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 <libdrm/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 // General pts/dts strategy:
 596 //
 597 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 598 // bit tricky. We assume first of all that the frame rate never goes _above_
 599 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 600 // in at least this speed, as long at dts <= pts (the frame is not attempted
 601 // presented before it is decoded). Furthermore, we never have longer chains of
 602 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 603 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 604 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 605 // frame, with an order O <= C.)
 606 //
 607 // Many strategies are possible, but we establish these rules:
 608 //
 609 //  - Tip frames have dts = pts - (C-O)*N.
 610 //  - Non-tip frames have dts = dts_last + N.
 611 //
 612 // An example, with C=2 and N=10 and the data flow showed with arrows:
 613 //
 614 //        I  B  P  B  B  P
 615 //   pts: 30 40 50 60 70 80
 616 //        ↓  ↓     ↓
 617 //   dts: 10 30 20 60 50←40
 618 //         |  |  ↑        ↑
 619 //         `--|--'        |
 620 //             `----------'
 621 //
 622 // To show that this works fine also with irregular spacings, let's say that
 623 // the third frame is delayed a bit (something earlier was dropped). Now the
 624 // situation looks like this:
 625 //
 626 //        I  B  P  B  B   P
 627 //   pts: 30 40 80 90 100 110
 628 //        ↓  ↓     ↓
 629 //   dts: 10 30 20 90 50←40
 630 //         |  |  ↑        ↑
 631 //         `--|--'        |
 632 //             `----------'
 633 //
 634 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 635 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 636 //
 637 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 638 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 639 // a tip frame and should be given a dts based on the previous one.
 640 #define FRAME_P 0
 641 #define FRAME_B 1
 642 #define FRAME_I 2
 643 #define FRAME_IDR 7
 644 void encoding2display_order(
 645     int encoding_order, int intra_period,
 646     int intra_idr_period, int ip_period,
 647     int *displaying_order,
 648     int *frame_type, int *pts_lag)
 649 {
 650     int encoding_order_gop = 0;
 651
 652     *pts_lag = 0;
 653
 654     if (intra_period == 1) { /* all are I/IDR frames */
 655         *displaying_order = encoding_order;
 656         if (intra_idr_period == 0)
 657             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 658         else
 659             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 660         return;
 661     }
 662
 663     if (intra_period == 0)
 664         intra_idr_period = 0;
 665
 666     if (ip_period == 1) {
 667         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 668         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 669         *displaying_order = encoding_order;
 670
 671         if (encoding_order_gop == 0) { /* the first frame */
 672             *frame_type = FRAME_IDR;
 673         } else if (intra_period != 0 && /* have I frames */
 674                    encoding_order_gop >= 2 &&
 675                    (encoding_order_gop % intra_period == 0)) {
 676             *frame_type = FRAME_I;
 677         } else {
 678             *frame_type = FRAME_P;
 679         }
 680         return;
 681     }
 682
 683     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 684     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 685     *pts_lag = -1;  // Most frames are not tip frames.
 686
 687     if (encoding_order_gop == 0) { /* the first frame */
 688         *frame_type = FRAME_IDR;
 689         *displaying_order = encoding_order;
 690         // IDR frames are a special case; I honestly can't find the logic behind
 691         // why this is the right thing, but it seems to line up nicely in practice :-)
 692         *pts_lag = TIMEBASE / MAX_FPS;
 693     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 694         *frame_type = FRAME_B;
 695         *displaying_order = encoding_order - 1;
 696         if ((encoding_order_gop % ip_period) == 0) {
 697             *pts_lag = 0;  // Last B-frame.
 698         }
 699     } else if (intra_period != 0 && /* have I frames */
 700                encoding_order_gop >= 2 &&
 701                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 702         *frame_type = FRAME_I;
 703         *displaying_order = encoding_order + ip_period - 1;
 704     } else {
 705         *frame_type = FRAME_P;
 706         *displaying_order = encoding_order + ip_period - 1;
 707     }
 708 }
 709
 710
 711 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 712 {
 713         if (global_flags.x264_video_to_disk) {
 714                 // Quick Sync is entirely disabled.
 715                 use_zerocopy = false;
 716         } else if (global_flags.uncompressed_video_to_http) {
 717                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 718                 use_zerocopy = false;
 719         } else if (global_flags.x264_video_to_http) {
 720                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 721                 use_zerocopy = false;
 722         } else if (!global_flags.v4l_output_device.empty()) {
 723                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --v4l-output.\n");
 724                 use_zerocopy = false;
 725         } else {
 726                 use_zerocopy = true;
 727         }
 728         global_flags.use_zerocopy = use_zerocopy;
 729 }
 730
 731 VADisplayWithCleanup::~VADisplayWithCleanup()
 732 {
 733         if (va_dpy != nullptr) {
 734                 vaTerminate(va_dpy);
 735         }
 736         if (x11_display != nullptr) {
 737                 XCloseDisplay(x11_display);
 738         }
 739         if (drm_fd != -1) {
 740                 close(drm_fd);
 741         }
 742 }
 743
 744 unique_ptr<VADisplayWithCleanup> va_open_display(const string &va_display)
 745 {
 746         if (va_display.empty() || va_display[0] != '/') {  // An X display.
 747                 Display *x11_display = XOpenDisplay(va_display.empty() ? nullptr : va_display.c_str());
 748                 if (x11_display == nullptr) {
 749                         fprintf(stderr, "error: can't connect to X server!\n");
 750                         return nullptr;
 751                 }
 752
 753                 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
 754                 ret->x11_display = x11_display;
 755                 ret->can_use_zerocopy = true;
 756                 ret->va_dpy = vaGetDisplay(x11_display);
 757                 if (ret->va_dpy == nullptr) {
 758                         return nullptr;
 759                 }
 760                 return ret;
 761         } else {  // A DRM node on the filesystem (e.g. /dev/dri/renderD128).
 762                 int drm_fd = open(va_display.c_str(), O_RDWR);
 763                 if (drm_fd == -1) {
 764                         perror(va_display.c_str());
 765                         return NULL;
 766                 }
 767                 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
 768                 ret->drm_fd = drm_fd;
 769                 ret->can_use_zerocopy = false;
 770                 ret->va_dpy = vaGetDisplayDRM(drm_fd);
 771                 if (ret->va_dpy == nullptr) {
 772                         return nullptr;
 773                 }
 774                 return ret;
 775         }
 776 }
 777
 778 unique_ptr<VADisplayWithCleanup> try_open_va(const string &va_display, VAProfile *h264_profile, string *error)
 779 {
 780         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
 781         if (va_dpy == nullptr) {
 782                 if (error) *error = "Opening VA display failed";
 783                 return nullptr;
 784         }
 785         int major_ver, minor_ver;
 786         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
 787         if (va_status != VA_STATUS_SUCCESS) {
 788                 char buf[256];
 789                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
 790                 if (error != nullptr) *error = buf;
 791                 return nullptr;
 792         }
 793
 794         int num_entrypoints = vaMaxNumEntrypoints(va_dpy->va_dpy);
 795         unique_ptr<VAEntrypoint[]> entrypoints(new VAEntrypoint[num_entrypoints]);
 796         if (entrypoints == nullptr) {
 797                 if (error != nullptr) *error = "Failed to allocate memory for VA entry points";
 798                 return nullptr;
 799         }
 800
 801         // Try the profiles from highest to lowest until we find one that can be encoded.
 802         constexpr VAProfile profile_list[] = { VAProfileH264High, VAProfileH264Main, VAProfileH264ConstrainedBaseline };
 803         for (unsigned i = 0; i < sizeof(profile_list) / sizeof(profile_list[0]); ++i) {
 804                 vaQueryConfigEntrypoints(va_dpy->va_dpy, profile_list[i], entrypoints.get(), &num_entrypoints);
 805                 for (int slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 806                         if (entrypoints[slice_entrypoint] != VAEntrypointEncSlice) {
 807                                 continue;
 808                         }
 809
 810                         // We found a usable encoder, so return it.
 811                         if (h264_profile != nullptr) {
 812                                 *h264_profile = profile_list[i];
 813                         }
 814                         return va_dpy;
 815                 }
 816         }
 817
 818         if (error != nullptr) *error = "Can't find VAEntrypointEncSlice for H264 profiles";
 819         return nullptr;
 820 }
 821
 822 int QuickSyncEncoderImpl::init_va(const string &va_display)
 823 {
 824     string error;
 825     va_dpy = try_open_va(va_display, &h264_profile, &error);
 826     if (va_dpy == nullptr) {
 827         fprintf(stderr, "error: %s\n", error.c_str());
 828         abort();
 829     }
 830     if (!va_dpy->can_use_zerocopy) {
 831         use_zerocopy = false;
 832     }
 833
 834     switch (h264_profile) {
 835         case VAProfileH264ConstrainedBaseline:
 836             constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 837             ip_period = 1;
 838             break;
 839
 840         case VAProfileH264Main:
 841             constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 842             break;
 843
 844         case VAProfileH264High:
 845             constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 846             break;
 847         default:
 848             h264_profile = VAProfileH264ConstrainedBaseline;
 849             ip_period = 1;
 850             constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 851             break;
 852     }
 853
 854     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 855
 856     /* find out the format for the render target, and rate control mode */
 857     for (unsigned i = 0; i < VAConfigAttribTypeMax; i++)
 858         attrib[i].type = (VAConfigAttribType)i;
 859
 860     VAStatus va_status = vaGetConfigAttributes(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
 861                                       &attrib[0], VAConfigAttribTypeMax);
 862     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 863     /* check the interested configattrib */
 864     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 865         printf("Not find desired YUV420 RT format\n");
 866         abort();
 867     } else {
 868         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 869         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 870         config_attrib_num++;
 871     }
 872
 873     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 874         if (!(attrib[VAConfigAttribRateControl].value & VA_RC_CQP)) {
 875             fprintf(stderr, "ERROR: VA-API encoder does not support CQP mode.\n");
 876             abort();
 877         }
 878
 879         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 880         config_attrib[config_attrib_num].value = VA_RC_CQP;
 881         config_attrib_num++;
 882     }
 883
 884
 885     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 886         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 887
 888         h264_packedheader = 1;
 889         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 890         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 891
 892         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 893             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 894         }
 895
 896         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 897             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 898         }
 899
 900         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 901             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 902         }
 903
 904         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
 905             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
 906         }
 907
 908         enc_packed_header_idx = config_attrib_num;
 909         config_attrib_num++;
 910     }
 911
 912     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
 913         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
 914         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 915         config_attrib_num++;
 916     }
 917
 918     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
 919         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
 920     }
 921
 922     return 0;
 923 }
 924
 925 int QuickSyncEncoderImpl::setup_encode()
 926 {
 927         if (!global_flags.x264_video_to_disk) {
 928                 VAStatus va_status;
 929                 VASurfaceID *tmp_surfaceid;
 930                 int codedbuf_size;
 931                 VASurfaceID src_surface[SURFACE_NUM];
 932                 VASurfaceID ref_surface[SURFACE_NUM];
 933
 934                 va_status = vaCreateConfig(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
 935                                 &config_attrib[0], config_attrib_num, &config_id);
 936                 CHECK_VASTATUS(va_status, "vaCreateConfig");
 937
 938                 /* create source surfaces */
 939                 va_status = vaCreateSurfaces(va_dpy->va_dpy,
 940                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 941                                 &src_surface[0], SURFACE_NUM,
 942                                 NULL, 0);
 943                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 944
 945                 /* create reference surfaces */
 946                 va_status = vaCreateSurfaces(va_dpy->va_dpy,
 947                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 948                                 &ref_surface[0], SURFACE_NUM,
 949                                 NULL, 0);
 950                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 951
 952                 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
 953                 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
 954                 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
 955
 956                 for (int i = 0; i < SURFACE_NUM; i++) {
 957                         gl_surfaces[i].src_surface = src_surface[i];
 958                         gl_surfaces[i].ref_surface = ref_surface[i];
 959                 }
 960
 961                 /* Create a context for this encode pipe */
 962                 va_status = vaCreateContext(va_dpy->va_dpy, config_id,
 963                                 frame_width_mbaligned, frame_height_mbaligned,
 964                                 VA_PROGRESSIVE,
 965                                 tmp_surfaceid, 2 * SURFACE_NUM,
 966                                 &context_id);
 967                 CHECK_VASTATUS(va_status, "vaCreateContext");
 968                 free(tmp_surfaceid);
 969
 970                 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
 971
 972                 for (int i = 0; i < SURFACE_NUM; i++) {
 973                         /* create coded buffer once for all
 974                          * other VA buffers which won't be used again after vaRenderPicture.
 975                          * so APP can always vaCreateBuffer for every frame
 976                          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
 977                          * so VA won't maintain the coded buffer
 978                          */
 979                         va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncCodedBufferType,
 980                                         codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
 981                         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 982                 }
 983         }
 984
 985         /* create OpenGL objects */
 986         for (int i = 0; i < SURFACE_NUM; i++) {
 987                 if (use_zerocopy) {
 988                         gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
 989                         gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
 990                 } else {
 991                         size_t bytes_per_pixel = (global_flags.x264_bit_depth > 8) ? 2 : 1;
 992
 993                         // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
 994                         // buffers, due to potentially differing pitch.
 995                         glGenBuffers(1, &gl_surfaces[i].pbo);
 996                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
 997                         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);
 998                         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);
 999                         gl_surfaces[i].y_offset = 0;
1000                         gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
1001                         gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1002                         gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1003                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1004                 }
1005         }
1006
1007         return 0;
1008 }
1009
1010 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1011 //
1012 //   2 1 0 [3] 4 8 9
1013 template<class T, class C>
1014 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1015 {
1016         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1017         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1018         sort(middle, end, less_than);
1019 }
1020
1021 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
1022 {
1023     if (frame_type == FRAME_B)
1024         return;
1025
1026     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1027
1028     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1029     lock_guard<mutex> lock(storage_task_queue_mutex);
1030
1031     // Insert the new frame at the start of the reference queue.
1032     reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
1033
1034     if (reference_frames.size() > num_ref_frames)
1035     {
1036         // The back frame frame is no longer in use as a reference.
1037         int display_frame_num = reference_frames.back().display_number;
1038         assert(surface_for_frame.count(display_frame_num));
1039         release_gl_surface(display_frame_num);
1040         reference_frames.pop_back();
1041     }
1042
1043     // Mark this frame in use as a reference.
1044     assert(surface_for_frame.count(current_display_frame));
1045     ++surface_for_frame[current_display_frame]->refcount;
1046
1047     current_ref_frame_num++;
1048     if (current_ref_frame_num > MaxFrameNum)
1049         current_ref_frame_num = 0;
1050 }
1051
1052
1053 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
1054 {
1055     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1056         return a.frame_idx > b.frame_idx;
1057     };
1058
1059     for (size_t i = 0; i < reference_frames.size(); ++i) {
1060         RefPicList0_P[i] = reference_frames[i].pic;
1061     }
1062     sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
1063 }
1064
1065 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
1066 {
1067     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1068         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1069     };
1070     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1071         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1072     };
1073
1074     for (size_t i = 0; i < reference_frames.size(); ++i) {
1075         RefPicList0_B[i] = reference_frames[i].pic;
1076         RefPicList1_B[i] = reference_frames[i].pic;
1077     }
1078     sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1079     sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1080 }
1081
1082
1083 int QuickSyncEncoderImpl::render_sequence()
1084 {
1085     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1086     VAStatus va_status;
1087     VAEncMiscParameterBuffer *misc_param;
1088     VAEncMiscParameterRateControl *misc_rate_ctrl;
1089
1090     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1091     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1092     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1093     seq_param.bits_per_second = frame_bitrate;
1094
1095     seq_param.intra_period = intra_period;
1096     seq_param.intra_idr_period = intra_idr_period;
1097     seq_param.ip_period = ip_period;
1098
1099     seq_param.max_num_ref_frames = num_ref_frames;
1100     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1101     seq_param.time_scale = TIMEBASE * 2;
1102     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1103     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1104     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1105     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1106     seq_param.seq_fields.bits.chroma_format_idc = 1;
1107     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1108
1109     if (frame_width != frame_width_mbaligned ||
1110         frame_height != frame_height_mbaligned) {
1111         seq_param.frame_cropping_flag = 1;
1112         seq_param.frame_crop_left_offset = 0;
1113         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1114         seq_param.frame_crop_top_offset = 0;
1115         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1116     }
1117
1118     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1119                                VAEncSequenceParameterBufferType,
1120                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1121     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1122
1123     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1124                                VAEncMiscParameterBufferType,
1125                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1126                                1, NULL, &rc_param_buf);
1127     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1128
1129     vaMapBuffer(va_dpy->va_dpy, rc_param_buf, (void **)&misc_param);
1130     misc_param->type = VAEncMiscParameterTypeRateControl;
1131     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1132     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1133     misc_rate_ctrl->bits_per_second = frame_bitrate;
1134     misc_rate_ctrl->target_percentage = 66;
1135     misc_rate_ctrl->window_size = 1000;
1136     misc_rate_ctrl->initial_qp = initial_qp;
1137     misc_rate_ctrl->min_qp = minimal_qp;
1138     misc_rate_ctrl->basic_unit_size = 0;
1139     vaUnmapBuffer(va_dpy->va_dpy, rc_param_buf);
1140
1141     render_id[0] = seq_param_buf;
1142     render_id[1] = rc_param_buf;
1143
1144     render_picture_and_delete(va_dpy->va_dpy, context_id, &render_id[0], 2);
1145
1146     return 0;
1147 }
1148
1149 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1150 {
1151     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1152     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1153     int PicOrderCntMsb, TopFieldOrderCnt;
1154
1155     if (frame_type == FRAME_IDR)
1156         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1157     else {
1158         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1159         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1160     }
1161
1162     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1163         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1164         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1165     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1166              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1167         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1168     else
1169         PicOrderCntMsb = prevPicOrderCntMsb;
1170
1171     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1172
1173     if (frame_type != FRAME_B) {
1174         PicOrderCntMsb_ref = PicOrderCntMsb;
1175         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1176     }
1177
1178     return TopFieldOrderCnt;
1179 }
1180
1181 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1182 {
1183     VABufferID pic_param_buf;
1184     VAStatus va_status;
1185     size_t i = 0;
1186
1187     pic_param.CurrPic.picture_id = surf->ref_surface;
1188     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1189     pic_param.CurrPic.flags = 0;
1190     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1191     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1192     CurrentCurrPic = pic_param.CurrPic;
1193
1194     for (i = 0; i < reference_frames.size(); i++) {
1195         pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1196     }
1197     for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1198         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1199         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1200     }
1201
1202     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1203     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1204     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1205     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1206     pic_param.frame_num = current_ref_frame_num;  // FIXME: is this correct?
1207     pic_param.coded_buf = surf->coded_buf;
1208     pic_param.last_picture = false;  // FIXME
1209     pic_param.pic_init_qp = initial_qp;
1210
1211     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncPictureParameterBufferType,
1212                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1213     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1214
1215     render_picture_and_delete(va_dpy->va_dpy, context_id, &pic_param_buf, 1);
1216
1217     return 0;
1218 }
1219
1220 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1221 {
1222     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1223     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1224     unsigned int length_in_bits;
1225     unsigned char *packedseq_buffer = NULL;
1226     VAStatus va_status;
1227
1228     length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1229
1230     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1231
1232     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1233     packedheader_param_buffer.has_emulation_bytes = 0;
1234     va_status = vaCreateBuffer(va_dpy->va_dpy,
1235                                context_id,
1236                                VAEncPackedHeaderParameterBufferType,
1237                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1238                                &packedseq_para_bufid);
1239     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1240
1241     va_status = vaCreateBuffer(va_dpy->va_dpy,
1242                                context_id,
1243                                VAEncPackedHeaderDataBufferType,
1244                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1245                                &packedseq_data_bufid);
1246     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1247
1248     render_id[0] = packedseq_para_bufid;
1249     render_id[1] = packedseq_data_bufid;
1250     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1251
1252     free(packedseq_buffer);
1253
1254     return 0;
1255 }
1256
1257
1258 int QuickSyncEncoderImpl::render_packedpicture()
1259 {
1260     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1261     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1262     unsigned int length_in_bits;
1263     unsigned char *packedpic_buffer = NULL;
1264     VAStatus va_status;
1265
1266     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1267     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1268     packedheader_param_buffer.bit_length = length_in_bits;
1269     packedheader_param_buffer.has_emulation_bytes = 0;
1270
1271     va_status = vaCreateBuffer(va_dpy->va_dpy,
1272                                context_id,
1273                                VAEncPackedHeaderParameterBufferType,
1274                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1275                                &packedpic_para_bufid);
1276     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1277
1278     va_status = vaCreateBuffer(va_dpy->va_dpy,
1279                                context_id,
1280                                VAEncPackedHeaderDataBufferType,
1281                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1282                                &packedpic_data_bufid);
1283     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1284
1285     render_id[0] = packedpic_para_bufid;
1286     render_id[1] = packedpic_data_bufid;
1287     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1288
1289     free(packedpic_buffer);
1290
1291     return 0;
1292 }
1293
1294 void QuickSyncEncoderImpl::render_packedslice()
1295 {
1296     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1297     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1298     unsigned int length_in_bits;
1299     unsigned char *packedslice_buffer = NULL;
1300     VAStatus va_status;
1301
1302     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1303     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1304     packedheader_param_buffer.bit_length = length_in_bits;
1305     packedheader_param_buffer.has_emulation_bytes = 0;
1306
1307     va_status = vaCreateBuffer(va_dpy->va_dpy,
1308                                context_id,
1309                                VAEncPackedHeaderParameterBufferType,
1310                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1311                                &packedslice_para_bufid);
1312     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1313
1314     va_status = vaCreateBuffer(va_dpy->va_dpy,
1315                                context_id,
1316                                VAEncPackedHeaderDataBufferType,
1317                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1318                                &packedslice_data_bufid);
1319     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1320
1321     render_id[0] = packedslice_para_bufid;
1322     render_id[1] = packedslice_data_bufid;
1323     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1324
1325     free(packedslice_buffer);
1326 }
1327
1328 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1329 {
1330     VABufferID slice_param_buf;
1331     VAStatus va_status;
1332     int i;
1333
1334     /* one frame, one slice */
1335     slice_param.macroblock_address = 0;
1336     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1337     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1338     if (frame_type == FRAME_IDR) {
1339         if (encoding_frame_num != 0)
1340             ++slice_param.idr_pic_id;
1341     } else if (frame_type == FRAME_P) {
1342         VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1343         update_RefPicList_P(RefPicList0_P);
1344
1345         int refpiclist0_max = h264_maxref & 0xffff;
1346         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1347
1348         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1349             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1350             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1351         }
1352     } else if (frame_type == FRAME_B) {
1353         VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1354         update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1355
1356         int refpiclist0_max = h264_maxref & 0xffff;
1357         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1358
1359         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1360         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1361             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1362             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1363         }
1364
1365         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1366         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1367             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1368             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1369         }
1370     }
1371
1372     slice_param.slice_alpha_c0_offset_div2 = 0;
1373     slice_param.slice_beta_offset_div2 = 0;
1374     slice_param.direct_spatial_mv_pred_flag = 1;
1375     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1376
1377
1378     if (h264_packedheader &&
1379         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1380         render_packedslice();
1381
1382     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncSliceParameterBufferType,
1383                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1384     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1385
1386     render_picture_and_delete(va_dpy->va_dpy, context_id, &slice_param_buf, 1);
1387
1388     return 0;
1389 }
1390
1391
1392
1393 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1394 {
1395         VACodedBufferSegment *buf_list = NULL;
1396         VAStatus va_status;
1397
1398         string data;
1399
1400         va_status = vaMapBuffer(va_dpy->va_dpy, surf->coded_buf, (void **)(&buf_list));
1401         CHECK_VASTATUS(va_status, "vaMapBuffer");
1402         while (buf_list != NULL) {
1403                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1404                 buf_list = (VACodedBufferSegment *) buf_list->next;
1405         }
1406         vaUnmapBuffer(va_dpy->va_dpy, surf->coded_buf);
1407
1408         static int frameno = 0;
1409         print_latency("Current Quick Sync latency (video inputs → disk mux):",
1410                 task.received_ts, (task.frame_type == FRAME_B), &frameno, &qs_latency_histogram);
1411
1412         {
1413                 // Add video.
1414                 AVPacket pkt;
1415                 memset(&pkt, 0, sizeof(pkt));
1416                 pkt.buf = nullptr;
1417                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1418                 pkt.size = data.size();
1419                 pkt.stream_index = 0;
1420                 if (task.frame_type == FRAME_IDR) {
1421                         pkt.flags = AV_PKT_FLAG_KEY;
1422                 } else {
1423                         pkt.flags = 0;
1424                 }
1425                 pkt.duration = task.duration;
1426                 if (file_mux) {
1427                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1428                 }
1429                 if (!global_flags.uncompressed_video_to_http &&
1430                     !global_flags.x264_video_to_http) {
1431                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1432                 }
1433         }
1434 }
1435
1436
1437 // this is weird. but it seems to put a new frame onto the queue
1438 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1439 {
1440         lock_guard<mutex> lock(storage_task_queue_mutex);
1441         storage_task_queue.push(move(task));
1442         storage_task_queue_changed.notify_all();
1443 }
1444
1445 void QuickSyncEncoderImpl::storage_task_thread()
1446 {
1447         pthread_setname_np(pthread_self(), "QS_Storage");
1448         for ( ;; ) {
1449                 storage_task current;
1450                 GLSurface *surf;
1451                 {
1452                         // wait until there's an encoded frame
1453                         unique_lock<mutex> lock(storage_task_queue_mutex);
1454                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1455                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1456                         current = move(storage_task_queue.front());
1457                         storage_task_queue.pop();
1458                         surf = surface_for_frame[current.display_order];
1459                         assert(surf != nullptr);
1460                 }
1461
1462                 VAStatus va_status;
1463
1464                 size_t display_order = current.display_order;
1465                 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1466
1467                 // waits for data, then saves it to disk.
1468                 va_status = vaSyncSurface(va_dpy->va_dpy, surf->src_surface);
1469                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1470                 save_codeddata(surf, move(current));
1471
1472                 // Unlock the frame, and all its references.
1473                 {
1474                         lock_guard<mutex> lock(storage_task_queue_mutex);
1475                         release_gl_surface(display_order);
1476
1477                         for (size_t frame_num : ref_display_frame_numbers) {
1478                                 release_gl_surface(frame_num);
1479                         }
1480                 }
1481         }
1482 }
1483
1484 void QuickSyncEncoderImpl::release_encode()
1485 {
1486         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1487                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, gl_surfaces[i].coded_buf);
1488                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
1489                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].src_surface, 1);
1490                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1491                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].ref_surface, 1);
1492                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1493         }
1494
1495         VAStatus va_status = vaDestroyContext(va_dpy->va_dpy, context_id);
1496         CHECK_VASTATUS(va_status, "vaDestroyContext");
1497         va_status = vaDestroyConfig(va_dpy->va_dpy, config_id);
1498         CHECK_VASTATUS(va_status, "vaDestroyConfig");
1499 }
1500
1501 void QuickSyncEncoderImpl::release_gl_resources()
1502 {
1503         assert(is_shutdown);
1504         if (has_released_gl_resources) {
1505                 return;
1506         }
1507
1508         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1509                 if (use_zerocopy) {
1510                         resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1511                         resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1512                 } else {
1513                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1514                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1515                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1516                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1517                 }
1518         }
1519
1520         has_released_gl_resources = true;
1521 }
1522
1523 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1524         : current_storage_frame(0), resource_pool(resource_pool), surface(surface), x264_encoder(x264_encoder), frame_width(width), frame_height(height), disk_space_estimator(disk_space_estimator)
1525 {
1526         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1527         open_output_file(filename);
1528         file_audio_encoder->add_mux(file_mux.get());
1529
1530         frame_width_mbaligned = (frame_width + 15) & (~15);
1531         frame_height_mbaligned = (frame_height + 15) & (~15);
1532
1533         //print_input();
1534
1535         if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1536                 assert(x264_encoder != nullptr);
1537         } else {
1538                 assert(x264_encoder == nullptr);
1539         }
1540
1541         enable_zerocopy_if_possible();
1542         if (!global_flags.x264_video_to_disk) {
1543                 init_va(va_display);
1544         }
1545         setup_encode();
1546
1547         if (!global_flags.x264_video_to_disk) {
1548                 memset(&seq_param, 0, sizeof(seq_param));
1549                 memset(&pic_param, 0, sizeof(pic_param));
1550                 memset(&slice_param, 0, sizeof(slice_param));
1551         }
1552
1553         if (!global_flags.v4l_output_device.empty()) {
1554                 v4l_output.reset(new V4LOutput(global_flags.v4l_output_device.c_str(), width, height));
1555         }
1556
1557         call_once(quick_sync_metrics_inited, [](){
1558                 mixer_latency_histogram.init("mixer");
1559                 qs_latency_histogram.init("quick_sync");
1560                 current_file_mux_metrics.init({{ "destination", "current_file" }});
1561                 total_mux_metrics.init({{ "destination", "files_total" }});
1562                 global_metrics.add("current_file_start_time_seconds", &metric_current_file_start_time_seconds, Metrics::TYPE_GAUGE);
1563                 global_metrics.add("quick_sync_stalled_frames", &metric_quick_sync_stalled_frames);
1564         });
1565
1566         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1567
1568         encode_thread = thread([this]{
1569                 QOpenGLContext *context = create_context(this->surface);
1570                 eglBindAPI(EGL_OPENGL_API);
1571                 if (!make_current(context, this->surface)) {
1572                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1573                                 eglGetError());
1574                         abort();
1575                 }
1576                 encode_thread_func();
1577                 delete_context(context);
1578         });
1579 }
1580
1581 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1582 {
1583         shutdown();
1584         release_gl_resources();
1585 }
1586
1587 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1588 {
1589         for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1590                 if (gl_surfaces[i].refcount == 0) {
1591                         ++gl_surfaces[i].refcount;
1592                         return &gl_surfaces[i];
1593                 }
1594         }
1595         return nullptr;
1596 }
1597
1598 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1599 {
1600         assert(surface_for_frame.count(display_frame_num));
1601         QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1602         if (--surf->refcount == 0) {
1603                 assert(surface_for_frame.count(display_frame_num));
1604                 surface_for_frame.erase(display_frame_num);
1605                 storage_task_queue_changed.notify_all();
1606         }
1607 }
1608
1609 bool QuickSyncEncoderImpl::is_zerocopy() const
1610 {
1611         return use_zerocopy;
1612 }
1613
1614 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1615 {
1616         assert(!is_shutdown);
1617         GLSurface *surf = nullptr;
1618         {
1619                 // Wait until this frame slot is done encoding.
1620                 unique_lock<mutex> lock(storage_task_queue_mutex);
1621                 surf = allocate_gl_surface();
1622                 if (surf == nullptr) {
1623                         fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1624                                 current_storage_frame);
1625                         ++metric_quick_sync_stalled_frames;
1626                         storage_task_queue_changed.wait(lock, [this, &surf]{
1627                                 if (storage_thread_should_quit)
1628                                         return true;
1629                                 surf = allocate_gl_surface();
1630                                 return surf != nullptr;
1631                         });
1632                 }
1633                 if (storage_thread_should_quit) return false;
1634                 assert(surf != nullptr);
1635                 surface_for_frame[current_storage_frame] = surf;
1636         }
1637
1638         if (use_zerocopy) {
1639                 *y_tex = surf->y_tex;
1640                 *cbcr_tex = surf->cbcr_tex;
1641         } else {
1642                 surf->y_tex = *y_tex;
1643                 surf->cbcr_tex = *cbcr_tex;
1644         }
1645
1646         if (!global_flags.x264_video_to_disk) {
1647                 VAStatus va_status = vaDeriveImage(va_dpy->va_dpy, surf->src_surface, &surf->surface_image);
1648                 CHECK_VASTATUS(va_status, "vaDeriveImage");
1649
1650                 if (use_zerocopy) {
1651                         VABufferInfo buf_info;
1652                         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1653                         va_status = vaAcquireBufferHandle(va_dpy->va_dpy, surf->surface_image.buf, &buf_info);
1654                         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1655
1656                         // Create Y image.
1657                         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1658                         EGLint y_attribs[] = {
1659                                 EGL_WIDTH, frame_width,
1660                                 EGL_HEIGHT, frame_height,
1661                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1662                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1663                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1664                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1665                                 EGL_NONE
1666                         };
1667
1668                         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1669                         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1670
1671                         // Associate Y image to a texture.
1672                         glBindTexture(GL_TEXTURE_2D, *y_tex);
1673                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1674
1675                         // Create CbCr image.
1676                         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1677                         EGLint cbcr_attribs[] = {
1678                                 EGL_WIDTH, frame_width / 2,
1679                                 EGL_HEIGHT, frame_height / 2,
1680                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1681                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1682                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1683                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1684                                 EGL_NONE
1685                         };
1686
1687                         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1688                         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1689
1690                         // Associate CbCr image to a texture.
1691                         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1692                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1693                 }
1694         }
1695
1696         current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1697
1698         return true;
1699 }
1700
1701 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1702 {
1703         lock_guard<mutex> lock(file_audio_encoder_mutex);
1704         assert(!is_shutdown);
1705         file_audio_encoder->encode_audio(audio, pts + global_delay());
1706 }
1707
1708 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1709 {
1710         assert(!is_shutdown);
1711
1712         if (!use_zerocopy) {
1713                 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1714                 GLSurface *surf;
1715                 {
1716                         lock_guard<mutex> lock(storage_task_queue_mutex);
1717                         surf = surface_for_frame[current_storage_frame];
1718                         assert(surf != nullptr);
1719                 }
1720
1721                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1722                 check_error();
1723
1724                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1725                 check_error();
1726
1727                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1728                 check_error();
1729                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1730                 check_error();
1731
1732                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1733                 check_error();
1734                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1735                 check_error();
1736
1737                 // We don't own these; the caller does.
1738                 surf->y_tex = surf->cbcr_tex = 0;
1739
1740                 glBindTexture(GL_TEXTURE_2D, 0);
1741                 check_error();
1742                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1743                 check_error();
1744
1745                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1746                 check_error();
1747         }
1748
1749         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1750         check_error();
1751         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1752         check_error();
1753
1754         {
1755                 lock_guard<mutex> lock(frame_queue_mutex);
1756                 current_video_frame.fence = fence;
1757                 pending_video_frames.push(move(current_video_frame));
1758                 ++current_storage_frame;
1759         }
1760         frame_queue_nonempty.notify_all();
1761         return fence;
1762 }
1763
1764 void QuickSyncEncoderImpl::shutdown()
1765 {
1766         if (is_shutdown) {
1767                 return;
1768         }
1769
1770         {
1771                 lock_guard<mutex> lock(frame_queue_mutex);
1772                 encode_thread_should_quit = true;
1773                 frame_queue_nonempty.notify_all();
1774         }
1775         encode_thread.join();
1776         {
1777                 lock_guard<mutex> lock(storage_task_queue_mutex);
1778                 storage_thread_should_quit = true;
1779                 frame_queue_nonempty.notify_all();
1780                 storage_task_queue_changed.notify_all();
1781         }
1782         storage_thread.join();
1783
1784         // Encode any leftover audio in the queues, and also any delayed frames.
1785         {
1786                 lock_guard<mutex> lock(file_audio_encoder_mutex);
1787                 file_audio_encoder->encode_last_audio();
1788         }
1789
1790         if (!global_flags.x264_video_to_disk) {
1791                 release_encode();
1792                 va_dpy.reset();
1793         }
1794         is_shutdown = true;
1795 }
1796
1797 void QuickSyncEncoderImpl::close_file()
1798 {
1799         file_mux.reset();
1800         metric_current_file_start_time_seconds = 0.0 / 0.0;
1801 }
1802
1803 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1804 {
1805         AVFormatContext *avctx = avformat_alloc_context();
1806         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1807         avctx->url = strdup(filename.c_str());
1808
1809         string url = "file:" + filename;
1810         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1811         if (ret < 0) {
1812                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1813                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1814                 abort();
1815         }
1816
1817         string video_extradata;  // FIXME: See other comment about global headers.
1818         if (global_flags.x264_video_to_disk) {
1819                 video_extradata = x264_encoder->get_global_headers();
1820         }
1821
1822         current_file_mux_metrics.reset();
1823
1824         {
1825                 lock_guard<mutex> lock(file_audio_encoder_mutex);
1826                 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1827                 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,
1828                         std::bind(&DiskSpaceEstimator::report_append, disk_space_estimator, filename, _1),
1829                         Mux::WRITE_BACKGROUND,
1830                         { &current_file_mux_metrics, &total_mux_metrics }));
1831         }
1832         metric_current_file_start_time_seconds = get_timestamp_for_metrics();
1833
1834         if (global_flags.x264_video_to_disk) {
1835                 x264_encoder->add_mux(file_mux.get());
1836         }
1837 }
1838
1839 void QuickSyncEncoderImpl::encode_thread_func()
1840 {
1841         pthread_setname_np(pthread_self(), "QS_Encode");
1842
1843         int64_t last_dts = -1;
1844         int gop_start_display_frame_num = 0;
1845         for (int display_frame_num = 0; ; ++display_frame_num) {
1846                 // Wait for the frame to be in the queue. Note that this only means
1847                 // we started rendering it.
1848                 PendingFrame frame;
1849                 {
1850                         unique_lock<mutex> lock(frame_queue_mutex);
1851                         frame_queue_nonempty.wait(lock, [this]{
1852                                 return encode_thread_should_quit || !pending_video_frames.empty();
1853                         });
1854                         if (encode_thread_should_quit && pending_video_frames.empty()) {
1855                                 // We may have queued frames left in the reorder buffer
1856                                 // that were supposed to be B-frames, but have no P-frame
1857                                 // to be encoded against. If so, encode them all as
1858                                 // P-frames instead. Note that this happens under the mutex,
1859                                 // but nobody else uses it at this point, since we're shutting down,
1860                                 // so there's no contention.
1861                                 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1862                                 return;
1863                         } else {
1864                                 frame = move(pending_video_frames.front());
1865                                 pending_video_frames.pop();
1866                         }
1867                 }
1868
1869                 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1870                 // Note that this implicitly waits for the frame to be done rendering.
1871                 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1872
1873                 if (global_flags.x264_video_to_disk) {
1874                         lock_guard<mutex> lock(storage_task_queue_mutex);
1875                         release_gl_surface(display_frame_num);
1876                         continue;
1877                 }
1878
1879                 reorder_buffer[display_frame_num] = move(frame);
1880
1881                 // Now encode as many QuickSync frames as we can using the frames we have available.
1882                 // (It could be zero, or it could be multiple.) FIXME: make a function.
1883                 for ( ;; ) {
1884                         int pts_lag;
1885                         int frame_type, quicksync_display_frame_num;
1886                         encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1887                                                &quicksync_display_frame_num, &frame_type, &pts_lag);
1888                         if (!reorder_buffer.count(quicksync_display_frame_num)) {
1889                                 break;
1890                         }
1891                         frame = move(reorder_buffer[quicksync_display_frame_num]);
1892                         reorder_buffer.erase(quicksync_display_frame_num);
1893
1894                         if (frame_type == FRAME_IDR) {
1895                                 // Release any reference frames from the previous GOP.
1896                                 {
1897                                         lock_guard<mutex> lock(storage_task_queue_mutex);
1898                                         for (const ReferenceFrame &frame : reference_frames) {
1899                                                 release_gl_surface(frame.display_number);
1900                                         }
1901                                 }
1902                                 reference_frames.clear();
1903                                 current_ref_frame_num = 0;
1904                                 gop_start_display_frame_num = quicksync_display_frame_num;
1905                         }
1906
1907                         // Determine the dts of this frame.
1908                         int64_t dts;
1909                         if (pts_lag == -1) {
1910                                 assert(last_dts != -1);
1911                                 dts = last_dts + (TIMEBASE / MAX_FPS);
1912                         } else {
1913                                 dts = frame.pts - pts_lag;
1914                         }
1915                         last_dts = dts;
1916
1917                         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);
1918                         ++quicksync_encoding_frame_num;
1919                 }
1920         }
1921 }
1922
1923 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1924 {
1925         if (reorder_buffer.empty()) {
1926                 return;
1927         }
1928
1929         for (auto &pending_frame : reorder_buffer) {
1930                 int display_frame_num = pending_frame.first;
1931                 assert(display_frame_num > 0);
1932                 PendingFrame frame = move(pending_frame.second);
1933                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1934                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1935                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1936                 last_dts = dts;
1937         }
1938 }
1939
1940 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1941 {
1942         AVPacket pkt;
1943         memset(&pkt, 0, sizeof(pkt));
1944         pkt.buf = nullptr;
1945         pkt.data = const_cast<uint8_t *>(data);
1946         pkt.size = frame_width * frame_height * 2;
1947         pkt.stream_index = 0;
1948         pkt.flags = AV_PKT_FLAG_KEY;
1949         pkt.duration = duration;
1950         stream_mux->add_packet(pkt, pts, pts);
1951 }
1952
1953 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1954 {
1955         if (src_width == dst_pitch) {
1956                 memcpy(dst, src, src_width * height);
1957         } else {
1958                 for (size_t y = 0; y < height; ++y) {
1959                         const uint8_t *sptr = src + y * src_width;
1960                         uint8_t *dptr = dst + y * dst_pitch;
1961                         memcpy(dptr, sptr, src_width);
1962                 }
1963         }
1964 }
1965
1966 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1967 {
1968         // Wait for the GPU to be done with the frame.
1969         GLenum sync_status;
1970         do {
1971                 sync_status = glClientWaitSync(frame.fence.get(), 0, 0);
1972                 check_error();
1973                 if (sync_status == GL_TIMEOUT_EXPIRED) {
1974                         // NVIDIA likes to busy-wait; yield instead.
1975                         this_thread::sleep_for(milliseconds(1));
1976                 }
1977         } while (sync_status == GL_TIMEOUT_EXPIRED);
1978         assert(sync_status != GL_WAIT_FAILED);
1979
1980         ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1981         static int frameno = 0;
1982         print_latency("Current mixer latency (video inputs → ready for encode):",
1983                 received_ts, false, &frameno, &mixer_latency_histogram);
1984
1985         // Release back any input frames we needed to render this frame.
1986         frame.input_frames.clear();
1987
1988         GLSurface *surf;
1989         {
1990                 lock_guard<mutex> lock(storage_task_queue_mutex);
1991                 surf = surface_for_frame[display_frame_num];
1992                 assert(surf != nullptr);
1993         }
1994         uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1995         if (global_flags.uncompressed_video_to_http) {
1996                 add_packet_for_uncompressed_frame(pts, duration, data);
1997         } else if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1998                 x264_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1999         }
2000
2001         if (v4l_output != nullptr) {
2002                 v4l_output->send_frame(data);
2003         }
2004 }
2005
2006 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2007                                         int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
2008 {
2009         const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
2010
2011         GLSurface *surf;
2012         {
2013                 lock_guard<mutex> lock(storage_task_queue_mutex);
2014                 surf = surface_for_frame[display_frame_num];
2015                 assert(surf != nullptr);
2016         }
2017         VAStatus va_status;
2018
2019         if (use_zerocopy) {
2020                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2021                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2022                 va_status = vaReleaseBufferHandle(va_dpy->va_dpy, surf->surface_image.buf);
2023                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2024         } else {
2025                 // Upload the frame to VA-API.
2026                 unsigned char *surface_p = nullptr;
2027                 vaMapBuffer(va_dpy->va_dpy, surf->surface_image.buf, (void **)&surface_p);
2028
2029                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2030                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2031
2032                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2033                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2034
2035                 va_status = vaUnmapBuffer(va_dpy->va_dpy, surf->surface_image.buf);
2036                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2037         }
2038
2039         va_status = vaDestroyImage(va_dpy->va_dpy, surf->surface_image.image_id);
2040         CHECK_VASTATUS(va_status, "vaDestroyImage");
2041
2042         // Schedule the frame for encoding.
2043         VASurfaceID va_surface = surf->src_surface;
2044         va_status = vaBeginPicture(va_dpy->va_dpy, context_id, va_surface);
2045         CHECK_VASTATUS(va_status, "vaBeginPicture");
2046
2047         if (frame_type == FRAME_IDR) {
2048                 // FIXME: If the mux wants global headers, we should not put the
2049                 // SPS/PPS before each IDR frame, but rather put it into the
2050                 // codec extradata (formatted differently?).
2051                 //
2052                 // NOTE: If we change ycbcr_coefficients, it will not take effect
2053                 // before the next IDR frame. This is acceptable, as it should only
2054                 // happen on a mode change, which is rare.
2055                 render_sequence();
2056                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2057                 if (h264_packedheader) {
2058                         render_packedsequence(ycbcr_coefficients);
2059                         render_packedpicture();
2060                 }
2061         } else {
2062                 //render_sequence();
2063                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2064         }
2065         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2066
2067         va_status = vaEndPicture(va_dpy->va_dpy, context_id);
2068         CHECK_VASTATUS(va_status, "vaEndPicture");
2069
2070         update_ReferenceFrames(display_frame_num, frame_type);
2071
2072         vector<size_t> ref_display_frame_numbers;
2073
2074         // Lock the references for this frame; otherwise, they could be
2075         // rendered to before this frame is done encoding.
2076         {
2077                 lock_guard<mutex> lock(storage_task_queue_mutex);
2078                 for (const ReferenceFrame &frame : reference_frames) {
2079                         assert(surface_for_frame.count(frame.display_number));
2080                         ++surface_for_frame[frame.display_number]->refcount;
2081                         ref_display_frame_numbers.push_back(frame.display_number);
2082                 }
2083         }
2084
2085         // so now the data is done encoding (well, async job kicked off)...
2086         // we send that to the storage thread
2087         storage_task tmp;
2088         tmp.display_order = display_frame_num;
2089         tmp.frame_type = frame_type;
2090         tmp.pts = pts;
2091         tmp.dts = dts;
2092         tmp.duration = duration;
2093         tmp.ycbcr_coefficients = ycbcr_coefficients;
2094         tmp.received_ts = received_ts;
2095         tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2096         storage_task_enqueue(move(tmp));
2097 }
2098
2099 // Proxy object.
2100 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
2101         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2102
2103 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2104 QuickSyncEncoder::~QuickSyncEncoder() {}
2105
2106 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2107 {
2108         impl->add_audio(pts, audio);
2109 }
2110
2111 bool QuickSyncEncoder::is_zerocopy() const
2112 {
2113         return impl->is_zerocopy();
2114 }
2115
2116 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2117 {
2118         return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2119 }
2120
2121 RefCountedGLsync QuickSyncEncoder::end_frame()
2122 {
2123         return impl->end_frame();
2124 }
2125
2126 void QuickSyncEncoder::shutdown()
2127 {
2128         impl->shutdown();
2129 }
2130
2131 void QuickSyncEncoder::close_file()
2132 {
2133         impl->shutdown();
2134 }
2135
2136 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2137 {
2138         impl->set_stream_mux(mux);
2139 }
2140
2141 int64_t QuickSyncEncoder::global_delay() const {
2142         return impl->global_delay();
2143 }
2144
2145 string QuickSyncEncoder::get_usable_va_display()
2146 {
2147         // Reduce the amount of chatter while probing,
2148         // unless the user has specified otherwise.
2149         bool need_env_reset = false;
2150         if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
2151                 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
2152                 need_env_reset = true;
2153         }
2154
2155         // First try the default (ie., whatever $DISPLAY is set to).
2156         unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va("", nullptr, nullptr);
2157         if (va_dpy != nullptr) {
2158                 if (need_env_reset) {
2159                         unsetenv("LIBVA_MESSAGING_LEVEL");
2160                 }
2161                 return "";
2162         }
2163
2164         fprintf(stderr, "No --va-display was given, and the X11 display did not expose a VA-API H.264 encoder.\n");
2165
2166         // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
2167         glob_t g;
2168         int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
2169         if (err != 0) {
2170                 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
2171         } else {
2172                 for (size_t i = 0; i < g.gl_pathc; ++i) {
2173                         string path = g.gl_pathv[i];
2174                         va_dpy = try_open_va(path, nullptr, nullptr);
2175                         if (va_dpy != nullptr) {
2176                                 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
2177                                         path.c_str());
2178                                 globfree(&g);
2179                                 if (need_env_reset) {
2180                                         unsetenv("LIBVA_MESSAGING_LEVEL");
2181                                 }
2182                                 return path;
2183                         }
2184                 }
2185         }
2186
2187         fprintf(stderr, "No suitable VA-API H.264 encoders were found in /dev/dri; giving up.\n");
2188         fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
2189         fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
2190         fprintf(stderr, "to expose Quick Sync. Alternatively, you can use --record-x264-video\n");
2191         fprintf(stderr, "to use software instead of hardware H.264 encoding, at the expense\n");
2192         fprintf(stderr, "of increased CPU usage and possibly bit rate.\n");
2193         abort();
2194 }