git.sesse.net Git - nageru/blob - nageru/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 <glob.h>
  13 #include <pthread.h>
  14 #include <stdio.h>
  15 #include <stdlib.h>
  16 #include <string.h>
  17 #include <unistd.h>
  18 #include <va/va.h>
  19 #include <va/va_drm.h>
  20 #include <va/va_drmcommon.h>
  21 #include <va/va_enc_h264.h>
  22 #include <va/va_x11.h>
  23 #include <algorithm>
  24 #include <chrono>
  25 #include <condition_variable>
  26 #include <cstddef>
  27 #include <cstdint>
  28 #include <functional>
  29 #include <map>
  30 #include <memory>
  31 #include <mutex>
  32 #include <queue>
  33 #include <stack>
  34 #include <string>
  35 #include <thread>
  36 #include <utility>
  37
  38 extern "C" {
  39
  40 #include <libavcodec/avcodec.h>
  41 #include <libavformat/avio.h>
  42 #include <libavutil/error.h>
  43 #include <drm_fourcc.h>
  44
  45 }  // namespace
  46
  47 #include "audio_encoder.h"
  48 #include "shared/context.h"
  49 #include "defs.h"
  50 #include "shared/disk_space_estimator.h"
  51 #include "shared/ffmpeg_raii.h"
  52 #include "flags.h"
  53 #include "shared/mux.h"
  54 #include "print_latency.h"
  55 #include "quicksync_encoder_impl.h"
  56 #include "ref_counted_frame.h"
  57 #include "shared/timebase.h"
  58 #include "x264_encoder.h"
  59
  60 using namespace movit;
  61 using namespace std;
  62 using namespace std::chrono;
  63 using namespace std::placeholders;
  64
  65 class QOpenGLContext;
  66 class QSurface;
  67
  68 namespace {
  69
  70 // These need to survive several QuickSyncEncoderImpl instances,
  71 // so they are outside.
  72 once_flag quick_sync_metrics_inited;
  73 LatencyHistogram mixer_latency_histogram, qs_latency_histogram;
  74 MuxMetrics current_file_mux_metrics, total_mux_metrics;
  75 std::atomic<double> metric_current_file_start_time_seconds{0.0 / 0.0};
  76 std::atomic<int64_t> metric_quick_sync_stalled_frames{0};
  77
  78 }  // namespace
  79
  80 #define CHECK_VASTATUS(va_status, func)                                 \
  81     if (va_status != VA_STATUS_SUCCESS) {                               \
  82         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  83         abort();                                                        \
  84     }
  85
  86 #undef BUFFER_OFFSET
  87 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  88
  89 //#include "loadsurface.h"
  90
  91 #define NAL_REF_IDC_NONE        0
  92 #define NAL_REF_IDC_LOW         1
  93 #define NAL_REF_IDC_MEDIUM      2
  94 #define NAL_REF_IDC_HIGH        3
  95
  96 #define NAL_NON_IDR             1
  97 #define NAL_IDR                 5
  98 #define NAL_SPS                 7
  99 #define NAL_PPS                 8
 100 #define NAL_SEI                 6
 101
 102 #define SLICE_TYPE_P            0
 103 #define SLICE_TYPE_B            1
 104 #define SLICE_TYPE_I            2
 105 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
 106 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
 107 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
 108
 109
 110 #define ENTROPY_MODE_CAVLC      0
 111 #define ENTROPY_MODE_CABAC      1
 112
 113 #define PROFILE_IDC_BASELINE    66
 114 #define PROFILE_IDC_MAIN        77
 115 #define PROFILE_IDC_HIGH        100
 116
 117 #define BITSTREAM_ALLOCATE_STEPPING     4096
 118
 119 static constexpr unsigned int MaxFrameNum = (2<<16);
 120 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
 121 static constexpr unsigned int Log2MaxFrameNum = 16;
 122 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
 123
 124 using namespace std;
 125
 126 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 127 // but if we don't delete it here, we get leaks. The GStreamer implementation
 128 // does the same.
 129 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 130 {
 131     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 132     CHECK_VASTATUS(va_status, "vaRenderPicture");
 133
 134     for (int i = 0; i < num_buffers; ++i) {
 135         va_status = vaDestroyBuffer(dpy, buffers[i]);
 136         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 137     }
 138 }
 139
 140 static unsigned int
 141 va_swap32(unsigned int val)
 142 {
 143     unsigned char *pval = (unsigned char *)&val;
 144
 145     return ((pval[0] << 24)     |
 146             (pval[1] << 16)     |
 147             (pval[2] << 8)      |
 148             (pval[3] << 0));
 149 }
 150
 151 static void
 152 bitstream_start(bitstream *bs)
 153 {
 154     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 155     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 156     bs->bit_offset = 0;
 157 }
 158
 159 static void
 160 bitstream_end(bitstream *bs)
 161 {
 162     int pos = (bs->bit_offset >> 5);
 163     int bit_offset = (bs->bit_offset & 0x1f);
 164     int bit_left = 32 - bit_offset;
 165
 166     if (bit_offset) {
 167         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 168     }
 169 }
 170
 171 static void
 172 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 173 {
 174     int pos = (bs->bit_offset >> 5);
 175     int bit_offset = (bs->bit_offset & 0x1f);
 176     int bit_left = 32 - bit_offset;
 177
 178     if (!size_in_bits)
 179         return;
 180
 181     bs->bit_offset += size_in_bits;
 182
 183     if (bit_left > size_in_bits) {
 184         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 185     } else {
 186         size_in_bits -= bit_left;
 187         if (bit_left >= 32) {
 188             bs->buffer[pos] = (val >> size_in_bits);
 189         } else {
 190             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 191         }
 192         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 193
 194         if (pos + 1 == bs->max_size_in_dword) {
 195             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 196             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 197         }
 198
 199         bs->buffer[pos + 1] = val;
 200     }
 201 }
 202
 203 static void
 204 bitstream_put_ue(bitstream *bs, unsigned int val)
 205 {
 206     int size_in_bits = 0;
 207     int tmp_val = ++val;
 208
 209     while (tmp_val) {
 210         tmp_val >>= 1;
 211         size_in_bits++;
 212     }
 213
 214     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 215     bitstream_put_ui(bs, val, size_in_bits);
 216 }
 217
 218 static void
 219 bitstream_put_se(bitstream *bs, int val)
 220 {
 221     unsigned int new_val;
 222
 223     if (val <= 0)
 224         new_val = -2 * val;
 225     else
 226         new_val = 2 * val - 1;
 227
 228     bitstream_put_ue(bs, new_val);
 229 }
 230
 231 static void
 232 bitstream_byte_aligning(bitstream *bs, int bit)
 233 {
 234     int bit_offset = (bs->bit_offset & 0x7);
 235     int bit_left = 8 - bit_offset;
 236     int new_val;
 237
 238     if (!bit_offset)
 239         return;
 240
 241     assert(bit == 0 || bit == 1);
 242
 243     if (bit)
 244         new_val = (1 << bit_left) - 1;
 245     else
 246         new_val = 0;
 247
 248     bitstream_put_ui(bs, new_val, bit_left);
 249 }
 250
 251 static void
 252 rbsp_trailing_bits(bitstream *bs)
 253 {
 254     bitstream_put_ui(bs, 1, 1);
 255     bitstream_byte_aligning(bs, 0);
 256 }
 257
 258 static void nal_start_code_prefix(bitstream *bs)
 259 {
 260     bitstream_put_ui(bs, 0x00000001, 32);
 261 }
 262
 263 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 264 {
 265     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 266     bitstream_put_ui(bs, nal_ref_idc, 2);
 267     bitstream_put_ui(bs, nal_unit_type, 5);
 268 }
 269
 270 void QuickSyncEncoderImpl::sps_rbsp(YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs)
 271 {
 272     int profile_idc = PROFILE_IDC_BASELINE;
 273
 274     if (h264_profile  == VAProfileH264High)
 275         profile_idc = PROFILE_IDC_HIGH;
 276     else if (h264_profile  == VAProfileH264Main)
 277         profile_idc = PROFILE_IDC_MAIN;
 278
 279     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 280     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 281     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 282     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 283     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 284     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 285     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 286     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 287
 288     if ( profile_idc == PROFILE_IDC_HIGH) {
 289         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 290         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 291         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 292         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 293         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 294     }
 295
 296     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 297     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 298
 299     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 300         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 301     else {
 302         assert(0);
 303     }
 304
 305     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 306     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 307
 308     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 309     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 310     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 311
 312     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 313         assert(0);
 314     }
 315
 316     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 317     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 318
 319     if (seq_param.frame_cropping_flag) {
 320         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 321         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 322         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 323         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 324     }
 325
 326     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 327     if ( false ) {
 328         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 329     } else {
 330         // See H.264 annex E for the definition of this header.
 331         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 332         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 333         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 334         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 335         {
 336             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 337             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 338             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 339             {
 340                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 341                 bitstream_put_ui(bs, 13, 8);  /* transfer_characteristics (13 = sRGB) */
 342                 if (ycbcr_coefficients == YCBCR_REC_709) {
 343                     bitstream_put_ui(bs, 1, 8);  /* matrix_coefficients (1 = BT.709) */
 344                 } else {
 345                     assert(ycbcr_coefficients == YCBCR_REC_601);
 346                     bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 347                 }
 348             }
 349         }
 350         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 351         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 352         {
 353             bitstream_put_ui(bs, 1, 32);  // FPS
 354             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 355             bitstream_put_ui(bs, 1, 1);
 356         }
 357         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 358         {
 359             // hrd_parameters
 360             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 361             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 362             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 363
 364             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 365             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 366             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 367
 368             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 369             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 370             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 371             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 372         }
 373         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 374         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 375
 376         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 377         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 378     }
 379
 380     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 381 }
 382
 383
 384 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 385 {
 386     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 387     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 388
 389     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 390
 391     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 392
 393     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 394
 395     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 396     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 397
 398     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 399     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 400
 401     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 402     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 403     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 404
 405     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 406     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 407     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 408
 409     /* more_rbsp_data */
 410     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 411     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 412     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 413
 414     rbsp_trailing_bits(bs);
 415 }
 416
 417 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 418 {
 419     int first_mb_in_slice = slice_param.macroblock_address;
 420
 421     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 422     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 423     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 424     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 425
 426     /* frame_mbs_only_flag == 1 */
 427     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 428         /* FIXME: */
 429         assert(0);
 430     }
 431
 432     if (pic_param.pic_fields.bits.idr_pic_flag)
 433         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 434
 435     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 436         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 437         /* pic_order_present_flag == 0 */
 438     } else {
 439         /* FIXME: */
 440         assert(0);
 441     }
 442
 443     /* redundant_pic_cnt_present_flag == 0 */
 444     /* slice type */
 445     if (IS_P_SLICE(slice_param.slice_type)) {
 446         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 447
 448         if (slice_param.num_ref_idx_active_override_flag)
 449             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 450
 451         /* ref_pic_list_reordering */
 452         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 453     } else if (IS_B_SLICE(slice_param.slice_type)) {
 454         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 455
 456         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 457
 458         if (slice_param.num_ref_idx_active_override_flag) {
 459             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 460             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 461         }
 462
 463         /* ref_pic_list_reordering */
 464         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 465         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 466     }
 467
 468     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 469          IS_P_SLICE(slice_param.slice_type)) ||
 470         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 471          IS_B_SLICE(slice_param.slice_type))) {
 472         /* FIXME: fill weight/offset table */
 473         assert(0);
 474     }
 475
 476     /* dec_ref_pic_marking */
 477     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 478         unsigned char no_output_of_prior_pics_flag = 0;
 479         unsigned char long_term_reference_flag = 0;
 480         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 481
 482         if (pic_param.pic_fields.bits.idr_pic_flag) {
 483             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 484             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 485         } else {
 486             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 487         }
 488     }
 489
 490     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 491         !IS_I_SLICE(slice_param.slice_type))
 492         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 493
 494     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 495
 496     /* ignore for SP/SI */
 497
 498     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 499         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 500
 501         if (slice_param.disable_deblocking_filter_idc != 1) {
 502             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 503             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 504         }
 505     }
 506
 507     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 508         bitstream_byte_aligning(bs, 1);
 509     }
 510 }
 511
 512 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 513 {
 514     bitstream bs;
 515
 516     bitstream_start(&bs);
 517     nal_start_code_prefix(&bs);
 518     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 519     pps_rbsp(&bs);
 520     bitstream_end(&bs);
 521
 522     *header_buffer = (unsigned char *)bs.buffer;
 523     return bs.bit_offset;
 524 }
 525
 526 int
 527 QuickSyncEncoderImpl::build_packed_seq_buffer(YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer)
 528 {
 529     bitstream bs;
 530
 531     bitstream_start(&bs);
 532     nal_start_code_prefix(&bs);
 533     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 534     sps_rbsp(ycbcr_coefficients, &bs);
 535     bitstream_end(&bs);
 536
 537     *header_buffer = (unsigned char *)bs.buffer;
 538     return bs.bit_offset;
 539 }
 540
 541 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 542 {
 543     bitstream bs;
 544     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 545     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 546
 547     bitstream_start(&bs);
 548     nal_start_code_prefix(&bs);
 549
 550     if (IS_I_SLICE(slice_param.slice_type)) {
 551         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 552     } else if (IS_P_SLICE(slice_param.slice_type)) {
 553         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 554     } else {
 555         assert(IS_B_SLICE(slice_param.slice_type));
 556         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 557     }
 558
 559     slice_header(&bs);
 560     bitstream_end(&bs);
 561
 562     *header_buffer = (unsigned char *)bs.buffer;
 563     return bs.bit_offset;
 564 }
 565
 566
 567 /*
 568   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 569   1) period between Frame #X and Frame #N = #X - #N
 570   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 571   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 572   4) intra_period and intra_idr_period take precedence over ip_period
 573   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 574      of I/IDR frames, see bellow examples
 575   -------------------------------------------------------------------
 576   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 577   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 578   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 579   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 580   1            1                ignored    IDR IDR IDR IDR...
 581   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 582   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 583   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 584                                               (PBB)(IBB)(PBB)(IBB)...
 585   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 586                                            IDRPPPPP IPPPPP IPPPPP...
 587   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 588                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 589                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 590                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 591                                            {IDR(PBB)(PBB)}                     (6/6/3)
 592                                            {IDR(PBB)(PBB)}.
 593 */
 594
 595 #define FRAME_P 0
 596 #define FRAME_B 1
 597 #define FRAME_I 2
 598 #define FRAME_IDR 7
 599 void encoding2display_order(
 600     int encoding_order, int intra_period,
 601     int intra_idr_period, int ip_period,
 602     int *displaying_order,
 603     int *frame_type)
 604 {
 605     int encoding_order_gop = 0;
 606
 607     if (intra_period == 1) { /* all are I/IDR frames */
 608         *displaying_order = encoding_order;
 609         if (intra_idr_period == 0)
 610             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 611         else
 612             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 613         return;
 614     }
 615
 616     if (intra_period == 0)
 617         intra_idr_period = 0;
 618
 619     if (ip_period == 1) {
 620         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 621         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 622         *displaying_order = encoding_order;
 623
 624         if (encoding_order_gop == 0) { /* the first frame */
 625             *frame_type = FRAME_IDR;
 626         } else if (intra_period != 0 && /* have I frames */
 627                    encoding_order_gop >= 2 &&
 628                    (encoding_order_gop % intra_period == 0)) {
 629             *frame_type = FRAME_I;
 630         } else {
 631             *frame_type = FRAME_P;
 632         }
 633         return;
 634     }
 635
 636     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 637     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 638
 639     if (encoding_order_gop == 0) { /* the first frame */
 640         *frame_type = FRAME_IDR;
 641         *displaying_order = encoding_order;
 642     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 643         *frame_type = FRAME_B;
 644         *displaying_order = encoding_order - 1;
 645     } else if (intra_period != 0 && /* have I frames */
 646                encoding_order_gop >= 2 &&
 647                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 648         *frame_type = FRAME_I;
 649         *displaying_order = encoding_order + ip_period - 1;
 650     } else {
 651         *frame_type = FRAME_P;
 652         *displaying_order = encoding_order + ip_period - 1;
 653     }
 654 }
 655
 656 // General pts/dts strategy:
 657 //
 658 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 659 // bit tricky. This strategy roughly matches what x264 seems to do: We take in
 660 // the pts as the frames are encoded, and reuse that as dts in the same order,
 661 // slightly offset.
 662 //
 663 // If we don't have B-frames (only I and P), this means pts == dts always.
 664 // This is the simple case. Now consider the case with a single B-frame:
 665 //
 666 //        I  B  P  B  P
 667 //   pts: 30 40 50 60 70
 668 //
 669 // Since we always inherently encode P-frames before B-frames, this means that
 670 // we see them in this order, which we can _almost_ use for dts:
 671 //
 672 //   dts: 30 50 40 70 60
 673 //
 674 // the only problem here is that for the B-frames, pts < dts. We solve this by
 675 // priming the queue at the very start with some made-up dts:
 676 //
 677 //        I  B  P  B  P
 678 //   pts: 30 40 50 60 70
 679 //   dts: xx 30 50 40 70 60
 680 //
 681 // Now we have all the desirable properties: pts >= dts, successive dts delta
 682 // is never larger than the decoder can figure out (assuming, of course,
 683 // the pts has that property), and there's minimal lag between pts and dts.
 684 // For the made-up dts, we assume 1/60 sec per frame, which should generally
 685 // be reasonable. dts can go negative, but this is corrected using global_delay()
 686 // by delaying both pts and dts (although we probably don't need to).
 687 //
 688 // If there's more than one B-frame possible, we simply insert more of them
 689 // (here shown with some irregular spacing, assuming B-frames don't depend
 690 // on each other and simply go back-to-front):
 691 //
 692 //        I  B  B  B  P  B  B  B  P
 693 //   pts: 30 40 55 60 65 66 67 68 80
 694 //   dts: xx yy zz 30 65 60 55 40 80 68 67 66
 695 class DTSReorderer {
 696 public:
 697         DTSReorderer(int num_b_frames) : num_b_frames(num_b_frames) {}
 698
 699         void push_pts(int64_t pts)
 700         {
 701                 if (buf.empty() && num_b_frames > 0) {  // First frame.
 702                         int64_t base_dts = pts - num_b_frames * (TIMEBASE / TYPICAL_FPS);
 703                         for (int i = 0; i < num_b_frames; ++i) {
 704                                 buf.push(base_dts + i * (TIMEBASE / TYPICAL_FPS));
 705                         }
 706                 }
 707                 buf.push(pts);
 708         }
 709
 710         int64_t pop_dts()
 711         {
 712                 assert(!buf.empty());
 713                 int64_t dts = buf.front();
 714                 buf.pop();
 715                 return dts;
 716         }
 717
 718 private:
 719         const int num_b_frames;
 720         queue<int64_t> buf;
 721 };
 722
 723 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 724 {
 725         if (global_flags.x264_video_to_disk) {
 726                 // Quick Sync is entirely disabled.
 727                 use_zerocopy = false;
 728         } else if (global_flags.x264_video_to_http) {
 729                 use_zerocopy = false;
 730         } else if (global_flags.av1_video_to_http) {
 731                 use_zerocopy = false;
 732         } else if (!global_flags.v4l_output_device.empty()) {
 733                 use_zerocopy = false;
 734         } else {
 735                 use_zerocopy = true;
 736         }
 737         global_flags.use_zerocopy = use_zerocopy;
 738 }
 739
 740 static unique_ptr<VADisplayWithCleanup> try_open_va_h264(const string &va_display, VAProfile *h264_profile, string *error)
 741 {
 742         return try_open_va(va_display, { VAProfileH264High, VAProfileH264Main, VAProfileH264ConstrainedBaseline },
 743                 VAEntrypointEncSlice, /*desired_configs=*/{}, h264_profile, error);
 744 }
 745
 746 int QuickSyncEncoderImpl::init_va(const string &va_display)
 747 {
 748     string error;
 749     va_dpy = try_open_va_h264(va_display, &h264_profile, &error);
 750     if (va_dpy == nullptr) {
 751         fprintf(stderr, "error: %s\n", error.c_str());
 752         abort();
 753     }
 754     if (!va_dpy->can_use_zerocopy) {
 755         use_zerocopy = false;
 756     }
 757
 758     switch (h264_profile) {
 759         case VAProfileH264ConstrainedBaseline:
 760             constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 761             ip_period = 1;
 762             break;
 763
 764         case VAProfileH264Main:
 765             constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 766             break;
 767
 768         case VAProfileH264High:
 769             constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 770             break;
 771         default:
 772             h264_profile = VAProfileH264ConstrainedBaseline;
 773             ip_period = 1;
 774             constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 775             break;
 776     }
 777
 778     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 779
 780     /* find out the format for the render target, and rate control mode */
 781     for (unsigned i = 0; i < VAConfigAttribTypeMax; i++)
 782         attrib[i].type = (VAConfigAttribType)i;
 783
 784     VAStatus va_status = vaGetConfigAttributes(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
 785                                       &attrib[0], VAConfigAttribTypeMax);
 786     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 787     /* check the interested configattrib */
 788     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 789         printf("Not find desired YUV420 RT format\n");
 790         abort();
 791     } else {
 792         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 793         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 794         config_attrib_num++;
 795     }
 796
 797     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 798         if (!(attrib[VAConfigAttribRateControl].value & VA_RC_CQP)) {
 799             fprintf(stderr, "ERROR: VA-API encoder does not support CQP mode.\n");
 800             abort();
 801         }
 802
 803         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 804         config_attrib[config_attrib_num].value = VA_RC_CQP;
 805         config_attrib_num++;
 806     }
 807
 808
 809     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 810         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 811
 812         h264_packedheader = 1;
 813         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 814         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 815
 816         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 817             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 818         }
 819
 820         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 821             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 822         }
 823
 824         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 825             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 826         }
 827
 828         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
 829             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
 830         }
 831
 832         enc_packed_header_idx = config_attrib_num;
 833         config_attrib_num++;
 834     }
 835
 836     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
 837         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
 838         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 839         config_attrib_num++;
 840     }
 841
 842     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
 843         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
 844     }
 845
 846     return 0;
 847 }
 848
 849 int QuickSyncEncoderImpl::setup_encode()
 850 {
 851         if (!global_flags.x264_video_to_disk) {
 852                 VAStatus va_status;
 853                 VASurfaceID *tmp_surfaceid;
 854                 int codedbuf_size;
 855                 VASurfaceID src_surface[SURFACE_NUM];
 856                 VASurfaceID ref_surface[SURFACE_NUM];
 857
 858                 va_status = vaCreateConfig(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
 859                                 &config_attrib[0], config_attrib_num, &config_id);
 860                 CHECK_VASTATUS(va_status, "vaCreateConfig");
 861
 862                 /* create source surfaces */
 863                 va_status = vaCreateSurfaces(va_dpy->va_dpy,
 864                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 865                                 &src_surface[0], SURFACE_NUM,
 866                                 NULL, 0);
 867                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 868
 869                 /* create reference surfaces */
 870                 va_status = vaCreateSurfaces(va_dpy->va_dpy,
 871                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 872                                 &ref_surface[0], SURFACE_NUM,
 873                                 NULL, 0);
 874                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 875
 876                 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
 877                 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
 878                 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
 879
 880                 for (int i = 0; i < SURFACE_NUM; i++) {
 881                         gl_surfaces[i].src_surface = src_surface[i];
 882                         gl_surfaces[i].ref_surface = ref_surface[i];
 883                 }
 884
 885                 /* Create a context for this encode pipe */
 886                 va_status = vaCreateContext(va_dpy->va_dpy, config_id,
 887                                 frame_width_mbaligned, frame_height_mbaligned,
 888                                 VA_PROGRESSIVE,
 889                                 tmp_surfaceid, 2 * SURFACE_NUM,
 890                                 &context_id);
 891                 CHECK_VASTATUS(va_status, "vaCreateContext");
 892                 free(tmp_surfaceid);
 893
 894                 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
 895
 896                 for (int i = 0; i < SURFACE_NUM; i++) {
 897                         /* create coded buffer once for all
 898                          * other VA buffers which won't be used again after vaRenderPicture.
 899                          * so APP can always vaCreateBuffer for every frame
 900                          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
 901                          * so VA won't maintain the coded buffer
 902                          */
 903                         va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncCodedBufferType,
 904                                         codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
 905                         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 906                 }
 907         }
 908
 909         /* create OpenGL objects */
 910         for (int i = 0; i < SURFACE_NUM; i++) {
 911                 if (use_zerocopy) {
 912                         gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
 913                         gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
 914                 } else {
 915                         size_t bytes_per_pixel = (global_flags.bit_depth > 8) ? 2 : 1;
 916
 917                         // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
 918                         // buffers, due to potentially differing pitch.
 919                         glGenBuffers(1, &gl_surfaces[i].pbo);
 920                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
 921                         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);
 922                         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);
 923                         gl_surfaces[i].y_offset = 0;
 924                         gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
 925                         gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
 926                         gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
 927                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 928                 }
 929         }
 930
 931         return 0;
 932 }
 933
 934 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
 935 //
 936 //   2 1 0 [3] 4 8 9
 937 template<class T, class C>
 938 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
 939 {
 940         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
 941         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
 942         sort(middle, end, less_than);
 943 }
 944
 945 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
 946 {
 947     if (frame_type == FRAME_B)
 948         return;
 949
 950     pic_param.CurrPic.frame_idx = current_ref_frame_num;
 951
 952     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
 953     lock_guard<mutex> lock(storage_task_queue_mutex);
 954
 955     // Insert the new frame at the start of the reference queue.
 956     reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
 957
 958     if (reference_frames.size() > num_ref_frames)
 959     {
 960         // The back frame frame is no longer in use as a reference.
 961         int display_frame_num = reference_frames.back().display_number;
 962         assert(surface_for_frame.count(display_frame_num));
 963         release_gl_surface(display_frame_num);
 964         reference_frames.pop_back();
 965     }
 966
 967     // Mark this frame in use as a reference.
 968     assert(surface_for_frame.count(current_display_frame));
 969     ++surface_for_frame[current_display_frame]->refcount;
 970
 971     current_ref_frame_num++;
 972     if (current_ref_frame_num > MaxFrameNum)
 973         current_ref_frame_num = 0;
 974 }
 975
 976
 977 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
 978 {
 979     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
 980         return a.frame_idx > b.frame_idx;
 981     };
 982
 983     for (size_t i = 0; i < reference_frames.size(); ++i) {
 984         RefPicList0_P[i] = reference_frames[i].pic;
 985     }
 986     sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
 987 }
 988
 989 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
 990 {
 991     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
 992         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
 993     };
 994     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
 995         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
 996     };
 997
 998     for (size_t i = 0; i < reference_frames.size(); ++i) {
 999         RefPicList0_B[i] = reference_frames[i].pic;
1000         RefPicList1_B[i] = reference_frames[i].pic;
1001     }
1002     sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1003     sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1004 }
1005
1006
1007 int QuickSyncEncoderImpl::render_sequence()
1008 {
1009     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1010     VAStatus va_status;
1011     VAEncMiscParameterBuffer *misc_param;
1012     VAEncMiscParameterRateControl *misc_rate_ctrl;
1013
1014     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1015     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1016     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1017     seq_param.bits_per_second = frame_bitrate;
1018
1019     seq_param.intra_period = intra_period;
1020     seq_param.intra_idr_period = intra_idr_period;
1021     seq_param.ip_period = ip_period;
1022
1023     seq_param.max_num_ref_frames = num_ref_frames;
1024     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1025     seq_param.time_scale = TIMEBASE * 2;
1026     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1027     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1028     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1029     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1030     seq_param.seq_fields.bits.chroma_format_idc = 1;
1031     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1032
1033     if (frame_width != frame_width_mbaligned ||
1034         frame_height != frame_height_mbaligned) {
1035         seq_param.frame_cropping_flag = 1;
1036         seq_param.frame_crop_left_offset = 0;
1037         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1038         seq_param.frame_crop_top_offset = 0;
1039         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1040     }
1041
1042     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1043                                VAEncSequenceParameterBufferType,
1044                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1045     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1046
1047     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1048                                VAEncMiscParameterBufferType,
1049                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1050                                1, NULL, &rc_param_buf);
1051     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1052
1053     vaMapBuffer(va_dpy->va_dpy, rc_param_buf, (void **)&misc_param);
1054     misc_param->type = VAEncMiscParameterTypeRateControl;
1055     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1056     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1057     misc_rate_ctrl->bits_per_second = frame_bitrate;
1058     misc_rate_ctrl->target_percentage = 66;
1059     misc_rate_ctrl->window_size = 1000;
1060     misc_rate_ctrl->initial_qp = initial_qp;
1061     misc_rate_ctrl->min_qp = minimal_qp;
1062     misc_rate_ctrl->basic_unit_size = 0;
1063     vaUnmapBuffer(va_dpy->va_dpy, rc_param_buf);
1064
1065     render_id[0] = seq_param_buf;
1066     render_id[1] = rc_param_buf;
1067
1068     render_picture_and_delete(va_dpy->va_dpy, context_id, &render_id[0], 2);
1069
1070     return 0;
1071 }
1072
1073 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1074 {
1075     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1076     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1077     int PicOrderCntMsb, TopFieldOrderCnt;
1078
1079     if (frame_type == FRAME_IDR)
1080         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1081     else {
1082         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1083         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1084     }
1085
1086     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1087         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1088         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1089     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1090              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1091         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1092     else
1093         PicOrderCntMsb = prevPicOrderCntMsb;
1094
1095     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1096
1097     if (frame_type != FRAME_B) {
1098         PicOrderCntMsb_ref = PicOrderCntMsb;
1099         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1100     }
1101
1102     return TopFieldOrderCnt;
1103 }
1104
1105 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1106 {
1107     VABufferID pic_param_buf;
1108     VAStatus va_status;
1109     size_t i = 0;
1110
1111     pic_param.CurrPic.picture_id = surf->ref_surface;
1112     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1113     pic_param.CurrPic.flags = 0;
1114     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1115     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1116     CurrentCurrPic = pic_param.CurrPic;
1117
1118     for (i = 0; i < reference_frames.size(); i++) {
1119         pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1120     }
1121     for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1122         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1123         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1124     }
1125
1126     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1127     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1128     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1129     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1130     pic_param.frame_num = current_ref_frame_num;  // FIXME: is this correct?
1131     pic_param.coded_buf = surf->coded_buf;
1132     pic_param.last_picture = false;  // FIXME
1133     pic_param.pic_init_qp = initial_qp;
1134
1135     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncPictureParameterBufferType,
1136                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1137     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1138
1139     render_picture_and_delete(va_dpy->va_dpy, context_id, &pic_param_buf, 1);
1140
1141     return 0;
1142 }
1143
1144 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1145 {
1146     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1147     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1148     unsigned int length_in_bits;
1149     unsigned char *packedseq_buffer = NULL;
1150     VAStatus va_status;
1151
1152     length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1153
1154     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1155
1156     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1157     packedheader_param_buffer.has_emulation_bytes = 0;
1158     va_status = vaCreateBuffer(va_dpy->va_dpy,
1159                                context_id,
1160                                VAEncPackedHeaderParameterBufferType,
1161                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1162                                &packedseq_para_bufid);
1163     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1164
1165     va_status = vaCreateBuffer(va_dpy->va_dpy,
1166                                context_id,
1167                                VAEncPackedHeaderDataBufferType,
1168                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1169                                &packedseq_data_bufid);
1170     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1171
1172     render_id[0] = packedseq_para_bufid;
1173     render_id[1] = packedseq_data_bufid;
1174     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1175
1176     free(packedseq_buffer);
1177
1178     return 0;
1179 }
1180
1181
1182 int QuickSyncEncoderImpl::render_packedpicture()
1183 {
1184     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1185     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1186     unsigned int length_in_bits;
1187     unsigned char *packedpic_buffer = NULL;
1188     VAStatus va_status;
1189
1190     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1191     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1192     packedheader_param_buffer.bit_length = length_in_bits;
1193     packedheader_param_buffer.has_emulation_bytes = 0;
1194
1195     va_status = vaCreateBuffer(va_dpy->va_dpy,
1196                                context_id,
1197                                VAEncPackedHeaderParameterBufferType,
1198                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1199                                &packedpic_para_bufid);
1200     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1201
1202     va_status = vaCreateBuffer(va_dpy->va_dpy,
1203                                context_id,
1204                                VAEncPackedHeaderDataBufferType,
1205                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1206                                &packedpic_data_bufid);
1207     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1208
1209     render_id[0] = packedpic_para_bufid;
1210     render_id[1] = packedpic_data_bufid;
1211     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1212
1213     free(packedpic_buffer);
1214
1215     return 0;
1216 }
1217
1218 void QuickSyncEncoderImpl::render_packedslice()
1219 {
1220     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1221     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1222     unsigned int length_in_bits;
1223     unsigned char *packedslice_buffer = NULL;
1224     VAStatus va_status;
1225
1226     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1227     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1228     packedheader_param_buffer.bit_length = length_in_bits;
1229     packedheader_param_buffer.has_emulation_bytes = 0;
1230
1231     va_status = vaCreateBuffer(va_dpy->va_dpy,
1232                                context_id,
1233                                VAEncPackedHeaderParameterBufferType,
1234                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1235                                &packedslice_para_bufid);
1236     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1237
1238     va_status = vaCreateBuffer(va_dpy->va_dpy,
1239                                context_id,
1240                                VAEncPackedHeaderDataBufferType,
1241                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1242                                &packedslice_data_bufid);
1243     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1244
1245     render_id[0] = packedslice_para_bufid;
1246     render_id[1] = packedslice_data_bufid;
1247     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1248
1249     free(packedslice_buffer);
1250 }
1251
1252 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1253 {
1254     VABufferID slice_param_buf;
1255     VAStatus va_status;
1256     int i;
1257
1258     /* one frame, one slice */
1259     slice_param.macroblock_address = 0;
1260     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1261     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1262     if (frame_type == FRAME_IDR) {
1263         if (encoding_frame_num != 0)
1264             ++slice_param.idr_pic_id;
1265     } else if (frame_type == FRAME_P) {
1266         VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1267         update_RefPicList_P(RefPicList0_P);
1268
1269         int refpiclist0_max = h264_maxref & 0xffff;
1270         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1271
1272         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1273             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1274             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1275         }
1276     } else if (frame_type == FRAME_B) {
1277         VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1278         update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1279
1280         int refpiclist0_max = h264_maxref & 0xffff;
1281         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1282
1283         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1284         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1285             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1286             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1287         }
1288
1289         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1290         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1291             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1292             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1293         }
1294     }
1295
1296     slice_param.slice_alpha_c0_offset_div2 = 0;
1297     slice_param.slice_beta_offset_div2 = 0;
1298     slice_param.direct_spatial_mv_pred_flag = 1;
1299     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1300
1301
1302     if (h264_packedheader &&
1303         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1304         render_packedslice();
1305
1306     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncSliceParameterBufferType,
1307                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1308     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1309
1310     render_picture_and_delete(va_dpy->va_dpy, context_id, &slice_param_buf, 1);
1311
1312     return 0;
1313 }
1314
1315
1316
1317 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1318 {
1319         VACodedBufferSegment *buf_list = NULL;
1320         VAStatus va_status;
1321
1322         string data;
1323
1324         va_status = vaMapBuffer(va_dpy->va_dpy, surf->coded_buf, (void **)(&buf_list));
1325         CHECK_VASTATUS(va_status, "vaMapBuffer");
1326         while (buf_list != NULL) {
1327                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1328                 buf_list = (VACodedBufferSegment *) buf_list->next;
1329         }
1330         vaUnmapBuffer(va_dpy->va_dpy, surf->coded_buf);
1331
1332         static int frameno = 0;
1333         print_latency("Current Quick Sync latency (video inputs → disk mux):",
1334                 task.received_ts, (task.frame_type == FRAME_B), &frameno, &qs_latency_histogram);
1335
1336         {
1337                 // Add video.
1338                 AVPacket pkt;
1339                 memset(&pkt, 0, sizeof(pkt));
1340                 pkt.buf = nullptr;
1341                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1342                 pkt.size = data.size();
1343                 pkt.stream_index = 0;
1344                 if (task.frame_type == FRAME_IDR) {
1345                         pkt.flags = AV_PKT_FLAG_KEY;
1346                 } else {
1347                         pkt.flags = 0;
1348                 }
1349                 pkt.duration = task.duration;
1350                 if (file_mux) {
1351                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1352                 }
1353                 if (!global_flags.x264_video_to_http &&
1354                     !global_flags.av1_video_to_http) {
1355                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1356                 }
1357         }
1358 }
1359
1360
1361 // this is weird. but it seems to put a new frame onto the queue
1362 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1363 {
1364         assert(task.pts >= task.dts);
1365         lock_guard<mutex> lock(storage_task_queue_mutex);
1366         storage_task_queue.push(move(task));
1367         storage_task_queue_changed.notify_all();
1368 }
1369
1370 void QuickSyncEncoderImpl::storage_task_thread()
1371 {
1372         pthread_setname_np(pthread_self(), "QS_Storage");
1373         for ( ;; ) {
1374                 storage_task current;
1375                 GLSurface *surf;
1376                 {
1377                         // wait until there's an encoded frame
1378                         unique_lock<mutex> lock(storage_task_queue_mutex);
1379                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1380                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1381                         current = move(storage_task_queue.front());
1382                         storage_task_queue.pop();
1383                         surf = surface_for_frame[current.display_order];
1384                         assert(surf != nullptr);
1385                 }
1386
1387                 VAStatus va_status;
1388
1389                 size_t display_order = current.display_order;
1390                 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1391
1392                 // waits for data, then saves it to disk.
1393                 va_status = vaSyncSurface(va_dpy->va_dpy, surf->src_surface);
1394                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1395                 save_codeddata(surf, move(current));
1396
1397                 // Unlock the frame, and all its references.
1398                 {
1399                         lock_guard<mutex> lock(storage_task_queue_mutex);
1400                         release_gl_surface(display_order);
1401
1402                         for (size_t frame_num : ref_display_frame_numbers) {
1403                                 release_gl_surface(frame_num);
1404                         }
1405                 }
1406         }
1407 }
1408
1409 void QuickSyncEncoderImpl::release_encode()
1410 {
1411         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1412                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, gl_surfaces[i].coded_buf);
1413                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
1414                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].src_surface, 1);
1415                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1416                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].ref_surface, 1);
1417                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1418         }
1419
1420         VAStatus va_status = vaDestroyContext(va_dpy->va_dpy, context_id);
1421         CHECK_VASTATUS(va_status, "vaDestroyContext");
1422         va_status = vaDestroyConfig(va_dpy->va_dpy, config_id);
1423         CHECK_VASTATUS(va_status, "vaDestroyConfig");
1424 }
1425
1426 void QuickSyncEncoderImpl::release_gl_resources()
1427 {
1428         assert(is_shutdown);
1429         if (has_released_gl_resources) {
1430                 return;
1431         }
1432
1433         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1434                 if (use_zerocopy) {
1435                         resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1436                         resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1437                 } else {
1438                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1439                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1440                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1441                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1442                 }
1443         }
1444
1445         has_released_gl_resources = true;
1446 }
1447
1448 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, const AVOutputFormat *oformat, VideoCodecInterface *http_encoder, VideoCodecInterface *disk_encoder, DiskSpaceEstimator *disk_space_estimator)
1449         : current_storage_frame(0), resource_pool(resource_pool), surface(surface), http_encoder(http_encoder), disk_encoder(disk_encoder), frame_width(width), frame_height(height), disk_space_estimator(disk_space_estimator)
1450 {
1451         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1452         open_output_file(filename);
1453         file_audio_encoder->add_mux(file_mux.get());
1454
1455         frame_width_mbaligned = (frame_width + 15) & (~15);
1456         frame_height_mbaligned = (frame_height + 15) & (~15);
1457
1458         //print_input();
1459
1460         if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1461                 assert(http_encoder != nullptr);
1462                 assert(disk_encoder != nullptr);
1463         } else if (global_flags.av1_video_to_http) {
1464                 assert(http_encoder != nullptr);
1465         } else {
1466                 assert(http_encoder == nullptr);
1467                 assert(disk_encoder == nullptr);
1468         }
1469
1470         enable_zerocopy_if_possible();
1471         if (!global_flags.x264_video_to_disk) {
1472                 init_va(va_display);
1473         }
1474         setup_encode();
1475
1476         if (!global_flags.x264_video_to_disk) {
1477                 memset(&seq_param, 0, sizeof(seq_param));
1478                 memset(&pic_param, 0, sizeof(pic_param));
1479                 memset(&slice_param, 0, sizeof(slice_param));
1480         }
1481
1482         if (!global_flags.v4l_output_device.empty()) {
1483                 v4l_output.reset(new V4LOutput(global_flags.v4l_output_device.c_str(), width, height));
1484         }
1485
1486         call_once(quick_sync_metrics_inited, [](){
1487                 mixer_latency_histogram.init("mixer");
1488                 qs_latency_histogram.init("quick_sync");
1489                 current_file_mux_metrics.init({{ "destination", "current_file" }});
1490                 total_mux_metrics.init({{ "destination", "files_total" }});
1491                 global_metrics.add("current_file_start_time_seconds", &metric_current_file_start_time_seconds, Metrics::TYPE_GAUGE);
1492                 global_metrics.add("quick_sync_stalled_frames", &metric_quick_sync_stalled_frames);
1493         });
1494
1495         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1496
1497         encode_thread = thread([this]{
1498                 QOpenGLContext *context = create_context(this->surface);
1499                 eglBindAPI(EGL_OPENGL_API);
1500                 if (!make_current(context, this->surface)) {
1501                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1502                                 eglGetError());
1503                         abort();
1504                 }
1505                 encode_thread_func();
1506                 delete_context(context);
1507         });
1508 }
1509
1510 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1511 {
1512         shutdown();
1513         release_gl_resources();
1514 }
1515
1516 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1517 {
1518         for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1519                 if (gl_surfaces[i].refcount == 0) {
1520                         ++gl_surfaces[i].refcount;
1521                         return &gl_surfaces[i];
1522                 }
1523         }
1524         return nullptr;
1525 }
1526
1527 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1528 {
1529         assert(surface_for_frame.count(display_frame_num));
1530         QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1531         if (--surf->refcount == 0) {
1532                 assert(surface_for_frame.count(display_frame_num));
1533                 surface_for_frame.erase(display_frame_num);
1534                 storage_task_queue_changed.notify_all();
1535         }
1536 }
1537
1538 bool QuickSyncEncoderImpl::is_zerocopy() const
1539 {
1540         return use_zerocopy;
1541 }
1542
1543 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1544 {
1545         assert(!is_shutdown);
1546         GLSurface *surf = nullptr;
1547         {
1548                 // Wait until this frame slot is done encoding.
1549                 unique_lock<mutex> lock(storage_task_queue_mutex);
1550                 surf = allocate_gl_surface();
1551                 if (surf == nullptr) {
1552                         fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1553                                 current_storage_frame);
1554                         ++metric_quick_sync_stalled_frames;
1555                         storage_task_queue_changed.wait(lock, [this, &surf]{
1556                                 if (storage_thread_should_quit)
1557                                         return true;
1558                                 surf = allocate_gl_surface();
1559                                 return surf != nullptr;
1560                         });
1561                 }
1562                 if (storage_thread_should_quit) return false;
1563                 assert(surf != nullptr);
1564                 surface_for_frame[current_storage_frame] = surf;
1565         }
1566
1567         if (use_zerocopy) {
1568                 *y_tex = surf->y_tex;
1569                 *cbcr_tex = surf->cbcr_tex;
1570         } else {
1571                 surf->y_tex = *y_tex;
1572                 surf->cbcr_tex = *cbcr_tex;
1573         }
1574
1575         if (!global_flags.x264_video_to_disk) {
1576                 VAStatus va_status = vaDeriveImage(va_dpy->va_dpy, surf->src_surface, &surf->surface_image);
1577                 CHECK_VASTATUS(va_status, "vaDeriveImage");
1578
1579                 if (use_zerocopy) {
1580                         VABufferInfo buf_info;
1581                         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1582                         va_status = vaAcquireBufferHandle(va_dpy->va_dpy, surf->surface_image.buf, &buf_info);
1583                         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1584
1585                         // Create Y image.
1586                         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1587                         EGLint y_attribs[] = {
1588                                 EGL_WIDTH, frame_width,
1589                                 EGL_HEIGHT, frame_height,
1590                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1591                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1592                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1593                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1594                                 EGL_NONE
1595                         };
1596
1597                         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1598                         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1599
1600                         // Associate Y image to a texture.
1601                         glBindTexture(GL_TEXTURE_2D, *y_tex);
1602                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1603
1604                         // Create CbCr image.
1605                         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1606                         EGLint cbcr_attribs[] = {
1607                                 EGL_WIDTH, frame_width / 2,
1608                                 EGL_HEIGHT, frame_height / 2,
1609                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1610                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1611                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1612                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1613                                 EGL_NONE
1614                         };
1615
1616                         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1617                         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1618
1619                         // Associate CbCr image to a texture.
1620                         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1621                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1622                 }
1623         }
1624
1625         current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1626
1627         return true;
1628 }
1629
1630 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1631 {
1632         lock_guard<mutex> lock(file_audio_encoder_mutex);
1633         assert(!is_shutdown);
1634         file_audio_encoder->encode_audio(audio, pts + global_delay());
1635 }
1636
1637 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1638 {
1639         assert(!is_shutdown);
1640
1641         if (!use_zerocopy) {
1642                 GLenum type = global_flags.bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1643                 GLSurface *surf;
1644                 {
1645                         lock_guard<mutex> lock(storage_task_queue_mutex);
1646                         surf = surface_for_frame[current_storage_frame];
1647                         assert(surf != nullptr);
1648                 }
1649
1650                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1651                 check_error();
1652
1653                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1654                 check_error();
1655
1656                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1657                 check_error();
1658                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1659                 check_error();
1660
1661                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1662                 check_error();
1663                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1664                 check_error();
1665
1666                 // We don't own these; the caller does.
1667                 surf->y_tex = surf->cbcr_tex = 0;
1668
1669                 glBindTexture(GL_TEXTURE_2D, 0);
1670                 check_error();
1671                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1672                 check_error();
1673
1674                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1675                 check_error();
1676         }
1677
1678         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1679         check_error();
1680         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1681         check_error();
1682
1683         {
1684                 lock_guard<mutex> lock(frame_queue_mutex);
1685                 current_video_frame.fence = fence;
1686                 pending_video_frames.push(move(current_video_frame));
1687                 ++current_storage_frame;
1688         }
1689         frame_queue_nonempty.notify_all();
1690         return fence;
1691 }
1692
1693 void QuickSyncEncoderImpl::shutdown()
1694 {
1695         if (is_shutdown) {
1696                 return;
1697         }
1698
1699         {
1700                 lock_guard<mutex> lock(frame_queue_mutex);
1701                 encode_thread_should_quit = true;
1702                 frame_queue_nonempty.notify_all();
1703         }
1704         encode_thread.join();
1705         {
1706                 lock_guard<mutex> lock(storage_task_queue_mutex);
1707                 storage_thread_should_quit = true;
1708                 frame_queue_nonempty.notify_all();
1709                 storage_task_queue_changed.notify_all();
1710         }
1711         storage_thread.join();
1712
1713         // Encode any leftover audio in the queues, and also any delayed frames.
1714         {
1715                 lock_guard<mutex> lock(file_audio_encoder_mutex);
1716                 file_audio_encoder->encode_last_audio();
1717         }
1718
1719         if (!global_flags.x264_video_to_disk) {
1720                 release_encode();
1721                 va_dpy.reset();
1722         }
1723         is_shutdown = true;
1724 }
1725
1726 void QuickSyncEncoderImpl::close_file()
1727 {
1728         file_mux.reset();
1729         metric_current_file_start_time_seconds = 0.0 / 0.0;
1730 }
1731
1732 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1733 {
1734         AVFormatContext *avctx = avformat_alloc_context();
1735         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1736         avctx->url = strdup(filename.c_str());
1737
1738         string url = "file:" + filename;
1739         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1740         if (ret < 0) {
1741                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1742                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1743                 abort();
1744         }
1745
1746         string video_extradata;  // FIXME: See other comment about global headers.
1747         if (global_flags.x264_video_to_disk) {
1748                 video_extradata = disk_encoder->get_global_headers();
1749         }
1750
1751         current_file_mux_metrics.reset();
1752
1753         {
1754                 lock_guard<mutex> lock(file_audio_encoder_mutex);
1755                 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1756                 file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), get_color_space(global_flags.ycbcr_rec709_coefficients), TIMEBASE,
1757                         std::bind(&DiskSpaceEstimator::report_append, disk_space_estimator, filename, _1),
1758                         Mux::WRITE_BACKGROUND,
1759                         { &current_file_mux_metrics, &total_mux_metrics }));
1760         }
1761         metric_current_file_start_time_seconds = get_timestamp_for_metrics();
1762
1763         if (global_flags.x264_video_to_disk) {
1764                 disk_encoder->add_mux(file_mux.get());
1765         }
1766 }
1767
1768 void QuickSyncEncoderImpl::encode_thread_func()
1769 {
1770         pthread_setname_np(pthread_self(), "QS_Encode");
1771
1772         DTSReorderer dts_reorder_buf(ip_period - 1);
1773
1774         int64_t last_dts = -1;
1775         int gop_start_display_frame_num = 0;
1776         for (int display_frame_num = 0; ; ++display_frame_num) {
1777                 // Wait for the frame to be in the queue. Note that this only means
1778                 // we started rendering it.
1779                 PendingFrame frame;
1780                 {
1781                         unique_lock<mutex> lock(frame_queue_mutex);
1782                         frame_queue_nonempty.wait(lock, [this]{
1783                                 return encode_thread_should_quit || !pending_video_frames.empty();
1784                         });
1785                         if (encode_thread_should_quit && pending_video_frames.empty()) {
1786                                 // We may have queued frames left in the reorder buffer
1787                                 // that were supposed to be B-frames, but have no P-frame
1788                                 // to be encoded against. If so, encode them all as
1789                                 // P-frames instead. Note that this happens under the mutex,
1790                                 // but nobody else uses it at this point, since we're shutting down,
1791                                 // so there's no contention.
1792                                 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1793                                 return;
1794                         } else {
1795                                 frame = move(pending_video_frames.front());
1796                                 pending_video_frames.pop();
1797                         }
1798                 }
1799
1800                 dts_reorder_buf.push_pts(frame.pts);
1801
1802                 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1803                 // Note that this implicitly waits for the frame to be done rendering.
1804                 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1805
1806                 if (global_flags.x264_video_to_disk) {
1807                         lock_guard<mutex> lock(storage_task_queue_mutex);
1808                         release_gl_surface(display_frame_num);
1809                         continue;
1810                 }
1811
1812                 reorder_buffer[display_frame_num] = move(frame);
1813
1814                 // Now encode as many QuickSync frames as we can using the frames we have available.
1815                 // (It could be zero, or it could be multiple.) FIXME: make a function.
1816                 for ( ;; ) {
1817                         int frame_type, quicksync_display_frame_num;
1818                         encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1819                                                &quicksync_display_frame_num, &frame_type);
1820                         if (!reorder_buffer.count(quicksync_display_frame_num)) {
1821                                 break;
1822                         }
1823                         frame = move(reorder_buffer[quicksync_display_frame_num]);
1824                         reorder_buffer.erase(quicksync_display_frame_num);
1825
1826                         if (frame_type == FRAME_IDR) {
1827                                 // Release any reference frames from the previous GOP.
1828                                 {
1829                                         lock_guard<mutex> lock(storage_task_queue_mutex);
1830                                         for (const ReferenceFrame &frame : reference_frames) {
1831                                                 release_gl_surface(frame.display_number);
1832                                         }
1833                                 }
1834                                 reference_frames.clear();
1835                                 current_ref_frame_num = 0;
1836                                 gop_start_display_frame_num = quicksync_display_frame_num;
1837                         }
1838
1839                         const int64_t dts = dts_reorder_buf.pop_dts();
1840                         last_dts = dts;
1841
1842                         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);
1843                         ++quicksync_encoding_frame_num;
1844                 }
1845         }
1846 }
1847
1848 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1849 {
1850         if (reorder_buffer.empty()) {
1851                 return;
1852         }
1853
1854         for (auto &pending_frame : reorder_buffer) {
1855                 int display_frame_num = pending_frame.first;
1856                 assert(display_frame_num > 0);
1857                 PendingFrame frame = move(pending_frame.second);
1858                 int64_t dts = last_dts + (TIMEBASE / TYPICAL_FPS);
1859                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1860                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1861                 last_dts = dts;
1862         }
1863 }
1864
1865 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1866 {
1867         if (src_width == dst_pitch) {
1868                 memcpy(dst, src, src_width * height);
1869         } else {
1870                 for (size_t y = 0; y < height; ++y) {
1871                         const uint8_t *sptr = src + y * src_width;
1872                         uint8_t *dptr = dst + y * dst_pitch;
1873                         memcpy(dptr, sptr, src_width);
1874                 }
1875         }
1876 }
1877
1878 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1879 {
1880         // Wait for the GPU to be done with the frame.
1881         GLenum sync_status;
1882         do {
1883                 sync_status = glClientWaitSync(frame.fence.get(), 0, 0);
1884                 check_error();
1885                 if (sync_status == GL_TIMEOUT_EXPIRED) {
1886                         // NVIDIA likes to busy-wait; yield instead.
1887                         this_thread::sleep_for(milliseconds(1));
1888                 }
1889         } while (sync_status == GL_TIMEOUT_EXPIRED);
1890         assert(sync_status != GL_WAIT_FAILED);
1891
1892         ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1893         static int frameno = 0;
1894         print_latency("Current mixer latency (video inputs → ready for encode):",
1895                 received_ts, false, &frameno, &mixer_latency_histogram);
1896
1897         // Release back any input frames we needed to render this frame.
1898         frame.input_frames.clear();
1899
1900         GLSurface *surf;
1901         {
1902                 lock_guard<mutex> lock(storage_task_queue_mutex);
1903                 surf = surface_for_frame[display_frame_num];
1904                 assert(surf != nullptr);
1905         }
1906         uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1907         if (http_encoder != nullptr) {
1908                 http_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1909         } if (disk_encoder != nullptr && disk_encoder != http_encoder) {
1910                 disk_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1911         }
1912
1913         if (v4l_output != nullptr) {
1914                 v4l_output->send_frame(data);
1915         }
1916 }
1917
1918 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1919                                         int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
1920 {
1921         const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1922
1923         GLSurface *surf;
1924         {
1925                 lock_guard<mutex> lock(storage_task_queue_mutex);
1926                 surf = surface_for_frame[display_frame_num];
1927                 assert(surf != nullptr);
1928         }
1929         VAStatus va_status;
1930
1931         if (use_zerocopy) {
1932                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1933                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1934                 va_status = vaReleaseBufferHandle(va_dpy->va_dpy, surf->surface_image.buf);
1935                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1936         } else {
1937                 // Upload the frame to VA-API.
1938                 unsigned char *surface_p = nullptr;
1939                 vaMapBuffer(va_dpy->va_dpy, surf->surface_image.buf, (void **)&surface_p);
1940
1941                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1942                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1943
1944                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1945                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1946
1947                 va_status = vaUnmapBuffer(va_dpy->va_dpy, surf->surface_image.buf);
1948                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1949         }
1950
1951         va_status = vaDestroyImage(va_dpy->va_dpy, surf->surface_image.image_id);
1952         CHECK_VASTATUS(va_status, "vaDestroyImage");
1953
1954         // Schedule the frame for encoding.
1955         VASurfaceID va_surface = surf->src_surface;
1956         va_status = vaBeginPicture(va_dpy->va_dpy, context_id, va_surface);
1957         CHECK_VASTATUS(va_status, "vaBeginPicture");
1958
1959         if (frame_type == FRAME_IDR) {
1960                 // FIXME: If the mux wants global headers, we should not put the
1961                 // SPS/PPS before each IDR frame, but rather put it into the
1962                 // codec extradata (formatted differently?).
1963                 //
1964                 // NOTE: If we change ycbcr_coefficients, it will not take effect
1965                 // before the next IDR frame. This is acceptable, as it should only
1966                 // happen on a mode change, which is rare.
1967                 render_sequence();
1968                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
1969                 if (h264_packedheader) {
1970                         render_packedsequence(ycbcr_coefficients);
1971                         render_packedpicture();
1972                 }
1973         } else {
1974                 //render_sequence();
1975                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
1976         }
1977         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1978
1979         va_status = vaEndPicture(va_dpy->va_dpy, context_id);
1980         CHECK_VASTATUS(va_status, "vaEndPicture");
1981
1982         update_ReferenceFrames(display_frame_num, frame_type);
1983
1984         vector<size_t> ref_display_frame_numbers;
1985
1986         // Lock the references for this frame; otherwise, they could be
1987         // rendered to before this frame is done encoding.
1988         {
1989                 lock_guard<mutex> lock(storage_task_queue_mutex);
1990                 for (const ReferenceFrame &frame : reference_frames) {
1991                         assert(surface_for_frame.count(frame.display_number));
1992                         ++surface_for_frame[frame.display_number]->refcount;
1993                         ref_display_frame_numbers.push_back(frame.display_number);
1994                 }
1995         }
1996
1997         // so now the data is done encoding (well, async job kicked off)...
1998         // we send that to the storage thread
1999         storage_task tmp;
2000         tmp.display_order = display_frame_num;
2001         tmp.frame_type = frame_type;
2002         tmp.pts = pts;
2003         tmp.dts = dts;
2004         tmp.duration = duration;
2005         tmp.ycbcr_coefficients = ycbcr_coefficients;
2006         tmp.received_ts = received_ts;
2007         tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2008         storage_task_enqueue(move(tmp));
2009 }
2010
2011 // Proxy object.
2012 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, const AVOutputFormat *oformat, VideoCodecInterface *http_encoder, VideoCodecInterface *disk_encoder, DiskSpaceEstimator *disk_space_estimator)
2013         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, http_encoder, disk_encoder, disk_space_estimator)) {}
2014
2015 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2016 QuickSyncEncoder::~QuickSyncEncoder() {}
2017
2018 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2019 {
2020         impl->add_audio(pts, audio);
2021 }
2022
2023 bool QuickSyncEncoder::is_zerocopy() const
2024 {
2025         return impl->is_zerocopy();
2026 }
2027
2028 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2029 {
2030         return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2031 }
2032
2033 RefCountedGLsync QuickSyncEncoder::end_frame()
2034 {
2035         return impl->end_frame();
2036 }
2037
2038 void QuickSyncEncoder::shutdown()
2039 {
2040         impl->shutdown();
2041 }
2042
2043 void QuickSyncEncoder::close_file()
2044 {
2045         impl->shutdown();
2046 }
2047
2048 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2049 {
2050         impl->set_stream_mux(mux);
2051 }
2052
2053 int64_t QuickSyncEncoder::global_delay() const {
2054         return impl->global_delay();
2055 }
2056
2057 string QuickSyncEncoder::get_usable_va_display()
2058 {
2059         // Reduce the amount of chatter while probing,
2060         // unless the user has specified otherwise.
2061         bool need_env_reset = false;
2062         if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
2063                 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
2064                 need_env_reset = true;
2065         }
2066
2067         // First try the default (ie., whatever $DISPLAY is set to).
2068         unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va_h264("", nullptr, nullptr);
2069         if (va_dpy != nullptr) {
2070                 if (need_env_reset) {
2071                         unsetenv("LIBVA_MESSAGING_LEVEL");
2072                 }
2073                 return "";
2074         }
2075
2076         fprintf(stderr, "No --va-display was given, and the X11 display did not expose a VA-API H.264 encoder.\n");
2077
2078         // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
2079         glob_t g;
2080         int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
2081         if (err != 0) {
2082                 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
2083         } else {
2084                 for (size_t i = 0; i < g.gl_pathc; ++i) {
2085                         string path = g.gl_pathv[i];
2086                         va_dpy = try_open_va_h264(path, nullptr, nullptr);
2087                         if (va_dpy != nullptr) {
2088                                 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
2089                                         path.c_str());
2090                                 globfree(&g);
2091                                 if (need_env_reset) {
2092                                         unsetenv("LIBVA_MESSAGING_LEVEL");
2093                                 }
2094                                 return path;
2095                         }
2096                 }
2097         }
2098
2099         fprintf(stderr, "No suitable VA-API H.264 encoders were found in /dev/dri; giving up.\n");
2100         fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
2101         fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
2102         fprintf(stderr, "to expose Quick Sync. Alternatively, you can use --record-x264-video\n");
2103         fprintf(stderr, "to use software instead of hardware H.264 encoding, at the expense\n");
2104         fprintf(stderr, "of increased CPU usage and possibly bit rate.\n");
2105         abort();
2106 }