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