printf("-");
1670             break;
1671         case 3:
1672             printf("\\");
1673             break;
1674     }
1675     printf("%08lld", encode_order);
1676     printf("(%06d bytes coded)", coded_size);
1677 #endif
1678 }
1679
1680
1681 // this is weird. but it seems to put a new frame onto the queue
1682 void H264Encoder::storage_task_enqueue(storage_task task)
1683 {
1684         unique_lock<mutex> lock(storage_task_queue_mutex);
1685         storage_task_queue.push(move(task));
1686         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1687         storage_task_queue_changed.notify_all();
1688 }
1689
1690 void H264Encoder::storage_task_thread()
1691 {
1692         for ( ;; ) {
1693                 storage_task current;
1694                 {
1695                         // wait until there's an encoded frame
1696                         unique_lock<mutex> lock(storage_task_queue_mutex);
1697                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1698                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1699                         current = move(storage_task_queue.front());
1700                         storage_task_queue.pop();
1701                 }
1702
1703                 VAStatus va_status;
1704
1705                 // waits for data, then saves it to disk.
1706                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1707                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1708                 save_codeddata(move(current));
1709
1710                 {
1711                         unique_lock<mutex> lock(storage_task_queue_mutex);
1712                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1713                         storage_task_queue_changed.notify_all();
1714                 }
1715         }
1716 }
1717
1718 static int release_encode()
1719 {
1720     int i;
1721
1722     for (i = 0; i < SURFACE_NUM; i++) {
1723         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1724         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1725         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1726     }
1727
1728     vaDestroyContext(va_dpy, context_id);
1729     vaDestroyConfig(va_dpy, config_id);
1730
1731     return 0;
1732 }
1733
1734 static int deinit_va()
1735 {
1736     vaTerminate(va_dpy);
1737
1738     va_close_display(va_dpy);
1739
1740     return 0;
1741 }
1742
1743
1744 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1745         : current_storage_frame(0), surface(surface), httpd(httpd)
1746 {
1747         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1748         context_audio = avcodec_alloc_context3(codec_audio);
1749         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1750         context_audio->sample_rate = OUTPUT_FREQUENCY;
1751         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1752         context_audio->channels = 2;
1753         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1754         context_audio->time_base = AVRational{1, TIMEBASE};
1755         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1756                 fprintf(stderr, "Could not open codec\n");
1757                 exit(1);
1758         }
1759
1760         frame_width = width;
1761         frame_height = height;
1762         frame_width_mbaligned = (frame_width + 15) & (~15);
1763         frame_height_mbaligned = (frame_height + 15) & (~15);
1764         frame_bitrate = 15000000;  // / 60;
1765
1766         //print_input();
1767
1768         init_va();
1769         setup_encode();
1770
1771         // No frames are ready yet.
1772         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1773
1774         memset(&seq_param, 0, sizeof(seq_param));
1775         memset(&pic_param, 0, sizeof(pic_param));
1776         memset(&slice_param, 0, sizeof(slice_param));
1777
1778         storage_thread = thread(&H264Encoder::storage_task_thread, this);
1779
1780         encode_thread = thread([this]{
1781                 //SDL_GL_MakeCurrent(window, context);
1782                 QOpenGLContext *context = create_context(this->surface);
1783                 eglBindAPI(EGL_OPENGL_API);
1784                 if (!make_current(context, this->surface)) {
1785                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1786                                 eglGetError());
1787                         exit(1);
1788                 }
1789                 encode_thread_func();
1790         });
1791 }
1792
1793 H264Encoder::~H264Encoder()
1794 {
1795         {
1796                 unique_lock<mutex> lock(frame_queue_mutex);
1797                 encode_thread_should_quit = true;
1798                 frame_queue_nonempty.notify_all();
1799         }
1800         encode_thread.join();
1801         {
1802                 unique_lock<mutex> lock(storage_task_queue_mutex);
1803                 storage_thread_should_quit = true;
1804                 frame_queue_nonempty.notify_all();
1805                 storage_task_queue_changed.notify_all();
1806         }
1807         storage_thread.join();
1808
1809         release_encode();
1810         deinit_va();
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                 unique_lock<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::add_audio(int64_t pts, vector<float> audio)
1878 {
1879         {
1880                 unique_lock<mutex> lock(frame_queue_mutex);
1881                 pending_audio_frames[pts] = move(audio);
1882         }
1883         frame_queue_nonempty.notify_all();
1884 }
1885
1886 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1887 {
1888         {
1889                 unique_lock<mutex> lock(frame_queue_mutex);
1890                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1891                 ++current_storage_frame;
1892         }
1893         frame_queue_nonempty.notify_all();
1894 }
1895
1896 void H264Encoder::encode_thread_func()
1897 {
1898         int64_t last_dts = -1;
1899         int gop_start_display_frame_num = 0;
1900         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1901                 PendingFrame frame;
1902                 int pts_lag;
1903                 int frame_type, display_frame_num;
1904                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1905                                        &display_frame_num, &frame_type, &pts_lag);
1906                 if (frame_type == FRAME_IDR) {
1907                         numShortTerm = 0;
1908                         current_frame_num = 0;
1909                         gop_start_display_frame_num = display_frame_num;
1910                 }
1911
1912                 {
1913                         unique_lock<mutex> lock(frame_queue_mutex);
1914                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1915                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1916                         });
1917                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1918                                 // We have queued frames that were supposed to be B-frames,
1919                                 // but will be no P-frame to encode them against. Encode them all
1920                                 // as P-frames instead. Note that this happens under the mutex,
1921                                 // but nobody else uses it at this point, since we're shutting down,
1922                                 // so there's no contention.
1923                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1924                                 return;
1925                         } else {
1926                                 frame = move(pending_video_frames[display_frame_num]);
1927                                 pending_video_frames.erase(display_frame_num);
1928                         }
1929                 }
1930
1931                 // Determine the dts of this frame.
1932                 int64_t dts;
1933                 if (pts_lag == -1) {
1934                         assert(last_dts != -1);
1935                         dts = last_dts + (TIMEBASE / MAX_FPS);
1936                 } else {
1937                         dts = frame.pts - pts_lag;
1938                 }
1939                 last_dts = dts;
1940
1941                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
1942         }
1943 }
1944
1945 void H264Encoder::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1946 {
1947         if (pending_video_frames.empty()) {
1948                 return;
1949         }
1950
1951         for (auto &pending_frame : pending_video_frames) {
1952                 int display_frame_num = pending_frame.first;
1953                 assert(display_frame_num > 0);
1954                 PendingFrame frame = move(pending_frame.second);
1955                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1956                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1957                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
1958                 last_dts = dts;
1959         }
1960 }
1961
1962 void H264Encoder::encode_frame(H264Encoder::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1963                                int frame_type, int64_t pts, int64_t dts)
1964 {
1965         // Wait for the GPU to be done with the frame.
1966         glClientWaitSync(frame.fence.get(), 0, 0);
1967
1968         // Release back any input frames we needed to render this frame.
1969         frame.input_frames.clear();
1970
1971         // Unmap the image.
1972         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1973         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1974         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1975         VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1976         CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1977         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1978         CHECK_VASTATUS(va_status, "vaDestroyImage");
1979
1980         VASurfaceID surface = surf->src_surface;
1981
1982         // Schedule the frame for encoding.
1983         va_status = vaBeginPicture(va_dpy, context_id, surface);
1984         CHECK_VASTATUS(va_status, "vaBeginPicture");
1985
1986         if (frame_type == FRAME_IDR) {
1987                 render_sequence();
1988                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1989                 if (h264_packedheader) {
1990                         render_packedsequence();
1991                         render_packedpicture();
1992                 }
1993         } else {
1994                 //render_sequence();
1995                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1996         }
1997         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1998
1999         va_status = vaEndPicture(va_dpy, context_id);
2000         CHECK_VASTATUS(va_status, "vaEndPicture");
2001
2002         // so now the data is done encoding (well, async job kicked off)...
2003         // we send that to the storage thread
2004         storage_task tmp;
2005         tmp.display_order = display_frame_num;
2006         tmp.frame_type = frame_type;
2007         tmp.pts = pts;
2008         tmp.dts = dts;
2009         storage_task_enqueue(move(tmp));
2010
2011         update_ReferenceFrames(frame_type);
2012 }