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         // See H.264 annex E for the definition of this header.
 326         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 327         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 328         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 329         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 330         {
 331             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 332             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 333             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 334             {
 335                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 336                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 337                 if (global_flags.ycbcr_rec709_coefficients) {
 338                     bitstream_put_ui(bs, 1, 8);  /* matrix_coefficients (1 = BT.709) */
 339                 } else {
 340                     bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 341                 }
 342             }
 343         }
 344         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 345         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 346         {
 347             bitstream_put_ui(bs, 1, 32);  // FPS
 348             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 349             bitstream_put_ui(bs, 1, 1);
 350         }
 351         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 352         {
 353             // hrd_parameters
 354             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 355             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 356             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 357
 358             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 359             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 360             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 361
 362             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 363             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 364             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 365             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 366         }
 367         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 368         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 369
 370         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 371         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 372     }
 373
 374     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 375 }
 376
 377
 378 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 379 {
 380     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 381     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 382
 383     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 384
 385     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 386
 387     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 388
 389     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 390     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 391
 392     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 393     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 394
 395     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 396     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 397     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 398
 399     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 400     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 401     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 402
 403     /* more_rbsp_data */
 404     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 405     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 406     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 407
 408     rbsp_trailing_bits(bs);
 409 }
 410
 411 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 412 {
 413     int first_mb_in_slice = slice_param.macroblock_address;
 414
 415     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 416     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 417     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 418     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 419
 420     /* frame_mbs_only_flag == 1 */
 421     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 422         /* FIXME: */
 423         assert(0);
 424     }
 425
 426     if (pic_param.pic_fields.bits.idr_pic_flag)
 427         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 428
 429     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 430         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 431         /* pic_order_present_flag == 0 */
 432     } else {
 433         /* FIXME: */
 434         assert(0);
 435     }
 436
 437     /* redundant_pic_cnt_present_flag == 0 */
 438     /* slice type */
 439     if (IS_P_SLICE(slice_param.slice_type)) {
 440         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 441
 442         if (slice_param.num_ref_idx_active_override_flag)
 443             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 444
 445         /* ref_pic_list_reordering */
 446         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 447     } else if (IS_B_SLICE(slice_param.slice_type)) {
 448         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 449
 450         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 451
 452         if (slice_param.num_ref_idx_active_override_flag) {
 453             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 454             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 455         }
 456
 457         /* ref_pic_list_reordering */
 458         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 459         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 460     }
 461
 462     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 463          IS_P_SLICE(slice_param.slice_type)) ||
 464         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 465          IS_B_SLICE(slice_param.slice_type))) {
 466         /* FIXME: fill weight/offset table */
 467         assert(0);
 468     }
 469
 470     /* dec_ref_pic_marking */
 471     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 472         unsigned char no_output_of_prior_pics_flag = 0;
 473         unsigned char long_term_reference_flag = 0;
 474         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 475
 476         if (pic_param.pic_fields.bits.idr_pic_flag) {
 477             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 478             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 479         } else {
 480             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 481         }
 482     }
 483
 484     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 485         !IS_I_SLICE(slice_param.slice_type))
 486         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 487
 488     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 489
 490     /* ignore for SP/SI */
 491
 492     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 493         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 494
 495         if (slice_param.disable_deblocking_filter_idc != 1) {
 496             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 497             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 498         }
 499     }
 500
 501     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 502         bitstream_byte_aligning(bs, 1);
 503     }
 504 }
 505
 506 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 507 {
 508     bitstream bs;
 509
 510     bitstream_start(&bs);
 511     nal_start_code_prefix(&bs);
 512     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 513     pps_rbsp(&bs);
 514     bitstream_end(&bs);
 515
 516     *header_buffer = (unsigned char *)bs.buffer;
 517     return bs.bit_offset;
 518 }
 519
 520 int
 521 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 522 {
 523     bitstream bs;
 524
 525     bitstream_start(&bs);
 526     nal_start_code_prefix(&bs);
 527     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 528     sps_rbsp(&bs);
 529     bitstream_end(&bs);
 530
 531     *header_buffer = (unsigned char *)bs.buffer;
 532     return bs.bit_offset;
 533 }
 534
 535 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 536 {
 537     bitstream bs;
 538     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 539     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 540
 541     bitstream_start(&bs);
 542     nal_start_code_prefix(&bs);
 543
 544     if (IS_I_SLICE(slice_param.slice_type)) {
 545         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 546     } else if (IS_P_SLICE(slice_param.slice_type)) {
 547         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 548     } else {
 549         assert(IS_B_SLICE(slice_param.slice_type));
 550         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 551     }
 552
 553     slice_header(&bs);
 554     bitstream_end(&bs);
 555
 556     *header_buffer = (unsigned char *)bs.buffer;
 557     return bs.bit_offset;
 558 }
 559
 560
 561 /*
 562   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 563   1) period between Frame #X and Frame #N = #X - #N
 564   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 565   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 566   4) intra_period and intra_idr_period take precedence over ip_period
 567   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 568      of I/IDR frames, see bellow examples
 569   -------------------------------------------------------------------
 570   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 571   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 572   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 573   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 574   1            1                ignored    IDR IDR IDR IDR...
 575   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 576   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 577   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 578                                               (PBB)(IBB)(PBB)(IBB)...
 579   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 580                                            IDRPPPPP IPPPPP IPPPPP...
 581   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 582                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 583                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 584                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 585                                            {IDR(PBB)(PBB)}                     (6/6/3)
 586                                            {IDR(PBB)(PBB)}.
 587 */
 588
 589 // General pts/dts strategy:
 590 //
 591 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 592 // bit tricky. We assume first of all that the frame rate never goes _above_
 593 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 594 // in at least this speed, as long at dts <= pts (the frame is not attempted
 595 // presented before it is decoded). Furthermore, we never have longer chains of
 596 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 597 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 598 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 599 // frame, with an order O <= C.)
 600 //
 601 // Many strategies are possible, but we establish these rules:
 602 //
 603 //  - Tip frames have dts = pts - (C-O)*N.
 604 //  - Non-tip frames have dts = dts_last + N.
 605 //
 606 // An example, with C=2 and N=10 and the data flow showed with arrows:
 607 //
 608 //        I  B  P  B  B  P
 609 //   pts: 30 40 50 60 70 80
 610 //        ↓  ↓     ↓
 611 //   dts: 10 30 20 60 50←40
 612 //         |  |  ↑        ↑
 613 //         `--|--'        |
 614 //             `----------'
 615 //
 616 // To show that this works fine also with irregular spacings, let's say that
 617 // the third frame is delayed a bit (something earlier was dropped). Now the
 618 // situation looks like this:
 619 //
 620 //        I  B  P  B  B   P
 621 //   pts: 30 40 80 90 100 110
 622 //        ↓  ↓     ↓
 623 //   dts: 10 30 20 90 50←40
 624 //         |  |  ↑        ↑
 625 //         `--|--'        |
 626 //             `----------'
 627 //
 628 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 629 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 630 //
 631 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 632 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 633 // a tip frame and should be given a dts based on the previous one.
 634 #define FRAME_P 0
 635 #define FRAME_B 1
 636 #define FRAME_I 2
 637 #define FRAME_IDR 7
 638 void encoding2display_order(
 639     int encoding_order, int intra_period,
 640     int intra_idr_period, int ip_period,
 641     int *displaying_order,
 642     int *frame_type, int *pts_lag)
 643 {
 644     int encoding_order_gop = 0;
 645
 646     *pts_lag = 0;
 647
 648     if (intra_period == 1) { /* all are I/IDR frames */
 649         *displaying_order = encoding_order;
 650         if (intra_idr_period == 0)
 651             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 652         else
 653             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 654         return;
 655     }
 656
 657     if (intra_period == 0)
 658         intra_idr_period = 0;
 659
 660     if (ip_period == 1) {
 661         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 662         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 663         *displaying_order = encoding_order;
 664
 665         if (encoding_order_gop == 0) { /* the first frame */
 666             *frame_type = FRAME_IDR;
 667         } else if (intra_period != 0 && /* have I frames */
 668                    encoding_order_gop >= 2 &&
 669                    (encoding_order_gop % intra_period == 0)) {
 670             *frame_type = FRAME_I;
 671         } else {
 672             *frame_type = FRAME_P;
 673         }
 674         return;
 675     }
 676
 677     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 678     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 679     *pts_lag = -1;  // Most frames are not tip frames.
 680
 681     if (encoding_order_gop == 0) { /* the first frame */
 682         *frame_type = FRAME_IDR;
 683         *displaying_order = encoding_order;
 684         // IDR frames are a special case; I honestly can't find the logic behind
 685         // why this is the right thing, but it seems to line up nicely in practice :-)
 686         *pts_lag = TIMEBASE / MAX_FPS;
 687     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 688         *frame_type = FRAME_B;
 689         *displaying_order = encoding_order - 1;
 690         if ((encoding_order_gop % ip_period) == 0) {
 691             *pts_lag = 0;  // Last B-frame.
 692         }
 693     } else if (intra_period != 0 && /* have I frames */
 694                encoding_order_gop >= 2 &&
 695                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 696         *frame_type = FRAME_I;
 697         *displaying_order = encoding_order + ip_period - 1;
 698     } else {
 699         *frame_type = FRAME_P;
 700         *displaying_order = encoding_order + ip_period - 1;
 701     }
 702 }
 703
 704
 705 static const char *rc_to_string(int rc_mode)
 706 {
 707     switch (rc_mode) {
 708     case VA_RC_NONE:
 709         return "NONE";
 710     case VA_RC_CBR:
 711         return "CBR";
 712     case VA_RC_VBR:
 713         return "VBR";
 714     case VA_RC_VCM:
 715         return "VCM";
 716     case VA_RC_CQP:
 717         return "CQP";
 718     case VA_RC_VBR_CONSTRAINED:
 719         return "VBR_CONSTRAINED";
 720     default:
 721         return "Unknown";
 722     }
 723 }
 724
 725 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 726 {
 727         if (global_flags.uncompressed_video_to_http) {
 728                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 729                 use_zerocopy = false;
 730         } else if (global_flags.x264_video_to_http) {
 731                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 732                 use_zerocopy = false;
 733         } else {
 734                 use_zerocopy = true;
 735         }
 736 }
 737
 738 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 739 {
 740         if (va_display.empty()) {
 741                 x11_display = XOpenDisplay(NULL);
 742                 if (!x11_display) {
 743                         fprintf(stderr, "error: can't connect to X server!\n");
 744                         return NULL;
 745                 }
 746                 enable_zerocopy_if_possible();
 747                 return vaGetDisplay(x11_display);
 748         } else if (va_display[0] != '/') {
 749                 x11_display = XOpenDisplay(va_display.c_str());
 750                 if (!x11_display) {
 751                         fprintf(stderr, "error: can't connect to X server!\n");
 752                         return NULL;
 753                 }
 754                 enable_zerocopy_if_possible();
 755                 return vaGetDisplay(x11_display);
 756         } else {
 757                 drm_fd = open(va_display.c_str(), O_RDWR);
 758                 if (drm_fd == -1) {
 759                         perror(va_display.c_str());
 760                         return NULL;
 761                 }
 762                 use_zerocopy = false;
 763                 return vaGetDisplayDRM(drm_fd);
 764         }
 765 }
 766
 767 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
 768 {
 769         if (x11_display) {
 770                 XCloseDisplay(x11_display);
 771                 x11_display = nullptr;
 772         }
 773         if (drm_fd != -1) {
 774                 close(drm_fd);
 775         }
 776 }
 777
 778 int QuickSyncEncoderImpl::init_va(const string &va_display)
 779 {
 780     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 781     VAEntrypoint *entrypoints;
 782     int num_entrypoints, slice_entrypoint;
 783     int support_encode = 0;
 784     int major_ver, minor_ver;
 785     VAStatus va_status;
 786     unsigned int i;
 787
 788     va_dpy = va_open_display(va_display);
 789     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 790     CHECK_VASTATUS(va_status, "vaInitialize");
 791
 792     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 793     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 794     if (!entrypoints) {
 795         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 796         exit(1);
 797     }
 798
 799     /* use the highest profile */
 800     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 801         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 802             continue;
 803
 804         h264_profile = profile_list[i];
 805         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 806         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 807             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 808                 support_encode = 1;
 809                 break;
 810             }
 811         }
 812         if (support_encode == 1)
 813             break;
 814     }
 815
 816     if (support_encode == 0) {
 817         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
 818         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
 819         printf("to use VA-API against DRM instead of X11.\n");
 820         exit(1);
 821     } else {
 822         switch (h264_profile) {
 823             case VAProfileH264Baseline:
 824                 ip_period = 1;
 825                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 826                 h264_entropy_mode = 0;
 827                 break;
 828             case VAProfileH264ConstrainedBaseline:
 829                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 830                 ip_period = 1;
 831                 break;
 832
 833             case VAProfileH264Main:
 834                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 835                 break;
 836
 837             case VAProfileH264High:
 838                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 839                 break;
 840             default:
 841                 h264_profile = VAProfileH264Baseline;
 842                 ip_period = 1;
 843                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 844                 break;
 845         }
 846     }
 847
 848     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 849
 850     /* find out the format for the render target, and rate control mode */
 851     for (i = 0; i < VAConfigAttribTypeMax; i++)
 852         attrib[i].type = (VAConfigAttribType)i;
 853
 854     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 855                                       &attrib[0], VAConfigAttribTypeMax);
 856     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 857     /* check the interested configattrib */
 858     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 859         printf("Not find desired YUV420 RT format\n");
 860         exit(1);
 861     } else {
 862         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 863         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 864         config_attrib_num++;
 865     }
 866
 867     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 868         int tmp = attrib[VAConfigAttribRateControl].value;
 869
 870         if (rc_mode == -1 || !(rc_mode & tmp))  {
 871             if (rc_mode != -1) {
 872                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
 873             }
 874
 875             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
 876                 if (rc_default_modes[i] & tmp) {
 877                     rc_mode = rc_default_modes[i];
 878                     break;
 879                 }
 880             }
 881         }
 882
 883         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 884         config_attrib[config_attrib_num].value = rc_mode;
 885         config_attrib_num++;
 886     }
 887
 888
 889     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 890         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 891
 892         h264_packedheader = 1;
 893         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 894         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 895
 896         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 897             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 898         }
 899
 900         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 901             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 902         }
 903
 904         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 905             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 906         }
 907
 908         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
 909             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
 910         }
 911
 912         enc_packed_header_idx = config_attrib_num;
 913         config_attrib_num++;
 914     }
 915
 916     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
 917         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
 918         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 919         config_attrib_num++;
 920     }
 921
 922     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
 923         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
 924     }
 925
 926     free(entrypoints);
 927     return 0;
 928 }
 929
 930 int QuickSyncEncoderImpl::setup_encode()
 931 {
 932     VAStatus va_status;
 933     VASurfaceID *tmp_surfaceid;
 934     int codedbuf_size, i;
 935     VASurfaceID src_surface[SURFACE_NUM];
 936     VASurfaceID ref_surface[SURFACE_NUM];
 937
 938     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
 939             &config_attrib[0], config_attrib_num, &config_id);
 940     CHECK_VASTATUS(va_status, "vaCreateConfig");
 941
 942     /* create source surfaces */
 943     va_status = vaCreateSurfaces(va_dpy,
 944                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 945                                  &src_surface[0], SURFACE_NUM,
 946                                  NULL, 0);
 947     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 948
 949     /* create reference surfaces */
 950     va_status = vaCreateSurfaces(va_dpy,
 951                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
 952                                  &ref_surface[0], SURFACE_NUM,
 953                                  NULL, 0);
 954     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 955
 956     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
 957     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
 958     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
 959
 960     /* Create a context for this encode pipe */
 961     va_status = vaCreateContext(va_dpy, config_id,
 962                                 frame_width_mbaligned, frame_height_mbaligned,
 963                                 VA_PROGRESSIVE,
 964                                 tmp_surfaceid, 2 * SURFACE_NUM,
 965                                 &context_id);
 966     CHECK_VASTATUS(va_status, "vaCreateContext");
 967     free(tmp_surfaceid);
 968
 969     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
 970
 971     for (i = 0; i < SURFACE_NUM; i++) {
 972         /* create coded buffer once for all
 973          * other VA buffers which won't be used again after vaRenderPicture.
 974          * so APP can always vaCreateBuffer for every frame
 975          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
 976          * so VA won't maintain the coded buffer
 977          */
 978         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
 979                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
 980         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 981     }
 982
 983     /* create OpenGL objects */
 984     //glGenFramebuffers(SURFACE_NUM, fbos);
 985
 986     for (i = 0; i < SURFACE_NUM; i++) {
 987         if (use_zerocopy) {
 988             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
 989             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
 990         } else {
 991             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
 992             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
 993
 994             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
 995             // buffers, due to potentially differing pitch.
 996             glGenBuffers(1, &gl_surfaces[i].pbo);
 997             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
 998             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
 999             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1000             gl_surfaces[i].y_offset = 0;
1001             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1002             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1003             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1004             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1005         }
1006     }
1007
1008     for (i = 0; i < SURFACE_NUM; i++) {
1009         gl_surfaces[i].src_surface = src_surface[i];
1010         gl_surfaces[i].ref_surface = ref_surface[i];
1011     }
1012
1013     return 0;
1014 }
1015
1016 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1017 //
1018 //   2 1 0 [3] 4 8 9
1019 template<class T, class C>
1020 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1021 {
1022         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1023         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1024         sort(middle, end, less_than);
1025 }
1026
1027 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1028 {
1029     int i;
1030
1031     if (frame_type == FRAME_B)
1032         return;
1033
1034     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1035     numShortTerm++;
1036     if (numShortTerm > num_ref_frames)
1037         numShortTerm = num_ref_frames;
1038     for (i=numShortTerm-1; i>0; i--)
1039         ReferenceFrames[i] = ReferenceFrames[i-1];
1040     ReferenceFrames[0] = CurrentCurrPic;
1041
1042     current_frame_num++;
1043     if (current_frame_num > MaxFrameNum)
1044         current_frame_num = 0;
1045 }
1046
1047
1048 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1049 {
1050     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1051         return a.frame_idx > b.frame_idx;
1052     };
1053     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1054         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1055     };
1056     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1057         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1058     };
1059
1060     if (frame_type == FRAME_P) {
1061         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1062         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1063     } else if (frame_type == FRAME_B) {
1064         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1065         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1066
1067         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1068         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1069     }
1070
1071     return 0;
1072 }
1073
1074
1075 int QuickSyncEncoderImpl::render_sequence()
1076 {
1077     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1078     VAStatus va_status;
1079     VAEncMiscParameterBuffer *misc_param;
1080     VAEncMiscParameterRateControl *misc_rate_ctrl;
1081
1082     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1083     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1084     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1085     seq_param.bits_per_second = frame_bitrate;
1086
1087     seq_param.intra_period = intra_period;
1088     seq_param.intra_idr_period = intra_idr_period;
1089     seq_param.ip_period = ip_period;
1090
1091     seq_param.max_num_ref_frames = num_ref_frames;
1092     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1093     seq_param.time_scale = TIMEBASE * 2;
1094     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1095     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1096     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1097     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1098     seq_param.seq_fields.bits.chroma_format_idc = 1;
1099     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1100
1101     if (frame_width != frame_width_mbaligned ||
1102         frame_height != frame_height_mbaligned) {
1103         seq_param.frame_cropping_flag = 1;
1104         seq_param.frame_crop_left_offset = 0;
1105         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1106         seq_param.frame_crop_top_offset = 0;
1107         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1108     }
1109
1110     va_status = vaCreateBuffer(va_dpy, context_id,
1111                                VAEncSequenceParameterBufferType,
1112                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1113     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1114
1115     va_status = vaCreateBuffer(va_dpy, context_id,
1116                                VAEncMiscParameterBufferType,
1117                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1118                                1, NULL, &rc_param_buf);
1119     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1120
1121     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1122     misc_param->type = VAEncMiscParameterTypeRateControl;
1123     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1124     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1125     misc_rate_ctrl->bits_per_second = frame_bitrate;
1126     misc_rate_ctrl->target_percentage = 66;
1127     misc_rate_ctrl->window_size = 1000;
1128     misc_rate_ctrl->initial_qp = initial_qp;
1129     misc_rate_ctrl->min_qp = minimal_qp;
1130     misc_rate_ctrl->basic_unit_size = 0;
1131     vaUnmapBuffer(va_dpy, rc_param_buf);
1132
1133     render_id[0] = seq_param_buf;
1134     render_id[1] = rc_param_buf;
1135
1136     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1137
1138     return 0;
1139 }
1140
1141 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1142 {
1143     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1144     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1145     int PicOrderCntMsb, TopFieldOrderCnt;
1146
1147     if (frame_type == FRAME_IDR)
1148         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1149     else {
1150         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1151         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1152     }
1153
1154     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1155         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1156         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1157     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1158              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1159         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1160     else
1161         PicOrderCntMsb = prevPicOrderCntMsb;
1162
1163     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1164
1165     if (frame_type != FRAME_B) {
1166         PicOrderCntMsb_ref = PicOrderCntMsb;
1167         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1168     }
1169
1170     return TopFieldOrderCnt;
1171 }
1172
1173 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1174 {
1175     VABufferID pic_param_buf;
1176     VAStatus va_status;
1177     int i = 0;
1178
1179     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1180     pic_param.CurrPic.frame_idx = current_frame_num;
1181     pic_param.CurrPic.flags = 0;
1182     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1183     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1184     CurrentCurrPic = pic_param.CurrPic;
1185
1186     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1187     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1188         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1189         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1190     }
1191
1192     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1193     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1194     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1195     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1196     pic_param.frame_num = current_frame_num;
1197     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1198     pic_param.last_picture = false;  // FIXME
1199     pic_param.pic_init_qp = initial_qp;
1200
1201     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1202                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1203     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1204
1205     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1206
1207     return 0;
1208 }
1209
1210 int QuickSyncEncoderImpl::render_packedsequence()
1211 {
1212     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1213     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1214     unsigned int length_in_bits;
1215     unsigned char *packedseq_buffer = NULL;
1216     VAStatus va_status;
1217
1218     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1219
1220     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1221
1222     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1223     packedheader_param_buffer.has_emulation_bytes = 0;
1224     va_status = vaCreateBuffer(va_dpy,
1225                                context_id,
1226                                VAEncPackedHeaderParameterBufferType,
1227                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1228                                &packedseq_para_bufid);
1229     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1230
1231     va_status = vaCreateBuffer(va_dpy,
1232                                context_id,
1233                                VAEncPackedHeaderDataBufferType,
1234                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1235                                &packedseq_data_bufid);
1236     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1237
1238     render_id[0] = packedseq_para_bufid;
1239     render_id[1] = packedseq_data_bufid;
1240     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1241
1242     free(packedseq_buffer);
1243
1244     return 0;
1245 }
1246
1247
1248 int QuickSyncEncoderImpl::render_packedpicture()
1249 {
1250     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1251     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1252     unsigned int length_in_bits;
1253     unsigned char *packedpic_buffer = NULL;
1254     VAStatus va_status;
1255
1256     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1257     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1258     packedheader_param_buffer.bit_length = length_in_bits;
1259     packedheader_param_buffer.has_emulation_bytes = 0;
1260
1261     va_status = vaCreateBuffer(va_dpy,
1262                                context_id,
1263                                VAEncPackedHeaderParameterBufferType,
1264                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1265                                &packedpic_para_bufid);
1266     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1267
1268     va_status = vaCreateBuffer(va_dpy,
1269                                context_id,
1270                                VAEncPackedHeaderDataBufferType,
1271                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1272                                &packedpic_data_bufid);
1273     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1274
1275     render_id[0] = packedpic_para_bufid;
1276     render_id[1] = packedpic_data_bufid;
1277     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1278
1279     free(packedpic_buffer);
1280
1281     return 0;
1282 }
1283
1284 void QuickSyncEncoderImpl::render_packedslice()
1285 {
1286     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1287     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1288     unsigned int length_in_bits;
1289     unsigned char *packedslice_buffer = NULL;
1290     VAStatus va_status;
1291
1292     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1293     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1294     packedheader_param_buffer.bit_length = length_in_bits;
1295     packedheader_param_buffer.has_emulation_bytes = 0;
1296
1297     va_status = vaCreateBuffer(va_dpy,
1298                                context_id,
1299                                VAEncPackedHeaderParameterBufferType,
1300                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1301                                &packedslice_para_bufid);
1302     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1303
1304     va_status = vaCreateBuffer(va_dpy,
1305                                context_id,
1306                                VAEncPackedHeaderDataBufferType,
1307                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1308                                &packedslice_data_bufid);
1309     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1310
1311     render_id[0] = packedslice_para_bufid;
1312     render_id[1] = packedslice_data_bufid;
1313     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1314
1315     free(packedslice_buffer);
1316 }
1317
1318 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1319 {
1320     VABufferID slice_param_buf;
1321     VAStatus va_status;
1322     int i;
1323
1324     update_RefPicList(frame_type);
1325
1326     /* one frame, one slice */
1327     slice_param.macroblock_address = 0;
1328     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1329     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1330     if (frame_type == FRAME_IDR) {
1331         if (encoding_frame_num != 0)
1332             ++slice_param.idr_pic_id;
1333     } else if (frame_type == FRAME_P) {
1334         int refpiclist0_max = h264_maxref & 0xffff;
1335         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1336
1337         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1338             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1339             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1340         }
1341     } else if (frame_type == FRAME_B) {
1342         int refpiclist0_max = h264_maxref & 0xffff;
1343         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1344
1345         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1346         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1347             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1348             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1349         }
1350
1351         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1352         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1353             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1354             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1355         }
1356     }
1357
1358     slice_param.slice_alpha_c0_offset_div2 = 0;
1359     slice_param.slice_beta_offset_div2 = 0;
1360     slice_param.direct_spatial_mv_pred_flag = 1;
1361     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1362
1363
1364     if (h264_packedheader &&
1365         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1366         render_packedslice();
1367
1368     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1369                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1370     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1371
1372     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1373
1374     return 0;
1375 }
1376
1377
1378
1379 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1380 {
1381         VACodedBufferSegment *buf_list = NULL;
1382         VAStatus va_status;
1383
1384         string data;
1385
1386         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1387         CHECK_VASTATUS(va_status, "vaMapBuffer");
1388         while (buf_list != NULL) {
1389                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1390                 buf_list = (VACodedBufferSegment *) buf_list->next;
1391         }
1392         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1393
1394         static int frameno = 0;
1395         print_latency("Current QuickSync latency (video inputs → disk mux):",
1396                 task.received_ts, (task.frame_type == FRAME_B), &frameno);
1397
1398         {
1399                 // Add video.
1400                 AVPacket pkt;
1401                 memset(&pkt, 0, sizeof(pkt));
1402                 pkt.buf = nullptr;
1403                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1404                 pkt.size = data.size();
1405                 pkt.stream_index = 0;
1406                 if (task.frame_type == FRAME_IDR) {
1407                         pkt.flags = AV_PKT_FLAG_KEY;
1408                 } else {
1409                         pkt.flags = 0;
1410                 }
1411                 pkt.duration = task.duration;
1412                 if (file_mux) {
1413                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1414                 }
1415                 if (!global_flags.uncompressed_video_to_http &&
1416                     !global_flags.x264_video_to_http) {
1417                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1418                 }
1419         }
1420 }
1421
1422
1423 // this is weird. but it seems to put a new frame onto the queue
1424 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1425 {
1426         unique_lock<mutex> lock(storage_task_queue_mutex);
1427         storage_task_queue.push(move(task));
1428         storage_task_queue_changed.notify_all();
1429 }
1430
1431 void QuickSyncEncoderImpl::storage_task_thread()
1432 {
1433         for ( ;; ) {
1434                 storage_task current;
1435                 {
1436                         // wait until there's an encoded frame
1437                         unique_lock<mutex> lock(storage_task_queue_mutex);
1438                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1439                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1440                         current = move(storage_task_queue.front());
1441                         storage_task_queue.pop();
1442                 }
1443
1444                 VAStatus va_status;
1445
1446                 // waits for data, then saves it to disk.
1447                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1448                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1449                 save_codeddata(move(current));
1450
1451                 {
1452                         unique_lock<mutex> lock(storage_task_queue_mutex);
1453                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1454                         storage_task_queue_changed.notify_all();
1455                 }
1456         }
1457 }
1458
1459 void QuickSyncEncoderImpl::release_encode()
1460 {
1461         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1462                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1463                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1464                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1465         }
1466
1467         vaDestroyContext(va_dpy, context_id);
1468         vaDestroyConfig(va_dpy, config_id);
1469 }
1470
1471 void QuickSyncEncoderImpl::release_gl_resources()
1472 {
1473         assert(is_shutdown);
1474         if (has_released_gl_resources) {
1475                 return;
1476         }
1477
1478         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1479                 if (!use_zerocopy) {
1480                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1481                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1482                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1483                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1484                 }
1485                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1486                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1487         }
1488
1489         has_released_gl_resources = true;
1490 }
1491
1492 int QuickSyncEncoderImpl::deinit_va()
1493 {
1494     vaTerminate(va_dpy);
1495
1496     va_close_display(va_dpy);
1497
1498     return 0;
1499 }
1500
1501 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)
1502         : 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)
1503 {
1504         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1505         open_output_file(filename);
1506         file_audio_encoder->add_mux(file_mux.get());
1507
1508         frame_width_mbaligned = (frame_width + 15) & (~15);
1509         frame_height_mbaligned = (frame_height + 15) & (~15);
1510
1511         //print_input();
1512
1513         if (global_flags.x264_video_to_http) {
1514                 assert(x264_encoder != nullptr);
1515         } else {
1516                 assert(x264_encoder == nullptr);
1517         }
1518
1519         init_va(va_display);
1520         setup_encode();
1521
1522         // No frames are ready yet.
1523         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1524
1525         memset(&seq_param, 0, sizeof(seq_param));
1526         memset(&pic_param, 0, sizeof(pic_param));
1527         memset(&slice_param, 0, sizeof(slice_param));
1528
1529         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1530
1531         encode_thread = thread([this]{
1532                 //SDL_GL_MakeCurrent(window, context);
1533                 QOpenGLContext *context = create_context(this->surface);
1534                 eglBindAPI(EGL_OPENGL_API);
1535                 if (!make_current(context, this->surface)) {
1536                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1537                                 eglGetError());
1538                         exit(1);
1539                 }
1540                 encode_thread_func();
1541                 delete_context(context);
1542         });
1543 }
1544
1545 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1546 {
1547         shutdown();
1548         release_gl_resources();
1549 }
1550
1551 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1552 {
1553         assert(!is_shutdown);
1554         {
1555                 // Wait until this frame slot is done encoding.
1556                 unique_lock<mutex> lock(storage_task_queue_mutex);
1557                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1558                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1559                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1560                 }
1561                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1562                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1563                 if (storage_thread_should_quit) return false;
1564         }
1565
1566         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1567         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1568         *y_tex = surf->y_tex;
1569         *cbcr_tex = surf->cbcr_tex;
1570
1571         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1572         CHECK_VASTATUS(va_status, "vaDeriveImage");
1573
1574         if (use_zerocopy) {
1575                 VABufferInfo buf_info;
1576                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1577                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1578                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1579
1580                 // Create Y image.
1581                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1582                 EGLint y_attribs[] = {
1583                         EGL_WIDTH, frame_width,
1584                         EGL_HEIGHT, frame_height,
1585                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1586                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1587                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1588                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1589                         EGL_NONE
1590                 };
1591
1592                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1593                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1594
1595                 // Associate Y image to a texture.
1596                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1597                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1598
1599                 // Create CbCr image.
1600                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1601                 EGLint cbcr_attribs[] = {
1602                         EGL_WIDTH, frame_width,
1603                         EGL_HEIGHT, frame_height,
1604                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1605                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1606                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1607                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1608                         EGL_NONE
1609                 };
1610
1611                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1612                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1613
1614                 // Associate CbCr image to a texture.
1615                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1616                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1617         }
1618
1619         return true;
1620 }
1621
1622 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1623 {
1624         assert(!is_shutdown);
1625         file_audio_encoder->encode_audio(audio, pts + global_delay());
1626 }
1627
1628 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1629 {
1630         assert(!is_shutdown);
1631
1632         if (!use_zerocopy) {
1633                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1634
1635                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1636                 check_error();
1637
1638                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1639                 check_error();
1640
1641                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1642                 check_error();
1643                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1644                 check_error();
1645
1646                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1647                 check_error();
1648                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1649                 check_error();
1650
1651                 glBindTexture(GL_TEXTURE_2D, 0);
1652                 check_error();
1653                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1654                 check_error();
1655
1656                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1657                 check_error();
1658         }
1659
1660         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1661         check_error();
1662         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1663         check_error();
1664
1665         {
1666                 unique_lock<mutex> lock(frame_queue_mutex);
1667                 pending_video_frames.push(PendingFrame{ fence, input_frames, pts, duration });
1668                 ++current_storage_frame;
1669         }
1670         frame_queue_nonempty.notify_all();
1671         return fence;
1672 }
1673
1674 void QuickSyncEncoderImpl::shutdown()
1675 {
1676         if (is_shutdown) {
1677                 return;
1678         }
1679
1680         {
1681                 unique_lock<mutex> lock(frame_queue_mutex);
1682                 encode_thread_should_quit = true;
1683                 frame_queue_nonempty.notify_all();
1684         }
1685         encode_thread.join();
1686         {
1687                 unique_lock<mutex> lock(storage_task_queue_mutex);
1688                 storage_thread_should_quit = true;
1689                 frame_queue_nonempty.notify_all();
1690                 storage_task_queue_changed.notify_all();
1691         }
1692         storage_thread.join();
1693
1694         // Encode any leftover audio in the queues, and also any delayed frames.
1695         file_audio_encoder->encode_last_audio();
1696
1697         release_encode();
1698         deinit_va();
1699         file_mux.reset();
1700         is_shutdown = true;
1701 }
1702
1703 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1704 {
1705         AVFormatContext *avctx = avformat_alloc_context();
1706         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1707         assert(filename.size() < sizeof(avctx->filename) - 1);
1708         strcpy(avctx->filename, filename.c_str());
1709
1710         string url = "file:" + filename;
1711         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1712         if (ret < 0) {
1713                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1714                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1715                 exit(1);
1716         }
1717
1718         string video_extradata = "";  // FIXME: See other comment about global headers.
1719         AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1720         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1721                 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1)));
1722 }
1723
1724 void QuickSyncEncoderImpl::encode_thread_func()
1725 {
1726         int64_t last_dts = -1;
1727         int gop_start_display_frame_num = 0;
1728         for (int display_frame_num = 0; ; ++display_frame_num) {
1729                 // Wait for the frame to be in the queue. Note that this only means
1730                 // we started rendering it.
1731                 PendingFrame frame;
1732                 {
1733                         unique_lock<mutex> lock(frame_queue_mutex);
1734                         frame_queue_nonempty.wait(lock, [this]{
1735                                 return encode_thread_should_quit || !pending_video_frames.empty();
1736                         });
1737                         if (encode_thread_should_quit && pending_video_frames.empty()) {
1738                                 // We may have queued frames left in the reorder buffer
1739                                 // that were supposed to be B-frames, but have no P-frame
1740                                 // to be encoded against. If so, encode them all as
1741                                 // P-frames instead. Note that this happens under the mutex,
1742                                 // but nobody else uses it at this point, since we're shutting down,
1743                                 // so there's no contention.
1744                                 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1745                                 return;
1746                         } else {
1747                                 frame = move(pending_video_frames.front());
1748                                 pending_video_frames.pop();
1749                         }
1750                 }
1751
1752                 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1753                 // Note that this implicitly waits for the frame to be done rendering.
1754                 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1755                 reorder_buffer[display_frame_num] = move(frame);
1756
1757                 // Now encode as many QuickSync frames as we can using the frames we have available.
1758                 // (It could be zero, or it could be multiple.) FIXME: make a function.
1759                 for ( ;; ) {
1760                         int pts_lag;
1761                         int frame_type, quicksync_display_frame_num;
1762                         encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1763                                                &quicksync_display_frame_num, &frame_type, &pts_lag);
1764                         if (!reorder_buffer.count(quicksync_display_frame_num)) {
1765                                 break;
1766                         }
1767                         frame = move(reorder_buffer[quicksync_display_frame_num]);
1768                         reorder_buffer.erase(quicksync_display_frame_num);
1769
1770                         if (frame_type == FRAME_IDR) {
1771                                 numShortTerm = 0;
1772                                 current_frame_num = 0;
1773                                 gop_start_display_frame_num = quicksync_display_frame_num;
1774                         }
1775
1776                         // Determine the dts of this frame.
1777                         int64_t dts;
1778                         if (pts_lag == -1) {
1779                                 assert(last_dts != -1);
1780                                 dts = last_dts + (TIMEBASE / MAX_FPS);
1781                         } else {
1782                                 dts = frame.pts - pts_lag;
1783                         }
1784                         last_dts = dts;
1785
1786                         encode_frame(frame, quicksync_encoding_frame_num, quicksync_display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
1787                         ++quicksync_encoding_frame_num;
1788                 }
1789         }
1790 }
1791
1792 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1793 {
1794         if (reorder_buffer.empty()) {
1795                 return;
1796         }
1797
1798         for (auto &pending_frame : reorder_buffer) {
1799                 int display_frame_num = pending_frame.first;
1800                 assert(display_frame_num > 0);
1801                 PendingFrame frame = move(pending_frame.second);
1802                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1803                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1804                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
1805                 last_dts = dts;
1806         }
1807 }
1808
1809 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1810 {
1811         AVPacket pkt;
1812         memset(&pkt, 0, sizeof(pkt));
1813         pkt.buf = nullptr;
1814         pkt.data = const_cast<uint8_t *>(data);
1815         pkt.size = frame_width * frame_height * 2;
1816         pkt.stream_index = 0;
1817         pkt.flags = AV_PKT_FLAG_KEY;
1818         pkt.duration = duration;
1819         stream_mux->add_packet(pkt, pts, pts);
1820 }
1821
1822 namespace {
1823
1824 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1825 {
1826         if (src_width == dst_pitch) {
1827                 memcpy(dst, src, src_width * height);
1828         } else {
1829                 for (size_t y = 0; y < height; ++y) {
1830                         const uint8_t *sptr = src + y * src_width;
1831                         uint8_t *dptr = dst + y * dst_pitch;
1832                         memcpy(dptr, sptr, src_width);
1833                 }
1834         }
1835 }
1836
1837 }  // namespace
1838
1839 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1840 {
1841         // Wait for the GPU to be done with the frame.
1842         GLenum sync_status;
1843         do {
1844                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1845                 check_error();
1846         } while (sync_status == GL_TIMEOUT_EXPIRED);
1847         assert(sync_status != GL_WAIT_FAILED);
1848
1849         ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1850         static int frameno = 0;
1851         print_latency("Current mixer latency (video inputs → ready for encode):",
1852                 received_ts, false, &frameno);
1853
1854         // Release back any input frames we needed to render this frame.
1855         frame.input_frames.clear();
1856
1857         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1858         uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1859         if (global_flags.uncompressed_video_to_http) {
1860                 add_packet_for_uncompressed_frame(pts, duration, data);
1861         } else if (global_flags.x264_video_to_http) {
1862                 x264_encoder->add_frame(pts, duration, data, received_ts);
1863         }
1864 }
1865
1866 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1867                                         int frame_type, int64_t pts, int64_t dts, int64_t duration)
1868 {
1869         const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1870
1871         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1872         VAStatus va_status;
1873
1874         if (use_zerocopy) {
1875                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1876                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1877                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1878                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1879         } else {
1880                 // Upload the frame to VA-API.
1881                 unsigned char *surface_p = nullptr;
1882                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
1883
1884                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1885                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1886
1887                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1888                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1889
1890                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
1891                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1892         }
1893
1894         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1895         CHECK_VASTATUS(va_status, "vaDestroyImage");
1896
1897         // Schedule the frame for encoding.
1898         VASurfaceID va_surface = surf->src_surface;
1899         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
1900         CHECK_VASTATUS(va_status, "vaBeginPicture");
1901
1902         if (frame_type == FRAME_IDR) {
1903                 // FIXME: If the mux wants global headers, we should not put the
1904                 // SPS/PPS before each IDR frame, but rather put it into the
1905                 // codec extradata (formatted differently?).
1906                 render_sequence();
1907                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1908                 if (h264_packedheader) {
1909                         render_packedsequence();
1910                         render_packedpicture();
1911                 }
1912         } else {
1913                 //render_sequence();
1914                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1915         }
1916         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1917
1918         va_status = vaEndPicture(va_dpy, context_id);
1919         CHECK_VASTATUS(va_status, "vaEndPicture");
1920
1921         // so now the data is done encoding (well, async job kicked off)...
1922         // we send that to the storage thread
1923         storage_task tmp;
1924         tmp.display_order = display_frame_num;
1925         tmp.frame_type = frame_type;
1926         tmp.pts = pts;
1927         tmp.dts = dts;
1928         tmp.duration = duration;
1929         tmp.received_ts = received_ts;
1930         storage_task_enqueue(move(tmp));
1931
1932         update_ReferenceFrames(frame_type);
1933 }
1934
1935 // Proxy object.
1936 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)
1937         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
1938
1939 // Must be defined here because unique_ptr<> destructor needs to know the impl.
1940 QuickSyncEncoder::~QuickSyncEncoder() {}
1941
1942 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
1943 {
1944         impl->add_audio(pts, audio);
1945 }
1946
1947 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1948 {
1949         return impl->begin_frame(y_tex, cbcr_tex);
1950 }
1951
1952 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1953 {
1954         return impl->end_frame(pts, duration, input_frames);
1955 }
1956
1957 void QuickSyncEncoder::shutdown()
1958 {
1959         impl->shutdown();
1960 }
1961
1962 void QuickSyncEncoder::set_stream_mux(Mux *mux)
1963 {
1964         impl->set_stream_mux(mux);
1965 }
1966
1967 int64_t QuickSyncEncoder::global_delay() const {
1968         return impl->global_delay();
1969 }