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

   1 #include "quicksync_encoder.h"
   2
   3 #include <movit/image_format.h>
   4 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
   5 #include <movit/util.h>
   6
   7 #include <EGL/eglplatform.h>
   8 #include <X11/Xlib.h>
   9 #include <assert.h>
  10 #include <epoxy/egl.h>
  11 #include <fcntl.h>
  12 #include <pthread.h>
  13 #include <stdio.h>
  14 #include <stdlib.h>
  15 #include <string.h>
  16 #include <unistd.h>
  17 #include <va/va.h>
  18 #include <va/va_drm.h>
  19 #include <va/va_drmcommon.h>
  20 #include <va/va_enc_h264.h>
  21 #include <va/va_x11.h>
  22 #include <algorithm>
  23 #include <chrono>
  24 #include <condition_variable>
  25 #include <cstddef>
  26 #include <cstdint>
  27 #include <functional>
  28 #include <map>
  29 #include <memory>
  30 #include <mutex>
  31 #include <queue>
  32 #include <stack>
  33 #include <string>
  34 #include <thread>
  35 #include <utility>
  36
  37 extern "C" {
  38
  39 #include <libavcodec/avcodec.h>
  40 #include <libavformat/avio.h>
  41 #include <libavutil/error.h>
  42 #include <libdrm/drm_fourcc.h>
  43
  44 }  // namespace
  45
  46 #include "audio_encoder.h"
  47 #include "context.h"
  48 #include "defs.h"
  49 #include "disk_space_estimator.h"
  50 #include "ffmpeg_raii.h"
  51 #include "flags.h"
  52 #include "mux.h"
  53 #include "print_latency.h"
  54 #include "quicksync_encoder_impl.h"
  55 #include "ref_counted_frame.h"
  56 #include "timebase.h"
  57 #include "x264_encoder.h"
  58
  59 using namespace movit;
  60 using namespace std;
  61 using namespace std::chrono;
  62 using namespace std::placeholders;
  63
  64 class QOpenGLContext;
  65 class QSurface;
  66
  67 namespace {
  68
  69 // These need to survive several QuickSyncEncoderImpl instances,
  70 // so they are outside.
  71 bool quick_sync_metrics_inited = false;
  72 LatencyHistogram mixer_latency_histogram, qs_latency_histogram;
  73 MuxMetrics current_file_mux_metrics, total_mux_metrics;
  74 std::atomic<double> metric_current_file_start_time_seconds{0.0 / 0.0};
  75 std::atomic<int64_t> metric_quick_sync_stalled_frames{0};
  76
  77 }  // namespace
  78
  79 #define CHECK_VASTATUS(va_status, func)                                 \
  80     if (va_status != VA_STATUS_SUCCESS) {                               \
  81         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  82         exit(1);                                                        \
  83     }
  84
  85 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  86
  87 //#include "loadsurface.h"
  88
  89 #define NAL_REF_IDC_NONE        0
  90 #define NAL_REF_IDC_LOW         1
  91 #define NAL_REF_IDC_MEDIUM      2
  92 #define NAL_REF_IDC_HIGH        3
  93
  94 #define NAL_NON_IDR             1
  95 #define NAL_IDR                 5
  96 #define NAL_SPS                 7
  97 #define NAL_PPS                 8
  98 #define NAL_SEI                 6
  99
 100 #define SLICE_TYPE_P            0
 101 #define SLICE_TYPE_B            1
 102 #define SLICE_TYPE_I            2
 103 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
 104 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
 105 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
 106
 107
 108 #define ENTROPY_MODE_CAVLC      0
 109 #define ENTROPY_MODE_CABAC      1
 110
 111 #define PROFILE_IDC_BASELINE    66
 112 #define PROFILE_IDC_MAIN        77
 113 #define PROFILE_IDC_HIGH        100
 114
 115 #define BITSTREAM_ALLOCATE_STEPPING     4096
 116
 117 static constexpr unsigned int MaxFrameNum = (2<<16);
 118 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
 119 static constexpr unsigned int Log2MaxFrameNum = 16;
 120 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
 121 static constexpr int rc_default_modes[] = {  // Priority list of modes.
 122     VA_RC_VBR,
 123     VA_RC_CQP,
 124     VA_RC_VBR_CONSTRAINED,
 125     VA_RC_CBR,
 126     VA_RC_VCM,
 127     VA_RC_NONE,
 128 };
 129
 130 using namespace std;
 131
 132 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 133 // but if we don't delete it here, we get leaks. The GStreamer implementation
 134 // does the same.
 135 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 136 {
 137     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 138     CHECK_VASTATUS(va_status, "vaRenderPicture");
 139
 140     for (int i = 0; i < num_buffers; ++i) {
 141         va_status = vaDestroyBuffer(dpy, buffers[i]);
 142         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 143     }
 144 }
 145
 146 static unsigned int
 147 va_swap32(unsigned int val)
 148 {
 149     unsigned char *pval = (unsigned char *)&val;
 150
 151     return ((pval[0] << 24)     |
 152             (pval[1] << 16)     |
 153             (pval[2] << 8)      |
 154             (pval[3] << 0));
 155 }
 156
 157 static void
 158 bitstream_start(bitstream *bs)
 159 {
 160     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 161     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 162     bs->bit_offset = 0;
 163 }
 164
 165 static void
 166 bitstream_end(bitstream *bs)
 167 {
 168     int pos = (bs->bit_offset >> 5);
 169     int bit_offset = (bs->bit_offset & 0x1f);
 170     int bit_left = 32 - bit_offset;
 171
 172     if (bit_offset) {
 173         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 174     }
 175 }
 176
 177 static void
 178 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 179 {
 180     int pos = (bs->bit_offset >> 5);
 181     int bit_offset = (bs->bit_offset & 0x1f);
 182     int bit_left = 32 - bit_offset;
 183
 184     if (!size_in_bits)
 185         return;
 186
 187     bs->bit_offset += size_in_bits;
 188
 189     if (bit_left > size_in_bits) {
 190         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 191     } else {
 192         size_in_bits -= bit_left;
 193         if (bit_left >= 32) {
 194             bs->buffer[pos] = (val >> size_in_bits);
 195         } else {
 196             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 197         }
 198         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 199
 200         if (pos + 1 == bs->max_size_in_dword) {
 201             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 202             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 203         }
 204
 205         bs->buffer[pos + 1] = val;
 206     }
 207 }
 208
 209 static void
 210 bitstream_put_ue(bitstream *bs, unsigned int val)
 211 {
 212     int size_in_bits = 0;
 213     int tmp_val = ++val;
 214
 215     while (tmp_val) {
 216         tmp_val >>= 1;
 217         size_in_bits++;
 218     }
 219
 220     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 221     bitstream_put_ui(bs, val, size_in_bits);
 222 }
 223
 224 static void
 225 bitstream_put_se(bitstream *bs, int val)
 226 {
 227     unsigned int new_val;
 228
 229     if (val <= 0)
 230         new_val = -2 * val;
 231     else
 232         new_val = 2 * val - 1;
 233
 234     bitstream_put_ue(bs, new_val);
 235 }
 236
 237 static void
 238 bitstream_byte_aligning(bitstream *bs, int bit)
 239 {
 240     int bit_offset = (bs->bit_offset & 0x7);
 241     int bit_left = 8 - bit_offset;
 242     int new_val;
 243
 244     if (!bit_offset)
 245         return;
 246
 247     assert(bit == 0 || bit == 1);
 248
 249     if (bit)
 250         new_val = (1 << bit_left) - 1;
 251     else
 252         new_val = 0;
 253
 254     bitstream_put_ui(bs, new_val, bit_left);
 255 }
 256
 257 static void
 258 rbsp_trailing_bits(bitstream *bs)
 259 {
 260     bitstream_put_ui(bs, 1, 1);
 261     bitstream_byte_aligning(bs, 0);
 262 }
 263
 264 static void nal_start_code_prefix(bitstream *bs)
 265 {
 266     bitstream_put_ui(bs, 0x00000001, 32);
 267 }
 268
 269 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 270 {
 271     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 272     bitstream_put_ui(bs, nal_ref_idc, 2);
 273     bitstream_put_ui(bs, nal_unit_type, 5);
 274 }
 275
 276 void QuickSyncEncoderImpl::sps_rbsp(YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs)
 277 {
 278     int profile_idc = PROFILE_IDC_BASELINE;
 279
 280     if (h264_profile  == VAProfileH264High)
 281         profile_idc = PROFILE_IDC_HIGH;
 282     else if (h264_profile  == VAProfileH264Main)
 283         profile_idc = PROFILE_IDC_MAIN;
 284
 285     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 286     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 287     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 288     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 289     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 290     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 291     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 292     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 293
 294     if ( profile_idc == PROFILE_IDC_HIGH) {
 295         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 296         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 297         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 298         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 299         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 300     }
 301
 302     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 303     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 304
 305     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 306         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 307     else {
 308         assert(0);
 309     }
 310
 311     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 312     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 313
 314     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 315     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 316     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 317
 318     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 319         assert(0);
 320     }
 321
 322     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 323     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 324
 325     if (seq_param.frame_cropping_flag) {
 326         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 327         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 328         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 329         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 330     }
 331
 332     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 333     if ( false ) {
 334         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 335     } else {
 336         // See H.264 annex E for the definition of this header.
 337         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 338         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 339         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 340         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 341         {
 342             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 343             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 344             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 345             {
 346                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 347                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 348                 if (ycbcr_coefficients == YCBCR_REC_709) {
 349                     bitstream_put_ui(bs, 1, 8);  /* matrix_coefficients (1 = BT.709) */
 350                 } else {
 351                     assert(ycbcr_coefficients == YCBCR_REC_601);
 352                     bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 353                 }
 354             }
 355         }
 356         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 357         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 358         {
 359             bitstream_put_ui(bs, 1, 32);  // FPS
 360             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 361             bitstream_put_ui(bs, 1, 1);
 362         }
 363         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 364         {
 365             // hrd_parameters
 366             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 367             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 368             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 369
 370             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 371             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 372             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 373
 374             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 375             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 376             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 377             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 378         }
 379         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 380         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 381
 382         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 383         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 384     }
 385
 386     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 387 }
 388
 389
 390 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 391 {
 392     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 393     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 394
 395     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 396
 397     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 398
 399     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 400
 401     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 402     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 403
 404     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 405     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 406
 407     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 408     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 409     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 410
 411     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 412     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 413     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 414
 415     /* more_rbsp_data */
 416     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 417     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 418     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 419
 420     rbsp_trailing_bits(bs);
 421 }
 422
 423 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 424 {
 425     int first_mb_in_slice = slice_param.macroblock_address;
 426
 427     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 428     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 429     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 430     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 431
 432     /* frame_mbs_only_flag == 1 */
 433     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 434         /* FIXME: */
 435         assert(0);
 436     }
 437
 438     if (pic_param.pic_fields.bits.idr_pic_flag)
 439         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 440
 441     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 442         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 443         /* pic_order_present_flag == 0 */
 444     } else {
 445         /* FIXME: */
 446         assert(0);
 447     }
 448
 449     /* redundant_pic_cnt_present_flag == 0 */
 450     /* slice type */
 451     if (IS_P_SLICE(slice_param.slice_type)) {
 452         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 453
 454         if (slice_param.num_ref_idx_active_override_flag)
 455             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 456
 457         /* ref_pic_list_reordering */
 458         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 459     } else if (IS_B_SLICE(slice_param.slice_type)) {
 460         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 461
 462         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 463
 464         if (slice_param.num_ref_idx_active_override_flag) {
 465             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 466             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 467         }
 468
 469         /* ref_pic_list_reordering */
 470         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 471         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 472     }
 473
 474     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 475          IS_P_SLICE(slice_param.slice_type)) ||
 476         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 477          IS_B_SLICE(slice_param.slice_type))) {
 478         /* FIXME: fill weight/offset table */
 479         assert(0);
 480     }
 481
 482     /* dec_ref_pic_marking */
 483     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 484         unsigned char no_output_of_prior_pics_flag = 0;
 485         unsigned char long_term_reference_flag = 0;
 486         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 487
 488         if (pic_param.pic_fields.bits.idr_pic_flag) {
 489             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 490             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 491         } else {
 492             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 493         }
 494     }
 495
 496     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 497         !IS_I_SLICE(slice_param.slice_type))
 498         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 499
 500     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 501
 502     /* ignore for SP/SI */
 503
 504     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 505         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 506
 507         if (slice_param.disable_deblocking_filter_idc != 1) {
 508             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 509             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 510         }
 511     }
 512
 513     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 514         bitstream_byte_aligning(bs, 1);
 515     }
 516 }
 517
 518 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 519 {
 520     bitstream bs;
 521
 522     bitstream_start(&bs);
 523     nal_start_code_prefix(&bs);
 524     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 525     pps_rbsp(&bs);
 526     bitstream_end(&bs);
 527
 528     *header_buffer = (unsigned char *)bs.buffer;
 529     return bs.bit_offset;
 530 }
 531
 532 int
 533 QuickSyncEncoderImpl::build_packed_seq_buffer(YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer)
 534 {
 535     bitstream bs;
 536
 537     bitstream_start(&bs);
 538     nal_start_code_prefix(&bs);
 539     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 540     sps_rbsp(ycbcr_coefficients, &bs);
 541     bitstream_end(&bs);
 542
 543     *header_buffer = (unsigned char *)bs.buffer;
 544     return bs.bit_offset;
 545 }
 546
 547 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 548 {
 549     bitstream bs;
 550     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 551     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 552
 553     bitstream_start(&bs);
 554     nal_start_code_prefix(&bs);
 555
 556     if (IS_I_SLICE(slice_param.slice_type)) {
 557         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 558     } else if (IS_P_SLICE(slice_param.slice_type)) {
 559         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 560     } else {
 561         assert(IS_B_SLICE(slice_param.slice_type));
 562         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 563     }
 564
 565     slice_header(&bs);
 566     bitstream_end(&bs);
 567
 568     *header_buffer = (unsigned char *)bs.buffer;
 569     return bs.bit_offset;
 570 }
 571
 572
 573 /*
 574   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 575   1) period between Frame #X and Frame #N = #X - #N
 576   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 577   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 578   4) intra_period and intra_idr_period take precedence over ip_period
 579   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 580      of I/IDR frames, see bellow examples
 581   -------------------------------------------------------------------
 582   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 583   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 584   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 585   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 586   1            1                ignored    IDR IDR IDR IDR...
 587   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 588   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 589   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 590                                               (PBB)(IBB)(PBB)(IBB)...
 591   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 592                                            IDRPPPPP IPPPPP IPPPPP...
 593   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 594                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 595                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 596                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 597                                            {IDR(PBB)(PBB)}                     (6/6/3)
 598                                            {IDR(PBB)(PBB)}.
 599 */
 600
 601 // General pts/dts strategy:
 602 //
 603 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 604 // bit tricky. We assume first of all that the frame rate never goes _above_
 605 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 606 // in at least this speed, as long at dts <= pts (the frame is not attempted
 607 // presented before it is decoded). Furthermore, we never have longer chains of
 608 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 609 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 610 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 611 // frame, with an order O <= C.)
 612 //
 613 // Many strategies are possible, but we establish these rules:
 614 //
 615 //  - Tip frames have dts = pts - (C-O)*N.
 616 //  - Non-tip frames have dts = dts_last + N.
 617 //
 618 // An example, with C=2 and N=10 and the data flow showed with arrows:
 619 //
 620 //        I  B  P  B  B  P
 621 //   pts: 30 40 50 60 70 80
 622 //        ↓  ↓     ↓
 623 //   dts: 10 30 20 60 50←40
 624 //         |  |  ↑        ↑
 625 //         `--|--'        |
 626 //             `----------'
 627 //
 628 // To show that this works fine also with irregular spacings, let's say that
 629 // the third frame is delayed a bit (something earlier was dropped). Now the
 630 // situation looks like this:
 631 //
 632 //        I  B  P  B  B   P
 633 //   pts: 30 40 80 90 100 110
 634 //        ↓  ↓     ↓
 635 //   dts: 10 30 20 90 50←40
 636 //         |  |  ↑        ↑
 637 //         `--|--'        |
 638 //             `----------'
 639 //
 640 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 641 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 642 //
 643 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 644 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 645 // a tip frame and should be given a dts based on the previous one.
 646 #define FRAME_P 0
 647 #define FRAME_B 1
 648 #define FRAME_I 2
 649 #define FRAME_IDR 7
 650 void encoding2display_order(
 651     int encoding_order, int intra_period,
 652     int intra_idr_period, int ip_period,
 653     int *displaying_order,
 654     int *frame_type, int *pts_lag)
 655 {
 656     int encoding_order_gop = 0;
 657
 658     *pts_lag = 0;
 659
 660     if (intra_period == 1) { /* all are I/IDR frames */
 661         *displaying_order = encoding_order;
 662         if (intra_idr_period == 0)
 663             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 664         else
 665             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 666         return;
 667     }
 668
 669     if (intra_period == 0)
 670         intra_idr_period = 0;
 671
 672     if (ip_period == 1) {
 673         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 674         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 675         *displaying_order = encoding_order;
 676
 677         if (encoding_order_gop == 0) { /* the first frame */
 678             *frame_type = FRAME_IDR;
 679         } else if (intra_period != 0 && /* have I frames */
 680                    encoding_order_gop >= 2 &&
 681                    (encoding_order_gop % intra_period == 0)) {
 682             *frame_type = FRAME_I;
 683         } else {
 684             *frame_type = FRAME_P;
 685         }
 686         return;
 687     }
 688
 689     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 690     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 691     *pts_lag = -1;  // Most frames are not tip frames.
 692
 693     if (encoding_order_gop == 0) { /* the first frame */
 694         *frame_type = FRAME_IDR;
 695         *displaying_order = encoding_order;
 696         // IDR frames are a special case; I honestly can't find the logic behind
 697         // why this is the right thing, but it seems to line up nicely in practice :-)
 698         *pts_lag = TIMEBASE / MAX_FPS;
 699     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 700         *frame_type = FRAME_B;
 701         *displaying_order = encoding_order - 1;
 702         if ((encoding_order_gop % ip_period) == 0) {
 703             *pts_lag = 0;  // Last B-frame.
 704         }
 705     } else if (intra_period != 0 && /* have I frames */
 706                encoding_order_gop >= 2 &&
 707                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 708         *frame_type = FRAME_I;
 709         *displaying_order = encoding_order + ip_period - 1;
 710     } else {
 711         *frame_type = FRAME_P;
 712         *displaying_order = encoding_order + ip_period - 1;
 713     }
 714 }
 715
 716
 717 static const char *rc_to_string(int rc_mode)
 718 {
 719     switch (rc_mode) {
 720     case VA_RC_NONE:
 721         return "NONE";
 722     case VA_RC_CBR:
 723         return "CBR";
 724     case VA_RC_VBR:
 725         return "VBR";
 726     case VA_RC_VCM:
 727         return "VCM";
 728     case VA_RC_CQP:
 729         return "CQP";
 730     case VA_RC_VBR_CONSTRAINED:
 731         return "VBR_CONSTRAINED";
 732     default:
 733         return "Unknown";
 734     }
 735 }
 736
 737 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 738 {
 739         if (global_flags.x264_video_to_disk) {
 740                 // Quick Sync is entirely disabled.
 741                 use_zerocopy = false;
 742         } else if (global_flags.uncompressed_video_to_http) {
 743                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 744                 use_zerocopy = false;
 745         } else if (global_flags.x264_video_to_http) {
 746                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 747                 use_zerocopy = false;
 748         } else {
 749                 use_zerocopy = true;
 750         }
 751         global_flags.use_zerocopy = use_zerocopy;
 752 }
 753
 754 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 755 {
 756         if (va_display.empty()) {
 757                 x11_display = XOpenDisplay(NULL);
 758                 if (!x11_display) {
 759                         fprintf(stderr, "error: can't connect to X server!\n");
 760                         return NULL;
 761                 }
 762                 return vaGetDisplay(x11_display);
 763         } else if (va_display[0] != '/') {
 764                 x11_display = XOpenDisplay(va_display.c_str());
 765                 if (!x11_display) {
 766                         fprintf(stderr, "error: can't connect to X server!\n");
 767                         return NULL;
 768                 }
 769                 return vaGetDisplay(x11_display);
 770         } else {
 771                 drm_fd = open(va_display.c_str(), O_RDWR);
 772                 if (drm_fd == -1) {
 773                         perror(va_display.c_str());
 774                         return NULL;
 775                 }
 776                 use_zerocopy = false;
 777                 return vaGetDisplayDRM(drm_fd);
 778         }
 779 }
 780
 781 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
 782 {
 783         if (x11_display) {
 784                 XCloseDisplay(x11_display);
 785                 x11_display = nullptr;
 786         }
 787         if (drm_fd != -1) {
 788                 close(drm_fd);
 789         }
 790 }
 791
 792 int QuickSyncEncoderImpl::init_va(const string &va_display)
 793 {
 794     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 795     VAEntrypoint *entrypoints;
 796     int num_entrypoints, slice_entrypoint;
 797     int support_encode = 0;
 798     int major_ver, minor_ver;
 799     VAStatus va_status;
 800     unsigned int i;
 801
 802     va_dpy = va_open_display(va_display);
 803     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 804     CHECK_VASTATUS(va_status, "vaInitialize");
 805
 806     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 807     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 808     if (!entrypoints) {
 809         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 810         exit(1);
 811     }
 812
 813     /* use the highest profile */
 814     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 815         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 816             continue;
 817
 818         h264_profile = profile_list[i];
 819         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 820         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 821             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 822                 support_encode = 1;
 823                 break;
 824             }
 825         }
 826         if (support_encode == 1)
 827             break;
 828     }
 829
 830     if (support_encode == 0) {
 831         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
 832         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
 833         printf("to use VA-API against DRM instead of X11.\n");
 834         exit(1);
 835     } else {
 836         switch (h264_profile) {
 837             case VAProfileH264Baseline:
 838                 ip_period = 1;
 839                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 840                 h264_entropy_mode = 0;
 841                 break;
 842             case VAProfileH264ConstrainedBaseline:
 843                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 844                 ip_period = 1;
 845                 break;
 846
 847             case VAProfileH264Main:
 848                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 849                 break;
 850
 851             case VAProfileH264High:
 852                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 853                 break;
 854             default:
 855                 h264_profile = VAProfileH264Baseline;
 856                 ip_period = 1;
 857                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 858                 break;
 859         }
 860     }
 861
 862     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 863
 864     /* find out the format for the render target, and rate control mode */
 865     for (i = 0; i < VAConfigAttribTypeMax; i++)
 866         attrib[i].type = (VAConfigAttribType)i;
 867
 868     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 869                                       &attrib[0], VAConfigAttribTypeMax);
 870     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 871     /* check the interested configattrib */
 872     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 873         printf("Not find desired YUV420 RT format\n");
 874         exit(1);
 875     } else {
 876         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 877         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 878         config_attrib_num++;
 879     }
 880
 881     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 882         int tmp = attrib[VAConfigAttribRateControl].value;
 883
 884         if (rc_mode == -1 || !(rc_mode & tmp))  {
 885             if (rc_mode != -1) {
 886                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
 887             }
 888
 889             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
 890                 if (rc_default_modes[i] & tmp) {
 891                     rc_mode = rc_default_modes[i];
 892                     break;
 893                 }
 894             }
 895         }
 896
 897         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 898         config_attrib[config_attrib_num].value = rc_mode;
 899         config_attrib_num++;
 900     }
 901
 902
 903     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 904         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 905
 906         h264_packedheader = 1;
 907         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 908         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 909
 910         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 911             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 912         }
 913
 914         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 915             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 916         }
 917
 918         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 919             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 920         }
 921
 922         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
 923             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
 924         }
 925
 926         enc_packed_header_idx = config_attrib_num;
 927         config_attrib_num++;
 928     }
 929
 930     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
 931         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
 932         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 933         config_attrib_num++;
 934     }
 935
 936     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
 937         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
 938     }
 939
 940     free(entrypoints);
 941     return 0;
 942 }
 943
 944 int QuickSyncEncoderImpl::setup_encode()
 945 {
 946         if (!global_flags.x264_video_to_disk) {
 947                 VAStatus va_status;
 948                 VASurfaceID *tmp_surfaceid;
 949                 int codedbuf_size;
 950                 VASurfaceID src_surface[SURFACE_NUM];
 951                 VASurfaceID ref_surface[SURFACE_NUM];
 952
 953                 va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
 954                                 &config_attrib[0], config_attrib_num, &config_id);
 955                 CHECK_VASTATUS(va_status, "vaCreateConfig");
 956
 957                 /* create source surfaces */
 958                 va_status = vaCreateSurfaces(va_dpy,
 959                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 960                                 &src_surface[0], SURFACE_NUM,
 961                                 NULL, 0);
 962                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 963
 964                 /* create reference surfaces */
 965                 va_status = vaCreateSurfaces(va_dpy,
 966                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 967                                 &ref_surface[0], SURFACE_NUM,
 968                                 NULL, 0);
 969                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 970
 971                 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
 972                 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
 973                 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
 974
 975                 for (int i = 0; i < SURFACE_NUM; i++) {
 976                         gl_surfaces[i].src_surface = src_surface[i];
 977                         gl_surfaces[i].ref_surface = ref_surface[i];
 978                 }
 979
 980                 /* Create a context for this encode pipe */
 981                 va_status = vaCreateContext(va_dpy, config_id,
 982                                 frame_width_mbaligned, frame_height_mbaligned,
 983                                 VA_PROGRESSIVE,
 984                                 tmp_surfaceid, 2 * SURFACE_NUM,
 985                                 &context_id);
 986                 CHECK_VASTATUS(va_status, "vaCreateContext");
 987                 free(tmp_surfaceid);
 988
 989                 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
 990
 991                 for (int i = 0; i < SURFACE_NUM; i++) {
 992                         /* create coded buffer once for all
 993                          * other VA buffers which won't be used again after vaRenderPicture.
 994                          * so APP can always vaCreateBuffer for every frame
 995                          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
 996                          * so VA won't maintain the coded buffer
 997                          */
 998                         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
 999                                         codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1000                         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1001                 }
1002         }
1003
1004         /* create OpenGL objects */
1005         for (int i = 0; i < SURFACE_NUM; i++) {
1006                 if (use_zerocopy) {
1007                         gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
1008                         gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
1009                 } else {
1010                         size_t bytes_per_pixel = (global_flags.x264_bit_depth > 8) ? 2 : 1;
1011
1012                         // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1013                         // buffers, due to potentially differing pitch.
1014                         glGenBuffers(1, &gl_surfaces[i].pbo);
1015                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1016                         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);
1017                         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);
1018                         gl_surfaces[i].y_offset = 0;
1019                         gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
1020                         gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1021                         gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1022                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1023                 }
1024         }
1025
1026         return 0;
1027 }
1028
1029 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1030 //
1031 //   2 1 0 [3] 4 8 9
1032 template<class T, class C>
1033 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1034 {
1035         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1036         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1037         sort(middle, end, less_than);
1038 }
1039
1040 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
1041 {
1042     if (frame_type == FRAME_B)
1043         return;
1044
1045     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1046
1047     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1048     unique_lock<mutex> lock(storage_task_queue_mutex);
1049
1050     // Insert the new frame at the start of the reference queue.
1051     reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
1052
1053     if (reference_frames.size() > num_ref_frames)
1054     {
1055         // The back frame frame is no longer in use as a reference.
1056         int display_frame_num = reference_frames.back().display_number;
1057         assert(surface_for_frame.count(display_frame_num));
1058         release_gl_surface(display_frame_num);
1059         reference_frames.pop_back();
1060     }
1061
1062     // Mark this frame in use as a reference.
1063     assert(surface_for_frame.count(current_display_frame));
1064     ++surface_for_frame[current_display_frame]->refcount;
1065
1066     current_ref_frame_num++;
1067     if (current_ref_frame_num > MaxFrameNum)
1068         current_ref_frame_num = 0;
1069 }
1070
1071
1072 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
1073 {
1074     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1075         return a.frame_idx > b.frame_idx;
1076     };
1077
1078     for (size_t i = 0; i < reference_frames.size(); ++i) {
1079         RefPicList0_P[i] = reference_frames[i].pic;
1080     }
1081     sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
1082 }
1083
1084 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
1085 {
1086     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1087         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1088     };
1089     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1090         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1091     };
1092
1093     for (size_t i = 0; i < reference_frames.size(); ++i) {
1094         RefPicList0_B[i] = reference_frames[i].pic;
1095         RefPicList1_B[i] = reference_frames[i].pic;
1096     }
1097     sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1098     sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1099 }
1100
1101
1102 int QuickSyncEncoderImpl::render_sequence()
1103 {
1104     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1105     VAStatus va_status;
1106     VAEncMiscParameterBuffer *misc_param;
1107     VAEncMiscParameterRateControl *misc_rate_ctrl;
1108
1109     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1110     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1111     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1112     seq_param.bits_per_second = frame_bitrate;
1113
1114     seq_param.intra_period = intra_period;
1115     seq_param.intra_idr_period = intra_idr_period;
1116     seq_param.ip_period = ip_period;
1117
1118     seq_param.max_num_ref_frames = num_ref_frames;
1119     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1120     seq_param.time_scale = TIMEBASE * 2;
1121     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1122     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1123     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1124     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1125     seq_param.seq_fields.bits.chroma_format_idc = 1;
1126     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1127
1128     if (frame_width != frame_width_mbaligned ||
1129         frame_height != frame_height_mbaligned) {
1130         seq_param.frame_cropping_flag = 1;
1131         seq_param.frame_crop_left_offset = 0;
1132         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1133         seq_param.frame_crop_top_offset = 0;
1134         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1135     }
1136
1137     va_status = vaCreateBuffer(va_dpy, context_id,
1138                                VAEncSequenceParameterBufferType,
1139                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1140     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1141
1142     va_status = vaCreateBuffer(va_dpy, context_id,
1143                                VAEncMiscParameterBufferType,
1144                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1145                                1, NULL, &rc_param_buf);
1146     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1147
1148     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1149     misc_param->type = VAEncMiscParameterTypeRateControl;
1150     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1151     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1152     misc_rate_ctrl->bits_per_second = frame_bitrate;
1153     misc_rate_ctrl->target_percentage = 66;
1154     misc_rate_ctrl->window_size = 1000;
1155     misc_rate_ctrl->initial_qp = initial_qp;
1156     misc_rate_ctrl->min_qp = minimal_qp;
1157     misc_rate_ctrl->basic_unit_size = 0;
1158     vaUnmapBuffer(va_dpy, rc_param_buf);
1159
1160     render_id[0] = seq_param_buf;
1161     render_id[1] = rc_param_buf;
1162
1163     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1164
1165     return 0;
1166 }
1167
1168 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1169 {
1170     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1171     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1172     int PicOrderCntMsb, TopFieldOrderCnt;
1173
1174     if (frame_type == FRAME_IDR)
1175         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1176     else {
1177         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1178         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1179     }
1180
1181     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1182         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1183         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1184     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1185              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1186         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1187     else
1188         PicOrderCntMsb = prevPicOrderCntMsb;
1189
1190     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1191
1192     if (frame_type != FRAME_B) {
1193         PicOrderCntMsb_ref = PicOrderCntMsb;
1194         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1195     }
1196
1197     return TopFieldOrderCnt;
1198 }
1199
1200 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1201 {
1202     VABufferID pic_param_buf;
1203     VAStatus va_status;
1204     size_t i = 0;
1205
1206     pic_param.CurrPic.picture_id = surf->ref_surface;
1207     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1208     pic_param.CurrPic.flags = 0;
1209     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1210     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1211     CurrentCurrPic = pic_param.CurrPic;
1212
1213     for (i = 0; i < reference_frames.size(); i++) {
1214         pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1215     }
1216     for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1217         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1218         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1219     }
1220
1221     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1222     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1223     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1224     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1225     pic_param.frame_num = current_ref_frame_num;  // FIXME: is this correct?
1226     pic_param.coded_buf = surf->coded_buf;
1227     pic_param.last_picture = false;  // FIXME
1228     pic_param.pic_init_qp = initial_qp;
1229
1230     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1231                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1232     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1233
1234     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1235
1236     return 0;
1237 }
1238
1239 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1240 {
1241     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1242     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1243     unsigned int length_in_bits;
1244     unsigned char *packedseq_buffer = NULL;
1245     VAStatus va_status;
1246
1247     length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1248
1249     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1250
1251     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1252     packedheader_param_buffer.has_emulation_bytes = 0;
1253     va_status = vaCreateBuffer(va_dpy,
1254                                context_id,
1255                                VAEncPackedHeaderParameterBufferType,
1256                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1257                                &packedseq_para_bufid);
1258     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1259
1260     va_status = vaCreateBuffer(va_dpy,
1261                                context_id,
1262                                VAEncPackedHeaderDataBufferType,
1263                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1264                                &packedseq_data_bufid);
1265     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1266
1267     render_id[0] = packedseq_para_bufid;
1268     render_id[1] = packedseq_data_bufid;
1269     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1270
1271     free(packedseq_buffer);
1272
1273     return 0;
1274 }
1275
1276
1277 int QuickSyncEncoderImpl::render_packedpicture()
1278 {
1279     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1280     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1281     unsigned int length_in_bits;
1282     unsigned char *packedpic_buffer = NULL;
1283     VAStatus va_status;
1284
1285     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1286     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1287     packedheader_param_buffer.bit_length = length_in_bits;
1288     packedheader_param_buffer.has_emulation_bytes = 0;
1289
1290     va_status = vaCreateBuffer(va_dpy,
1291                                context_id,
1292                                VAEncPackedHeaderParameterBufferType,
1293                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1294                                &packedpic_para_bufid);
1295     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1296
1297     va_status = vaCreateBuffer(va_dpy,
1298                                context_id,
1299                                VAEncPackedHeaderDataBufferType,
1300                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1301                                &packedpic_data_bufid);
1302     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1303
1304     render_id[0] = packedpic_para_bufid;
1305     render_id[1] = packedpic_data_bufid;
1306     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1307
1308     free(packedpic_buffer);
1309
1310     return 0;
1311 }
1312
1313 void QuickSyncEncoderImpl::render_packedslice()
1314 {
1315     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1316     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1317     unsigned int length_in_bits;
1318     unsigned char *packedslice_buffer = NULL;
1319     VAStatus va_status;
1320
1321     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1322     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1323     packedheader_param_buffer.bit_length = length_in_bits;
1324     packedheader_param_buffer.has_emulation_bytes = 0;
1325
1326     va_status = vaCreateBuffer(va_dpy,
1327                                context_id,
1328                                VAEncPackedHeaderParameterBufferType,
1329                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1330                                &packedslice_para_bufid);
1331     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1332
1333     va_status = vaCreateBuffer(va_dpy,
1334                                context_id,
1335                                VAEncPackedHeaderDataBufferType,
1336                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1337                                &packedslice_data_bufid);
1338     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1339
1340     render_id[0] = packedslice_para_bufid;
1341     render_id[1] = packedslice_data_bufid;
1342     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1343
1344     free(packedslice_buffer);
1345 }
1346
1347 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1348 {
1349     VABufferID slice_param_buf;
1350     VAStatus va_status;
1351     int i;
1352
1353     /* one frame, one slice */
1354     slice_param.macroblock_address = 0;
1355     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1356     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1357     if (frame_type == FRAME_IDR) {
1358         if (encoding_frame_num != 0)
1359             ++slice_param.idr_pic_id;
1360     } else if (frame_type == FRAME_P) {
1361         VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1362         update_RefPicList_P(RefPicList0_P);
1363
1364         int refpiclist0_max = h264_maxref & 0xffff;
1365         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1366
1367         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1368             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1369             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1370         }
1371     } else if (frame_type == FRAME_B) {
1372         VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1373         update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1374
1375         int refpiclist0_max = h264_maxref & 0xffff;
1376         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1377
1378         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1379         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1380             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1381             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1382         }
1383
1384         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1385         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1386             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1387             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1388         }
1389     }
1390
1391     slice_param.slice_alpha_c0_offset_div2 = 0;
1392     slice_param.slice_beta_offset_div2 = 0;
1393     slice_param.direct_spatial_mv_pred_flag = 1;
1394     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1395
1396
1397     if (h264_packedheader &&
1398         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1399         render_packedslice();
1400
1401     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1402                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1403     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1404
1405     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1406
1407     return 0;
1408 }
1409
1410
1411
1412 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1413 {
1414         VACodedBufferSegment *buf_list = NULL;
1415         VAStatus va_status;
1416
1417         string data;
1418
1419         va_status = vaMapBuffer(va_dpy, surf->coded_buf, (void **)(&buf_list));
1420         CHECK_VASTATUS(va_status, "vaMapBuffer");
1421         while (buf_list != NULL) {
1422                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1423                 buf_list = (VACodedBufferSegment *) buf_list->next;
1424         }
1425         vaUnmapBuffer(va_dpy, surf->coded_buf);
1426
1427         static int frameno = 0;
1428         print_latency("Current Quick Sync latency (video inputs → disk mux):",
1429                 task.received_ts, (task.frame_type == FRAME_B), &frameno, &qs_latency_histogram);
1430
1431         {
1432                 // Add video.
1433                 AVPacket pkt;
1434                 memset(&pkt, 0, sizeof(pkt));
1435                 pkt.buf = nullptr;
1436                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1437                 pkt.size = data.size();
1438                 pkt.stream_index = 0;
1439                 if (task.frame_type == FRAME_IDR) {
1440                         pkt.flags = AV_PKT_FLAG_KEY;
1441                 } else {
1442                         pkt.flags = 0;
1443                 }
1444                 pkt.duration = task.duration;
1445                 if (file_mux) {
1446                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1447                 }
1448                 if (!global_flags.uncompressed_video_to_http &&
1449                     !global_flags.x264_video_to_http) {
1450                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1451                 }
1452         }
1453 }
1454
1455
1456 // this is weird. but it seems to put a new frame onto the queue
1457 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1458 {
1459         unique_lock<mutex> lock(storage_task_queue_mutex);
1460         storage_task_queue.push(move(task));
1461         storage_task_queue_changed.notify_all();
1462 }
1463
1464 void QuickSyncEncoderImpl::storage_task_thread()
1465 {
1466         pthread_setname_np(pthread_self(), "QS_Storage");
1467         for ( ;; ) {
1468                 storage_task current;
1469                 GLSurface *surf;
1470                 {
1471                         // wait until there's an encoded frame
1472                         unique_lock<mutex> lock(storage_task_queue_mutex);
1473                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1474                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1475                         current = move(storage_task_queue.front());
1476                         storage_task_queue.pop();
1477                         surf = surface_for_frame[current.display_order];
1478                         assert(surf != nullptr);
1479                 }
1480
1481                 VAStatus va_status;
1482
1483                 size_t display_order = current.display_order;
1484                 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1485
1486                 // waits for data, then saves it to disk.
1487                 va_status = vaSyncSurface(va_dpy, surf->src_surface);
1488                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1489                 save_codeddata(surf, move(current));
1490
1491                 // Unlock the frame, and all its references.
1492                 {
1493                         unique_lock<mutex> lock(storage_task_queue_mutex);
1494                         release_gl_surface(display_order);
1495
1496                         for (size_t frame_num : ref_display_frame_numbers) {
1497                                 release_gl_surface(frame_num);
1498                         }
1499                 }
1500         }
1501 }
1502
1503 void QuickSyncEncoderImpl::release_encode()
1504 {
1505         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1506                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1507                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1508                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1509         }
1510
1511         vaDestroyContext(va_dpy, context_id);
1512         vaDestroyConfig(va_dpy, config_id);
1513 }
1514
1515 void QuickSyncEncoderImpl::release_gl_resources()
1516 {
1517         assert(is_shutdown);
1518         if (has_released_gl_resources) {
1519                 return;
1520         }
1521
1522         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1523                 if (use_zerocopy) {
1524                         resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1525                         resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1526                 } else {
1527                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1528                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1529                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1530                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1531                 }
1532         }
1533
1534         has_released_gl_resources = true;
1535 }
1536
1537 int QuickSyncEncoderImpl::deinit_va()
1538 {
1539     vaTerminate(va_dpy);
1540
1541     va_close_display(va_dpy);
1542
1543     return 0;
1544 }
1545
1546 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)
1547         : 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)
1548 {
1549         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1550         open_output_file(filename);
1551         file_audio_encoder->add_mux(file_mux.get());
1552
1553         frame_width_mbaligned = (frame_width + 15) & (~15);
1554         frame_height_mbaligned = (frame_height + 15) & (~15);
1555
1556         //print_input();
1557
1558         if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1559                 assert(x264_encoder != nullptr);
1560         } else {
1561                 assert(x264_encoder == nullptr);
1562         }
1563
1564         enable_zerocopy_if_possible();
1565         if (!global_flags.x264_video_to_disk) {
1566                 init_va(va_display);
1567         }
1568         setup_encode();
1569
1570         if (!global_flags.x264_video_to_disk) {
1571                 memset(&seq_param, 0, sizeof(seq_param));
1572                 memset(&pic_param, 0, sizeof(pic_param));
1573                 memset(&slice_param, 0, sizeof(slice_param));
1574         }
1575
1576         if (!quick_sync_metrics_inited) {
1577                 mixer_latency_histogram.init("mixer");
1578                 qs_latency_histogram.init("quick_sync");
1579                 current_file_mux_metrics.init({{ "destination", "current_file" }});
1580                 total_mux_metrics.init({{ "destination", "files_total" }});
1581                 global_metrics.add("current_file_start_time_seconds", &metric_current_file_start_time_seconds, Metrics::TYPE_GAUGE);
1582                 global_metrics.add("quick_sync_stalled_frames", &metric_quick_sync_stalled_frames);
1583                 quick_sync_metrics_inited = true;
1584         }
1585
1586         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1587
1588         encode_thread = thread([this]{
1589                 QOpenGLContext *context = create_context(this->surface);
1590                 eglBindAPI(EGL_OPENGL_API);
1591                 if (!make_current(context, this->surface)) {
1592                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1593                                 eglGetError());
1594                         exit(1);
1595                 }
1596                 encode_thread_func();
1597                 delete_context(context);
1598         });
1599 }
1600
1601 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1602 {
1603         shutdown();
1604         release_gl_resources();
1605 }
1606
1607 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1608 {
1609         for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1610                 if (gl_surfaces[i].refcount == 0) {
1611                         ++gl_surfaces[i].refcount;
1612                         return &gl_surfaces[i];
1613                 }
1614         }
1615         return nullptr;
1616 }
1617
1618 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1619 {
1620         assert(surface_for_frame.count(display_frame_num));
1621         QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1622         if (--surf->refcount == 0) {
1623                 assert(surface_for_frame.count(display_frame_num));
1624                 surface_for_frame.erase(display_frame_num);
1625                 storage_task_queue_changed.notify_all();
1626         }
1627 }
1628
1629 bool QuickSyncEncoderImpl::is_zerocopy() const
1630 {
1631         return use_zerocopy;
1632 }
1633
1634 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1635 {
1636         assert(!is_shutdown);
1637         GLSurface *surf = nullptr;
1638         {
1639                 // Wait until this frame slot is done encoding.
1640                 unique_lock<mutex> lock(storage_task_queue_mutex);
1641                 surf = allocate_gl_surface();
1642                 if (surf == nullptr) {
1643                         fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1644                                 current_storage_frame);
1645                         ++metric_quick_sync_stalled_frames;
1646                         storage_task_queue_changed.wait(lock, [this, &surf]{
1647                                 if (storage_thread_should_quit)
1648                                         return true;
1649                                 surf = allocate_gl_surface();
1650                                 return surf != nullptr;
1651                         });
1652                 }
1653                 if (storage_thread_should_quit) return false;
1654                 assert(surf != nullptr);
1655                 surface_for_frame[current_storage_frame] = surf;
1656         }
1657
1658         if (use_zerocopy) {
1659                 *y_tex = surf->y_tex;
1660                 *cbcr_tex = surf->cbcr_tex;
1661         } else {
1662                 surf->y_tex = *y_tex;
1663                 surf->cbcr_tex = *cbcr_tex;
1664         }
1665
1666         if (!global_flags.x264_video_to_disk) {
1667                 VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1668                 CHECK_VASTATUS(va_status, "vaDeriveImage");
1669
1670                 if (use_zerocopy) {
1671                         VABufferInfo buf_info;
1672                         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1673                         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1674                         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1675
1676                         // Create Y image.
1677                         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1678                         EGLint y_attribs[] = {
1679                                 EGL_WIDTH, frame_width,
1680                                 EGL_HEIGHT, frame_height,
1681                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1682                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1683                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1684                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1685                                 EGL_NONE
1686                         };
1687
1688                         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1689                         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1690
1691                         // Associate Y image to a texture.
1692                         glBindTexture(GL_TEXTURE_2D, *y_tex);
1693                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1694
1695                         // Create CbCr image.
1696                         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1697                         EGLint cbcr_attribs[] = {
1698                                 EGL_WIDTH, frame_width,
1699                                 EGL_HEIGHT, frame_height,
1700                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1701                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1702                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1703                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1704                                 EGL_NONE
1705                         };
1706
1707                         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1708                         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1709
1710                         // Associate CbCr image to a texture.
1711                         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1712                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1713                 }
1714         }
1715
1716         current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1717
1718         return true;
1719 }
1720
1721 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1722 {
1723         assert(!is_shutdown);
1724         file_audio_encoder->encode_audio(audio, pts + global_delay());
1725 }
1726
1727 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1728 {
1729         assert(!is_shutdown);
1730
1731         if (!use_zerocopy) {
1732                 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1733                 GLSurface *surf;
1734                 {
1735                         unique_lock<mutex> lock(storage_task_queue_mutex);
1736                         surf = surface_for_frame[current_storage_frame];
1737                         assert(surf != nullptr);
1738                 }
1739
1740                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1741                 check_error();
1742
1743                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1744                 check_error();
1745
1746                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1747                 check_error();
1748                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1749                 check_error();
1750
1751                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1752                 check_error();
1753                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1754                 check_error();
1755
1756                 // We don't own these; the caller does.
1757                 surf->y_tex = surf->cbcr_tex = 0;
1758
1759                 glBindTexture(GL_TEXTURE_2D, 0);
1760                 check_error();
1761                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1762                 check_error();
1763
1764                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1765                 check_error();
1766         }
1767
1768         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1769         check_error();
1770         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1771         check_error();
1772
1773         {
1774                 unique_lock<mutex> lock(frame_queue_mutex);
1775                 current_video_frame.fence = fence;
1776                 pending_video_frames.push(move(current_video_frame));
1777                 ++current_storage_frame;
1778         }
1779         frame_queue_nonempty.notify_all();
1780         return fence;
1781 }
1782
1783 void QuickSyncEncoderImpl::shutdown()
1784 {
1785         if (is_shutdown) {
1786                 return;
1787         }
1788
1789         {
1790                 unique_lock<mutex> lock(frame_queue_mutex);
1791                 encode_thread_should_quit = true;
1792                 frame_queue_nonempty.notify_all();
1793         }
1794         encode_thread.join();
1795         {
1796                 unique_lock<mutex> lock(storage_task_queue_mutex);
1797                 storage_thread_should_quit = true;
1798                 frame_queue_nonempty.notify_all();
1799                 storage_task_queue_changed.notify_all();
1800         }
1801         storage_thread.join();
1802
1803         // Encode any leftover audio in the queues, and also any delayed frames.
1804         file_audio_encoder->encode_last_audio();
1805
1806         if (!global_flags.x264_video_to_disk) {
1807                 release_encode();
1808                 deinit_va();
1809         }
1810         is_shutdown = true;
1811 }
1812
1813 void QuickSyncEncoderImpl::close_file()
1814 {
1815         file_mux.reset();
1816         metric_current_file_start_time_seconds = 0.0 / 0.0;
1817 }
1818
1819 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1820 {
1821         AVFormatContext *avctx = avformat_alloc_context();
1822         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1823         assert(filename.size() < sizeof(avctx->filename) - 1);
1824         strcpy(avctx->filename, filename.c_str());
1825
1826         string url = "file:" + filename;
1827         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1828         if (ret < 0) {
1829                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1830                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1831                 exit(1);
1832         }
1833
1834         string video_extradata;  // FIXME: See other comment about global headers.
1835         if (global_flags.x264_video_to_disk) {
1836                 video_extradata = x264_encoder->get_global_headers();
1837         }
1838
1839         current_file_mux_metrics.reset();
1840
1841         AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1842         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1843                 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1),
1844                 { &current_file_mux_metrics, &total_mux_metrics }));
1845         metric_current_file_start_time_seconds = get_timestamp_for_metrics();
1846
1847         if (global_flags.x264_video_to_disk) {
1848                 x264_encoder->add_mux(file_mux.get());
1849         }
1850 }
1851
1852 void QuickSyncEncoderImpl::encode_thread_func()
1853 {
1854         pthread_setname_np(pthread_self(), "QS_Encode");
1855
1856         int64_t last_dts = -1;
1857         int gop_start_display_frame_num = 0;
1858         for (int display_frame_num = 0; ; ++display_frame_num) {
1859                 // Wait for the frame to be in the queue. Note that this only means
1860                 // we started rendering it.
1861                 PendingFrame frame;
1862                 {
1863                         unique_lock<mutex> lock(frame_queue_mutex);
1864                         frame_queue_nonempty.wait(lock, [this]{
1865                                 return encode_thread_should_quit || !pending_video_frames.empty();
1866                         });
1867                         if (encode_thread_should_quit && pending_video_frames.empty()) {
1868                                 // We may have queued frames left in the reorder buffer
1869                                 // that were supposed to be B-frames, but have no P-frame
1870                                 // to be encoded against. If so, encode them all as
1871                                 // P-frames instead. Note that this happens under the mutex,
1872                                 // but nobody else uses it at this point, since we're shutting down,
1873                                 // so there's no contention.
1874                                 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1875                                 return;
1876                         } else {
1877                                 frame = move(pending_video_frames.front());
1878                                 pending_video_frames.pop();
1879                         }
1880                 }
1881
1882                 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1883                 // Note that this implicitly waits for the frame to be done rendering.
1884                 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1885
1886                 if (global_flags.x264_video_to_disk) {
1887                         unique_lock<mutex> lock(storage_task_queue_mutex);
1888                         release_gl_surface(display_frame_num);
1889                         continue;
1890                 }
1891
1892                 reorder_buffer[display_frame_num] = move(frame);
1893
1894                 // Now encode as many QuickSync frames as we can using the frames we have available.
1895                 // (It could be zero, or it could be multiple.) FIXME: make a function.
1896                 for ( ;; ) {
1897                         int pts_lag;
1898                         int frame_type, quicksync_display_frame_num;
1899                         encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1900                                                &quicksync_display_frame_num, &frame_type, &pts_lag);
1901                         if (!reorder_buffer.count(quicksync_display_frame_num)) {
1902                                 break;
1903                         }
1904                         frame = move(reorder_buffer[quicksync_display_frame_num]);
1905                         reorder_buffer.erase(quicksync_display_frame_num);
1906
1907                         if (frame_type == FRAME_IDR) {
1908                                 // Release any reference frames from the previous GOP.
1909                                 for (const ReferenceFrame &frame : reference_frames) {
1910                                         release_gl_surface(frame.display_number);
1911                                 }
1912                                 reference_frames.clear();
1913                                 current_ref_frame_num = 0;
1914                                 gop_start_display_frame_num = quicksync_display_frame_num;
1915                         }
1916
1917                         // Determine the dts of this frame.
1918                         int64_t dts;
1919                         if (pts_lag == -1) {
1920                                 assert(last_dts != -1);
1921                                 dts = last_dts + (TIMEBASE / MAX_FPS);
1922                         } else {
1923                                 dts = frame.pts - pts_lag;
1924                         }
1925                         last_dts = dts;
1926
1927                         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);
1928                         ++quicksync_encoding_frame_num;
1929                 }
1930         }
1931 }
1932
1933 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1934 {
1935         if (reorder_buffer.empty()) {
1936                 return;
1937         }
1938
1939         for (auto &pending_frame : reorder_buffer) {
1940                 int display_frame_num = pending_frame.first;
1941                 assert(display_frame_num > 0);
1942                 PendingFrame frame = move(pending_frame.second);
1943                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1944                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1945                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1946                 last_dts = dts;
1947         }
1948 }
1949
1950 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1951 {
1952         AVPacket pkt;
1953         memset(&pkt, 0, sizeof(pkt));
1954         pkt.buf = nullptr;
1955         pkt.data = const_cast<uint8_t *>(data);
1956         pkt.size = frame_width * frame_height * 2;
1957         pkt.stream_index = 0;
1958         pkt.flags = AV_PKT_FLAG_KEY;
1959         pkt.duration = duration;
1960         stream_mux->add_packet(pkt, pts, pts);
1961 }
1962
1963 namespace {
1964
1965 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1966 {
1967         if (src_width == dst_pitch) {
1968                 memcpy(dst, src, src_width * height);
1969         } else {
1970                 for (size_t y = 0; y < height; ++y) {
1971                         const uint8_t *sptr = src + y * src_width;
1972                         uint8_t *dptr = dst + y * dst_pitch;
1973                         memcpy(dptr, sptr, src_width);
1974                 }
1975         }
1976 }
1977
1978 }  // namespace
1979
1980 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1981 {
1982         // Wait for the GPU to be done with the frame.
1983         GLenum sync_status;
1984         do {
1985                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1986                 check_error();
1987         } while (sync_status == GL_TIMEOUT_EXPIRED);
1988         assert(sync_status != GL_WAIT_FAILED);
1989
1990         ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1991         static int frameno = 0;
1992         print_latency("Current mixer latency (video inputs → ready for encode):",
1993                 received_ts, false, &frameno, &mixer_latency_histogram);
1994
1995         // Release back any input frames we needed to render this frame.
1996         frame.input_frames.clear();
1997
1998         GLSurface *surf;
1999         {
2000                 unique_lock<mutex> lock(storage_task_queue_mutex);
2001                 surf = surface_for_frame[display_frame_num];
2002                 assert(surf != nullptr);
2003         }
2004         uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
2005         if (global_flags.uncompressed_video_to_http) {
2006                 add_packet_for_uncompressed_frame(pts, duration, data);
2007         } else if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
2008                 x264_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
2009         }
2010 }
2011
2012 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2013                                         int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
2014 {
2015         const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
2016
2017         GLSurface *surf;
2018         {
2019                 unique_lock<mutex> lock(storage_task_queue_mutex);
2020                 surf = surface_for_frame[display_frame_num];
2021                 assert(surf != nullptr);
2022         }
2023         VAStatus va_status;
2024
2025         if (use_zerocopy) {
2026                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2027                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2028                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2029                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2030         } else {
2031                 // Upload the frame to VA-API.
2032                 unsigned char *surface_p = nullptr;
2033                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2034
2035                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2036                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2037
2038                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2039                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2040
2041                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2042                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2043         }
2044
2045         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2046         CHECK_VASTATUS(va_status, "vaDestroyImage");
2047
2048         // Schedule the frame for encoding.
2049         VASurfaceID va_surface = surf->src_surface;
2050         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2051         CHECK_VASTATUS(va_status, "vaBeginPicture");
2052
2053         if (frame_type == FRAME_IDR) {
2054                 // FIXME: If the mux wants global headers, we should not put the
2055                 // SPS/PPS before each IDR frame, but rather put it into the
2056                 // codec extradata (formatted differently?).
2057                 //
2058                 // NOTE: If we change ycbcr_coefficients, it will not take effect
2059                 // before the next IDR frame. This is acceptable, as it should only
2060                 // happen on a mode change, which is rare.
2061                 render_sequence();
2062                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2063                 if (h264_packedheader) {
2064                         render_packedsequence(ycbcr_coefficients);
2065                         render_packedpicture();
2066                 }
2067         } else {
2068                 //render_sequence();
2069                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2070         }
2071         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2072
2073         va_status = vaEndPicture(va_dpy, context_id);
2074         CHECK_VASTATUS(va_status, "vaEndPicture");
2075
2076         update_ReferenceFrames(display_frame_num, frame_type);
2077
2078         vector<size_t> ref_display_frame_numbers;
2079
2080         // Lock the references for this frame; otherwise, they could be
2081         // rendered to before this frame is done encoding.
2082         {
2083                 unique_lock<mutex> lock(storage_task_queue_mutex);
2084                 for (const ReferenceFrame &frame : reference_frames) {
2085                         assert(surface_for_frame.count(frame.display_number));
2086                         ++surface_for_frame[frame.display_number]->refcount;
2087                         ref_display_frame_numbers.push_back(frame.display_number);
2088                 }
2089         }
2090
2091         // so now the data is done encoding (well, async job kicked off)...
2092         // we send that to the storage thread
2093         storage_task tmp;
2094         tmp.display_order = display_frame_num;
2095         tmp.frame_type = frame_type;
2096         tmp.pts = pts;
2097         tmp.dts = dts;
2098         tmp.duration = duration;
2099         tmp.ycbcr_coefficients = ycbcr_coefficients;
2100         tmp.received_ts = received_ts;
2101         tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2102         storage_task_enqueue(move(tmp));
2103 }
2104
2105 // Proxy object.
2106 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)
2107         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2108
2109 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2110 QuickSyncEncoder::~QuickSyncEncoder() {}
2111
2112 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2113 {
2114         impl->add_audio(pts, audio);
2115 }
2116
2117 bool QuickSyncEncoder::is_zerocopy() const
2118 {
2119         return impl->is_zerocopy();
2120 }
2121
2122 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2123 {
2124         return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2125 }
2126
2127 RefCountedGLsync QuickSyncEncoder::end_frame()
2128 {
2129         return impl->end_frame();
2130 }
2131
2132 void QuickSyncEncoder::shutdown()
2133 {
2134         impl->shutdown();
2135 }
2136
2137 void QuickSyncEncoder::close_file()
2138 {
2139         impl->shutdown();
2140 }
2141
2142 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2143 {
2144         impl->set_stream_mux(mux);
2145 }
2146
2147 int64_t QuickSyncEncoder::global_delay() const {
2148         return impl->global_delay();
2149 }