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

   1 #include "quicksync_encoder.h"
   2
   3 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
   4 #include <movit/util.h>
   5
   6 #include <EGL/eglplatform.h>
   7 #include <X11/Xlib.h>
   8 #include <assert.h>
   9 #include <epoxy/egl.h>
  10 #include <fcntl.h>
  11 #include <stdio.h>
  12 #include <stdlib.h>
  13 #include <string.h>
  14 #include <unistd.h>
  15 #include <va/va.h>
  16 #include <va/va_drm.h>
  17 #include <va/va_drmcommon.h>
  18 #include <va/va_enc_h264.h>
  19 #include <va/va_x11.h>
  20 #include <algorithm>
  21 #include <chrono>
  22 #include <condition_variable>
  23 #include <cstddef>
  24 #include <cstdint>
  25 #include <functional>
  26 #include <map>
  27 #include <memory>
  28 #include <mutex>
  29 #include <queue>
  30 #include <stack>
  31 #include <string>
  32 #include <thread>
  33 #include <utility>
  34
  35 extern "C" {
  36
  37 #include <libavcodec/avcodec.h>
  38 #include <libavformat/avio.h>
  39 #include <libavutil/error.h>
  40 #include <libdrm/drm_fourcc.h>
  41
  42 }  // namespace
  43
  44 #include "audio_encoder.h"
  45 #include "context.h"
  46 #include "defs.h"
  47 #include "disk_space_estimator.h"
  48 #include "ffmpeg_raii.h"
  49 #include "flags.h"
  50 #include "mux.h"
  51 #include "print_latency.h"
  52 #include "quicksync_encoder_impl.h"
  53 #include "ref_counted_frame.h"
  54 #include "timebase.h"
  55 #include "x264_encoder.h"
  56
  57 using namespace std;
  58 using namespace std::chrono;
  59 using namespace std::placeholders;
  60
  61 class QOpenGLContext;
  62 class QSurface;
  63
  64 #define CHECK_VASTATUS(va_status, func)                                 \
  65     if (va_status != VA_STATUS_SUCCESS) {                               \
  66         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  67         exit(1);                                                        \
  68     }
  69
  70 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  71
  72 //#include "loadsurface.h"
  73
  74 #define NAL_REF_IDC_NONE        0
  75 #define NAL_REF_IDC_LOW         1
  76 #define NAL_REF_IDC_MEDIUM      2
  77 #define NAL_REF_IDC_HIGH        3
  78
  79 #define NAL_NON_IDR             1
  80 #define NAL_IDR                 5
  81 #define NAL_SPS                 7
  82 #define NAL_PPS                 8
  83 #define NAL_SEI                 6
  84
  85 #define SLICE_TYPE_P            0
  86 #define SLICE_TYPE_B            1
  87 #define SLICE_TYPE_I            2
  88 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  89 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  90 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  91
  92
  93 #define ENTROPY_MODE_CAVLC      0
  94 #define ENTROPY_MODE_CABAC      1
  95
  96 #define PROFILE_IDC_BASELINE    66
  97 #define PROFILE_IDC_MAIN        77
  98 #define PROFILE_IDC_HIGH        100
  99
 100 #define BITSTREAM_ALLOCATE_STEPPING     4096
 101
 102 static constexpr unsigned int MaxFrameNum = (2<<16);
 103 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
 104 static constexpr unsigned int Log2MaxFrameNum = 16;
 105 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
 106 static constexpr int rc_default_modes[] = {  // Priority list of modes.
 107     VA_RC_VBR,
 108     VA_RC_CQP,
 109     VA_RC_VBR_CONSTRAINED,
 110     VA_RC_CBR,
 111     VA_RC_VCM,
 112     VA_RC_NONE,
 113 };
 114
 115 /* thread to save coded data */
 116 #define SRC_SURFACE_FREE        0
 117 #define SRC_SURFACE_IN_ENCODING 1
 118
 119 using namespace std;
 120
 121 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 122 // but if we don't delete it here, we get leaks. The GStreamer implementation
 123 // does the same.
 124 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 125 {
 126     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 127     CHECK_VASTATUS(va_status, "vaRenderPicture");
 128
 129     for (int i = 0; i < num_buffers; ++i) {
 130         va_status = vaDestroyBuffer(dpy, buffers[i]);
 131         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 132     }
 133 }
 134
 135 static unsigned int
 136 va_swap32(unsigned int val)
 137 {
 138     unsigned char *pval = (unsigned char *)&val;
 139
 140     return ((pval[0] << 24)     |
 141             (pval[1] << 16)     |
 142             (pval[2] << 8)      |
 143             (pval[3] << 0));
 144 }
 145
 146 static void
 147 bitstream_start(bitstream *bs)
 148 {
 149     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 150     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 151     bs->bit_offset = 0;
 152 }
 153
 154 static void
 155 bitstream_end(bitstream *bs)
 156 {
 157     int pos = (bs->bit_offset >> 5);
 158     int bit_offset = (bs->bit_offset & 0x1f);
 159     int bit_left = 32 - bit_offset;
 160
 161     if (bit_offset) {
 162         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 163     }
 164 }
 165
 166 static void
 167 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 168 {
 169     int pos = (bs->bit_offset >> 5);
 170     int bit_offset = (bs->bit_offset & 0x1f);
 171     int bit_left = 32 - bit_offset;
 172
 173     if (!size_in_bits)
 174         return;
 175
 176     bs->bit_offset += size_in_bits;
 177
 178     if (bit_left > size_in_bits) {
 179         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 180     } else {
 181         size_in_bits -= bit_left;
 182         if (bit_left >= 32) {
 183             bs->buffer[pos] = (val >> size_in_bits);
 184         } else {
 185             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 186         }
 187         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 188
 189         if (pos + 1 == bs->max_size_in_dword) {
 190             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 191             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 192         }
 193
 194         bs->buffer[pos + 1] = val;
 195     }
 196 }
 197
 198 static void
 199 bitstream_put_ue(bitstream *bs, unsigned int val)
 200 {
 201     int size_in_bits = 0;
 202     int tmp_val = ++val;
 203
 204     while (tmp_val) {
 205         tmp_val >>= 1;
 206         size_in_bits++;
 207     }
 208
 209     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 210     bitstream_put_ui(bs, val, size_in_bits);
 211 }
 212
 213 static void
 214 bitstream_put_se(bitstream *bs, int val)
 215 {
 216     unsigned int new_val;
 217
 218     if (val <= 0)
 219         new_val = -2 * val;
 220     else
 221         new_val = 2 * val - 1;
 222
 223     bitstream_put_ue(bs, new_val);
 224 }
 225
 226 static void
 227 bitstream_byte_aligning(bitstream *bs, int bit)
 228 {
 229     int bit_offset = (bs->bit_offset & 0x7);
 230     int bit_left = 8 - bit_offset;
 231     int new_val;
 232
 233     if (!bit_offset)
 234         return;
 235
 236     assert(bit == 0 || bit == 1);
 237
 238     if (bit)
 239         new_val = (1 << bit_left) - 1;
 240     else
 241         new_val = 0;
 242
 243     bitstream_put_ui(bs, new_val, bit_left);
 244 }
 245
 246 static void
 247 rbsp_trailing_bits(bitstream *bs)
 248 {
 249     bitstream_put_ui(bs, 1, 1);
 250     bitstream_byte_aligning(bs, 0);
 251 }
 252
 253 static void nal_start_code_prefix(bitstream *bs)
 254 {
 255     bitstream_put_ui(bs, 0x00000001, 32);
 256 }
 257
 258 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 259 {
 260     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 261     bitstream_put_ui(bs, nal_ref_idc, 2);
 262     bitstream_put_ui(bs, nal_unit_type, 5);
 263 }
 264
 265 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 266 {
 267     int profile_idc = PROFILE_IDC_BASELINE;
 268
 269     if (h264_profile  == VAProfileH264High)
 270         profile_idc = PROFILE_IDC_HIGH;
 271     else if (h264_profile  == VAProfileH264Main)
 272         profile_idc = PROFILE_IDC_MAIN;
 273
 274     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 275     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 276     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 277     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 278     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 279     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 280     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 281     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 282
 283     if ( profile_idc == PROFILE_IDC_HIGH) {
 284         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 285         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 286         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 287         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 288         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 289     }
 290
 291     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 292     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 293
 294     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 295         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 296     else {
 297         assert(0);
 298     }
 299
 300     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 301     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 302
 303     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 304     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 305     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 306
 307     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 308         assert(0);
 309     }
 310
 311     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 312     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 313
 314     if (seq_param.frame_cropping_flag) {
 315         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 316         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 317         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 318         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 319     }
 320
 321     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 322     if ( false ) {
 323         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 324     } else {
 325         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 326         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 327         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 328         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 329         {
 330             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 331             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 332             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 333             {
 334                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 335                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 336                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 337             }
 338         }
 339         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 340         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 341         {
 342             bitstream_put_ui(bs, 1, 32);  // FPS
 343             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 344             bitstream_put_ui(bs, 1, 1);
 345         }
 346         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 347         {
 348             // hrd_parameters
 349             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 350             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 351             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 352
 353             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 354             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 355             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 356
 357             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 358             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 359             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 360             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 361         }
 362         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 363         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 364
 365         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 366         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 367     }
 368
 369     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 370 }
 371
 372
 373 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 374 {
 375     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 376     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 377
 378     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 379
 380     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 381
 382     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 383
 384     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 385     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 386
 387     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 388     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 389
 390     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 391     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 392     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 393
 394     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 395     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 396     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 397
 398     /* more_rbsp_data */
 399     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 400     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 401     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 402
 403     rbsp_trailing_bits(bs);
 404 }
 405
 406 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 407 {
 408     int first_mb_in_slice = slice_param.macroblock_address;
 409
 410     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 411     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 412     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 413     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 414
 415     /* frame_mbs_only_flag == 1 */
 416     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 417         /* FIXME: */
 418         assert(0);
 419     }
 420
 421     if (pic_param.pic_fields.bits.idr_pic_flag)
 422         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 423
 424     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 425         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 426         /* pic_order_present_flag == 0 */
 427     } else {
 428         /* FIXME: */
 429         assert(0);
 430     }
 431
 432     /* redundant_pic_cnt_present_flag == 0 */
 433     /* slice type */
 434     if (IS_P_SLICE(slice_param.slice_type)) {
 435         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 436
 437         if (slice_param.num_ref_idx_active_override_flag)
 438             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 439
 440         /* ref_pic_list_reordering */
 441         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 442     } else if (IS_B_SLICE(slice_param.slice_type)) {
 443         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 444
 445         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 446
 447         if (slice_param.num_ref_idx_active_override_flag) {
 448             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 449             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 450         }
 451
 452         /* ref_pic_list_reordering */
 453         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 454         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 455     }
 456
 457     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 458          IS_P_SLICE(slice_param.slice_type)) ||
 459         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 460          IS_B_SLICE(slice_param.slice_type))) {
 461         /* FIXME: fill weight/offset table */
 462         assert(0);
 463     }
 464
 465     /* dec_ref_pic_marking */
 466     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 467         unsigned char no_output_of_prior_pics_flag = 0;
 468         unsigned char long_term_reference_flag = 0;
 469         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 470
 471         if (pic_param.pic_fields.bits.idr_pic_flag) {
 472             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 473             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 474         } else {
 475             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 476         }
 477     }
 478
 479     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 480         !IS_I_SLICE(slice_param.slice_type))
 481         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 482
 483     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 484
 485     /* ignore for SP/SI */
 486
 487     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 488         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 489
 490         if (slice_param.disable_deblocking_filter_idc != 1) {
 491             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 492             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 493         }
 494     }
 495
 496     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 497         bitstream_byte_aligning(bs, 1);
 498     }
 499 }
 500
 501 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 502 {
 503     bitstream bs;
 504
 505     bitstream_start(&bs);
 506     nal_start_code_prefix(&bs);
 507     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 508     pps_rbsp(&bs);
 509     bitstream_end(&bs);
 510
 511     *header_buffer = (unsigned char *)bs.buffer;
 512     return bs.bit_offset;
 513 }
 514
 515 int
 516 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 517 {
 518     bitstream bs;
 519
 520     bitstream_start(&bs);
 521     nal_start_code_prefix(&bs);
 522     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 523     sps_rbsp(&bs);
 524     bitstream_end(&bs);
 525
 526     *header_buffer = (unsigned char *)bs.buffer;
 527     return bs.bit_offset;
 528 }
 529
 530 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 531 {
 532     bitstream bs;
 533     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 534     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 535
 536     bitstream_start(&bs);
 537     nal_start_code_prefix(&bs);
 538
 539     if (IS_I_SLICE(slice_param.slice_type)) {
 540         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 541     } else if (IS_P_SLICE(slice_param.slice_type)) {
 542         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 543     } else {
 544         assert(IS_B_SLICE(slice_param.slice_type));
 545         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 546     }
 547
 548     slice_header(&bs);
 549     bitstream_end(&bs);
 550
 551     *header_buffer = (unsigned char *)bs.buffer;
 552     return bs.bit_offset;
 553 }
 554
 555
 556 /*
 557   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 558   1) period between Frame #X and Frame #N = #X - #N
 559   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 560   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 561   4) intra_period and intra_idr_period take precedence over ip_period
 562   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 563      of I/IDR frames, see bellow examples
 564   -------------------------------------------------------------------
 565   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 566   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 567   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 568   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 569   1            1                ignored    IDR IDR IDR IDR...
 570   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 571   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 572   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 573                                               (PBB)(IBB)(PBB)(IBB)...
 574   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 575                                            IDRPPPPP IPPPPP IPPPPP...
 576   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 577                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 578                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 579                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 580                                            {IDR(PBB)(PBB)}                     (6/6/3)
 581                                            {IDR(PBB)(PBB)}.
 582 */
 583
 584 // General pts/dts strategy:
 585 //
 586 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 587 // bit tricky. We assume first of all that the frame rate never goes _above_
 588 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 589 // in at least this speed, as long at dts <= pts (the frame is not attempted
 590 // presented before it is decoded). Furthermore, we never have longer chains of
 591 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 592 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 593 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 594 // frame, with an order O <= C.)
 595 //
 596 // Many strategies are possible, but we establish these rules:
 597 //
 598 //  - Tip frames have dts = pts - (C-O)*N.
 599 //  - Non-tip frames have dts = dts_last + N.
 600 //
 601 // An example, with C=2 and N=10 and the data flow showed with arrows:
 602 //
 603 //        I  B  P  B  B  P
 604 //   pts: 30 40 50 60 70 80
 605 //        ↓  ↓     ↓
 606 //   dts: 10 30 20 60 50←40
 607 //         |  |  ↑        ↑
 608 //         `--|--'        |
 609 //             `----------'
 610 //
 611 // To show that this works fine also with irregular spacings, let's say that
 612 // the third frame is delayed a bit (something earlier was dropped). Now the
 613 // situation looks like this:
 614 //
 615 //        I  B  P  B  B   P
 616 //   pts: 30 40 80 90 100 110
 617 //        ↓  ↓     ↓
 618 //   dts: 10 30 20 90 50←40
 619 //         |  |  ↑        ↑
 620 //         `--|--'        |
 621 //             `----------'
 622 //
 623 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 624 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 625 //
 626 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 627 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 628 // a tip frame and should be given a dts based on the previous one.
 629 #define FRAME_P 0
 630 #define FRAME_B 1
 631 #define FRAME_I 2
 632 #define FRAME_IDR 7
 633 void encoding2display_order(
 634     int encoding_order, int intra_period,
 635     int intra_idr_period, int ip_period,
 636     int *displaying_order,
 637     int *frame_type, int *pts_lag)
 638 {
 639     int encoding_order_gop = 0;
 640
 641     *pts_lag = 0;
 642
 643     if (intra_period == 1) { /* all are I/IDR frames */
 644         *displaying_order = encoding_order;
 645         if (intra_idr_period == 0)
 646             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 647         else
 648             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 649         return;
 650     }
 651
 652     if (intra_period == 0)
 653         intra_idr_period = 0;
 654
 655     if (ip_period == 1) {
 656         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 657         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 658         *displaying_order = encoding_order;
 659
 660         if (encoding_order_gop == 0) { /* the first frame */
 661             *frame_type = FRAME_IDR;
 662         } else if (intra_period != 0 && /* have I frames */
 663                    encoding_order_gop >= 2 &&
 664                    (encoding_order_gop % intra_period == 0)) {
 665             *frame_type = FRAME_I;
 666         } else {
 667             *frame_type = FRAME_P;
 668         }
 669         return;
 670     }
 671
 672     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 673     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 674     *pts_lag = -1;  // Most frames are not tip frames.
 675
 676     if (encoding_order_gop == 0) { /* the first frame */
 677         *frame_type = FRAME_IDR;
 678         *displaying_order = encoding_order;
 679         // IDR frames are a special case; I honestly can't find the logic behind
 680         // why this is the right thing, but it seems to line up nicely in practice :-)
 681         *pts_lag = TIMEBASE / MAX_FPS;
 682     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 683         *frame_type = FRAME_B;
 684         *displaying_order = encoding_order - 1;
 685         if ((encoding_order_gop % ip_period) == 0) {
 686             *pts_lag = 0;  // Last B-frame.
 687         }
 688     } else if (intra_period != 0 && /* have I frames */
 689                encoding_order_gop >= 2 &&
 690                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 691         *frame_type = FRAME_I;
 692         *displaying_order = encoding_order + ip_period - 1;
 693     } else {
 694         *frame_type = FRAME_P;
 695         *displaying_order = encoding_order + ip_period - 1;
 696     }
 697 }
 698
 699
 700 static const char *rc_to_string(int rc_mode)
 701 {
 702     switch (rc_mode) {
 703     case VA_RC_NONE:
 704         return "NONE";
 705     case VA_RC_CBR:
 706         return "CBR";
 707     case VA_RC_VBR:
 708         return "VBR";
 709     case VA_RC_VCM:
 710         return "VCM";
 711     case VA_RC_CQP:
 712         return "CQP";
 713     case VA_RC_VBR_CONSTRAINED:
 714         return "VBR_CONSTRAINED";
 715     default:
 716         return "Unknown";
 717     }
 718 }
 719
 720 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 721 {
 722         if (global_flags.uncompressed_video_to_http) {
 723                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 724                 use_zerocopy = false;
 725         } else if (global_flags.x264_video_to_http) {
 726                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 727                 use_zerocopy = false;
 728         } else {
 729                 use_zerocopy = true;
 730         }
 731 }
 732
 733 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 734 {
 735         if (va_display.empty()) {
 736                 x11_display = XOpenDisplay(NULL);
 737                 if (!x11_display) {
 738                         fprintf(stderr, "error: can't connect to X server!\n");
 739                         return NULL;
 740                 }
 741                 enable_zerocopy_if_possible();
 742                 return vaGetDisplay(x11_display);
 743         } else if (va_display[0] != '/') {
 744                 x11_display = XOpenDisplay(va_display.c_str());
 745                 if (!x11_display) {
 746                         fprintf(stderr, "error: can't connect to X server!\n");
 747                         return NULL;
 748                 }
 749                 enable_zerocopy_if_possible();
 750                 return vaGetDisplay(x11_display);
 751         } else {
 752                 drm_fd = open(va_display.c_str(), O_RDWR);
 753                 if (drm_fd == -1) {
 754                         perror(va_display.c_str());
 755                         return NULL;
 756                 }
 757                 use_zerocopy = false;
 758                 return vaGetDisplayDRM(drm_fd);
 759         }
 760 }
 761
 762 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
 763 {
 764         if (x11_display) {
 765                 XCloseDisplay(x11_display);
 766                 x11_display = nullptr;
 767         }
 768         if (drm_fd != -1) {
 769                 close(drm_fd);
 770         }
 771 }
 772
 773 int QuickSyncEncoderImpl::init_va(const string &va_display)
 774 {
 775     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 776     VAEntrypoint *entrypoints;
 777     int num_entrypoints, slice_entrypoint;
 778     int support_encode = 0;
 779     int major_ver, minor_ver;
 780     VAStatus va_status;
 781     unsigned int i;
 782
 783     va_dpy = va_open_display(va_display);
 784     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 785     CHECK_VASTATUS(va_status, "vaInitialize");
 786
 787     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 788     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 789     if (!entrypoints) {
 790         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 791         exit(1);
 792     }
 793
 794     /* use the highest profile */
 795     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 796         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 797             continue;
 798
 799         h264_profile = profile_list[i];
 800         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 801         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 802             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 803                 support_encode = 1;
 804                 break;
 805             }
 806         }
 807         if (support_encode == 1)
 808             break;
 809     }
 810
 811     if (support_encode == 0) {
 812         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
 813         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
 814         printf("to use VA-API against DRM instead of X11.\n");
 815         exit(1);
 816     } else {
 817         switch (h264_profile) {
 818             case VAProfileH264Baseline:
 819                 ip_period = 1;
 820                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 821                 h264_entropy_mode = 0;
 822                 break;
 823             case VAProfileH264ConstrainedBaseline:
 824                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 825                 ip_period = 1;
 826                 break;
 827
 828             case VAProfileH264Main:
 829                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 830                 break;
 831
 832             case VAProfileH264High:
 833                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 834                 break;
 835             default:
 836                 h264_profile = VAProfileH264Baseline;
 837                 ip_period = 1;
 838                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 839                 break;
 840         }
 841     }
 842
 843     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 844
 845     /* find out the format for the render target, and rate control mode */
 846     for (i = 0; i < VAConfigAttribTypeMax; i++)
 847         attrib[i].type = (VAConfigAttribType)i;
 848
 849     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 850                                       &attrib[0], VAConfigAttribTypeMax);
 851     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 852     /* check the interested configattrib */
 853     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 854         printf("Not find desired YUV420 RT format\n");
 855         exit(1);
 856     } else {
 857         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 858         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 859         config_attrib_num++;
 860     }
 861
 862     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 863         int tmp = attrib[VAConfigAttribRateControl].value;
 864
 865         if (rc_mode == -1 || !(rc_mode & tmp))  {
 866             if (rc_mode != -1) {
 867                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
 868             }
 869
 870             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
 871                 if (rc_default_modes[i] & tmp) {
 872                     rc_mode = rc_default_modes[i];
 873                     break;
 874                 }
 875             }
 876         }
 877
 878         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 879         config_attrib[config_attrib_num].value = rc_mode;
 880         config_attrib_num++;
 881     }
 882
 883
 884     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 885         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 886
 887         h264_packedheader = 1;
 888         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 889         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 890
 891         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 892             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 893         }
 894
 895         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 896             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 897         }
 898
 899         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 900             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 901         }
 902
 903         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
 904             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
 905         }
 906
 907         enc_packed_header_idx = config_attrib_num;
 908         config_attrib_num++;
 909     }
 910
 911     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
 912         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
 913         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 914         config_attrib_num++;
 915     }
 916
 917     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
 918         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
 919     }
 920
 921     free(entrypoints);
 922     return 0;
 923 }
 924
 925 int QuickSyncEncoderImpl::setup_encode()
 926 {
 927     VAStatus va_status;
 928     VASurfaceID *tmp_surfaceid;
 929     int codedbuf_size, i;
 930     VASurfaceID src_surface[SURFACE_NUM];
 931     VASurfaceID ref_surface[SURFACE_NUM];
 932
 933     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
 934             &config_attrib[0], config_attrib_num, &config_id);
 935     CHECK_VASTATUS(va_status, "vaCreateConfig");
 936
 937     /* create source surfaces */
 938     va_status = vaCreateSurfaces(va_dpy,
 939                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 940                                  &src_surface[0], SURFACE_NUM,
 941                                  NULL, 0);
 942     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 943
 944     /* create reference surfaces */
 945     va_status = vaCreateSurfaces(va_dpy,
 946                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 947                                  &ref_surface[0], SURFACE_NUM,
 948                                  NULL, 0);
 949     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 950
 951     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
 952     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
 953     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
 954
 955     /* Create a context for this encode pipe */
 956     va_status = vaCreateContext(va_dpy, config_id,
 957                                 frame_width_mbaligned, frame_height_mbaligned,
 958                                 VA_PROGRESSIVE,
 959                                 tmp_surfaceid, 2 * SURFACE_NUM,
 960                                 &context_id);
 961     CHECK_VASTATUS(va_status, "vaCreateContext");
 962     free(tmp_surfaceid);
 963
 964     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
 965
 966     for (i = 0; i < SURFACE_NUM; i++) {
 967         /* create coded buffer once for all
 968          * other VA buffers which won't be used again after vaRenderPicture.
 969          * so APP can always vaCreateBuffer for every frame
 970          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
 971          * so VA won't maintain the coded buffer
 972          */
 973         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
 974                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
 975         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 976     }
 977
 978     /* create OpenGL objects */
 979     //glGenFramebuffers(SURFACE_NUM, fbos);
 980
 981     for (i = 0; i < SURFACE_NUM; i++) {
 982         if (use_zerocopy) {
 983             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
 984             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
 985         } else {
 986             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
 987             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
 988
 989             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
 990             // buffers, due to potentially differing pitch.
 991             glGenBuffers(1, &gl_surfaces[i].pbo);
 992             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
 993             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
 994             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 995             gl_surfaces[i].y_offset = 0;
 996             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
 997             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
 998             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
 999             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1000         }
1001     }
1002
1003     for (i = 0; i < SURFACE_NUM; i++) {
1004         gl_surfaces[i].src_surface = src_surface[i];
1005         gl_surfaces[i].ref_surface = ref_surface[i];
1006     }
1007
1008     return 0;
1009 }
1010
1011 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1012 //
1013 //   2 1 0 [3] 4 8 9
1014 template<class T, class C>
1015 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1016 {
1017         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1018         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1019         sort(middle, end, less_than);
1020 }
1021
1022 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1023 {
1024     int i;
1025
1026     if (frame_type == FRAME_B)
1027         return;
1028
1029     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1030     numShortTerm++;
1031     if (numShortTerm > num_ref_frames)
1032         numShortTerm = num_ref_frames;
1033     for (i=numShortTerm-1; i>0; i--)
1034         ReferenceFrames[i] = ReferenceFrames[i-1];
1035     ReferenceFrames[0] = CurrentCurrPic;
1036
1037     current_frame_num++;
1038     if (current_frame_num > MaxFrameNum)
1039         current_frame_num = 0;
1040 }
1041
1042
1043 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1044 {
1045     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1046         return a.frame_idx > b.frame_idx;
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     if (frame_type == FRAME_P) {
1056         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1057         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1058     } else if (frame_type == FRAME_B) {
1059         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1060         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1061
1062         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1063         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1064     }
1065
1066     return 0;
1067 }
1068
1069
1070 int QuickSyncEncoderImpl::render_sequence()
1071 {
1072     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1073     VAStatus va_status;
1074     VAEncMiscParameterBuffer *misc_param;
1075     VAEncMiscParameterRateControl *misc_rate_ctrl;
1076
1077     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1078     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1079     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1080     seq_param.bits_per_second = frame_bitrate;
1081
1082     seq_param.intra_period = intra_period;
1083     seq_param.intra_idr_period = intra_idr_period;
1084     seq_param.ip_period = ip_period;
1085
1086     seq_param.max_num_ref_frames = num_ref_frames;
1087     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1088     seq_param.time_scale = TIMEBASE * 2;
1089     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1090     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1091     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1092     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1093     seq_param.seq_fields.bits.chroma_format_idc = 1;
1094     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1095
1096     if (frame_width != frame_width_mbaligned ||
1097         frame_height != frame_height_mbaligned) {
1098         seq_param.frame_cropping_flag = 1;
1099         seq_param.frame_crop_left_offset = 0;
1100         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1101         seq_param.frame_crop_top_offset = 0;
1102         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1103     }
1104
1105     va_status = vaCreateBuffer(va_dpy, context_id,
1106                                VAEncSequenceParameterBufferType,
1107                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1108     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1109
1110     va_status = vaCreateBuffer(va_dpy, context_id,
1111                                VAEncMiscParameterBufferType,
1112                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1113                                1, NULL, &rc_param_buf);
1114     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1115
1116     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1117     misc_param->type = VAEncMiscParameterTypeRateControl;
1118     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1119     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1120     misc_rate_ctrl->bits_per_second = frame_bitrate;
1121     misc_rate_ctrl->target_percentage = 66;
1122     misc_rate_ctrl->window_size = 1000;
1123     misc_rate_ctrl->initial_qp = initial_qp;
1124     misc_rate_ctrl->min_qp = minimal_qp;
1125     misc_rate_ctrl->basic_unit_size = 0;
1126     vaUnmapBuffer(va_dpy, rc_param_buf);
1127
1128     render_id[0] = seq_param_buf;
1129     render_id[1] = rc_param_buf;
1130
1131     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1132
1133     return 0;
1134 }
1135
1136 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1137 {
1138     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1139     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1140     int PicOrderCntMsb, TopFieldOrderCnt;
1141
1142     if (frame_type == FRAME_IDR)
1143         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1144     else {
1145         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1146         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1147     }
1148
1149     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1150         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1151         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1152     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1153              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1154         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1155     else
1156         PicOrderCntMsb = prevPicOrderCntMsb;
1157
1158     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1159
1160     if (frame_type != FRAME_B) {
1161         PicOrderCntMsb_ref = PicOrderCntMsb;
1162         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1163     }
1164
1165     return TopFieldOrderCnt;
1166 }
1167
1168 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1169 {
1170     VABufferID pic_param_buf;
1171     VAStatus va_status;
1172     int i = 0;
1173
1174     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1175     pic_param.CurrPic.frame_idx = current_frame_num;
1176     pic_param.CurrPic.flags = 0;
1177     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1178     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1179     CurrentCurrPic = pic_param.CurrPic;
1180
1181     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1182     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1183         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1184         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1185     }
1186
1187     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1188     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1189     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1190     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1191     pic_param.frame_num = current_frame_num;
1192     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1193     pic_param.last_picture = false;  // FIXME
1194     pic_param.pic_init_qp = initial_qp;
1195
1196     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1197                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1198     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1199
1200     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1201
1202     return 0;
1203 }
1204
1205 int QuickSyncEncoderImpl::render_packedsequence()
1206 {
1207     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1208     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1209     unsigned int length_in_bits;
1210     unsigned char *packedseq_buffer = NULL;
1211     VAStatus va_status;
1212
1213     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1214
1215     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1216
1217     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1218     packedheader_param_buffer.has_emulation_bytes = 0;
1219     va_status = vaCreateBuffer(va_dpy,
1220                                context_id,
1221                                VAEncPackedHeaderParameterBufferType,
1222                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1223                                &packedseq_para_bufid);
1224     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1225
1226     va_status = vaCreateBuffer(va_dpy,
1227                                context_id,
1228                                VAEncPackedHeaderDataBufferType,
1229                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1230                                &packedseq_data_bufid);
1231     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1232
1233     render_id[0] = packedseq_para_bufid;
1234     render_id[1] = packedseq_data_bufid;
1235     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1236
1237     free(packedseq_buffer);
1238
1239     return 0;
1240 }
1241
1242
1243 int QuickSyncEncoderImpl::render_packedpicture()
1244 {
1245     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1246     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1247     unsigned int length_in_bits;
1248     unsigned char *packedpic_buffer = NULL;
1249     VAStatus va_status;
1250
1251     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1252     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1253     packedheader_param_buffer.bit_length = length_in_bits;
1254     packedheader_param_buffer.has_emulation_bytes = 0;
1255
1256     va_status = vaCreateBuffer(va_dpy,
1257                                context_id,
1258                                VAEncPackedHeaderParameterBufferType,
1259                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1260                                &packedpic_para_bufid);
1261     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1262
1263     va_status = vaCreateBuffer(va_dpy,
1264                                context_id,
1265                                VAEncPackedHeaderDataBufferType,
1266                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1267                                &packedpic_data_bufid);
1268     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1269
1270     render_id[0] = packedpic_para_bufid;
1271     render_id[1] = packedpic_data_bufid;
1272     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1273
1274     free(packedpic_buffer);
1275
1276     return 0;
1277 }
1278
1279 void QuickSyncEncoderImpl::render_packedslice()
1280 {
1281     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1282     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1283     unsigned int length_in_bits;
1284     unsigned char *packedslice_buffer = NULL;
1285     VAStatus va_status;
1286
1287     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1288     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1289     packedheader_param_buffer.bit_length = length_in_bits;
1290     packedheader_param_buffer.has_emulation_bytes = 0;
1291
1292     va_status = vaCreateBuffer(va_dpy,
1293                                context_id,
1294                                VAEncPackedHeaderParameterBufferType,
1295                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1296                                &packedslice_para_bufid);
1297     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1298
1299     va_status = vaCreateBuffer(va_dpy,
1300                                context_id,
1301                                VAEncPackedHeaderDataBufferType,
1302                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1303                                &packedslice_data_bufid);
1304     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1305
1306     render_id[0] = packedslice_para_bufid;
1307     render_id[1] = packedslice_data_bufid;
1308     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1309
1310     free(packedslice_buffer);
1311 }
1312
1313 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1314 {
1315     VABufferID slice_param_buf;
1316     VAStatus va_status;
1317     int i;
1318
1319     update_RefPicList(frame_type);
1320
1321     /* one frame, one slice */
1322     slice_param.macroblock_address = 0;
1323     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1324     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1325     if (frame_type == FRAME_IDR) {
1326         if (encoding_frame_num != 0)
1327             ++slice_param.idr_pic_id;
1328     } else if (frame_type == FRAME_P) {
1329         int refpiclist0_max = h264_maxref & 0xffff;
1330         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1331
1332         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1333             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1334             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1335         }
1336     } else if (frame_type == FRAME_B) {
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(storage_task task)
1375 {
1376         VACodedBufferSegment *buf_list = NULL;
1377         VAStatus va_status;
1378
1379         string data;
1380
1381         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].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, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1388
1389         static int frameno = 0;
1390         print_latency("Current QuickSync latency (video inputs → disk mux):",
1391                 task.received_ts, (task.frame_type == FRAME_B), &frameno);
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         for ( ;; ) {
1429                 storage_task current;
1430                 {
1431                         // wait until there's an encoded frame
1432                         unique_lock<mutex> lock(storage_task_queue_mutex);
1433                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1434                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1435                         current = move(storage_task_queue.front());
1436                         storage_task_queue.pop();
1437                 }
1438
1439                 VAStatus va_status;
1440
1441                 // waits for data, then saves it to disk.
1442                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1443                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1444                 save_codeddata(move(current));
1445
1446                 {
1447                         unique_lock<mutex> lock(storage_task_queue_mutex);
1448                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1449                         storage_task_queue_changed.notify_all();
1450                 }
1451         }
1452 }
1453
1454 void QuickSyncEncoderImpl::release_encode()
1455 {
1456         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1457                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1458                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1459                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1460         }
1461
1462         vaDestroyContext(va_dpy, context_id);
1463         vaDestroyConfig(va_dpy, config_id);
1464 }
1465
1466 void QuickSyncEncoderImpl::release_gl_resources()
1467 {
1468         assert(is_shutdown);
1469         if (has_released_gl_resources) {
1470                 return;
1471         }
1472
1473         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1474                 if (!use_zerocopy) {
1475                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1476                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1477                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1478                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1479                 }
1480                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1481                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1482         }
1483
1484         has_released_gl_resources = true;
1485 }
1486
1487 int QuickSyncEncoderImpl::deinit_va()
1488 {
1489     vaTerminate(va_dpy);
1490
1491     va_close_display(va_dpy);
1492
1493     return 0;
1494 }
1495
1496 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1497         : 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)
1498 {
1499         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1500         open_output_file(filename);
1501         file_audio_encoder->add_mux(file_mux.get());
1502
1503         frame_width_mbaligned = (frame_width + 15) & (~15);
1504         frame_height_mbaligned = (frame_height + 15) & (~15);
1505
1506         //print_input();
1507
1508         if (global_flags.x264_video_to_http) {
1509                 assert(x264_encoder != nullptr);
1510         } else {
1511                 assert(x264_encoder == nullptr);
1512         }
1513
1514         init_va(va_display);
1515         setup_encode();
1516
1517         // No frames are ready yet.
1518         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1519
1520         memset(&seq_param, 0, sizeof(seq_param));
1521         memset(&pic_param, 0, sizeof(pic_param));
1522         memset(&slice_param, 0, sizeof(slice_param));
1523
1524         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1525
1526         encode_thread = thread([this]{
1527                 //SDL_GL_MakeCurrent(window, context);
1528                 QOpenGLContext *context = create_context(this->surface);
1529                 eglBindAPI(EGL_OPENGL_API);
1530                 if (!make_current(context, this->surface)) {
1531                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1532                                 eglGetError());
1533                         exit(1);
1534                 }
1535                 encode_thread_func();
1536                 delete_context(context);
1537         });
1538 }
1539
1540 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1541 {
1542         shutdown();
1543         release_gl_resources();
1544 }
1545
1546 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1547 {
1548         assert(!is_shutdown);
1549         {
1550                 // Wait until this frame slot is done encoding.
1551                 unique_lock<mutex> lock(storage_task_queue_mutex);
1552                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1553                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1554                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1555                 }
1556                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1557                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1558                 if (storage_thread_should_quit) return false;
1559         }
1560
1561         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1562         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1563         *y_tex = surf->y_tex;
1564         *cbcr_tex = surf->cbcr_tex;
1565
1566         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1567         CHECK_VASTATUS(va_status, "vaDeriveImage");
1568
1569         if (use_zerocopy) {
1570                 VABufferInfo buf_info;
1571                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1572                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1573                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1574
1575                 // Create Y image.
1576                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1577                 EGLint y_attribs[] = {
1578                         EGL_WIDTH, frame_width,
1579                         EGL_HEIGHT, frame_height,
1580                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1581                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1582                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1583                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1584                         EGL_NONE
1585                 };
1586
1587                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1588                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1589
1590                 // Associate Y image to a texture.
1591                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1592                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1593
1594                 // Create CbCr image.
1595                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1596                 EGLint cbcr_attribs[] = {
1597                         EGL_WIDTH, frame_width,
1598                         EGL_HEIGHT, frame_height,
1599                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1600                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1601                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1602                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1603                         EGL_NONE
1604                 };
1605
1606                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1607                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1608
1609                 // Associate CbCr image to a texture.
1610                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1611                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1612         }
1613
1614         return true;
1615 }
1616
1617 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1618 {
1619         assert(!is_shutdown);
1620         file_audio_encoder->encode_audio(audio, pts + global_delay());
1621 }
1622
1623 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1624 {
1625         assert(!is_shutdown);
1626
1627         if (!use_zerocopy) {
1628                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1629
1630                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1631                 check_error();
1632
1633                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1634                 check_error();
1635
1636                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1637                 check_error();
1638                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1639                 check_error();
1640
1641                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1642                 check_error();
1643                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1644                 check_error();
1645
1646                 glBindTexture(GL_TEXTURE_2D, 0);
1647                 check_error();
1648                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1649                 check_error();
1650
1651                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1652                 check_error();
1653         }
1654
1655         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1656         check_error();
1657         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1658         check_error();
1659
1660         {
1661                 unique_lock<mutex> lock(frame_queue_mutex);
1662                 pending_video_frames.push(PendingFrame{ fence, input_frames, pts, duration });
1663                 ++current_storage_frame;
1664         }
1665         frame_queue_nonempty.notify_all();
1666         return fence;
1667 }
1668
1669 void QuickSyncEncoderImpl::shutdown()
1670 {
1671         if (is_shutdown) {
1672                 return;
1673         }
1674
1675         {
1676                 unique_lock<mutex> lock(frame_queue_mutex);
1677                 encode_thread_should_quit = true;
1678                 frame_queue_nonempty.notify_all();
1679         }
1680         encode_thread.join();
1681         {
1682                 unique_lock<mutex> lock(storage_task_queue_mutex);
1683                 storage_thread_should_quit = true;
1684                 frame_queue_nonempty.notify_all();
1685                 storage_task_queue_changed.notify_all();
1686         }
1687         storage_thread.join();
1688
1689         // Encode any leftover audio in the queues, and also any delayed frames.
1690         file_audio_encoder->encode_last_audio();
1691
1692         release_encode();
1693         deinit_va();
1694         file_mux.reset();
1695         is_shutdown = true;
1696 }
1697
1698 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1699 {
1700         AVFormatContext *avctx = avformat_alloc_context();
1701         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1702         assert(filename.size() < sizeof(avctx->filename) - 1);
1703         strcpy(avctx->filename, filename.c_str());
1704
1705         string url = "file:" + filename;
1706         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1707         if (ret < 0) {
1708                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1709                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1710                 exit(1);
1711         }
1712
1713         string video_extradata = "";  // FIXME: See other comment about global headers.
1714         AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1715         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1716                 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1)));
1717 }
1718
1719 void QuickSyncEncoderImpl::encode_thread_func()
1720 {
1721         int64_t last_dts = -1;
1722         int gop_start_display_frame_num = 0;
1723         for (int display_frame_num = 0; ; ++display_frame_num) {
1724                 // Wait for the frame to be in the queue. Note that this only means
1725                 // we started rendering it.
1726                 PendingFrame frame;
1727                 {
1728                         unique_lock<mutex> lock(frame_queue_mutex);
1729                         frame_queue_nonempty.wait(lock, [this]{
1730                                 return encode_thread_should_quit || !pending_video_frames.empty();
1731                         });
1732                         if (encode_thread_should_quit && pending_video_frames.empty()) {
1733                                 // We may have queued frames left in the reorder buffer
1734                                 // that were supposed to be B-frames, but have no P-frame
1735                                 // to be encoded against. If so, encode them all as
1736                                 // P-frames instead. Note that this happens under the mutex,
1737                                 // but nobody else uses it at this point, since we're shutting down,
1738                                 // so there's no contention.
1739                                 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1740                                 return;
1741                         } else {
1742                                 frame = move(pending_video_frames.front());
1743                                 pending_video_frames.pop();
1744                         }
1745                 }
1746
1747                 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1748                 // Note that this implicitly waits for the frame to be done rendering.
1749                 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1750                 reorder_buffer[display_frame_num] = move(frame);
1751
1752                 // Now encode as many QuickSync frames as we can using the frames we have available.
1753                 // (It could be zero, or it could be multiple.) FIXME: make a function.
1754                 for ( ;; ) {
1755                         int pts_lag;
1756                         int frame_type, quicksync_display_frame_num;
1757                         encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1758                                                &quicksync_display_frame_num, &frame_type, &pts_lag);
1759                         if (!reorder_buffer.count(quicksync_display_frame_num)) {
1760                                 break;
1761                         }
1762                         frame = move(reorder_buffer[quicksync_display_frame_num]);
1763                         reorder_buffer.erase(quicksync_display_frame_num);
1764
1765                         if (frame_type == FRAME_IDR) {
1766                                 numShortTerm = 0;
1767                                 current_frame_num = 0;
1768                                 gop_start_display_frame_num = quicksync_display_frame_num;
1769                         }
1770
1771                         // Determine the dts of this frame.
1772                         int64_t dts;
1773                         if (pts_lag == -1) {
1774                                 assert(last_dts != -1);
1775                                 dts = last_dts + (TIMEBASE / MAX_FPS);
1776                         } else {
1777                                 dts = frame.pts - pts_lag;
1778                         }
1779                         last_dts = dts;
1780
1781                         encode_frame(frame, quicksync_encoding_frame_num, quicksync_display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
1782                         ++quicksync_encoding_frame_num;
1783                 }
1784         }
1785 }
1786
1787 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1788 {
1789         if (reorder_buffer.empty()) {
1790                 return;
1791         }
1792
1793         for (auto &pending_frame : reorder_buffer) {
1794                 int display_frame_num = pending_frame.first;
1795                 assert(display_frame_num > 0);
1796                 PendingFrame frame = move(pending_frame.second);
1797                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1798                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1799                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
1800                 last_dts = dts;
1801         }
1802 }
1803
1804 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1805 {
1806         AVPacket pkt;
1807         memset(&pkt, 0, sizeof(pkt));
1808         pkt.buf = nullptr;
1809         pkt.data = const_cast<uint8_t *>(data);
1810         pkt.size = frame_width * frame_height * 2;
1811         pkt.stream_index = 0;
1812         pkt.flags = AV_PKT_FLAG_KEY;
1813         pkt.duration = duration;
1814         stream_mux->add_packet(pkt, pts, pts);
1815 }
1816
1817 namespace {
1818
1819 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1820 {
1821         if (src_width == dst_pitch) {
1822                 memcpy(dst, src, src_width * height);
1823         } else {
1824                 for (size_t y = 0; y < height; ++y) {
1825                         const uint8_t *sptr = src + y * src_width;
1826                         uint8_t *dptr = dst + y * dst_pitch;
1827                         memcpy(dptr, sptr, src_width);
1828                 }
1829         }
1830 }
1831
1832 ReceivedTimestamps find_received_timestamp(const vector<RefCountedFrame> &input_frames)
1833 {
1834         // Find min and max timestamp of all input frames that have a timestamp.
1835         steady_clock::time_point min_ts = steady_clock::time_point::max(), max_ts = steady_clock::time_point::min();
1836         for (const RefCountedFrame &input_frame : input_frames) {
1837                 if (input_frame && input_frame->received_timestamp > steady_clock::time_point::min()) {
1838                         min_ts = min(min_ts, input_frame->received_timestamp);
1839                         max_ts = max(max_ts, input_frame->received_timestamp);
1840                 }
1841         }
1842         return { min_ts, max_ts };
1843 }
1844
1845 }  // namespace
1846
1847 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1848 {
1849         // Wait for the GPU to be done with the frame.
1850         GLenum sync_status;
1851         do {
1852                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1853                 check_error();
1854         } while (sync_status == GL_TIMEOUT_EXPIRED);
1855         assert(sync_status != GL_WAIT_FAILED);
1856
1857         ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1858         static int frameno = 0;
1859         print_latency("Current mixer latency (video inputs → ready for encode):",
1860                 received_ts, false, &frameno);
1861
1862         // Release back any input frames we needed to render this frame.
1863         frame.input_frames.clear();
1864
1865         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1866         uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1867         if (global_flags.uncompressed_video_to_http) {
1868                 add_packet_for_uncompressed_frame(pts, duration, data);
1869         } else if (global_flags.x264_video_to_http) {
1870                 x264_encoder->add_frame(pts, duration, data, received_ts);
1871         }
1872 }
1873
1874 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1875                                         int frame_type, int64_t pts, int64_t dts, int64_t duration)
1876 {
1877         const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1878
1879         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1880         VAStatus va_status;
1881
1882         if (use_zerocopy) {
1883                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1884                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1885                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1886                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1887         } else {
1888                 // Upload the frame to VA-API.
1889                 unsigned char *surface_p = nullptr;
1890                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
1891
1892                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1893                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1894
1895                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1896                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1897
1898                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
1899                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1900         }
1901
1902         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1903         CHECK_VASTATUS(va_status, "vaDestroyImage");
1904
1905         // Schedule the frame for encoding.
1906         VASurfaceID va_surface = surf->src_surface;
1907         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
1908         CHECK_VASTATUS(va_status, "vaBeginPicture");
1909
1910         if (frame_type == FRAME_IDR) {
1911                 // FIXME: If the mux wants global headers, we should not put the
1912                 // SPS/PPS before each IDR frame, but rather put it into the
1913                 // codec extradata (formatted differently?).
1914                 render_sequence();
1915                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1916                 if (h264_packedheader) {
1917                         render_packedsequence();
1918                         render_packedpicture();
1919                 }
1920         } else {
1921                 //render_sequence();
1922                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1923         }
1924         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1925
1926         va_status = vaEndPicture(va_dpy, context_id);
1927         CHECK_VASTATUS(va_status, "vaEndPicture");
1928
1929         // so now the data is done encoding (well, async job kicked off)...
1930         // we send that to the storage thread
1931         storage_task tmp;
1932         tmp.display_order = display_frame_num;
1933         tmp.frame_type = frame_type;
1934         tmp.pts = pts;
1935         tmp.dts = dts;
1936         tmp.duration = duration;
1937         tmp.received_ts = received_ts;
1938         storage_task_enqueue(move(tmp));
1939
1940         update_ReferenceFrames(frame_type);
1941 }
1942
1943 // Proxy object.
1944 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1945         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
1946
1947 // Must be defined here because unique_ptr<> destructor needs to know the impl.
1948 QuickSyncEncoder::~QuickSyncEncoder() {}
1949
1950 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
1951 {
1952         impl->add_audio(pts, audio);
1953 }
1954
1955 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1956 {
1957         return impl->begin_frame(y_tex, cbcr_tex);
1958 }
1959
1960 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1961 {
1962         return impl->end_frame(pts, duration, input_frames);
1963 }
1964
1965 void QuickSyncEncoder::shutdown()
1966 {
1967         impl->shutdown();
1968 }
1969
1970 void QuickSyncEncoder::set_stream_mux(Mux *mux)
1971 {
1972         impl->set_stream_mux(mux);
1973 }
1974
1975 int64_t QuickSyncEncoder::global_delay() const {
1976         return impl->global_delay();
1977 }