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

   1 #include "quicksync_encoder.h"
   2
   3 #include <movit/image_format.h>
   4 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
   5 #include <movit/util.h>
   6
   7 #include <EGL/eglplatform.h>
   8 #include <X11/Xlib.h>
   9 #include <assert.h>
  10 #include <epoxy/egl.h>
  11 #include <fcntl.h>
  12 #include <pthread.h>
  13 #include <stdio.h>
  14 #include <stdlib.h>
  15 #include <string.h>
  16 #include <unistd.h>
  17 #include <va/va.h>
  18 #include <va/va_drm.h>
  19 #include <va/va_drmcommon.h>
  20 #include <va/va_enc_h264.h>
  21 #include <va/va_x11.h>
  22 #include <algorithm>
  23 #include <chrono>
  24 #include <condition_variable>
  25 #include <cstddef>
  26 #include <cstdint>
  27 #include <functional>
  28 #include <map>
  29 #include <memory>
  30 #include <mutex>
  31 #include <queue>
  32 #include <stack>
  33 #include <string>
  34 #include <thread>
  35 #include <utility>
  36
  37 extern "C" {
  38
  39 #include <libavcodec/avcodec.h>
  40 #include <libavformat/avio.h>
  41 #include <libavutil/error.h>
  42 #include <libdrm/drm_fourcc.h>
  43
  44 }  // namespace
  45
  46 #include "audio_encoder.h"
  47 #include "context.h"
  48 #include "defs.h"
  49 #include "disk_space_estimator.h"
  50 #include "ffmpeg_raii.h"
  51 #include "flags.h"
  52 #include "mux.h"
  53 #include "print_latency.h"
  54 #include "quicksync_encoder_impl.h"
  55 #include "ref_counted_frame.h"
  56 #include "timebase.h"
  57 #include "x264_encoder.h"
  58
  59 using namespace movit;
  60 using namespace std;
  61 using namespace std::chrono;
  62 using namespace std::placeholders;
  63
  64 class QOpenGLContext;
  65 class QSurface;
  66
  67 #define CHECK_VASTATUS(va_status, func)                                 \
  68     if (va_status != VA_STATUS_SUCCESS) {                               \
  69         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  70         exit(1);                                                        \
  71     }
  72
  73 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  74
  75 //#include "loadsurface.h"
  76
  77 #define NAL_REF_IDC_NONE        0
  78 #define NAL_REF_IDC_LOW         1
  79 #define NAL_REF_IDC_MEDIUM      2
  80 #define NAL_REF_IDC_HIGH        3
  81
  82 #define NAL_NON_IDR             1
  83 #define NAL_IDR                 5
  84 #define NAL_SPS                 7
  85 #define NAL_PPS                 8
  86 #define NAL_SEI                 6
  87
  88 #define SLICE_TYPE_P            0
  89 #define SLICE_TYPE_B            1
  90 #define SLICE_TYPE_I            2
  91 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  92 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  93 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  94
  95
  96 #define ENTROPY_MODE_CAVLC      0
  97 #define ENTROPY_MODE_CABAC      1
  98
  99 #define PROFILE_IDC_BASELINE    66
 100 #define PROFILE_IDC_MAIN        77
 101 #define PROFILE_IDC_HIGH        100
 102
 103 #define BITSTREAM_ALLOCATE_STEPPING     4096
 104
 105 static constexpr unsigned int MaxFrameNum = (2<<16);
 106 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
 107 static constexpr unsigned int Log2MaxFrameNum = 16;
 108 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
 109 static constexpr int rc_default_modes[] = {  // Priority list of modes.
 110     VA_RC_VBR,
 111     VA_RC_CQP,
 112     VA_RC_VBR_CONSTRAINED,
 113     VA_RC_CBR,
 114     VA_RC_VCM,
 115     VA_RC_NONE,
 116 };
 117
 118 using namespace std;
 119
 120 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 121 // but if we don't delete it here, we get leaks. The GStreamer implementation
 122 // does the same.
 123 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 124 {
 125     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 126     CHECK_VASTATUS(va_status, "vaRenderPicture");
 127
 128     for (int i = 0; i < num_buffers; ++i) {
 129         va_status = vaDestroyBuffer(dpy, buffers[i]);
 130         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 131     }
 132 }
 133
 134 static unsigned int
 135 va_swap32(unsigned int val)
 136 {
 137     unsigned char *pval = (unsigned char *)&val;
 138
 139     return ((pval[0] << 24)     |
 140             (pval[1] << 16)     |
 141             (pval[2] << 8)      |
 142             (pval[3] << 0));
 143 }
 144
 145 static void
 146 bitstream_start(bitstream *bs)
 147 {
 148     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 149     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 150     bs->bit_offset = 0;
 151 }
 152
 153 static void
 154 bitstream_end(bitstream *bs)
 155 {
 156     int pos = (bs->bit_offset >> 5);
 157     int bit_offset = (bs->bit_offset & 0x1f);
 158     int bit_left = 32 - bit_offset;
 159
 160     if (bit_offset) {
 161         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 162     }
 163 }
 164
 165 static void
 166 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 167 {
 168     int pos = (bs->bit_offset >> 5);
 169     int bit_offset = (bs->bit_offset & 0x1f);
 170     int bit_left = 32 - bit_offset;
 171
 172     if (!size_in_bits)
 173         return;
 174
 175     bs->bit_offset += size_in_bits;
 176
 177     if (bit_left > size_in_bits) {
 178         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 179     } else {
 180         size_in_bits -= bit_left;
 181         if (bit_left >= 32) {
 182             bs->buffer[pos] = (val >> size_in_bits);
 183         } else {
 184             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 185         }
 186         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 187
 188         if (pos + 1 == bs->max_size_in_dword) {
 189             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 190             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 191         }
 192
 193         bs->buffer[pos + 1] = val;
 194     }
 195 }
 196
 197 static void
 198 bitstream_put_ue(bitstream *bs, unsigned int val)
 199 {
 200     int size_in_bits = 0;
 201     int tmp_val = ++val;
 202
 203     while (tmp_val) {
 204         tmp_val >>= 1;
 205         size_in_bits++;
 206     }
 207
 208     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 209     bitstream_put_ui(bs, val, size_in_bits);
 210 }
 211
 212 static void
 213 bitstream_put_se(bitstream *bs, int val)
 214 {
 215     unsigned int new_val;
 216
 217     if (val <= 0)
 218         new_val = -2 * val;
 219     else
 220         new_val = 2 * val - 1;
 221
 222     bitstream_put_ue(bs, new_val);
 223 }
 224
 225 static void
 226 bitstream_byte_aligning(bitstream *bs, int bit)
 227 {
 228     int bit_offset = (bs->bit_offset & 0x7);
 229     int bit_left = 8 - bit_offset;
 230     int new_val;
 231
 232     if (!bit_offset)
 233         return;
 234
 235     assert(bit == 0 || bit == 1);
 236
 237     if (bit)
 238         new_val = (1 << bit_left) - 1;
 239     else
 240         new_val = 0;
 241
 242     bitstream_put_ui(bs, new_val, bit_left);
 243 }
 244
 245 static void
 246 rbsp_trailing_bits(bitstream *bs)
 247 {
 248     bitstream_put_ui(bs, 1, 1);
 249     bitstream_byte_aligning(bs, 0);
 250 }
 251
 252 static void nal_start_code_prefix(bitstream *bs)
 253 {
 254     bitstream_put_ui(bs, 0x00000001, 32);
 255 }
 256
 257 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 258 {
 259     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 260     bitstream_put_ui(bs, nal_ref_idc, 2);
 261     bitstream_put_ui(bs, nal_unit_type, 5);
 262 }
 263
 264 void QuickSyncEncoderImpl::sps_rbsp(YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs)
 265 {
 266     int profile_idc = PROFILE_IDC_BASELINE;
 267
 268     if (h264_profile  == VAProfileH264High)
 269         profile_idc = PROFILE_IDC_HIGH;
 270     else if (h264_profile  == VAProfileH264Main)
 271         profile_idc = PROFILE_IDC_MAIN;
 272
 273     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 274     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 275     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 276     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 277     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 278     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 279     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 280     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 281
 282     if ( profile_idc == PROFILE_IDC_HIGH) {
 283         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 284         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 285         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 286         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 287         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 288     }
 289
 290     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 291     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 292
 293     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 294         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 295     else {
 296         assert(0);
 297     }
 298
 299     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 300     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 301
 302     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 303     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 304     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 305
 306     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 307         assert(0);
 308     }
 309
 310     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 311     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 312
 313     if (seq_param.frame_cropping_flag) {
 314         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 315         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 316         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 317         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 318     }
 319
 320     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 321     if ( false ) {
 322         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 323     } else {
 324         // See H.264 annex E for the definition of this header.
 325         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 326         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 327         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 328         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 329         {
 330             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 331             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 332             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 333             {
 334                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 335                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 336                 if (ycbcr_coefficients == YCBCR_REC_709) {
 337                     bitstream_put_ui(bs, 1, 8);  /* matrix_coefficients (1 = BT.709) */
 338                 } else {
 339                     assert(ycbcr_coefficients == YCBCR_REC_601);
 340                     bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 341                 }
 342             }
 343         }
 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 void QuickSyncEncoderImpl::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 void QuickSyncEncoderImpl::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 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 507 {
 508     bitstream bs;
 509
 510     bitstream_start(&bs);
 511     nal_start_code_prefix(&bs);
 512     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 513     pps_rbsp(&bs);
 514     bitstream_end(&bs);
 515
 516     *header_buffer = (unsigned char *)bs.buffer;
 517     return bs.bit_offset;
 518 }
 519
 520 int
 521 QuickSyncEncoderImpl::build_packed_seq_buffer(YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer)
 522 {
 523     bitstream bs;
 524
 525     bitstream_start(&bs);
 526     nal_start_code_prefix(&bs);
 527     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 528     sps_rbsp(ycbcr_coefficients, &bs);
 529     bitstream_end(&bs);
 530
 531     *header_buffer = (unsigned char *)bs.buffer;
 532     return bs.bit_offset;
 533 }
 534
 535 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 536 {
 537     bitstream bs;
 538     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 539     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 540
 541     bitstream_start(&bs);
 542     nal_start_code_prefix(&bs);
 543
 544     if (IS_I_SLICE(slice_param.slice_type)) {
 545         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 546     } else if (IS_P_SLICE(slice_param.slice_type)) {
 547         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 548     } else {
 549         assert(IS_B_SLICE(slice_param.slice_type));
 550         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 551     }
 552
 553     slice_header(&bs);
 554     bitstream_end(&bs);
 555
 556     *header_buffer = (unsigned char *)bs.buffer;
 557     return bs.bit_offset;
 558 }
 559
 560
 561 /*
 562   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 563   1) period between Frame #X and Frame #N = #X - #N
 564   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 565   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 566   4) intra_period and intra_idr_period take precedence over ip_period
 567   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 568      of I/IDR frames, see bellow examples
 569   -------------------------------------------------------------------
 570   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 571   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 572   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 573   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 574   1            1                ignored    IDR IDR IDR IDR...
 575   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 576   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 577   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 578                                               (PBB)(IBB)(PBB)(IBB)...
 579   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 580                                            IDRPPPPP IPPPPP IPPPPP...
 581   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 582                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 583                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 584                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 585                                            {IDR(PBB)(PBB)}                     (6/6/3)
 586                                            {IDR(PBB)(PBB)}.
 587 */
 588
 589 // General pts/dts strategy:
 590 //
 591 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 592 // bit tricky. We assume first of all that the frame rate never goes _above_
 593 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 594 // in at least this speed, as long at dts <= pts (the frame is not attempted
 595 // presented before it is decoded). Furthermore, we never have longer chains of
 596 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 597 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 598 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 599 // frame, with an order O <= C.)
 600 //
 601 // Many strategies are possible, but we establish these rules:
 602 //
 603 //  - Tip frames have dts = pts - (C-O)*N.
 604 //  - Non-tip frames have dts = dts_last + N.
 605 //
 606 // An example, with C=2 and N=10 and the data flow showed with arrows:
 607 //
 608 //        I  B  P  B  B  P
 609 //   pts: 30 40 50 60 70 80
 610 //        ↓  ↓     ↓
 611 //   dts: 10 30 20 60 50←40
 612 //         |  |  ↑        ↑
 613 //         `--|--'        |
 614 //             `----------'
 615 //
 616 // To show that this works fine also with irregular spacings, let's say that
 617 // the third frame is delayed a bit (something earlier was dropped). Now the
 618 // situation looks like this:
 619 //
 620 //        I  B  P  B  B   P
 621 //   pts: 30 40 80 90 100 110
 622 //        ↓  ↓     ↓
 623 //   dts: 10 30 20 90 50←40
 624 //         |  |  ↑        ↑
 625 //         `--|--'        |
 626 //             `----------'
 627 //
 628 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 629 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 630 //
 631 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 632 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 633 // a tip frame and should be given a dts based on the previous one.
 634 #define FRAME_P 0
 635 #define FRAME_B 1
 636 #define FRAME_I 2
 637 #define FRAME_IDR 7
 638 void encoding2display_order(
 639     int encoding_order, int intra_period,
 640     int intra_idr_period, int ip_period,
 641     int *displaying_order,
 642     int *frame_type, int *pts_lag)
 643 {
 644     int encoding_order_gop = 0;
 645
 646     *pts_lag = 0;
 647
 648     if (intra_period == 1) { /* all are I/IDR frames */
 649         *displaying_order = encoding_order;
 650         if (intra_idr_period == 0)
 651             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 652         else
 653             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 654         return;
 655     }
 656
 657     if (intra_period == 0)
 658         intra_idr_period = 0;
 659
 660     if (ip_period == 1) {
 661         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 662         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 663         *displaying_order = encoding_order;
 664
 665         if (encoding_order_gop == 0) { /* the first frame */
 666             *frame_type = FRAME_IDR;
 667         } else if (intra_period != 0 && /* have I frames */
 668                    encoding_order_gop >= 2 &&
 669                    (encoding_order_gop % intra_period == 0)) {
 670             *frame_type = FRAME_I;
 671         } else {
 672             *frame_type = FRAME_P;
 673         }
 674         return;
 675     }
 676
 677     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 678     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 679     *pts_lag = -1;  // Most frames are not tip frames.
 680
 681     if (encoding_order_gop == 0) { /* the first frame */
 682         *frame_type = FRAME_IDR;
 683         *displaying_order = encoding_order;
 684         // IDR frames are a special case; I honestly can't find the logic behind
 685         // why this is the right thing, but it seems to line up nicely in practice :-)
 686         *pts_lag = TIMEBASE / MAX_FPS;
 687     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 688         *frame_type = FRAME_B;
 689         *displaying_order = encoding_order - 1;
 690         if ((encoding_order_gop % ip_period) == 0) {
 691             *pts_lag = 0;  // Last B-frame.
 692         }
 693     } else if (intra_period != 0 && /* have I frames */
 694                encoding_order_gop >= 2 &&
 695                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 696         *frame_type = FRAME_I;
 697         *displaying_order = encoding_order + ip_period - 1;
 698     } else {
 699         *frame_type = FRAME_P;
 700         *displaying_order = encoding_order + ip_period - 1;
 701     }
 702 }
 703
 704
 705 static const char *rc_to_string(int rc_mode)
 706 {
 707     switch (rc_mode) {
 708     case VA_RC_NONE:
 709         return "NONE";
 710     case VA_RC_CBR:
 711         return "CBR";
 712     case VA_RC_VBR:
 713         return "VBR";
 714     case VA_RC_VCM:
 715         return "VCM";
 716     case VA_RC_CQP:
 717         return "CQP";
 718     case VA_RC_VBR_CONSTRAINED:
 719         return "VBR_CONSTRAINED";
 720     default:
 721         return "Unknown";
 722     }
 723 }
 724
 725 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 726 {
 727         if (global_flags.x264_video_to_disk) {
 728                 // Quick Sync is entirely disabled.
 729                 use_zerocopy = false;
 730         } else if (global_flags.uncompressed_video_to_http) {
 731                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 732                 use_zerocopy = false;
 733         } else if (global_flags.x264_video_to_http) {
 734                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 735                 use_zerocopy = false;
 736         } else {
 737                 use_zerocopy = true;
 738         }
 739         global_flags.use_zerocopy = use_zerocopy;
 740 }
 741
 742 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 743 {
 744         if (va_display.empty()) {
 745                 x11_display = XOpenDisplay(NULL);
 746                 if (!x11_display) {
 747                         fprintf(stderr, "error: can't connect to X server!\n");
 748                         return NULL;
 749                 }
 750                 return vaGetDisplay(x11_display);
 751         } else if (va_display[0] != '/') {
 752                 x11_display = XOpenDisplay(va_display.c_str());
 753                 if (!x11_display) {
 754                         fprintf(stderr, "error: can't connect to X server!\n");
 755                         return NULL;
 756                 }
 757                 return vaGetDisplay(x11_display);
 758         } else {
 759                 drm_fd = open(va_display.c_str(), O_RDWR);
 760                 if (drm_fd == -1) {
 761                         perror(va_display.c_str());
 762                         return NULL;
 763                 }
 764                 use_zerocopy = false;
 765                 return vaGetDisplayDRM(drm_fd);
 766         }
 767 }
 768
 769 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
 770 {
 771         if (x11_display) {
 772                 XCloseDisplay(x11_display);
 773                 x11_display = nullptr;
 774         }
 775         if (drm_fd != -1) {
 776                 close(drm_fd);
 777         }
 778 }
 779
 780 int QuickSyncEncoderImpl::init_va(const string &va_display)
 781 {
 782     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 783     VAEntrypoint *entrypoints;
 784     int num_entrypoints, slice_entrypoint;
 785     int support_encode = 0;
 786     int major_ver, minor_ver;
 787     VAStatus va_status;
 788     unsigned int i;
 789
 790     va_dpy = va_open_display(va_display);
 791     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 792     CHECK_VASTATUS(va_status, "vaInitialize");
 793
 794     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 795     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 796     if (!entrypoints) {
 797         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 798         exit(1);
 799     }
 800
 801     /* use the highest profile */
 802     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 803         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 804             continue;
 805
 806         h264_profile = profile_list[i];
 807         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 808         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 809             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 810                 support_encode = 1;
 811                 break;
 812             }
 813         }
 814         if (support_encode == 1)
 815             break;
 816     }
 817
 818     if (support_encode == 0) {
 819         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
 820         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
 821         printf("to use VA-API against DRM instead of X11.\n");
 822         exit(1);
 823     } else {
 824         switch (h264_profile) {
 825             case VAProfileH264Baseline:
 826                 ip_period = 1;
 827                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 828                 h264_entropy_mode = 0;
 829                 break;
 830             case VAProfileH264ConstrainedBaseline:
 831                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 832                 ip_period = 1;
 833                 break;
 834
 835             case VAProfileH264Main:
 836                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 837                 break;
 838
 839             case VAProfileH264High:
 840                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 841                 break;
 842             default:
 843                 h264_profile = VAProfileH264Baseline;
 844                 ip_period = 1;
 845                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 846                 break;
 847         }
 848     }
 849
 850     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 851
 852     /* find out the format for the render target, and rate control mode */
 853     for (i = 0; i < VAConfigAttribTypeMax; i++)
 854         attrib[i].type = (VAConfigAttribType)i;
 855
 856     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 857                                       &attrib[0], VAConfigAttribTypeMax);
 858     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 859     /* check the interested configattrib */
 860     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 861         printf("Not find desired YUV420 RT format\n");
 862         exit(1);
 863     } else {
 864         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 865         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 866         config_attrib_num++;
 867     }
 868
 869     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 870         int tmp = attrib[VAConfigAttribRateControl].value;
 871
 872         if (rc_mode == -1 || !(rc_mode & tmp))  {
 873             if (rc_mode != -1) {
 874                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
 875             }
 876
 877             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
 878                 if (rc_default_modes[i] & tmp) {
 879                     rc_mode = rc_default_modes[i];
 880                     break;
 881                 }
 882             }
 883         }
 884
 885         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 886         config_attrib[config_attrib_num].value = rc_mode;
 887         config_attrib_num++;
 888     }
 889
 890
 891     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 892         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 893
 894         h264_packedheader = 1;
 895         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 896         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 897
 898         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 899             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 900         }
 901
 902         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 903             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 904         }
 905
 906         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 907             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 908         }
 909
 910         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
 911             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
 912         }
 913
 914         enc_packed_header_idx = config_attrib_num;
 915         config_attrib_num++;
 916     }
 917
 918     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
 919         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
 920         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 921         config_attrib_num++;
 922     }
 923
 924     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
 925         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
 926     }
 927
 928     free(entrypoints);
 929     return 0;
 930 }
 931
 932 int QuickSyncEncoderImpl::setup_encode()
 933 {
 934         if (!global_flags.x264_video_to_disk) {
 935                 VAStatus va_status;
 936                 VASurfaceID *tmp_surfaceid;
 937                 int codedbuf_size;
 938                 VASurfaceID src_surface[SURFACE_NUM];
 939                 VASurfaceID ref_surface[SURFACE_NUM];
 940
 941                 va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
 942                                 &config_attrib[0], config_attrib_num, &config_id);
 943                 CHECK_VASTATUS(va_status, "vaCreateConfig");
 944
 945                 /* create source surfaces */
 946                 va_status = vaCreateSurfaces(va_dpy,
 947                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 948                                 &src_surface[0], SURFACE_NUM,
 949                                 NULL, 0);
 950                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 951
 952                 /* create reference surfaces */
 953                 va_status = vaCreateSurfaces(va_dpy,
 954                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 955                                 &ref_surface[0], SURFACE_NUM,
 956                                 NULL, 0);
 957                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 958
 959                 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
 960                 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
 961                 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
 962
 963                 for (int i = 0; i < SURFACE_NUM; i++) {
 964                         gl_surfaces[i].src_surface = src_surface[i];
 965                         gl_surfaces[i].ref_surface = ref_surface[i];
 966                 }
 967
 968                 /* Create a context for this encode pipe */
 969                 va_status = vaCreateContext(va_dpy, config_id,
 970                                 frame_width_mbaligned, frame_height_mbaligned,
 971                                 VA_PROGRESSIVE,
 972                                 tmp_surfaceid, 2 * SURFACE_NUM,
 973                                 &context_id);
 974                 CHECK_VASTATUS(va_status, "vaCreateContext");
 975                 free(tmp_surfaceid);
 976
 977                 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
 978
 979                 for (int i = 0; i < SURFACE_NUM; i++) {
 980                         /* create coded buffer once for all
 981                          * other VA buffers which won't be used again after vaRenderPicture.
 982                          * so APP can always vaCreateBuffer for every frame
 983                          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
 984                          * so VA won't maintain the coded buffer
 985                          */
 986                         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
 987                                         codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
 988                         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 989                 }
 990         }
 991
 992         /* create OpenGL objects */
 993         for (int i = 0; i < SURFACE_NUM; i++) {
 994                 if (use_zerocopy) {
 995                         gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
 996                         gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
 997                 } else {
 998                         size_t bytes_per_pixel = (global_flags.x264_bit_depth > 8) ? 2 : 1;
 999
1000                         // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1001                         // buffers, due to potentially differing pitch.
1002                         glGenBuffers(1, &gl_surfaces[i].pbo);
1003                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1004                         glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2 * bytes_per_pixel, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1005                         uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2 * bytes_per_pixel, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1006                         gl_surfaces[i].y_offset = 0;
1007                         gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
1008                         gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1009                         gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1010                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1011                 }
1012         }
1013
1014         return 0;
1015 }
1016
1017 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1018 //
1019 //   2 1 0 [3] 4 8 9
1020 template<class T, class C>
1021 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1022 {
1023         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1024         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1025         sort(middle, end, less_than);
1026 }
1027
1028 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
1029 {
1030     if (frame_type == FRAME_B)
1031         return;
1032
1033     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1034
1035     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1036     unique_lock<mutex> lock(storage_task_queue_mutex);
1037
1038     // Insert the new frame at the start of the reference queue.
1039     reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
1040
1041     if (reference_frames.size() > num_ref_frames)
1042     {
1043         // The back frame frame is no longer in use as a reference.
1044         int display_frame_num = reference_frames.back().display_number;
1045         assert(surface_for_frame.count(display_frame_num));
1046         release_gl_surface(display_frame_num);
1047         reference_frames.pop_back();
1048     }
1049
1050     // Mark this frame in use as a reference.
1051     assert(surface_for_frame.count(current_display_frame));
1052     ++surface_for_frame[current_display_frame]->refcount;
1053
1054     current_ref_frame_num++;
1055     if (current_ref_frame_num > MaxFrameNum)
1056         current_ref_frame_num = 0;
1057 }
1058
1059
1060 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
1061 {
1062     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1063         return a.frame_idx > b.frame_idx;
1064     };
1065
1066     for (size_t i = 0; i < reference_frames.size(); ++i) {
1067         RefPicList0_P[i] = reference_frames[i].pic;
1068     }
1069     sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
1070 }
1071
1072 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
1073 {
1074     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1075         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1076     };
1077     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1078         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1079     };
1080
1081     for (size_t i = 0; i < reference_frames.size(); ++i) {
1082         RefPicList0_B[i] = reference_frames[i].pic;
1083         RefPicList1_B[i] = reference_frames[i].pic;
1084     }
1085     sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1086     sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1087 }
1088
1089
1090 int QuickSyncEncoderImpl::render_sequence()
1091 {
1092     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1093     VAStatus va_status;
1094     VAEncMiscParameterBuffer *misc_param;
1095     VAEncMiscParameterRateControl *misc_rate_ctrl;
1096
1097     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1098     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1099     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1100     seq_param.bits_per_second = frame_bitrate;
1101
1102     seq_param.intra_period = intra_period;
1103     seq_param.intra_idr_period = intra_idr_period;
1104     seq_param.ip_period = ip_period;
1105
1106     seq_param.max_num_ref_frames = num_ref_frames;
1107     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1108     seq_param.time_scale = TIMEBASE * 2;
1109     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1110     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1111     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1112     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1113     seq_param.seq_fields.bits.chroma_format_idc = 1;
1114     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1115
1116     if (frame_width != frame_width_mbaligned ||
1117         frame_height != frame_height_mbaligned) {
1118         seq_param.frame_cropping_flag = 1;
1119         seq_param.frame_crop_left_offset = 0;
1120         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1121         seq_param.frame_crop_top_offset = 0;
1122         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1123     }
1124
1125     va_status = vaCreateBuffer(va_dpy, context_id,
1126                                VAEncSequenceParameterBufferType,
1127                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1128     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1129
1130     va_status = vaCreateBuffer(va_dpy, context_id,
1131                                VAEncMiscParameterBufferType,
1132                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1133                                1, NULL, &rc_param_buf);
1134     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1135
1136     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1137     misc_param->type = VAEncMiscParameterTypeRateControl;
1138     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1139     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1140     misc_rate_ctrl->bits_per_second = frame_bitrate;
1141     misc_rate_ctrl->target_percentage = 66;
1142     misc_rate_ctrl->window_size = 1000;
1143     misc_rate_ctrl->initial_qp = initial_qp;
1144     misc_rate_ctrl->min_qp = minimal_qp;
1145     misc_rate_ctrl->basic_unit_size = 0;
1146     vaUnmapBuffer(va_dpy, rc_param_buf);
1147
1148     render_id[0] = seq_param_buf;
1149     render_id[1] = rc_param_buf;
1150
1151     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1152
1153     return 0;
1154 }
1155
1156 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1157 {
1158     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1159     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1160     int PicOrderCntMsb, TopFieldOrderCnt;
1161
1162     if (frame_type == FRAME_IDR)
1163         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1164     else {
1165         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1166         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1167     }
1168
1169     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1170         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1171         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1172     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1173              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1174         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1175     else
1176         PicOrderCntMsb = prevPicOrderCntMsb;
1177
1178     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1179
1180     if (frame_type != FRAME_B) {
1181         PicOrderCntMsb_ref = PicOrderCntMsb;
1182         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1183     }
1184
1185     return TopFieldOrderCnt;
1186 }
1187
1188 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1189 {
1190     VABufferID pic_param_buf;
1191     VAStatus va_status;
1192     size_t i = 0;
1193
1194     pic_param.CurrPic.picture_id = surf->ref_surface;
1195     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1196     pic_param.CurrPic.flags = 0;
1197     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1198     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1199     CurrentCurrPic = pic_param.CurrPic;
1200
1201     for (i = 0; i < reference_frames.size(); i++) {
1202         pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1203     }
1204     for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1205         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1206         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1207     }
1208
1209     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1210     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1211     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1212     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1213     pic_param.frame_num = current_ref_frame_num;  // FIXME: is this correct?
1214     pic_param.coded_buf = surf->coded_buf;
1215     pic_param.last_picture = false;  // FIXME
1216     pic_param.pic_init_qp = initial_qp;
1217
1218     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1219                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1220     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1221
1222     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1223
1224     return 0;
1225 }
1226
1227 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1228 {
1229     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1230     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1231     unsigned int length_in_bits;
1232     unsigned char *packedseq_buffer = NULL;
1233     VAStatus va_status;
1234
1235     length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1236
1237     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1238
1239     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1240     packedheader_param_buffer.has_emulation_bytes = 0;
1241     va_status = vaCreateBuffer(va_dpy,
1242                                context_id,
1243                                VAEncPackedHeaderParameterBufferType,
1244                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1245                                &packedseq_para_bufid);
1246     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1247
1248     va_status = vaCreateBuffer(va_dpy,
1249                                context_id,
1250                                VAEncPackedHeaderDataBufferType,
1251                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1252                                &packedseq_data_bufid);
1253     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1254
1255     render_id[0] = packedseq_para_bufid;
1256     render_id[1] = packedseq_data_bufid;
1257     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1258
1259     free(packedseq_buffer);
1260
1261     return 0;
1262 }
1263
1264
1265 int QuickSyncEncoderImpl::render_packedpicture()
1266 {
1267     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1268     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1269     unsigned int length_in_bits;
1270     unsigned char *packedpic_buffer = NULL;
1271     VAStatus va_status;
1272
1273     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1274     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1275     packedheader_param_buffer.bit_length = length_in_bits;
1276     packedheader_param_buffer.has_emulation_bytes = 0;
1277
1278     va_status = vaCreateBuffer(va_dpy,
1279                                context_id,
1280                                VAEncPackedHeaderParameterBufferType,
1281                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1282                                &packedpic_para_bufid);
1283     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1284
1285     va_status = vaCreateBuffer(va_dpy,
1286                                context_id,
1287                                VAEncPackedHeaderDataBufferType,
1288                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1289                                &packedpic_data_bufid);
1290     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1291
1292     render_id[0] = packedpic_para_bufid;
1293     render_id[1] = packedpic_data_bufid;
1294     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1295
1296     free(packedpic_buffer);
1297
1298     return 0;
1299 }
1300
1301 void QuickSyncEncoderImpl::render_packedslice()
1302 {
1303     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1304     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1305     unsigned int length_in_bits;
1306     unsigned char *packedslice_buffer = NULL;
1307     VAStatus va_status;
1308
1309     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1310     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1311     packedheader_param_buffer.bit_length = length_in_bits;
1312     packedheader_param_buffer.has_emulation_bytes = 0;
1313
1314     va_status = vaCreateBuffer(va_dpy,
1315                                context_id,
1316                                VAEncPackedHeaderParameterBufferType,
1317                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1318                                &packedslice_para_bufid);
1319     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1320
1321     va_status = vaCreateBuffer(va_dpy,
1322                                context_id,
1323                                VAEncPackedHeaderDataBufferType,
1324                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1325                                &packedslice_data_bufid);
1326     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1327
1328     render_id[0] = packedslice_para_bufid;
1329     render_id[1] = packedslice_data_bufid;
1330     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1331
1332     free(packedslice_buffer);
1333 }
1334
1335 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1336 {
1337     VABufferID slice_param_buf;
1338     VAStatus va_status;
1339     int i;
1340
1341     /* one frame, one slice */
1342     slice_param.macroblock_address = 0;
1343     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1344     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1345     if (frame_type == FRAME_IDR) {
1346         if (encoding_frame_num != 0)
1347             ++slice_param.idr_pic_id;
1348     } else if (frame_type == FRAME_P) {
1349         VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1350         update_RefPicList_P(RefPicList0_P);
1351
1352         int refpiclist0_max = h264_maxref & 0xffff;
1353         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1354
1355         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1356             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1357             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1358         }
1359     } else if (frame_type == FRAME_B) {
1360         VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1361         update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1362
1363         int refpiclist0_max = h264_maxref & 0xffff;
1364         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1365
1366         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1367         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1368             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1369             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1370         }
1371
1372         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1373         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1374             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1375             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1376         }
1377     }
1378
1379     slice_param.slice_alpha_c0_offset_div2 = 0;
1380     slice_param.slice_beta_offset_div2 = 0;
1381     slice_param.direct_spatial_mv_pred_flag = 1;
1382     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1383
1384
1385     if (h264_packedheader &&
1386         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1387         render_packedslice();
1388
1389     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1390                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1391     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1392
1393     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1394
1395     return 0;
1396 }
1397
1398
1399
1400 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1401 {
1402         VACodedBufferSegment *buf_list = NULL;
1403         VAStatus va_status;
1404
1405         string data;
1406
1407         va_status = vaMapBuffer(va_dpy, surf->coded_buf, (void **)(&buf_list));
1408         CHECK_VASTATUS(va_status, "vaMapBuffer");
1409         while (buf_list != NULL) {
1410                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1411                 buf_list = (VACodedBufferSegment *) buf_list->next;
1412         }
1413         vaUnmapBuffer(va_dpy, surf->coded_buf);
1414
1415         static int frameno = 0;
1416         print_latency("Current Quick Sync latency (video inputs → disk mux):",
1417                 task.received_ts, (task.frame_type == FRAME_B), &frameno);
1418
1419         {
1420                 // Add video.
1421                 AVPacket pkt;
1422                 memset(&pkt, 0, sizeof(pkt));
1423                 pkt.buf = nullptr;
1424                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1425                 pkt.size = data.size();
1426                 pkt.stream_index = 0;
1427                 if (task.frame_type == FRAME_IDR) {
1428                         pkt.flags = AV_PKT_FLAG_KEY;
1429                 } else {
1430                         pkt.flags = 0;
1431                 }
1432                 pkt.duration = task.duration;
1433                 if (file_mux) {
1434                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1435                 }
1436                 if (!global_flags.uncompressed_video_to_http &&
1437                     !global_flags.x264_video_to_http) {
1438                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1439                 }
1440         }
1441 }
1442
1443
1444 // this is weird. but it seems to put a new frame onto the queue
1445 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1446 {
1447         unique_lock<mutex> lock(storage_task_queue_mutex);
1448         storage_task_queue.push(move(task));
1449         storage_task_queue_changed.notify_all();
1450 }
1451
1452 void QuickSyncEncoderImpl::storage_task_thread()
1453 {
1454         pthread_setname_np(pthread_self(), "QS_Storage");
1455         for ( ;; ) {
1456                 storage_task current;
1457                 GLSurface *surf;
1458                 {
1459                         // wait until there's an encoded frame
1460                         unique_lock<mutex> lock(storage_task_queue_mutex);
1461                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1462                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1463                         current = move(storage_task_queue.front());
1464                         storage_task_queue.pop();
1465                         surf = surface_for_frame[current.display_order];
1466                         assert(surf != nullptr);
1467                 }
1468
1469                 VAStatus va_status;
1470
1471                 size_t display_order = current.display_order;
1472                 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1473
1474                 // waits for data, then saves it to disk.
1475                 va_status = vaSyncSurface(va_dpy, surf->src_surface);
1476                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1477                 save_codeddata(surf, move(current));
1478
1479                 // Unlock the frame, and all its references.
1480                 {
1481                         unique_lock<mutex> lock(storage_task_queue_mutex);
1482                         release_gl_surface(display_order);
1483
1484                         for (size_t frame_num : ref_display_frame_numbers) {
1485                                 release_gl_surface(frame_num);
1486                         }
1487                 }
1488         }
1489 }
1490
1491 void QuickSyncEncoderImpl::release_encode()
1492 {
1493         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1494                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1495                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1496                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1497         }
1498
1499         vaDestroyContext(va_dpy, context_id);
1500         vaDestroyConfig(va_dpy, config_id);
1501 }
1502
1503 void QuickSyncEncoderImpl::release_gl_resources()
1504 {
1505         assert(is_shutdown);
1506         if (has_released_gl_resources) {
1507                 return;
1508         }
1509
1510         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1511                 if (use_zerocopy) {
1512                         resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1513                         resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1514                 } else {
1515                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1516                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1517                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1518                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1519                 }
1520         }
1521
1522         has_released_gl_resources = true;
1523 }
1524
1525 int QuickSyncEncoderImpl::deinit_va()
1526 {
1527     vaTerminate(va_dpy);
1528
1529     va_close_display(va_dpy);
1530
1531     return 0;
1532 }
1533
1534 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1535         : current_storage_frame(0), resource_pool(resource_pool), surface(surface), x264_encoder(x264_encoder), frame_width(width), frame_height(height), disk_space_estimator(disk_space_estimator)
1536 {
1537         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1538         open_output_file(filename);
1539         file_audio_encoder->add_mux(file_mux.get());
1540
1541         frame_width_mbaligned = (frame_width + 15) & (~15);
1542         frame_height_mbaligned = (frame_height + 15) & (~15);
1543
1544         //print_input();
1545
1546         if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1547                 assert(x264_encoder != nullptr);
1548         } else {
1549                 assert(x264_encoder == nullptr);
1550         }
1551
1552         enable_zerocopy_if_possible();
1553         if (!global_flags.x264_video_to_disk) {
1554                 init_va(va_display);
1555         }
1556         setup_encode();
1557
1558         if (!global_flags.x264_video_to_disk) {
1559                 memset(&seq_param, 0, sizeof(seq_param));
1560                 memset(&pic_param, 0, sizeof(pic_param));
1561                 memset(&slice_param, 0, sizeof(slice_param));
1562         }
1563
1564         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1565
1566         encode_thread = thread([this]{
1567                 QOpenGLContext *context = create_context(this->surface);
1568                 eglBindAPI(EGL_OPENGL_API);
1569                 if (!make_current(context, this->surface)) {
1570                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1571                                 eglGetError());
1572                         exit(1);
1573                 }
1574                 encode_thread_func();
1575                 delete_context(context);
1576         });
1577 }
1578
1579 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1580 {
1581         shutdown();
1582         release_gl_resources();
1583 }
1584
1585 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1586 {
1587         for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1588                 if (gl_surfaces[i].refcount == 0) {
1589                         ++gl_surfaces[i].refcount;
1590                         return &gl_surfaces[i];
1591                 }
1592         }
1593         return nullptr;
1594 }
1595
1596 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1597 {
1598         assert(surface_for_frame.count(display_frame_num));
1599         QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1600         if (--surf->refcount == 0) {
1601                 assert(surface_for_frame.count(display_frame_num));
1602                 surface_for_frame.erase(display_frame_num);
1603                 storage_task_queue_changed.notify_all();
1604         }
1605 }
1606
1607 bool QuickSyncEncoderImpl::is_zerocopy() const
1608 {
1609         return use_zerocopy;
1610 }
1611
1612 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1613 {
1614         assert(!is_shutdown);
1615         GLSurface *surf = nullptr;
1616         {
1617                 // Wait until this frame slot is done encoding.
1618                 unique_lock<mutex> lock(storage_task_queue_mutex);
1619                 surf = allocate_gl_surface();
1620                 if (surf == nullptr) {
1621                         fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1622                                 current_storage_frame);
1623                         storage_task_queue_changed.wait(lock, [this, &surf]{
1624                                 if (storage_thread_should_quit)
1625                                         return true;
1626                                 surf = allocate_gl_surface();
1627                                 return surf != nullptr;
1628                         });
1629                 }
1630                 if (storage_thread_should_quit) return false;
1631                 assert(surf != nullptr);
1632                 surface_for_frame[current_storage_frame] = surf;
1633         }
1634
1635         if (use_zerocopy) {
1636                 *y_tex = surf->y_tex;
1637                 *cbcr_tex = surf->cbcr_tex;
1638         } else {
1639                 surf->y_tex = *y_tex;
1640                 surf->cbcr_tex = *cbcr_tex;
1641         }
1642
1643         if (!global_flags.x264_video_to_disk) {
1644                 VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1645                 CHECK_VASTATUS(va_status, "vaDeriveImage");
1646
1647                 if (use_zerocopy) {
1648                         VABufferInfo buf_info;
1649                         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1650                         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1651                         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1652
1653                         // Create Y image.
1654                         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1655                         EGLint y_attribs[] = {
1656                                 EGL_WIDTH, frame_width,
1657                                 EGL_HEIGHT, frame_height,
1658                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1659                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1660                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1661                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1662                                 EGL_NONE
1663                         };
1664
1665                         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1666                         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1667
1668                         // Associate Y image to a texture.
1669                         glBindTexture(GL_TEXTURE_2D, *y_tex);
1670                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1671
1672                         // Create CbCr image.
1673                         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1674                         EGLint cbcr_attribs[] = {
1675                                 EGL_WIDTH, frame_width,
1676                                 EGL_HEIGHT, frame_height,
1677                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1678                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1679                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1680                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1681                                 EGL_NONE
1682                         };
1683
1684                         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1685                         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1686
1687                         // Associate CbCr image to a texture.
1688                         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1689                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1690                 }
1691         }
1692
1693         current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1694
1695         return true;
1696 }
1697
1698 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1699 {
1700         assert(!is_shutdown);
1701         file_audio_encoder->encode_audio(audio, pts + global_delay());
1702 }
1703
1704 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1705 {
1706         assert(!is_shutdown);
1707
1708         if (!use_zerocopy) {
1709                 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1710                 GLSurface *surf;
1711                 {
1712                         unique_lock<mutex> lock(storage_task_queue_mutex);
1713                         surf = surface_for_frame[current_storage_frame];
1714                         assert(surf != nullptr);
1715                 }
1716
1717                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1718                 check_error();
1719
1720                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1721                 check_error();
1722
1723                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1724                 check_error();
1725                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1726                 check_error();
1727
1728                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1729                 check_error();
1730                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1731                 check_error();
1732
1733                 // We don't own these; the caller does.
1734                 surf->y_tex = surf->cbcr_tex = 0;
1735
1736                 glBindTexture(GL_TEXTURE_2D, 0);
1737                 check_error();
1738                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1739                 check_error();
1740
1741                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1742                 check_error();
1743         }
1744
1745         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1746         check_error();
1747         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1748         check_error();
1749
1750         {
1751                 unique_lock<mutex> lock(frame_queue_mutex);
1752                 current_video_frame.fence = fence;
1753                 pending_video_frames.push(move(current_video_frame));
1754                 ++current_storage_frame;
1755         }
1756         frame_queue_nonempty.notify_all();
1757         return fence;
1758 }
1759
1760 void QuickSyncEncoderImpl::shutdown()
1761 {
1762         if (is_shutdown) {
1763                 return;
1764         }
1765
1766         {
1767                 unique_lock<mutex> lock(frame_queue_mutex);
1768                 encode_thread_should_quit = true;
1769                 frame_queue_nonempty.notify_all();
1770         }
1771         encode_thread.join();
1772         {
1773                 unique_lock<mutex> lock(storage_task_queue_mutex);
1774                 storage_thread_should_quit = true;
1775                 frame_queue_nonempty.notify_all();
1776                 storage_task_queue_changed.notify_all();
1777         }
1778         storage_thread.join();
1779
1780         // Encode any leftover audio in the queues, and also any delayed frames.
1781         file_audio_encoder->encode_last_audio();
1782
1783         if (!global_flags.x264_video_to_disk) {
1784                 release_encode();
1785                 deinit_va();
1786         }
1787         is_shutdown = true;
1788 }
1789
1790 void QuickSyncEncoderImpl::close_file()
1791 {
1792         file_mux.reset();
1793 }
1794
1795 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1796 {
1797         AVFormatContext *avctx = avformat_alloc_context();
1798         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1799         assert(filename.size() < sizeof(avctx->filename) - 1);
1800         strcpy(avctx->filename, filename.c_str());
1801
1802         string url = "file:" + filename;
1803         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1804         if (ret < 0) {
1805                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1806                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1807                 exit(1);
1808         }
1809
1810         string video_extradata;  // FIXME: See other comment about global headers.
1811         if (global_flags.x264_video_to_disk) {
1812                 video_extradata = x264_encoder->get_global_headers();
1813         }
1814
1815         AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1816         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1817                 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1)));
1818
1819         if (global_flags.x264_video_to_disk) {
1820                 x264_encoder->add_mux(file_mux.get());
1821         }
1822 }
1823
1824 void QuickSyncEncoderImpl::encode_thread_func()
1825 {
1826         pthread_setname_np(pthread_self(), "QS_Encode");
1827
1828         int64_t last_dts = -1;
1829         int gop_start_display_frame_num = 0;
1830         for (int display_frame_num = 0; ; ++display_frame_num) {
1831                 // Wait for the frame to be in the queue. Note that this only means
1832                 // we started rendering it.
1833                 PendingFrame frame;
1834                 {
1835                         unique_lock<mutex> lock(frame_queue_mutex);
1836                         frame_queue_nonempty.wait(lock, [this]{
1837                                 return encode_thread_should_quit || !pending_video_frames.empty();
1838                         });
1839                         if (encode_thread_should_quit && pending_video_frames.empty()) {
1840                                 // We may have queued frames left in the reorder buffer
1841                                 // that were supposed to be B-frames, but have no P-frame
1842                                 // to be encoded against. If so, encode them all as
1843                                 // P-frames instead. Note that this happens under the mutex,
1844                                 // but nobody else uses it at this point, since we're shutting down,
1845                                 // so there's no contention.
1846                                 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1847                                 return;
1848                         } else {
1849                                 frame = move(pending_video_frames.front());
1850                                 pending_video_frames.pop();
1851                         }
1852                 }
1853
1854                 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1855                 // Note that this implicitly waits for the frame to be done rendering.
1856                 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1857
1858                 if (global_flags.x264_video_to_disk) {
1859                         unique_lock<mutex> lock(storage_task_queue_mutex);
1860                         release_gl_surface(display_frame_num);
1861                         continue;
1862                 }
1863
1864                 reorder_buffer[display_frame_num] = move(frame);
1865
1866                 // Now encode as many QuickSync frames as we can using the frames we have available.
1867                 // (It could be zero, or it could be multiple.) FIXME: make a function.
1868                 for ( ;; ) {
1869                         int pts_lag;
1870                         int frame_type, quicksync_display_frame_num;
1871                         encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1872                                                &quicksync_display_frame_num, &frame_type, &pts_lag);
1873                         if (!reorder_buffer.count(quicksync_display_frame_num)) {
1874                                 break;
1875                         }
1876                         frame = move(reorder_buffer[quicksync_display_frame_num]);
1877                         reorder_buffer.erase(quicksync_display_frame_num);
1878
1879                         if (frame_type == FRAME_IDR) {
1880                                 // Release any reference frames from the previous GOP.
1881                                 for (const ReferenceFrame &frame : reference_frames) {
1882                                         release_gl_surface(frame.display_number);
1883                                 }
1884                                 reference_frames.clear();
1885                                 current_ref_frame_num = 0;
1886                                 gop_start_display_frame_num = quicksync_display_frame_num;
1887                         }
1888
1889                         // Determine the dts of this frame.
1890                         int64_t dts;
1891                         if (pts_lag == -1) {
1892                                 assert(last_dts != -1);
1893                                 dts = last_dts + (TIMEBASE / MAX_FPS);
1894                         } else {
1895                                 dts = frame.pts - pts_lag;
1896                         }
1897                         last_dts = dts;
1898
1899                         encode_frame(frame, quicksync_encoding_frame_num, quicksync_display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1900                         ++quicksync_encoding_frame_num;
1901                 }
1902         }
1903 }
1904
1905 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1906 {
1907         if (reorder_buffer.empty()) {
1908                 return;
1909         }
1910
1911         for (auto &pending_frame : reorder_buffer) {
1912                 int display_frame_num = pending_frame.first;
1913                 assert(display_frame_num > 0);
1914                 PendingFrame frame = move(pending_frame.second);
1915                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1916                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1917                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1918                 last_dts = dts;
1919         }
1920 }
1921
1922 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1923 {
1924         AVPacket pkt;
1925         memset(&pkt, 0, sizeof(pkt));
1926         pkt.buf = nullptr;
1927         pkt.data = const_cast<uint8_t *>(data);
1928         pkt.size = frame_width * frame_height * 2;
1929         pkt.stream_index = 0;
1930         pkt.flags = AV_PKT_FLAG_KEY;
1931         pkt.duration = duration;
1932         stream_mux->add_packet(pkt, pts, pts);
1933 }
1934
1935 namespace {
1936
1937 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1938 {
1939         if (src_width == dst_pitch) {
1940                 memcpy(dst, src, src_width * height);
1941         } else {
1942                 for (size_t y = 0; y < height; ++y) {
1943                         const uint8_t *sptr = src + y * src_width;
1944                         uint8_t *dptr = dst + y * dst_pitch;
1945                         memcpy(dptr, sptr, src_width);
1946                 }
1947         }
1948 }
1949
1950 }  // namespace
1951
1952 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1953 {
1954         // Wait for the GPU to be done with the frame.
1955         GLenum sync_status;
1956         do {
1957                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1958                 check_error();
1959         } while (sync_status == GL_TIMEOUT_EXPIRED);
1960         assert(sync_status != GL_WAIT_FAILED);
1961
1962         ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1963         static int frameno = 0;
1964         print_latency("Current mixer latency (video inputs → ready for encode):",
1965                 received_ts, false, &frameno);
1966
1967         // Release back any input frames we needed to render this frame.
1968         frame.input_frames.clear();
1969
1970         GLSurface *surf;
1971         {
1972                 unique_lock<mutex> lock(storage_task_queue_mutex);
1973                 surf = surface_for_frame[display_frame_num];
1974                 assert(surf != nullptr);
1975         }
1976         uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1977         if (global_flags.uncompressed_video_to_http) {
1978                 add_packet_for_uncompressed_frame(pts, duration, data);
1979         } else if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1980                 x264_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1981         }
1982 }
1983
1984 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1985                                         int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
1986 {
1987         const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1988
1989         GLSurface *surf;
1990         {
1991                 unique_lock<mutex> lock(storage_task_queue_mutex);
1992                 surf = surface_for_frame[display_frame_num];
1993                 assert(surf != nullptr);
1994         }
1995         VAStatus va_status;
1996
1997         if (use_zerocopy) {
1998                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1999                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2000                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2001                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2002         } else {
2003                 // Upload the frame to VA-API.
2004                 unsigned char *surface_p = nullptr;
2005                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2006
2007                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2008                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2009
2010                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2011                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2012
2013                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2014                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2015         }
2016
2017         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2018         CHECK_VASTATUS(va_status, "vaDestroyImage");
2019
2020         // Schedule the frame for encoding.
2021         VASurfaceID va_surface = surf->src_surface;
2022         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2023         CHECK_VASTATUS(va_status, "vaBeginPicture");
2024
2025         if (frame_type == FRAME_IDR) {
2026                 // FIXME: If the mux wants global headers, we should not put the
2027                 // SPS/PPS before each IDR frame, but rather put it into the
2028                 // codec extradata (formatted differently?).
2029                 //
2030                 // NOTE: If we change ycbcr_coefficients, it will not take effect
2031                 // before the next IDR frame. This is acceptable, as it should only
2032                 // happen on a mode change, which is rare.
2033                 render_sequence();
2034                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2035                 if (h264_packedheader) {
2036                         render_packedsequence(ycbcr_coefficients);
2037                         render_packedpicture();
2038                 }
2039         } else {
2040                 //render_sequence();
2041                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2042         }
2043         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2044
2045         va_status = vaEndPicture(va_dpy, context_id);
2046         CHECK_VASTATUS(va_status, "vaEndPicture");
2047
2048         update_ReferenceFrames(display_frame_num, frame_type);
2049
2050         vector<size_t> ref_display_frame_numbers;
2051
2052         // Lock the references for this frame; otherwise, they could be
2053         // rendered to before this frame is done encoding.
2054         {
2055                 unique_lock<mutex> lock(storage_task_queue_mutex);
2056                 for (const ReferenceFrame &frame : reference_frames) {
2057                         assert(surface_for_frame.count(frame.display_number));
2058                         ++surface_for_frame[frame.display_number]->refcount;
2059                         ref_display_frame_numbers.push_back(frame.display_number);
2060                 }
2061         }
2062
2063         // so now the data is done encoding (well, async job kicked off)...
2064         // we send that to the storage thread
2065         storage_task tmp;
2066         tmp.display_order = display_frame_num;
2067         tmp.frame_type = frame_type;
2068         tmp.pts = pts;
2069         tmp.dts = dts;
2070         tmp.duration = duration;
2071         tmp.ycbcr_coefficients = ycbcr_coefficients;
2072         tmp.received_ts = received_ts;
2073         tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2074         storage_task_enqueue(move(tmp));
2075 }
2076
2077 // Proxy object.
2078 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
2079         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2080
2081 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2082 QuickSyncEncoder::~QuickSyncEncoder() {}
2083
2084 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2085 {
2086         impl->add_audio(pts, audio);
2087 }
2088
2089 bool QuickSyncEncoder::is_zerocopy() const
2090 {
2091         return impl->is_zerocopy();
2092 }
2093
2094 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2095 {
2096         return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2097 }
2098
2099 RefCountedGLsync QuickSyncEncoder::end_frame()
2100 {
2101         return impl->end_frame();
2102 }
2103
2104 void QuickSyncEncoder::shutdown()
2105 {
2106         impl->shutdown();
2107 }
2108
2109 void QuickSyncEncoder::close_file()
2110 {
2111         impl->shutdown();
2112 }
2113
2114 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2115 {
2116         impl->set_stream_mux(mux);
2117 }
2118
2119 int64_t QuickSyncEncoder::global_delay() const {
2120         return impl->global_delay();
2121 }