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 "ref_counted_frame.h"
  53 #include "timebase.h"
  54 #include "x264_encoder.h"
  55
  56 using namespace std;
  57 using namespace std::chrono;
  58 using namespace std::placeholders;
  59
  60 class QOpenGLContext;
  61 class QSurface;
  62
  63 #define CHECK_VASTATUS(va_status, func)                                 \
  64     if (va_status != VA_STATUS_SUCCESS) {                               \
  65         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  66         exit(1);                                                        \
  67     }
  68
  69 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  70
  71 //#include "loadsurface.h"
  72
  73 #define NAL_REF_IDC_NONE        0
  74 #define NAL_REF_IDC_LOW         1
  75 #define NAL_REF_IDC_MEDIUM      2
  76 #define NAL_REF_IDC_HIGH        3
  77
  78 #define NAL_NON_IDR             1
  79 #define NAL_IDR                 5
  80 #define NAL_SPS                 7
  81 #define NAL_PPS                 8
  82 #define NAL_SEI                 6
  83
  84 #define SLICE_TYPE_P            0
  85 #define SLICE_TYPE_B            1
  86 #define SLICE_TYPE_I            2
  87 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  88 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  89 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  90
  91
  92 #define ENTROPY_MODE_CAVLC      0
  93 #define ENTROPY_MODE_CABAC      1
  94
  95 #define PROFILE_IDC_BASELINE    66
  96 #define PROFILE_IDC_MAIN        77
  97 #define PROFILE_IDC_HIGH        100
  98
  99 #define BITSTREAM_ALLOCATE_STEPPING     4096
 100 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
 101 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
 102 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
 103
 104 static constexpr unsigned int MaxFrameNum = (2<<16);
 105 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
 106 static constexpr unsigned int Log2MaxFrameNum = 16;
 107 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
 108 static constexpr int rc_default_modes[] = {  // Priority list of modes.
 109     VA_RC_VBR,
 110     VA_RC_CQP,
 111     VA_RC_VBR_CONSTRAINED,
 112     VA_RC_CBR,
 113     VA_RC_VCM,
 114     VA_RC_NONE,
 115 };
 116
 117 /* thread to save coded data */
 118 #define SRC_SURFACE_FREE        0
 119 #define SRC_SURFACE_IN_ENCODING 1
 120
 121 struct __bitstream {
 122     unsigned int *buffer;
 123     int bit_offset;
 124     int max_size_in_dword;
 125 };
 126 typedef struct __bitstream bitstream;
 127
 128 using namespace std;
 129
 130 // H.264 video comes out in encoding order (e.g. with two B-frames:
 131 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 132 // come in the right order. Since we do everything, including waiting
 133 // for the frames to come out of OpenGL, in encoding order, we need
 134 // a reordering buffer for uncompressed frames so that they come out
 135 // correctly. We go the super-lazy way of not making it understand
 136 // anything about the true order (which introduces some extra latency,
 137 // though); we know that for N B-frames we need at most (N-1) frames
 138 // in the reorder buffer, and can just sort on that.
 139 //
 140 // The class also deals with keeping a freelist as needed.
 141 class FrameReorderer {
 142 public:
 143         FrameReorderer(unsigned queue_length, int width, int height);
 144
 145         struct Frame {
 146                 int64_t pts, duration;
 147                 uint8_t *data;
 148                 ReceivedTimestamps received_ts;
 149
 150                 // Invert to get the smallest pts first.
 151                 bool operator< (const Frame &other) const { return pts > other.pts; }
 152         };
 153
 154         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 155         // Does _not_ take ownership of data; a copy is taken if needed.
 156         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 157         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 158         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data, const ReceivedTimestamps &received_ts);
 159
 160         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 161         Frame get_first_frame();
 162
 163         bool empty() const { return frames.empty(); }
 164
 165 private:
 166         unsigned queue_length;
 167         int width, height;
 168
 169         priority_queue<Frame> frames;
 170         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 171
 172         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 173         // except priority_queue doesn't work well with movable-only types.
 174         vector<unique_ptr<uint8_t[]>> owner;
 175 };
 176
 177 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 178     : queue_length(queue_length), width(width), height(height)
 179 {
 180         for (unsigned i = 0; i < queue_length; ++i) {
 181                 owner.emplace_back(new uint8_t[width * height * 2]);
 182                 freelist.push(owner.back().get());
 183         }
 184 }
 185
 186 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data, const ReceivedTimestamps &received_ts)
 187 {
 188         if (queue_length == 0) {
 189                 return Frame{pts, duration, data, received_ts};
 190         }
 191
 192         assert(!freelist.empty());
 193         uint8_t *storage = freelist.top();
 194         freelist.pop();
 195         memcpy(storage, data, width * height * 2);
 196         frames.push(Frame{pts, duration, storage, received_ts});
 197
 198         if (frames.size() >= queue_length) {
 199                 return get_first_frame();
 200         } else {
 201                 return Frame{-1, -1, nullptr, steady_clock::time_point::min(), steady_clock::time_point::min()};
 202         }
 203 }
 204
 205 FrameReorderer::Frame FrameReorderer::get_first_frame()
 206 {
 207         assert(!frames.empty());
 208         Frame storage = frames.top();
 209         frames.pop();
 210         freelist.push(storage.data);
 211         return storage;
 212 }
 213
 214 class QuickSyncEncoderImpl {
 215 public:
 216         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);
 217         ~QuickSyncEncoderImpl();
 218         void add_audio(int64_t pts, vector<float> audio);
 219         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 220         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
 221         void shutdown();
 222         void release_gl_resources();
 223         void set_stream_mux(Mux *mux)
 224         {
 225                 stream_mux = mux;
 226         }
 227
 228         // So we never get negative dts.
 229         int64_t global_delay() const {
 230                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 231         }
 232
 233 private:
 234         struct storage_task {
 235                 unsigned long long display_order;
 236                 int frame_type;
 237                 vector<float> audio;
 238                 int64_t pts, dts, duration;
 239                 ReceivedTimestamps received_ts;
 240         };
 241         struct PendingFrame {
 242                 RefCountedGLsync fence;
 243                 vector<RefCountedFrame> input_frames;
 244                 int64_t pts, duration;
 245         };
 246
 247         void open_output_file(const std::string &filename);
 248         void encode_thread_func();
 249         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 250         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 251         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 252                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 253         void storage_task_thread();
 254         void storage_task_enqueue(storage_task task);
 255         void save_codeddata(storage_task task);
 256         int render_packedsequence();
 257         int render_packedpicture();
 258         void render_packedslice();
 259         int render_sequence();
 260         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 261         void sps_rbsp(bitstream *bs);
 262         void pps_rbsp(bitstream *bs);
 263         int build_packed_pic_buffer(unsigned char **header_buffer);
 264         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 265         void slice_header(bitstream *bs);
 266         int build_packed_seq_buffer(unsigned char **header_buffer);
 267         int build_packed_slice_buffer(unsigned char **header_buffer);
 268         int init_va(const string &va_display);
 269         int deinit_va();
 270         void enable_zerocopy_if_possible();
 271         VADisplay va_open_display(const string &va_display);
 272         void va_close_display(VADisplay va_dpy);
 273         int setup_encode();
 274         void release_encode();
 275         void update_ReferenceFrames(int frame_type);
 276         int update_RefPicList(int frame_type);
 277
 278         bool is_shutdown = false;
 279         bool has_released_gl_resources = false;
 280         bool use_zerocopy;
 281         int drm_fd = -1;
 282
 283         thread encode_thread, storage_thread;
 284
 285         mutex storage_task_queue_mutex;
 286         condition_variable storage_task_queue_changed;
 287         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 288         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 289         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 290
 291         mutex frame_queue_mutex;
 292         condition_variable frame_queue_nonempty;
 293         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 294
 295         int current_storage_frame;
 296
 297         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 298         movit::ResourcePool *resource_pool;
 299         QSurface *surface;
 300
 301         unique_ptr<AudioEncoder> file_audio_encoder;
 302
 303         unique_ptr<FrameReorderer> reorderer;
 304         X264Encoder *x264_encoder;  // nullptr if not using x264.
 305
 306         Mux* stream_mux = nullptr;  // To HTTP.
 307         unique_ptr<Mux> file_mux;  // To local disk.
 308
 309         Display *x11_display = nullptr;
 310
 311         // Encoder parameters
 312         VADisplay va_dpy;
 313         VAProfile h264_profile = (VAProfile)~0;
 314         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 315         int config_attrib_num = 0, enc_packed_header_idx;
 316
 317         struct GLSurface {
 318                 VASurfaceID src_surface, ref_surface;
 319                 VABufferID coded_buf;
 320
 321                 VAImage surface_image;
 322                 GLuint y_tex, cbcr_tex;
 323
 324                 // Only if use_zerocopy == true.
 325                 EGLImage y_egl_image, cbcr_egl_image;
 326
 327                 // Only if use_zerocopy == false.
 328                 GLuint pbo;
 329                 uint8_t *y_ptr, *cbcr_ptr;
 330                 size_t y_offset, cbcr_offset;
 331         };
 332         GLSurface gl_surfaces[SURFACE_NUM];
 333
 334         VAConfigID config_id;
 335         VAContextID context_id;
 336         VAEncSequenceParameterBufferH264 seq_param;
 337         VAEncPictureParameterBufferH264 pic_param;
 338         VAEncSliceParameterBufferH264 slice_param;
 339         VAPictureH264 CurrentCurrPic;
 340         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 341
 342         // Static quality settings.
 343         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 344         static constexpr unsigned int num_ref_frames = 2;
 345         static constexpr int initial_qp = 15;
 346         static constexpr int minimal_qp = 0;
 347         static constexpr int intra_period = 30;
 348         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 349
 350         // Quality settings that are meant to be static, but might be overridden
 351         // by the profile.
 352         int constraint_set_flag = 0;
 353         int h264_packedheader = 0; /* support pack header? */
 354         int h264_maxref = (1<<16|1);
 355         int h264_entropy_mode = 1; /* cabac */
 356         int ip_period = 3;
 357
 358         int rc_mode = -1;
 359         unsigned int current_frame_num = 0;
 360         unsigned int numShortTerm = 0;
 361
 362         int frame_width;
 363         int frame_height;
 364         int frame_width_mbaligned;
 365         int frame_height_mbaligned;
 366
 367         DiskSpaceEstimator *disk_space_estimator;
 368 };
 369
 370 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 371 // but if we don't delete it here, we get leaks. The GStreamer implementation
 372 // does the same.
 373 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 374 {
 375     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 376     CHECK_VASTATUS(va_status, "vaRenderPicture");
 377
 378     for (int i = 0; i < num_buffers; ++i) {
 379         va_status = vaDestroyBuffer(dpy, buffers[i]);
 380         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 381     }
 382 }
 383
 384 static unsigned int
 385 va_swap32(unsigned int val)
 386 {
 387     unsigned char *pval = (unsigned char *)&val;
 388
 389     return ((pval[0] << 24)     |
 390             (pval[1] << 16)     |
 391             (pval[2] << 8)      |
 392             (pval[3] << 0));
 393 }
 394
 395 static void
 396 bitstream_start(bitstream *bs)
 397 {
 398     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 399     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 400     bs->bit_offset = 0;
 401 }
 402
 403 static void
 404 bitstream_end(bitstream *bs)
 405 {
 406     int pos = (bs->bit_offset >> 5);
 407     int bit_offset = (bs->bit_offset & 0x1f);
 408     int bit_left = 32 - bit_offset;
 409
 410     if (bit_offset) {
 411         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 412     }
 413 }
 414
 415 static void
 416 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 417 {
 418     int pos = (bs->bit_offset >> 5);
 419     int bit_offset = (bs->bit_offset & 0x1f);
 420     int bit_left = 32 - bit_offset;
 421
 422     if (!size_in_bits)
 423         return;
 424
 425     bs->bit_offset += size_in_bits;
 426
 427     if (bit_left > size_in_bits) {
 428         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 429     } else {
 430         size_in_bits -= bit_left;
 431         if (bit_left >= 32) {
 432             bs->buffer[pos] = (val >> size_in_bits);
 433         } else {
 434             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 435         }
 436         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 437
 438         if (pos + 1 == bs->max_size_in_dword) {
 439             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 440             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 441         }
 442
 443         bs->buffer[pos + 1] = val;
 444     }
 445 }
 446
 447 static void
 448 bitstream_put_ue(bitstream *bs, unsigned int val)
 449 {
 450     int size_in_bits = 0;
 451     int tmp_val = ++val;
 452
 453     while (tmp_val) {
 454         tmp_val >>= 1;
 455         size_in_bits++;
 456     }
 457
 458     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 459     bitstream_put_ui(bs, val, size_in_bits);
 460 }
 461
 462 static void
 463 bitstream_put_se(bitstream *bs, int val)
 464 {
 465     unsigned int new_val;
 466
 467     if (val <= 0)
 468         new_val = -2 * val;
 469     else
 470         new_val = 2 * val - 1;
 471
 472     bitstream_put_ue(bs, new_val);
 473 }
 474
 475 static void
 476 bitstream_byte_aligning(bitstream *bs, int bit)
 477 {
 478     int bit_offset = (bs->bit_offset & 0x7);
 479     int bit_left = 8 - bit_offset;
 480     int new_val;
 481
 482     if (!bit_offset)
 483         return;
 484
 485     assert(bit == 0 || bit == 1);
 486
 487     if (bit)
 488         new_val = (1 << bit_left) - 1;
 489     else
 490         new_val = 0;
 491
 492     bitstream_put_ui(bs, new_val, bit_left);
 493 }
 494
 495 static void
 496 rbsp_trailing_bits(bitstream *bs)
 497 {
 498     bitstream_put_ui(bs, 1, 1);
 499     bitstream_byte_aligning(bs, 0);
 500 }
 501
 502 static void nal_start_code_prefix(bitstream *bs)
 503 {
 504     bitstream_put_ui(bs, 0x00000001, 32);
 505 }
 506
 507 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 508 {
 509     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 510     bitstream_put_ui(bs, nal_ref_idc, 2);
 511     bitstream_put_ui(bs, nal_unit_type, 5);
 512 }
 513
 514 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 515 {
 516     int profile_idc = PROFILE_IDC_BASELINE;
 517
 518     if (h264_profile  == VAProfileH264High)
 519         profile_idc = PROFILE_IDC_HIGH;
 520     else if (h264_profile  == VAProfileH264Main)
 521         profile_idc = PROFILE_IDC_MAIN;
 522
 523     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 524     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 525     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 526     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 527     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 528     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 529     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 530     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 531
 532     if ( profile_idc == PROFILE_IDC_HIGH) {
 533         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 534         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 535         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 536         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 537         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 538     }
 539
 540     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 541     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 542
 543     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 544         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 545     else {
 546         assert(0);
 547     }
 548
 549     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 550     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 551
 552     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 553     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 554     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 555
 556     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 557         assert(0);
 558     }
 559
 560     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 561     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 562
 563     if (seq_param.frame_cropping_flag) {
 564         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 565         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 566         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 567         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 568     }
 569
 570     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 571     if ( false ) {
 572         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 573     } else {
 574         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 575         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 576         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 577         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 578         {
 579             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 580             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 581             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 582             {
 583                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 584                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 585                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 586             }
 587         }
 588         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 589         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 590         {
 591             bitstream_put_ui(bs, 1, 32);  // FPS
 592             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 593             bitstream_put_ui(bs, 1, 1);
 594         }
 595         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 596         {
 597             // hrd_parameters
 598             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 599             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 600             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 601
 602             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 603             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 604             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 605
 606             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 607             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 608             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 609             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 610         }
 611         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 612         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 613
 614         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 615         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 616     }
 617
 618     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 619 }
 620
 621
 622 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 623 {
 624     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 625     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 626
 627     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 628
 629     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 630
 631     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 632
 633     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 634     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 635
 636     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 637     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 638
 639     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 640     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 641     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 642
 643     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 644     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 645     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 646
 647     /* more_rbsp_data */
 648     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 649     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 650     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 651
 652     rbsp_trailing_bits(bs);
 653 }
 654
 655 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 656 {
 657     int first_mb_in_slice = slice_param.macroblock_address;
 658
 659     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 660     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 661     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 662     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 663
 664     /* frame_mbs_only_flag == 1 */
 665     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 666         /* FIXME: */
 667         assert(0);
 668     }
 669
 670     if (pic_param.pic_fields.bits.idr_pic_flag)
 671         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 672
 673     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 674         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 675         /* pic_order_present_flag == 0 */
 676     } else {
 677         /* FIXME: */
 678         assert(0);
 679     }
 680
 681     /* redundant_pic_cnt_present_flag == 0 */
 682     /* slice type */
 683     if (IS_P_SLICE(slice_param.slice_type)) {
 684         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 685
 686         if (slice_param.num_ref_idx_active_override_flag)
 687             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 688
 689         /* ref_pic_list_reordering */
 690         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 691     } else if (IS_B_SLICE(slice_param.slice_type)) {
 692         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 693
 694         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 695
 696         if (slice_param.num_ref_idx_active_override_flag) {
 697             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 698             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 699         }
 700
 701         /* ref_pic_list_reordering */
 702         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 703         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 704     }
 705
 706     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 707          IS_P_SLICE(slice_param.slice_type)) ||
 708         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 709          IS_B_SLICE(slice_param.slice_type))) {
 710         /* FIXME: fill weight/offset table */
 711         assert(0);
 712     }
 713
 714     /* dec_ref_pic_marking */
 715     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 716         unsigned char no_output_of_prior_pics_flag = 0;
 717         unsigned char long_term_reference_flag = 0;
 718         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 719
 720         if (pic_param.pic_fields.bits.idr_pic_flag) {
 721             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 722             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 723         } else {
 724             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 725         }
 726     }
 727
 728     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 729         !IS_I_SLICE(slice_param.slice_type))
 730         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 731
 732     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 733
 734     /* ignore for SP/SI */
 735
 736     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 737         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 738
 739         if (slice_param.disable_deblocking_filter_idc != 1) {
 740             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 741             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 742         }
 743     }
 744
 745     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 746         bitstream_byte_aligning(bs, 1);
 747     }
 748 }
 749
 750 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 751 {
 752     bitstream bs;
 753
 754     bitstream_start(&bs);
 755     nal_start_code_prefix(&bs);
 756     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 757     pps_rbsp(&bs);
 758     bitstream_end(&bs);
 759
 760     *header_buffer = (unsigned char *)bs.buffer;
 761     return bs.bit_offset;
 762 }
 763
 764 int
 765 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 766 {
 767     bitstream bs;
 768
 769     bitstream_start(&bs);
 770     nal_start_code_prefix(&bs);
 771     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 772     sps_rbsp(&bs);
 773     bitstream_end(&bs);
 774
 775     *header_buffer = (unsigned char *)bs.buffer;
 776     return bs.bit_offset;
 777 }
 778
 779 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 780 {
 781     bitstream bs;
 782     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 783     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 784
 785     bitstream_start(&bs);
 786     nal_start_code_prefix(&bs);
 787
 788     if (IS_I_SLICE(slice_param.slice_type)) {
 789         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 790     } else if (IS_P_SLICE(slice_param.slice_type)) {
 791         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 792     } else {
 793         assert(IS_B_SLICE(slice_param.slice_type));
 794         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 795     }
 796
 797     slice_header(&bs);
 798     bitstream_end(&bs);
 799
 800     *header_buffer = (unsigned char *)bs.buffer;
 801     return bs.bit_offset;
 802 }
 803
 804
 805 /*
 806   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 807   1) period between Frame #X and Frame #N = #X - #N
 808   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 809   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 810   4) intra_period and intra_idr_period take precedence over ip_period
 811   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 812      of I/IDR frames, see bellow examples
 813   -------------------------------------------------------------------
 814   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 815   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 816   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 817   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 818   1            1                ignored    IDR IDR IDR IDR...
 819   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 820   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 821   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 822                                               (PBB)(IBB)(PBB)(IBB)...
 823   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 824                                            IDRPPPPP IPPPPP IPPPPP...
 825   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 826                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 827                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 828                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 829                                            {IDR(PBB)(PBB)}                     (6/6/3)
 830                                            {IDR(PBB)(PBB)}.
 831 */
 832
 833 // General pts/dts strategy:
 834 //
 835 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 836 // bit tricky. We assume first of all that the frame rate never goes _above_
 837 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 838 // in at least this speed, as long at dts <= pts (the frame is not attempted
 839 // presented before it is decoded). Furthermore, we never have longer chains of
 840 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 841 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 842 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 843 // frame, with an order O <= C.)
 844 //
 845 // Many strategies are possible, but we establish these rules:
 846 //
 847 //  - Tip frames have dts = pts - (C-O)*N.
 848 //  - Non-tip frames have dts = dts_last + N.
 849 //
 850 // An example, with C=2 and N=10 and the data flow showed with arrows:
 851 //
 852 //        I  B  P  B  B  P
 853 //   pts: 30 40 50 60 70 80
 854 //        ↓  ↓     ↓
 855 //   dts: 10 30 20 60 50←40
 856 //         |  |  ↑        ↑
 857 //         `--|--'        |
 858 //             `----------'
 859 //
 860 // To show that this works fine also with irregular spacings, let's say that
 861 // the third frame is delayed a bit (something earlier was dropped). Now the
 862 // situation looks like this:
 863 //
 864 //        I  B  P  B  B   P
 865 //   pts: 30 40 80 90 100 110
 866 //        ↓  ↓     ↓
 867 //   dts: 10 30 20 90 50←40
 868 //         |  |  ↑        ↑
 869 //         `--|--'        |
 870 //             `----------'
 871 //
 872 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 873 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 874 //
 875 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 876 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 877 // a tip frame and should be given a dts based on the previous one.
 878 #define FRAME_P 0
 879 #define FRAME_B 1
 880 #define FRAME_I 2
 881 #define FRAME_IDR 7
 882 void encoding2display_order(
 883     int encoding_order, int intra_period,
 884     int intra_idr_period, int ip_period,
 885     int *displaying_order,
 886     int *frame_type, int *pts_lag)
 887 {
 888     int encoding_order_gop = 0;
 889
 890     *pts_lag = 0;
 891
 892     if (intra_period == 1) { /* all are I/IDR frames */
 893         *displaying_order = encoding_order;
 894         if (intra_idr_period == 0)
 895             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 896         else
 897             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 898         return;
 899     }
 900
 901     if (intra_period == 0)
 902         intra_idr_period = 0;
 903
 904     if (ip_period == 1) {
 905         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 906         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 907         *displaying_order = encoding_order;
 908
 909         if (encoding_order_gop == 0) { /* the first frame */
 910             *frame_type = FRAME_IDR;
 911         } else if (intra_period != 0 && /* have I frames */
 912                    encoding_order_gop >= 2 &&
 913                    (encoding_order_gop % intra_period == 0)) {
 914             *frame_type = FRAME_I;
 915         } else {
 916             *frame_type = FRAME_P;
 917         }
 918         return;
 919     }
 920
 921     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 922     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 923     *pts_lag = -1;  // Most frames are not tip frames.
 924
 925     if (encoding_order_gop == 0) { /* the first frame */
 926         *frame_type = FRAME_IDR;
 927         *displaying_order = encoding_order;
 928         // IDR frames are a special case; I honestly can't find the logic behind
 929         // why this is the right thing, but it seems to line up nicely in practice :-)
 930         *pts_lag = TIMEBASE / MAX_FPS;
 931     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 932         *frame_type = FRAME_B;
 933         *displaying_order = encoding_order - 1;
 934         if ((encoding_order_gop % ip_period) == 0) {
 935             *pts_lag = 0;  // Last B-frame.
 936         }
 937     } else if (intra_period != 0 && /* have I frames */
 938                encoding_order_gop >= 2 &&
 939                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 940         *frame_type = FRAME_I;
 941         *displaying_order = encoding_order + ip_period - 1;
 942     } else {
 943         *frame_type = FRAME_P;
 944         *displaying_order = encoding_order + ip_period - 1;
 945     }
 946 }
 947
 948
 949 static const char *rc_to_string(int rc_mode)
 950 {
 951     switch (rc_mode) {
 952     case VA_RC_NONE:
 953         return "NONE";
 954     case VA_RC_CBR:
 955         return "CBR";
 956     case VA_RC_VBR:
 957         return "VBR";
 958     case VA_RC_VCM:
 959         return "VCM";
 960     case VA_RC_CQP:
 961         return "CQP";
 962     case VA_RC_VBR_CONSTRAINED:
 963         return "VBR_CONSTRAINED";
 964     default:
 965         return "Unknown";
 966     }
 967 }
 968
 969 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 970 {
 971         if (global_flags.uncompressed_video_to_http) {
 972                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 973                 use_zerocopy = false;
 974         } else if (global_flags.x264_video_to_http) {
 975                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 976                 use_zerocopy = false;
 977         } else {
 978                 use_zerocopy = true;
 979         }
 980 }
 981
 982 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 983 {
 984         if (va_display.empty()) {
 985                 x11_display = XOpenDisplay(NULL);
 986                 if (!x11_display) {
 987                         fprintf(stderr, "error: can't connect to X server!\n");
 988                         return NULL;
 989                 }
 990                 enable_zerocopy_if_possible();
 991                 return vaGetDisplay(x11_display);
 992         } else if (va_display[0] != '/') {
 993                 x11_display = XOpenDisplay(va_display.c_str());
 994                 if (!x11_display) {
 995                         fprintf(stderr, "error: can't connect to X server!\n");
 996                         return NULL;
 997                 }
 998                 enable_zerocopy_if_possible();
 999                 return vaGetDisplay(x11_display);
1000         } else {
1001                 drm_fd = open(va_display.c_str(), O_RDWR);
1002                 if (drm_fd == -1) {
1003                         perror(va_display.c_str());
1004                         return NULL;
1005                 }
1006                 use_zerocopy = false;
1007                 return vaGetDisplayDRM(drm_fd);
1008         }
1009 }
1010
1011 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
1012 {
1013         if (x11_display) {
1014                 XCloseDisplay(x11_display);
1015                 x11_display = nullptr;
1016         }
1017         if (drm_fd != -1) {
1018                 close(drm_fd);
1019         }
1020 }
1021
1022 int QuickSyncEncoderImpl::init_va(const string &va_display)
1023 {
1024     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1025     VAEntrypoint *entrypoints;
1026     int num_entrypoints, slice_entrypoint;
1027     int support_encode = 0;
1028     int major_ver, minor_ver;
1029     VAStatus va_status;
1030     unsigned int i;
1031
1032     va_dpy = va_open_display(va_display);
1033     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1034     CHECK_VASTATUS(va_status, "vaInitialize");
1035
1036     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1037     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1038     if (!entrypoints) {
1039         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1040         exit(1);
1041     }
1042
1043     /* use the highest profile */
1044     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1045         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1046             continue;
1047
1048         h264_profile = profile_list[i];
1049         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1050         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1051             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1052                 support_encode = 1;
1053                 break;
1054             }
1055         }
1056         if (support_encode == 1)
1057             break;
1058     }
1059
1060     if (support_encode == 0) {
1061         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1062         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1063         printf("to use VA-API against DRM instead of X11.\n");
1064         exit(1);
1065     } else {
1066         switch (h264_profile) {
1067             case VAProfileH264Baseline:
1068                 ip_period = 1;
1069                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1070                 h264_entropy_mode = 0;
1071                 break;
1072             case VAProfileH264ConstrainedBaseline:
1073                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1074                 ip_period = 1;
1075                 break;
1076
1077             case VAProfileH264Main:
1078                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1079                 break;
1080
1081             case VAProfileH264High:
1082                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1083                 break;
1084             default:
1085                 h264_profile = VAProfileH264Baseline;
1086                 ip_period = 1;
1087                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1088                 break;
1089         }
1090     }
1091
1092     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1093
1094     /* find out the format for the render target, and rate control mode */
1095     for (i = 0; i < VAConfigAttribTypeMax; i++)
1096         attrib[i].type = (VAConfigAttribType)i;
1097
1098     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1099                                       &attrib[0], VAConfigAttribTypeMax);
1100     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1101     /* check the interested configattrib */
1102     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1103         printf("Not find desired YUV420 RT format\n");
1104         exit(1);
1105     } else {
1106         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1107         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1108         config_attrib_num++;
1109     }
1110
1111     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1112         int tmp = attrib[VAConfigAttribRateControl].value;
1113
1114         if (rc_mode == -1 || !(rc_mode & tmp))  {
1115             if (rc_mode != -1) {
1116                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1117             }
1118
1119             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1120                 if (rc_default_modes[i] & tmp) {
1121                     rc_mode = rc_default_modes[i];
1122                     break;
1123                 }
1124             }
1125         }
1126
1127         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1128         config_attrib[config_attrib_num].value = rc_mode;
1129         config_attrib_num++;
1130     }
1131
1132
1133     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1134         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1135
1136         h264_packedheader = 1;
1137         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1138         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1139
1140         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1141             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1142         }
1143
1144         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1145             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1146         }
1147
1148         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1149             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1150         }
1151
1152         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1153             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1154         }
1155
1156         enc_packed_header_idx = config_attrib_num;
1157         config_attrib_num++;
1158     }
1159
1160     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1161         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1162         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1163         config_attrib_num++;
1164     }
1165
1166     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1167         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1168     }
1169
1170     free(entrypoints);
1171     return 0;
1172 }
1173
1174 int QuickSyncEncoderImpl::setup_encode()
1175 {
1176     VAStatus va_status;
1177     VASurfaceID *tmp_surfaceid;
1178     int codedbuf_size, i;
1179     VASurfaceID src_surface[SURFACE_NUM];
1180     VASurfaceID ref_surface[SURFACE_NUM];
1181
1182     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1183             &config_attrib[0], config_attrib_num, &config_id);
1184     CHECK_VASTATUS(va_status, "vaCreateConfig");
1185
1186     /* create source surfaces */
1187     va_status = vaCreateSurfaces(va_dpy,
1188                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1189                                  &src_surface[0], SURFACE_NUM,
1190                                  NULL, 0);
1191     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1192
1193     /* create reference surfaces */
1194     va_status = vaCreateSurfaces(va_dpy,
1195                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1196                                  &ref_surface[0], SURFACE_NUM,
1197                                  NULL, 0);
1198     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1199
1200     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1201     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1202     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1203
1204     /* Create a context for this encode pipe */
1205     va_status = vaCreateContext(va_dpy, config_id,
1206                                 frame_width_mbaligned, frame_height_mbaligned,
1207                                 VA_PROGRESSIVE,
1208                                 tmp_surfaceid, 2 * SURFACE_NUM,
1209                                 &context_id);
1210     CHECK_VASTATUS(va_status, "vaCreateContext");
1211     free(tmp_surfaceid);
1212
1213     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1214
1215     for (i = 0; i < SURFACE_NUM; i++) {
1216         /* create coded buffer once for all
1217          * other VA buffers which won't be used again after vaRenderPicture.
1218          * so APP can always vaCreateBuffer for every frame
1219          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1220          * so VA won't maintain the coded buffer
1221          */
1222         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1223                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1224         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1225     }
1226
1227     /* create OpenGL objects */
1228     //glGenFramebuffers(SURFACE_NUM, fbos);
1229
1230     for (i = 0; i < SURFACE_NUM; i++) {
1231         if (use_zerocopy) {
1232             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
1233             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
1234         } else {
1235             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
1236             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
1237
1238             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1239             // buffers, due to potentially differing pitch.
1240             glGenBuffers(1, &gl_surfaces[i].pbo);
1241             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1242             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1243             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1244             gl_surfaces[i].y_offset = 0;
1245             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1246             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1247             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1248             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1249         }
1250     }
1251
1252     for (i = 0; i < SURFACE_NUM; i++) {
1253         gl_surfaces[i].src_surface = src_surface[i];
1254         gl_surfaces[i].ref_surface = ref_surface[i];
1255     }
1256
1257     return 0;
1258 }
1259
1260 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1261 //
1262 //   2 1 0 [3] 4 8 9
1263 template<class T, class C>
1264 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1265 {
1266         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1267         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1268         sort(middle, end, less_than);
1269 }
1270
1271 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1272 {
1273     int i;
1274
1275     if (frame_type == FRAME_B)
1276         return;
1277
1278     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1279     numShortTerm++;
1280     if (numShortTerm > num_ref_frames)
1281         numShortTerm = num_ref_frames;
1282     for (i=numShortTerm-1; i>0; i--)
1283         ReferenceFrames[i] = ReferenceFrames[i-1];
1284     ReferenceFrames[0] = CurrentCurrPic;
1285
1286     current_frame_num++;
1287     if (current_frame_num > MaxFrameNum)
1288         current_frame_num = 0;
1289 }
1290
1291
1292 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1293 {
1294     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1295         return a.frame_idx > b.frame_idx;
1296     };
1297     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1298         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1299     };
1300     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1301         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1302     };
1303
1304     if (frame_type == FRAME_P) {
1305         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1306         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1307     } else if (frame_type == FRAME_B) {
1308         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1309         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1310
1311         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1312         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1313     }
1314
1315     return 0;
1316 }
1317
1318
1319 int QuickSyncEncoderImpl::render_sequence()
1320 {
1321     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1322     VAStatus va_status;
1323     VAEncMiscParameterBuffer *misc_param;
1324     VAEncMiscParameterRateControl *misc_rate_ctrl;
1325
1326     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1327     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1328     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1329     seq_param.bits_per_second = frame_bitrate;
1330
1331     seq_param.intra_period = intra_period;
1332     seq_param.intra_idr_period = intra_idr_period;
1333     seq_param.ip_period = ip_period;
1334
1335     seq_param.max_num_ref_frames = num_ref_frames;
1336     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1337     seq_param.time_scale = TIMEBASE * 2;
1338     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1339     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1340     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1341     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1342     seq_param.seq_fields.bits.chroma_format_idc = 1;
1343     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1344
1345     if (frame_width != frame_width_mbaligned ||
1346         frame_height != frame_height_mbaligned) {
1347         seq_param.frame_cropping_flag = 1;
1348         seq_param.frame_crop_left_offset = 0;
1349         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1350         seq_param.frame_crop_top_offset = 0;
1351         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1352     }
1353
1354     va_status = vaCreateBuffer(va_dpy, context_id,
1355                                VAEncSequenceParameterBufferType,
1356                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1357     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1358
1359     va_status = vaCreateBuffer(va_dpy, context_id,
1360                                VAEncMiscParameterBufferType,
1361                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1362                                1, NULL, &rc_param_buf);
1363     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1364
1365     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1366     misc_param->type = VAEncMiscParameterTypeRateControl;
1367     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1368     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1369     misc_rate_ctrl->bits_per_second = frame_bitrate;
1370     misc_rate_ctrl->target_percentage = 66;
1371     misc_rate_ctrl->window_size = 1000;
1372     misc_rate_ctrl->initial_qp = initial_qp;
1373     misc_rate_ctrl->min_qp = minimal_qp;
1374     misc_rate_ctrl->basic_unit_size = 0;
1375     vaUnmapBuffer(va_dpy, rc_param_buf);
1376
1377     render_id[0] = seq_param_buf;
1378     render_id[1] = rc_param_buf;
1379
1380     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1381
1382     return 0;
1383 }
1384
1385 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1386 {
1387     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1388     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1389     int PicOrderCntMsb, TopFieldOrderCnt;
1390
1391     if (frame_type == FRAME_IDR)
1392         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1393     else {
1394         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1395         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1396     }
1397
1398     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1399         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1400         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1401     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1402              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1403         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1404     else
1405         PicOrderCntMsb = prevPicOrderCntMsb;
1406
1407     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1408
1409     if (frame_type != FRAME_B) {
1410         PicOrderCntMsb_ref = PicOrderCntMsb;
1411         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1412     }
1413
1414     return TopFieldOrderCnt;
1415 }
1416
1417 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1418 {
1419     VABufferID pic_param_buf;
1420     VAStatus va_status;
1421     int i = 0;
1422
1423     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1424     pic_param.CurrPic.frame_idx = current_frame_num;
1425     pic_param.CurrPic.flags = 0;
1426     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1427     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1428     CurrentCurrPic = pic_param.CurrPic;
1429
1430     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1431     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1432         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1433         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1434     }
1435
1436     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1437     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1438     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1439     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1440     pic_param.frame_num = current_frame_num;
1441     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1442     pic_param.last_picture = false;  // FIXME
1443     pic_param.pic_init_qp = initial_qp;
1444
1445     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1446                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1447     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1448
1449     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1450
1451     return 0;
1452 }
1453
1454 int QuickSyncEncoderImpl::render_packedsequence()
1455 {
1456     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1457     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1458     unsigned int length_in_bits;
1459     unsigned char *packedseq_buffer = NULL;
1460     VAStatus va_status;
1461
1462     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1463
1464     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1465
1466     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1467     packedheader_param_buffer.has_emulation_bytes = 0;
1468     va_status = vaCreateBuffer(va_dpy,
1469                                context_id,
1470                                VAEncPackedHeaderParameterBufferType,
1471                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1472                                &packedseq_para_bufid);
1473     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1474
1475     va_status = vaCreateBuffer(va_dpy,
1476                                context_id,
1477                                VAEncPackedHeaderDataBufferType,
1478                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1479                                &packedseq_data_bufid);
1480     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1481
1482     render_id[0] = packedseq_para_bufid;
1483     render_id[1] = packedseq_data_bufid;
1484     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1485
1486     free(packedseq_buffer);
1487
1488     return 0;
1489 }
1490
1491
1492 int QuickSyncEncoderImpl::render_packedpicture()
1493 {
1494     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1495     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1496     unsigned int length_in_bits;
1497     unsigned char *packedpic_buffer = NULL;
1498     VAStatus va_status;
1499
1500     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1501     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1502     packedheader_param_buffer.bit_length = length_in_bits;
1503     packedheader_param_buffer.has_emulation_bytes = 0;
1504
1505     va_status = vaCreateBuffer(va_dpy,
1506                                context_id,
1507                                VAEncPackedHeaderParameterBufferType,
1508                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1509                                &packedpic_para_bufid);
1510     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1511
1512     va_status = vaCreateBuffer(va_dpy,
1513                                context_id,
1514                                VAEncPackedHeaderDataBufferType,
1515                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1516                                &packedpic_data_bufid);
1517     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1518
1519     render_id[0] = packedpic_para_bufid;
1520     render_id[1] = packedpic_data_bufid;
1521     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1522
1523     free(packedpic_buffer);
1524
1525     return 0;
1526 }
1527
1528 void QuickSyncEncoderImpl::render_packedslice()
1529 {
1530     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1531     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1532     unsigned int length_in_bits;
1533     unsigned char *packedslice_buffer = NULL;
1534     VAStatus va_status;
1535
1536     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1537     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1538     packedheader_param_buffer.bit_length = length_in_bits;
1539     packedheader_param_buffer.has_emulation_bytes = 0;
1540
1541     va_status = vaCreateBuffer(va_dpy,
1542                                context_id,
1543                                VAEncPackedHeaderParameterBufferType,
1544                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1545                                &packedslice_para_bufid);
1546     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1547
1548     va_status = vaCreateBuffer(va_dpy,
1549                                context_id,
1550                                VAEncPackedHeaderDataBufferType,
1551                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1552                                &packedslice_data_bufid);
1553     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1554
1555     render_id[0] = packedslice_para_bufid;
1556     render_id[1] = packedslice_data_bufid;
1557     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1558
1559     free(packedslice_buffer);
1560 }
1561
1562 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1563 {
1564     VABufferID slice_param_buf;
1565     VAStatus va_status;
1566     int i;
1567
1568     update_RefPicList(frame_type);
1569
1570     /* one frame, one slice */
1571     slice_param.macroblock_address = 0;
1572     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1573     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1574     if (frame_type == FRAME_IDR) {
1575         if (encoding_frame_num != 0)
1576             ++slice_param.idr_pic_id;
1577     } else if (frame_type == FRAME_P) {
1578         int refpiclist0_max = h264_maxref & 0xffff;
1579         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1580
1581         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1582             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1583             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1584         }
1585     } else if (frame_type == FRAME_B) {
1586         int refpiclist0_max = h264_maxref & 0xffff;
1587         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1588
1589         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1590         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1591             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1592             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1593         }
1594
1595         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1596         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1597             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1598             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1599         }
1600     }
1601
1602     slice_param.slice_alpha_c0_offset_div2 = 0;
1603     slice_param.slice_beta_offset_div2 = 0;
1604     slice_param.direct_spatial_mv_pred_flag = 1;
1605     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1606
1607
1608     if (h264_packedheader &&
1609         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1610         render_packedslice();
1611
1612     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1613                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1614     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1615
1616     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1617
1618     return 0;
1619 }
1620
1621
1622
1623 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1624 {
1625         VACodedBufferSegment *buf_list = NULL;
1626         VAStatus va_status;
1627
1628         string data;
1629
1630         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1631         CHECK_VASTATUS(va_status, "vaMapBuffer");
1632         while (buf_list != NULL) {
1633                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1634                 buf_list = (VACodedBufferSegment *) buf_list->next;
1635         }
1636         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1637
1638         static int frameno = 0;
1639         print_latency("Current QuickSync latency (video inputs → disk mux):",
1640                 task.received_ts, (task.frame_type == FRAME_B), &frameno);
1641
1642         {
1643                 // Add video.
1644                 AVPacket pkt;
1645                 memset(&pkt, 0, sizeof(pkt));
1646                 pkt.buf = nullptr;
1647                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1648                 pkt.size = data.size();
1649                 pkt.stream_index = 0;
1650                 if (task.frame_type == FRAME_IDR) {
1651                         pkt.flags = AV_PKT_FLAG_KEY;
1652                 } else {
1653                         pkt.flags = 0;
1654                 }
1655                 pkt.duration = task.duration;
1656                 if (file_mux) {
1657                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1658                 }
1659                 if (!global_flags.uncompressed_video_to_http &&
1660                     !global_flags.x264_video_to_http) {
1661                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1662                 }
1663         }
1664 }
1665
1666
1667 // this is weird. but it seems to put a new frame onto the queue
1668 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1669 {
1670         unique_lock<mutex> lock(storage_task_queue_mutex);
1671         storage_task_queue.push(move(task));
1672         storage_task_queue_changed.notify_all();
1673 }
1674
1675 void QuickSyncEncoderImpl::storage_task_thread()
1676 {
1677         for ( ;; ) {
1678                 storage_task current;
1679                 {
1680                         // wait until there's an encoded frame
1681                         unique_lock<mutex> lock(storage_task_queue_mutex);
1682                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1683                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1684                         current = move(storage_task_queue.front());
1685                         storage_task_queue.pop();
1686                 }
1687
1688                 VAStatus va_status;
1689
1690                 // waits for data, then saves it to disk.
1691                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1692                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1693                 save_codeddata(move(current));
1694
1695                 {
1696                         unique_lock<mutex> lock(storage_task_queue_mutex);
1697                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1698                         storage_task_queue_changed.notify_all();
1699                 }
1700         }
1701 }
1702
1703 void QuickSyncEncoderImpl::release_encode()
1704 {
1705         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1706                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1707                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1708                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1709         }
1710
1711         vaDestroyContext(va_dpy, context_id);
1712         vaDestroyConfig(va_dpy, config_id);
1713 }
1714
1715 void QuickSyncEncoderImpl::release_gl_resources()
1716 {
1717         assert(is_shutdown);
1718         if (has_released_gl_resources) {
1719                 return;
1720         }
1721
1722         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1723                 if (!use_zerocopy) {
1724                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1725                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1726                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1727                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1728                 }
1729                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1730                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1731         }
1732
1733         has_released_gl_resources = true;
1734 }
1735
1736 int QuickSyncEncoderImpl::deinit_va()
1737 {
1738     vaTerminate(va_dpy);
1739
1740     va_close_display(va_dpy);
1741
1742     return 0;
1743 }
1744
1745 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)
1746         : 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)
1747 {
1748         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1749         open_output_file(filename);
1750         file_audio_encoder->add_mux(file_mux.get());
1751
1752         frame_width_mbaligned = (frame_width + 15) & (~15);
1753         frame_height_mbaligned = (frame_height + 15) & (~15);
1754
1755         //print_input();
1756
1757         if (global_flags.uncompressed_video_to_http ||
1758             global_flags.x264_video_to_http) {
1759                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1760         }
1761         if (global_flags.x264_video_to_http) {
1762                 assert(x264_encoder != nullptr);
1763         } else {
1764                 assert(x264_encoder == nullptr);
1765         }
1766
1767         init_va(va_display);
1768         setup_encode();
1769
1770         // No frames are ready yet.
1771         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1772
1773         memset(&seq_param, 0, sizeof(seq_param));
1774         memset(&pic_param, 0, sizeof(pic_param));
1775         memset(&slice_param, 0, sizeof(slice_param));
1776
1777         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1778
1779         encode_thread = thread([this]{
1780                 //SDL_GL_MakeCurrent(window, context);
1781                 QOpenGLContext *context = create_context(this->surface);
1782                 eglBindAPI(EGL_OPENGL_API);
1783                 if (!make_current(context, this->surface)) {
1784                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1785                                 eglGetError());
1786                         exit(1);
1787                 }
1788                 encode_thread_func();
1789                 delete_context(context);
1790         });
1791 }
1792
1793 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1794 {
1795         shutdown();
1796         release_gl_resources();
1797 }
1798
1799 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1800 {
1801         assert(!is_shutdown);
1802         {
1803                 // Wait until this frame slot is done encoding.
1804                 unique_lock<mutex> lock(storage_task_queue_mutex);
1805                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1806                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1807                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1808                 }
1809                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1810                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1811                 if (storage_thread_should_quit) return false;
1812         }
1813
1814         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1815         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1816         *y_tex = surf->y_tex;
1817         *cbcr_tex = surf->cbcr_tex;
1818
1819         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1820         CHECK_VASTATUS(va_status, "vaDeriveImage");
1821
1822         if (use_zerocopy) {
1823                 VABufferInfo buf_info;
1824                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1825                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1826                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1827
1828                 // Create Y image.
1829                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1830                 EGLint y_attribs[] = {
1831                         EGL_WIDTH, frame_width,
1832                         EGL_HEIGHT, frame_height,
1833                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1834                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1835                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1836                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1837                         EGL_NONE
1838                 };
1839
1840                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1841                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1842
1843                 // Associate Y image to a texture.
1844                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1845                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1846
1847                 // Create CbCr image.
1848                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1849                 EGLint cbcr_attribs[] = {
1850                         EGL_WIDTH, frame_width,
1851                         EGL_HEIGHT, frame_height,
1852                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1853                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1854                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1855                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1856                         EGL_NONE
1857                 };
1858
1859                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1860                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1861
1862                 // Associate CbCr image to a texture.
1863                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1864                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1865         }
1866
1867         return true;
1868 }
1869
1870 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1871 {
1872         assert(!is_shutdown);
1873         file_audio_encoder->encode_audio(audio, pts + global_delay());
1874 }
1875
1876 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1877 {
1878         assert(!is_shutdown);
1879
1880         if (!use_zerocopy) {
1881                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1882
1883                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1884                 check_error();
1885
1886                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1887                 check_error();
1888
1889                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1890                 check_error();
1891                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1892                 check_error();
1893
1894                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1895                 check_error();
1896                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1897                 check_error();
1898
1899                 glBindTexture(GL_TEXTURE_2D, 0);
1900                 check_error();
1901                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1902                 check_error();
1903
1904                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1905                 check_error();
1906         }
1907
1908         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1909         check_error();
1910         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1911         check_error();
1912
1913         {
1914                 unique_lock<mutex> lock(frame_queue_mutex);
1915                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1916                 ++current_storage_frame;
1917         }
1918         frame_queue_nonempty.notify_all();
1919         return fence;
1920 }
1921
1922 void QuickSyncEncoderImpl::shutdown()
1923 {
1924         if (is_shutdown) {
1925                 return;
1926         }
1927
1928         {
1929                 unique_lock<mutex> lock(frame_queue_mutex);
1930                 encode_thread_should_quit = true;
1931                 frame_queue_nonempty.notify_all();
1932         }
1933         encode_thread.join();
1934         {
1935                 unique_lock<mutex> lock(storage_task_queue_mutex);
1936                 storage_thread_should_quit = true;
1937                 frame_queue_nonempty.notify_all();
1938                 storage_task_queue_changed.notify_all();
1939         }
1940         storage_thread.join();
1941
1942         // Encode any leftover audio in the queues, and also any delayed frames.
1943         file_audio_encoder->encode_last_audio();
1944
1945         release_encode();
1946         deinit_va();
1947         file_mux.reset();
1948         is_shutdown = true;
1949 }
1950
1951 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1952 {
1953         AVFormatContext *avctx = avformat_alloc_context();
1954         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1955         assert(filename.size() < sizeof(avctx->filename) - 1);
1956         strcpy(avctx->filename, filename.c_str());
1957
1958         string url = "file:" + filename;
1959         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1960         if (ret < 0) {
1961                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1962                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1963                 exit(1);
1964         }
1965
1966         string video_extradata = "";  // FIXME: See other comment about global headers.
1967         AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1968         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1969                 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1)));
1970 }
1971
1972 void QuickSyncEncoderImpl::encode_thread_func()
1973 {
1974         int64_t last_dts = -1;
1975         int gop_start_display_frame_num = 0;
1976         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1977                 PendingFrame frame;
1978                 int pts_lag;
1979                 int frame_type, display_frame_num;
1980                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1981                                        &display_frame_num, &frame_type, &pts_lag);
1982                 if (frame_type == FRAME_IDR) {
1983                         numShortTerm = 0;
1984                         current_frame_num = 0;
1985                         gop_start_display_frame_num = display_frame_num;
1986                 }
1987
1988                 {
1989                         unique_lock<mutex> lock(frame_queue_mutex);
1990                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1991                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1992                         });
1993                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1994                                 // We have queued frames that were supposed to be B-frames,
1995                                 // but will be no P-frame to encode them against. Encode them all
1996                                 // as P-frames instead. Note that this happens under the mutex,
1997                                 // but nobody else uses it at this point, since we're shutting down,
1998                                 // so there's no contention.
1999                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
2000                                 return;
2001                         } else {
2002                                 frame = move(pending_video_frames[display_frame_num]);
2003                                 pending_video_frames.erase(display_frame_num);
2004                         }
2005                 }
2006
2007                 // Determine the dts of this frame.
2008                 int64_t dts;
2009                 if (pts_lag == -1) {
2010                         assert(last_dts != -1);
2011                         dts = last_dts + (TIMEBASE / MAX_FPS);
2012                 } else {
2013                         dts = frame.pts - pts_lag;
2014                 }
2015                 last_dts = dts;
2016
2017                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2018         }
2019 }
2020
2021 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2022 {
2023         if (pending_video_frames.empty()) {
2024                 return;
2025         }
2026
2027         for (auto &pending_frame : pending_video_frames) {
2028                 int display_frame_num = pending_frame.first;
2029                 assert(display_frame_num > 0);
2030                 PendingFrame frame = move(pending_frame.second);
2031                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2032                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2033                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2034                 last_dts = dts;
2035         }
2036
2037         if (global_flags.uncompressed_video_to_http ||
2038             global_flags.x264_video_to_http) {
2039                 // Add frames left in reorderer.
2040                 while (!reorderer->empty()) {
2041                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2042                         if (global_flags.uncompressed_video_to_http) {
2043                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2044                         } else {
2045                                 assert(global_flags.x264_video_to_http);
2046                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data, output_frame.received_ts);
2047                         }
2048                 }
2049         }
2050 }
2051
2052 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2053 {
2054         AVPacket pkt;
2055         memset(&pkt, 0, sizeof(pkt));
2056         pkt.buf = nullptr;
2057         pkt.data = const_cast<uint8_t *>(data);
2058         pkt.size = frame_width * frame_height * 2;
2059         pkt.stream_index = 0;
2060         pkt.flags = AV_PKT_FLAG_KEY;
2061         pkt.duration = duration;
2062         stream_mux->add_packet(pkt, pts, pts);
2063 }
2064
2065 namespace {
2066
2067 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2068 {
2069         if (src_width == dst_pitch) {
2070                 memcpy(dst, src, src_width * height);
2071         } else {
2072                 for (size_t y = 0; y < height; ++y) {
2073                         const uint8_t *sptr = src + y * src_width;
2074                         uint8_t *dptr = dst + y * dst_pitch;
2075                         memcpy(dptr, sptr, src_width);
2076                 }
2077         }
2078 }
2079
2080 }  // namespace
2081
2082 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2083                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2084 {
2085         // Wait for the GPU to be done with the frame.
2086         GLenum sync_status;
2087         do {
2088                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2089                 check_error();
2090         } while (sync_status == GL_TIMEOUT_EXPIRED);
2091         assert(sync_status != GL_WAIT_FAILED);
2092
2093         // Find min and max timestamp of all input frames that have a timestamp.
2094         steady_clock::time_point min_ts = steady_clock::time_point::max(), max_ts = steady_clock::time_point::min();
2095         for (const RefCountedFrame &input_frame : frame.input_frames) {
2096                 if (input_frame && input_frame->received_timestamp > steady_clock::time_point::min()) {
2097                         min_ts = min(min_ts, input_frame->received_timestamp);
2098                         max_ts = max(max_ts, input_frame->received_timestamp);
2099                 }
2100         }
2101         const ReceivedTimestamps received_ts{ min_ts, max_ts };
2102
2103         static int frameno = 0;
2104         print_latency("Current mixer latency (video inputs → ready for encode):",
2105                 received_ts, (frame_type == FRAME_B), &frameno);
2106
2107         // Release back any input frames we needed to render this frame.
2108         frame.input_frames.clear();
2109
2110         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2111         VAStatus va_status;
2112
2113         if (use_zerocopy) {
2114                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2115                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2116                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2117                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2118         } else {
2119                 unsigned char *surface_p = nullptr;
2120                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2121
2122                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2123                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2124
2125                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2126                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2127
2128                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2129                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2130
2131                 if (global_flags.uncompressed_video_to_http ||
2132                     global_flags.x264_video_to_http) {
2133                         // Add uncompressed video. (Note that pts == dts here.)
2134                         // Delay needs to match audio.
2135                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr), received_ts);
2136                         if (output_frame.data != nullptr) {
2137                                 if (global_flags.uncompressed_video_to_http) {
2138                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2139                                 } else {
2140                                         assert(global_flags.x264_video_to_http);
2141                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data, output_frame.received_ts);
2142                                 }
2143                         }
2144                 }
2145         }
2146
2147         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2148         CHECK_VASTATUS(va_status, "vaDestroyImage");
2149
2150         // Schedule the frame for encoding.
2151         VASurfaceID va_surface = surf->src_surface;
2152         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2153         CHECK_VASTATUS(va_status, "vaBeginPicture");
2154
2155         if (frame_type == FRAME_IDR) {
2156                 // FIXME: If the mux wants global headers, we should not put the
2157                 // SPS/PPS before each IDR frame, but rather put it into the
2158                 // codec extradata (formatted differently?).
2159                 render_sequence();
2160                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2161                 if (h264_packedheader) {
2162                         render_packedsequence();
2163                         render_packedpicture();
2164                 }
2165         } else {
2166                 //render_sequence();
2167                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2168         }
2169         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2170
2171         va_status = vaEndPicture(va_dpy, context_id);
2172         CHECK_VASTATUS(va_status, "vaEndPicture");
2173
2174         // so now the data is done encoding (well, async job kicked off)...
2175         // we send that to the storage thread
2176         storage_task tmp;
2177         tmp.display_order = display_frame_num;
2178         tmp.frame_type = frame_type;
2179         tmp.pts = pts;
2180         tmp.dts = dts;
2181         tmp.duration = duration;
2182         tmp.received_ts = received_ts;
2183         storage_task_enqueue(move(tmp));
2184
2185         update_ReferenceFrames(frame_type);
2186 }
2187
2188 // Proxy object.
2189 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)
2190         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2191
2192 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2193 QuickSyncEncoder::~QuickSyncEncoder() {}
2194
2195 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2196 {
2197         impl->add_audio(pts, audio);
2198 }
2199
2200 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2201 {
2202         return impl->begin_frame(y_tex, cbcr_tex);
2203 }
2204
2205 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2206 {
2207         return impl->end_frame(pts, duration, input_frames);
2208 }
2209
2210 void QuickSyncEncoder::shutdown()
2211 {
2212         impl->shutdown();
2213 }
2214
2215 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2216 {
2217         impl->set_stream_mux(mux);
2218 }
2219
2220 int64_t QuickSyncEncoder::global_delay() const {
2221         return impl->global_delay();
2222 }