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