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