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