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