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

   1 //#include "sysdeps.h"
   2 #include "quicksync_encoder.h"
   3
   4 #include <movit/resource_pool.h>
   5 #include <movit/util.h>
   6 #include <EGL/eglplatform.h>
   7 #include <X11/X.h>
   8 #include <X11/Xlib.h>
   9 #include <assert.h>
  10 #include <epoxy/egl.h>
  11 #include <libdrm/drm_fourcc.h>
  12 #include <stdio.h>
  13 #include <stdlib.h>
  14 #include <string.h>
  15 #include <fcntl.h>
  16 #include <va/va.h>
  17 #include <va/va_drm.h>
  18 #include <va/va_drmcommon.h>
  19 #include <va/va_enc_h264.h>
  20 #include <va/va_x11.h>
  21 #include <algorithm>
  22 #include <condition_variable>
  23 #include <cstdint>
  24 #include <map>
  25 #include <memory>
  26 #include <mutex>
  27 #include <queue>
  28 #include <string>
  29 #include <thread>
  30 #include <utility>
  31
  32 #include "audio_encoder.h"
  33 #include "context.h"
  34 #include "defs.h"
  35 #include "flags.h"
  36 #include "mux.h"
  37 #include "timebase.h"
  38 #include "x264_encoder.h"
  39
  40 using namespace std;
  41
  42 class QOpenGLContext;
  43 class QSurface;
  44
  45 #define CHECK_VASTATUS(va_status, func)                                 \
  46     if (va_status != VA_STATUS_SUCCESS) {                               \
  47         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  48         exit(1);                                                        \
  49     }
  50
  51 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  52
  53 //#include "loadsurface.h"
  54
  55 #define NAL_REF_IDC_NONE        0
  56 #define NAL_REF_IDC_LOW         1
  57 #define NAL_REF_IDC_MEDIUM      2
  58 #define NAL_REF_IDC_HIGH        3
  59
  60 #define NAL_NON_IDR             1
  61 #define NAL_IDR                 5
  62 #define NAL_SPS                 7
  63 #define NAL_PPS                 8
  64 #define NAL_SEI                 6
  65
  66 #define SLICE_TYPE_P            0
  67 #define SLICE_TYPE_B            1
  68 #define SLICE_TYPE_I            2
  69 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  70 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  71 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  72
  73
  74 #define ENTROPY_MODE_CAVLC      0
  75 #define ENTROPY_MODE_CABAC      1
  76
  77 #define PROFILE_IDC_BASELINE    66
  78 #define PROFILE_IDC_MAIN        77
  79 #define PROFILE_IDC_HIGH        100
  80
  81 #define BITSTREAM_ALLOCATE_STEPPING     4096
  82 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  83 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  84 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  85
  86 static constexpr unsigned int MaxFrameNum = (2<<16);
  87 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  88 static constexpr unsigned int Log2MaxFrameNum = 16;
  89 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  90 static constexpr int rc_default_modes[] = {  // Priority list of modes.
  91     VA_RC_VBR,
  92     VA_RC_CQP,
  93     VA_RC_VBR_CONSTRAINED,
  94     VA_RC_CBR,
  95     VA_RC_VCM,
  96     VA_RC_NONE,
  97 };
  98
  99 /* thread to save coded data */
 100 #define SRC_SURFACE_FREE        0
 101 #define SRC_SURFACE_IN_ENCODING 1
 102
 103 struct __bitstream {
 104     unsigned int *buffer;
 105     int bit_offset;
 106     int max_size_in_dword;
 107 };
 108 typedef struct __bitstream bitstream;
 109
 110 using namespace std;
 111
 112 // H.264 video comes out in encoding order (e.g. with two B-frames:
 113 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 114 // come in the right order. Since we do everything, including waiting
 115 // for the frames to come out of OpenGL, in encoding order, we need
 116 // a reordering buffer for uncompressed frames so that they come out
 117 // correctly. We go the super-lazy way of not making it understand
 118 // anything about the true order (which introduces some extra latency,
 119 // though); we know that for N B-frames we need at most (N-1) frames
 120 // in the reorder buffer, and can just sort on that.
 121 //
 122 // The class also deals with keeping a freelist as needed.
 123 class FrameReorderer {
 124 public:
 125         FrameReorderer(unsigned queue_length, int width, int height);
 126
 127         struct Frame {
 128                 int64_t pts, duration;
 129                 uint8_t *data;
 130
 131                 // Invert to get the smallest pts first.
 132                 bool operator< (const Frame &other) const { return pts > other.pts; }
 133         };
 134
 135         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 136         // Does _not_ take ownership of data; a copy is taken if needed.
 137         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 138         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 139         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data);
 140
 141         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 142         Frame get_first_frame();
 143
 144         bool empty() const { return frames.empty(); }
 145
 146 private:
 147         unsigned queue_length;
 148         int width, height;
 149
 150         priority_queue<Frame> frames;
 151         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 152
 153         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 154         // except priority_queue doesn't work well with movable-only types.
 155         vector<unique_ptr<uint8_t[]>> owner;
 156 };
 157
 158 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 159     : queue_length(queue_length), width(width), height(height)
 160 {
 161         for (unsigned i = 0; i < queue_length; ++i) {
 162                 owner.emplace_back(new uint8_t[width * height * 2]);
 163                 freelist.push(owner.back().get());
 164         }
 165 }
 166
 167 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data)
 168 {
 169         if (queue_length == 0) {
 170                 return Frame{pts, duration, data};
 171         }
 172
 173         assert(!freelist.empty());
 174         uint8_t *storage = freelist.top();
 175         freelist.pop();
 176         memcpy(storage, data, width * height * 2);
 177         frames.push(Frame{pts, duration, storage});
 178
 179         if (frames.size() >= queue_length) {
 180                 return get_first_frame();
 181         } else {
 182                 return Frame{-1, -1, nullptr};
 183         }
 184 }
 185
 186 FrameReorderer::Frame FrameReorderer::get_first_frame()
 187 {
 188         assert(!frames.empty());
 189         Frame storage = frames.top();
 190         frames.pop();
 191         freelist.push(storage.data);
 192         return storage;
 193 }
 194
 195 class QuickSyncEncoderImpl {
 196 public:
 197         QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder);
 198         ~QuickSyncEncoderImpl();
 199         void add_audio(int64_t pts, vector<float> audio);
 200         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 201         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
 202         void shutdown();
 203         void set_stream_mux(Mux *mux)
 204         {
 205                 stream_mux = mux;
 206         }
 207
 208 private:
 209         struct storage_task {
 210                 unsigned long long display_order;
 211                 int frame_type;
 212                 vector<float> audio;
 213                 int64_t pts, dts, duration;
 214         };
 215         struct PendingFrame {
 216                 RefCountedGLsync fence;
 217                 vector<RefCountedFrame> input_frames;
 218                 int64_t pts, duration;
 219         };
 220
 221         // So we never get negative dts.
 222         int64_t global_delay() const {
 223                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 224         }
 225
 226         void open_output_file(const std::string &filename);
 227         void encode_thread_func();
 228         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 229         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 230         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 231                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 232         void storage_task_thread();
 233         void encode_remaining_audio();
 234         void storage_task_enqueue(storage_task task);
 235         void save_codeddata(storage_task task);
 236         int render_packedsequence();
 237         int render_packedpicture();
 238         void render_packedslice();
 239         int render_sequence();
 240         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 241         void sps_rbsp(bitstream *bs);
 242         void pps_rbsp(bitstream *bs);
 243         int build_packed_pic_buffer(unsigned char **header_buffer);
 244         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 245         void slice_header(bitstream *bs);
 246         int build_packed_seq_buffer(unsigned char **header_buffer);
 247         int build_packed_slice_buffer(unsigned char **header_buffer);
 248         int init_va(const string &va_display);
 249         int deinit_va();
 250         void enable_zerocopy_if_possible();
 251         VADisplay va_open_display(const string &va_display);
 252         void va_close_display(VADisplay va_dpy);
 253         int setup_encode();
 254         int release_encode();
 255         void update_ReferenceFrames(int frame_type);
 256         int update_RefPicList(int frame_type);
 257
 258         bool is_shutdown = false;
 259         bool use_zerocopy;
 260         int drm_fd = -1;
 261
 262         thread encode_thread, storage_thread;
 263
 264         mutex storage_task_queue_mutex;
 265         condition_variable storage_task_queue_changed;
 266         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 267         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 268         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 269
 270         mutex frame_queue_mutex;
 271         condition_variable frame_queue_nonempty;
 272         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 273
 274         int current_storage_frame;
 275
 276         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 277         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 278         movit::ResourcePool *resource_pool;
 279         QSurface *surface;
 280
 281         unique_ptr<AudioEncoder> file_audio_encoder;
 282         AudioEncoder *stream_audio_encoder;
 283
 284         unique_ptr<FrameReorderer> reorderer;
 285         X264Encoder *x264_encoder;  // nullptr if not using x264.
 286
 287         Mux* stream_mux = nullptr;  // To HTTP.
 288         unique_ptr<Mux> file_mux;  // To local disk.
 289
 290         Display *x11_display = nullptr;
 291
 292         // Encoder parameters
 293         VADisplay va_dpy;
 294         VAProfile h264_profile = (VAProfile)~0;
 295         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 296         int config_attrib_num = 0, enc_packed_header_idx;
 297
 298         struct GLSurface {
 299                 VASurfaceID src_surface, ref_surface;
 300                 VABufferID coded_buf;
 301
 302                 VAImage surface_image;
 303                 GLuint y_tex, cbcr_tex;
 304
 305                 // Only if use_zerocopy == true.
 306                 EGLImage y_egl_image, cbcr_egl_image;
 307
 308                 // Only if use_zerocopy == false.
 309                 GLuint pbo;
 310                 uint8_t *y_ptr, *cbcr_ptr;
 311                 size_t y_offset, cbcr_offset;
 312         };
 313         GLSurface gl_surfaces[SURFACE_NUM];
 314
 315         VAConfigID config_id;
 316         VAContextID context_id;
 317         VAEncSequenceParameterBufferH264 seq_param;
 318         VAEncPictureParameterBufferH264 pic_param;
 319         VAEncSliceParameterBufferH264 slice_param;
 320         VAPictureH264 CurrentCurrPic;
 321         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 322
 323         // Static quality settings.
 324         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 325         static constexpr unsigned int num_ref_frames = 2;
 326         static constexpr int initial_qp = 15;
 327         static constexpr int minimal_qp = 0;
 328         static constexpr int intra_period = 30;
 329         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 330
 331         // Quality settings that are meant to be static, but might be overridden
 332         // by the profile.
 333         int constraint_set_flag = 0;
 334         int h264_packedheader = 0; /* support pack header? */
 335         int h264_maxref = (1<<16|1);
 336         int h264_entropy_mode = 1; /* cabac */
 337         int ip_period = 3;
 338
 339         int rc_mode = -1;
 340         unsigned int current_frame_num = 0;
 341         unsigned int numShortTerm = 0;
 342
 343         int frame_width;
 344         int frame_height;
 345         int frame_width_mbaligned;
 346         int frame_height_mbaligned;
 347 };
 348
 349 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 350 // but if we don't delete it here, we get leaks. The GStreamer implementation
 351 // does the same.
 352 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 353 {
 354     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 355     CHECK_VASTATUS(va_status, "vaRenderPicture");
 356
 357     for (int i = 0; i < num_buffers; ++i) {
 358         va_status = vaDestroyBuffer(dpy, buffers[i]);
 359         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 360     }
 361 }
 362
 363 static unsigned int
 364 va_swap32(unsigned int val)
 365 {
 366     unsigned char *pval = (unsigned char *)&val;
 367
 368     return ((pval[0] << 24)     |
 369             (pval[1] << 16)     |
 370             (pval[2] << 8)      |
 371             (pval[3] << 0));
 372 }
 373
 374 static void
 375 bitstream_start(bitstream *bs)
 376 {
 377     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 378     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 379     bs->bit_offset = 0;
 380 }
 381
 382 static void
 383 bitstream_end(bitstream *bs)
 384 {
 385     int pos = (bs->bit_offset >> 5);
 386     int bit_offset = (bs->bit_offset & 0x1f);
 387     int bit_left = 32 - bit_offset;
 388
 389     if (bit_offset) {
 390         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 391     }
 392 }
 393
 394 static void
 395 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 396 {
 397     int pos = (bs->bit_offset >> 5);
 398     int bit_offset = (bs->bit_offset & 0x1f);
 399     int bit_left = 32 - bit_offset;
 400
 401     if (!size_in_bits)
 402         return;
 403
 404     bs->bit_offset += size_in_bits;
 405
 406     if (bit_left > size_in_bits) {
 407         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 408     } else {
 409         size_in_bits -= bit_left;
 410         if (bit_left >= 32) {
 411             bs->buffer[pos] = (val >> size_in_bits);
 412         } else {
 413             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 414         }
 415         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 416
 417         if (pos + 1 == bs->max_size_in_dword) {
 418             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 419             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 420         }
 421
 422         bs->buffer[pos + 1] = val;
 423     }
 424 }
 425
 426 static void
 427 bitstream_put_ue(bitstream *bs, unsigned int val)
 428 {
 429     int size_in_bits = 0;
 430     int tmp_val = ++val;
 431
 432     while (tmp_val) {
 433         tmp_val >>= 1;
 434         size_in_bits++;
 435     }
 436
 437     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 438     bitstream_put_ui(bs, val, size_in_bits);
 439 }
 440
 441 static void
 442 bitstream_put_se(bitstream *bs, int val)
 443 {
 444     unsigned int new_val;
 445
 446     if (val <= 0)
 447         new_val = -2 * val;
 448     else
 449         new_val = 2 * val - 1;
 450
 451     bitstream_put_ue(bs, new_val);
 452 }
 453
 454 static void
 455 bitstream_byte_aligning(bitstream *bs, int bit)
 456 {
 457     int bit_offset = (bs->bit_offset & 0x7);
 458     int bit_left = 8 - bit_offset;
 459     int new_val;
 460
 461     if (!bit_offset)
 462         return;
 463
 464     assert(bit == 0 || bit == 1);
 465
 466     if (bit)
 467         new_val = (1 << bit_left) - 1;
 468     else
 469         new_val = 0;
 470
 471     bitstream_put_ui(bs, new_val, bit_left);
 472 }
 473
 474 static void
 475 rbsp_trailing_bits(bitstream *bs)
 476 {
 477     bitstream_put_ui(bs, 1, 1);
 478     bitstream_byte_aligning(bs, 0);
 479 }
 480
 481 static void nal_start_code_prefix(bitstream *bs)
 482 {
 483     bitstream_put_ui(bs, 0x00000001, 32);
 484 }
 485
 486 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 487 {
 488     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 489     bitstream_put_ui(bs, nal_ref_idc, 2);
 490     bitstream_put_ui(bs, nal_unit_type, 5);
 491 }
 492
 493 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 494 {
 495     int profile_idc = PROFILE_IDC_BASELINE;
 496
 497     if (h264_profile  == VAProfileH264High)
 498         profile_idc = PROFILE_IDC_HIGH;
 499     else if (h264_profile  == VAProfileH264Main)
 500         profile_idc = PROFILE_IDC_MAIN;
 501
 502     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 503     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 504     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 505     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 506     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 507     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 508     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 509     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 510
 511     if ( profile_idc == PROFILE_IDC_HIGH) {
 512         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 513         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 514         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 515         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 516         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 517     }
 518
 519     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 520     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 521
 522     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 523         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 524     else {
 525         assert(0);
 526     }
 527
 528     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 529     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 530
 531     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 532     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 533     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 534
 535     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 536         assert(0);
 537     }
 538
 539     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 540     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 541
 542     if (seq_param.frame_cropping_flag) {
 543         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 544         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 545         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 546         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 547     }
 548
 549     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 550     if ( false ) {
 551         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 552     } else {
 553         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 554         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 555         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 556         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 557         {
 558             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 559             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 560             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 561             {
 562                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 563                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 564                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 565             }
 566         }
 567         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 568         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 569         {
 570             bitstream_put_ui(bs, 1, 32);  // FPS
 571             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 572             bitstream_put_ui(bs, 1, 1);
 573         }
 574         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 575         {
 576             // hrd_parameters
 577             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 578             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 579             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 580
 581             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 582             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 583             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 584
 585             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 586             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 587             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 588             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 589         }
 590         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 591         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 592
 593         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 594         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 595     }
 596
 597     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 598 }
 599
 600
 601 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 602 {
 603     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 604     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 605
 606     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 607
 608     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 609
 610     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 611
 612     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 613     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 614
 615     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 616     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 617
 618     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 619     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 620     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 621
 622     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 623     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 624     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 625
 626     /* more_rbsp_data */
 627     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 628     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 629     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 630
 631     rbsp_trailing_bits(bs);
 632 }
 633
 634 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 635 {
 636     int first_mb_in_slice = slice_param.macroblock_address;
 637
 638     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 639     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 640     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 641     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 642
 643     /* frame_mbs_only_flag == 1 */
 644     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 645         /* FIXME: */
 646         assert(0);
 647     }
 648
 649     if (pic_param.pic_fields.bits.idr_pic_flag)
 650         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 651
 652     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 653         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 654         /* pic_order_present_flag == 0 */
 655     } else {
 656         /* FIXME: */
 657         assert(0);
 658     }
 659
 660     /* redundant_pic_cnt_present_flag == 0 */
 661     /* slice type */
 662     if (IS_P_SLICE(slice_param.slice_type)) {
 663         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 664
 665         if (slice_param.num_ref_idx_active_override_flag)
 666             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 667
 668         /* ref_pic_list_reordering */
 669         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 670     } else if (IS_B_SLICE(slice_param.slice_type)) {
 671         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 672
 673         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 674
 675         if (slice_param.num_ref_idx_active_override_flag) {
 676             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 677             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 678         }
 679
 680         /* ref_pic_list_reordering */
 681         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 682         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 683     }
 684
 685     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 686          IS_P_SLICE(slice_param.slice_type)) ||
 687         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 688          IS_B_SLICE(slice_param.slice_type))) {
 689         /* FIXME: fill weight/offset table */
 690         assert(0);
 691     }
 692
 693     /* dec_ref_pic_marking */
 694     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 695         unsigned char no_output_of_prior_pics_flag = 0;
 696         unsigned char long_term_reference_flag = 0;
 697         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 698
 699         if (pic_param.pic_fields.bits.idr_pic_flag) {
 700             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 701             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 702         } else {
 703             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 704         }
 705     }
 706
 707     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 708         !IS_I_SLICE(slice_param.slice_type))
 709         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 710
 711     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 712
 713     /* ignore for SP/SI */
 714
 715     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 716         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 717
 718         if (slice_param.disable_deblocking_filter_idc != 1) {
 719             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 720             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 721         }
 722     }
 723
 724     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 725         bitstream_byte_aligning(bs, 1);
 726     }
 727 }
 728
 729 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 730 {
 731     bitstream bs;
 732
 733     bitstream_start(&bs);
 734     nal_start_code_prefix(&bs);
 735     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 736     pps_rbsp(&bs);
 737     bitstream_end(&bs);
 738
 739     *header_buffer = (unsigned char *)bs.buffer;
 740     return bs.bit_offset;
 741 }
 742
 743 int
 744 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 745 {
 746     bitstream bs;
 747
 748     bitstream_start(&bs);
 749     nal_start_code_prefix(&bs);
 750     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 751     sps_rbsp(&bs);
 752     bitstream_end(&bs);
 753
 754     *header_buffer = (unsigned char *)bs.buffer;
 755     return bs.bit_offset;
 756 }
 757
 758 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 759 {
 760     bitstream bs;
 761     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 762     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 763
 764     bitstream_start(&bs);
 765     nal_start_code_prefix(&bs);
 766
 767     if (IS_I_SLICE(slice_param.slice_type)) {
 768         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 769     } else if (IS_P_SLICE(slice_param.slice_type)) {
 770         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 771     } else {
 772         assert(IS_B_SLICE(slice_param.slice_type));
 773         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 774     }
 775
 776     slice_header(&bs);
 777     bitstream_end(&bs);
 778
 779     *header_buffer = (unsigned char *)bs.buffer;
 780     return bs.bit_offset;
 781 }
 782
 783
 784 /*
 785   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 786   1) period between Frame #X and Frame #N = #X - #N
 787   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 788   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 789   4) intra_period and intra_idr_period take precedence over ip_period
 790   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 791      of I/IDR frames, see bellow examples
 792   -------------------------------------------------------------------
 793   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 794   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 795   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 796   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 797   1            1                ignored    IDR IDR IDR IDR...
 798   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 799   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 800   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 801                                               (PBB)(IBB)(PBB)(IBB)...
 802   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 803                                            IDRPPPPP IPPPPP IPPPPP...
 804   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 805                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 806                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 807                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 808                                            {IDR(PBB)(PBB)}                     (6/6/3)
 809                                            {IDR(PBB)(PBB)}.
 810 */
 811
 812 // General pts/dts strategy:
 813 //
 814 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 815 // bit tricky. We assume first of all that the frame rate never goes _above_
 816 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 817 // in at least this speed, as long at dts <= pts (the frame is not attempted
 818 // presented before it is decoded). Furthermore, we never have longer chains of
 819 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 820 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 821 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 822 // frame, with an order O <= C.)
 823 //
 824 // Many strategies are possible, but we establish these rules:
 825 //
 826 //  - Tip frames have dts = pts - (C-O)*N.
 827 //  - Non-tip frames have dts = dts_last + N.
 828 //
 829 // An example, with C=2 and N=10 and the data flow showed with arrows:
 830 //
 831 //        I  B  P  B  B  P
 832 //   pts: 30 40 50 60 70 80
 833 //        ↓  ↓     ↓
 834 //   dts: 10 30 20 60 50←40
 835 //         |  |  ↑        ↑
 836 //         `--|--'        |
 837 //             `----------'
 838 //
 839 // To show that this works fine also with irregular spacings, let's say that
 840 // the third frame is delayed a bit (something earlier was dropped). Now the
 841 // situation looks like this:
 842 //
 843 //        I  B  P  B  B   P
 844 //   pts: 30 40 80 90 100 110
 845 //        ↓  ↓     ↓
 846 //   dts: 10 30 20 90 50←40
 847 //         |  |  ↑        ↑
 848 //         `--|--'        |
 849 //             `----------'
 850 //
 851 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 852 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 853 //
 854 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 855 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 856 // a tip frame and should be given a dts based on the previous one.
 857 #define FRAME_P 0
 858 #define FRAME_B 1
 859 #define FRAME_I 2
 860 #define FRAME_IDR 7
 861 void encoding2display_order(
 862     int encoding_order, int intra_period,
 863     int intra_idr_period, int ip_period,
 864     int *displaying_order,
 865     int *frame_type, int *pts_lag)
 866 {
 867     int encoding_order_gop = 0;
 868
 869     *pts_lag = 0;
 870
 871     if (intra_period == 1) { /* all are I/IDR frames */
 872         *displaying_order = encoding_order;
 873         if (intra_idr_period == 0)
 874             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 875         else
 876             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 877         return;
 878     }
 879
 880     if (intra_period == 0)
 881         intra_idr_period = 0;
 882
 883     if (ip_period == 1) {
 884         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 885         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 886         *displaying_order = encoding_order;
 887
 888         if (encoding_order_gop == 0) { /* the first frame */
 889             *frame_type = FRAME_IDR;
 890         } else if (intra_period != 0 && /* have I frames */
 891                    encoding_order_gop >= 2 &&
 892                    (encoding_order_gop % intra_period == 0)) {
 893             *frame_type = FRAME_I;
 894         } else {
 895             *frame_type = FRAME_P;
 896         }
 897         return;
 898     }
 899
 900     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 901     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 902     *pts_lag = -1;  // Most frames are not tip frames.
 903
 904     if (encoding_order_gop == 0) { /* the first frame */
 905         *frame_type = FRAME_IDR;
 906         *displaying_order = encoding_order;
 907         // IDR frames are a special case; I honestly can't find the logic behind
 908         // why this is the right thing, but it seems to line up nicely in practice :-)
 909         *pts_lag = TIMEBASE / MAX_FPS;
 910     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 911         *frame_type = FRAME_B;
 912         *displaying_order = encoding_order - 1;
 913         if ((encoding_order_gop % ip_period) == 0) {
 914             *pts_lag = 0;  // Last B-frame.
 915         }
 916     } else if (intra_period != 0 && /* have I frames */
 917                encoding_order_gop >= 2 &&
 918                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 919         *frame_type = FRAME_I;
 920         *displaying_order = encoding_order + ip_period - 1;
 921     } else {
 922         *frame_type = FRAME_P;
 923         *displaying_order = encoding_order + ip_period - 1;
 924     }
 925 }
 926
 927
 928 static const char *rc_to_string(int rc_mode)
 929 {
 930     switch (rc_mode) {
 931     case VA_RC_NONE:
 932         return "NONE";
 933     case VA_RC_CBR:
 934         return "CBR";
 935     case VA_RC_VBR:
 936         return "VBR";
 937     case VA_RC_VCM:
 938         return "VCM";
 939     case VA_RC_CQP:
 940         return "CQP";
 941     case VA_RC_VBR_CONSTRAINED:
 942         return "VBR_CONSTRAINED";
 943     default:
 944         return "Unknown";
 945     }
 946 }
 947
 948 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 949 {
 950         if (global_flags.uncompressed_video_to_http) {
 951                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 952                 use_zerocopy = false;
 953         } else if (global_flags.x264_video_to_http) {
 954                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 955                 use_zerocopy = false;
 956         } else {
 957                 use_zerocopy = true;
 958         }
 959 }
 960
 961 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 962 {
 963         if (va_display.empty()) {
 964                 x11_display = XOpenDisplay(NULL);
 965                 if (!x11_display) {
 966                         fprintf(stderr, "error: can't connect to X server!\n");
 967                         return NULL;
 968                 }
 969                 enable_zerocopy_if_possible();
 970                 return vaGetDisplay(x11_display);
 971         } else if (va_display[0] != '/') {
 972                 x11_display = XOpenDisplay(va_display.c_str());
 973                 if (!x11_display) {
 974                         fprintf(stderr, "error: can't connect to X server!\n");
 975                         return NULL;
 976                 }
 977                 enable_zerocopy_if_possible();
 978                 return vaGetDisplay(x11_display);
 979         } else {
 980                 drm_fd = open(va_display.c_str(), O_RDWR);
 981                 if (drm_fd == -1) {
 982                         perror(va_display.c_str());
 983                         return NULL;
 984                 }
 985                 use_zerocopy = false;
 986                 return vaGetDisplayDRM(drm_fd);
 987         }
 988 }
 989
 990 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
 991 {
 992         if (x11_display) {
 993                 XCloseDisplay(x11_display);
 994                 x11_display = nullptr;
 995         }
 996         if (drm_fd != -1) {
 997                 close(drm_fd);
 998         }
 999 }
1000
1001 int QuickSyncEncoderImpl::init_va(const string &va_display)
1002 {
1003     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1004     VAEntrypoint *entrypoints;
1005     int num_entrypoints, slice_entrypoint;
1006     int support_encode = 0;
1007     int major_ver, minor_ver;
1008     VAStatus va_status;
1009     unsigned int i;
1010
1011     va_dpy = va_open_display(va_display);
1012     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1013     CHECK_VASTATUS(va_status, "vaInitialize");
1014
1015     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1016     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1017     if (!entrypoints) {
1018         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1019         exit(1);
1020     }
1021
1022     /* use the highest profile */
1023     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1024         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1025             continue;
1026
1027         h264_profile = profile_list[i];
1028         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1029         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1030             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1031                 support_encode = 1;
1032                 break;
1033             }
1034         }
1035         if (support_encode == 1)
1036             break;
1037     }
1038
1039     if (support_encode == 0) {
1040         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1041         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1042         printf("to use VA-API against DRM instead of X11.\n");
1043         exit(1);
1044     } else {
1045         switch (h264_profile) {
1046             case VAProfileH264Baseline:
1047                 ip_period = 1;
1048                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1049                 h264_entropy_mode = 0;
1050                 break;
1051             case VAProfileH264ConstrainedBaseline:
1052                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1053                 ip_period = 1;
1054                 break;
1055
1056             case VAProfileH264Main:
1057                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1058                 break;
1059
1060             case VAProfileH264High:
1061                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1062                 break;
1063             default:
1064                 h264_profile = VAProfileH264Baseline;
1065                 ip_period = 1;
1066                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1067                 break;
1068         }
1069     }
1070
1071     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1072
1073     /* find out the format for the render target, and rate control mode */
1074     for (i = 0; i < VAConfigAttribTypeMax; i++)
1075         attrib[i].type = (VAConfigAttribType)i;
1076
1077     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1078                                       &attrib[0], VAConfigAttribTypeMax);
1079     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1080     /* check the interested configattrib */
1081     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1082         printf("Not find desired YUV420 RT format\n");
1083         exit(1);
1084     } else {
1085         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1086         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1087         config_attrib_num++;
1088     }
1089
1090     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1091         int tmp = attrib[VAConfigAttribRateControl].value;
1092
1093         if (rc_mode == -1 || !(rc_mode & tmp))  {
1094             if (rc_mode != -1) {
1095                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1096             }
1097
1098             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1099                 if (rc_default_modes[i] & tmp) {
1100                     rc_mode = rc_default_modes[i];
1101                     break;
1102                 }
1103             }
1104         }
1105
1106         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1107         config_attrib[config_attrib_num].value = rc_mode;
1108         config_attrib_num++;
1109     }
1110
1111
1112     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1113         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1114
1115         h264_packedheader = 1;
1116         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1117         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1118
1119         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1120             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1121         }
1122
1123         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1124             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1125         }
1126
1127         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1128             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1129         }
1130
1131         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1132             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1133         }
1134
1135         enc_packed_header_idx = config_attrib_num;
1136         config_attrib_num++;
1137     }
1138
1139     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1140         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1141         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1142         config_attrib_num++;
1143     }
1144
1145     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1146         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1147     }
1148
1149     free(entrypoints);
1150     return 0;
1151 }
1152
1153 int QuickSyncEncoderImpl::setup_encode()
1154 {
1155     VAStatus va_status;
1156     VASurfaceID *tmp_surfaceid;
1157     int codedbuf_size, i;
1158     VASurfaceID src_surface[SURFACE_NUM];
1159     VASurfaceID ref_surface[SURFACE_NUM];
1160
1161     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1162             &config_attrib[0], config_attrib_num, &config_id);
1163     CHECK_VASTATUS(va_status, "vaCreateConfig");
1164
1165     /* create source surfaces */
1166     va_status = vaCreateSurfaces(va_dpy,
1167                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1168                                  &src_surface[0], SURFACE_NUM,
1169                                  NULL, 0);
1170     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1171
1172     /* create reference surfaces */
1173     va_status = vaCreateSurfaces(va_dpy,
1174                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1175                                  &ref_surface[0], SURFACE_NUM,
1176                                  NULL, 0);
1177     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1178
1179     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1180     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1181     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1182
1183     /* Create a context for this encode pipe */
1184     va_status = vaCreateContext(va_dpy, config_id,
1185                                 frame_width_mbaligned, frame_height_mbaligned,
1186                                 VA_PROGRESSIVE,
1187                                 tmp_surfaceid, 2 * SURFACE_NUM,
1188                                 &context_id);
1189     CHECK_VASTATUS(va_status, "vaCreateContext");
1190     free(tmp_surfaceid);
1191
1192     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1193
1194     for (i = 0; i < SURFACE_NUM; i++) {
1195         /* create coded buffer once for all
1196          * other VA buffers which won't be used again after vaRenderPicture.
1197          * so APP can always vaCreateBuffer for every frame
1198          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1199          * so VA won't maintain the coded buffer
1200          */
1201         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1202                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1203         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1204     }
1205
1206     /* create OpenGL objects */
1207     //glGenFramebuffers(SURFACE_NUM, fbos);
1208
1209     for (i = 0; i < SURFACE_NUM; i++) {
1210         if (use_zerocopy) {
1211             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
1212             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
1213         } else {
1214             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
1215             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
1216
1217             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1218             // buffers, due to potentially differing pitch.
1219             glGenBuffers(1, &gl_surfaces[i].pbo);
1220             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1221             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1222             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1223             gl_surfaces[i].y_offset = 0;
1224             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1225             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1226             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1227             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1228         }
1229     }
1230
1231     for (i = 0; i < SURFACE_NUM; i++) {
1232         gl_surfaces[i].src_surface = src_surface[i];
1233         gl_surfaces[i].ref_surface = ref_surface[i];
1234     }
1235
1236     return 0;
1237 }
1238
1239 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1240 //
1241 //   2 1 0 [3] 4 8 9
1242 template<class T, class C>
1243 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1244 {
1245         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1246         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1247         sort(middle, end, less_than);
1248 }
1249
1250 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1251 {
1252     int i;
1253
1254     if (frame_type == FRAME_B)
1255         return;
1256
1257     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1258     numShortTerm++;
1259     if (numShortTerm > num_ref_frames)
1260         numShortTerm = num_ref_frames;
1261     for (i=numShortTerm-1; i>0; i--)
1262         ReferenceFrames[i] = ReferenceFrames[i-1];
1263     ReferenceFrames[0] = CurrentCurrPic;
1264
1265     current_frame_num++;
1266     if (current_frame_num > MaxFrameNum)
1267         current_frame_num = 0;
1268 }
1269
1270
1271 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1272 {
1273     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1274         return a.frame_idx > b.frame_idx;
1275     };
1276     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1277         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1278     };
1279     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1280         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1281     };
1282
1283     if (frame_type == FRAME_P) {
1284         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1285         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1286     } else if (frame_type == FRAME_B) {
1287         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1288         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1289
1290         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1291         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1292     }
1293
1294     return 0;
1295 }
1296
1297
1298 int QuickSyncEncoderImpl::render_sequence()
1299 {
1300     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1301     VAStatus va_status;
1302     VAEncMiscParameterBuffer *misc_param;
1303     VAEncMiscParameterRateControl *misc_rate_ctrl;
1304
1305     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1306     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1307     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1308     seq_param.bits_per_second = frame_bitrate;
1309
1310     seq_param.intra_period = intra_period;
1311     seq_param.intra_idr_period = intra_idr_period;
1312     seq_param.ip_period = ip_period;
1313
1314     seq_param.max_num_ref_frames = num_ref_frames;
1315     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1316     seq_param.time_scale = TIMEBASE * 2;
1317     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1318     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1319     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1320     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1321     seq_param.seq_fields.bits.chroma_format_idc = 1;
1322     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1323
1324     if (frame_width != frame_width_mbaligned ||
1325         frame_height != frame_height_mbaligned) {
1326         seq_param.frame_cropping_flag = 1;
1327         seq_param.frame_crop_left_offset = 0;
1328         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1329         seq_param.frame_crop_top_offset = 0;
1330         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1331     }
1332
1333     va_status = vaCreateBuffer(va_dpy, context_id,
1334                                VAEncSequenceParameterBufferType,
1335                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1336     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1337
1338     va_status = vaCreateBuffer(va_dpy, context_id,
1339                                VAEncMiscParameterBufferType,
1340                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1341                                1, NULL, &rc_param_buf);
1342     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1343
1344     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1345     misc_param->type = VAEncMiscParameterTypeRateControl;
1346     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1347     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1348     misc_rate_ctrl->bits_per_second = frame_bitrate;
1349     misc_rate_ctrl->target_percentage = 66;
1350     misc_rate_ctrl->window_size = 1000;
1351     misc_rate_ctrl->initial_qp = initial_qp;
1352     misc_rate_ctrl->min_qp = minimal_qp;
1353     misc_rate_ctrl->basic_unit_size = 0;
1354     vaUnmapBuffer(va_dpy, rc_param_buf);
1355
1356     render_id[0] = seq_param_buf;
1357     render_id[1] = rc_param_buf;
1358
1359     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1360
1361     return 0;
1362 }
1363
1364 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1365 {
1366     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1367     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1368     int PicOrderCntMsb, TopFieldOrderCnt;
1369
1370     if (frame_type == FRAME_IDR)
1371         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1372     else {
1373         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1374         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1375     }
1376
1377     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1378         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1379         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1380     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1381              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1382         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1383     else
1384         PicOrderCntMsb = prevPicOrderCntMsb;
1385
1386     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1387
1388     if (frame_type != FRAME_B) {
1389         PicOrderCntMsb_ref = PicOrderCntMsb;
1390         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1391     }
1392
1393     return TopFieldOrderCnt;
1394 }
1395
1396 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1397 {
1398     VABufferID pic_param_buf;
1399     VAStatus va_status;
1400     int i = 0;
1401
1402     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1403     pic_param.CurrPic.frame_idx = current_frame_num;
1404     pic_param.CurrPic.flags = 0;
1405     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1406     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1407     CurrentCurrPic = pic_param.CurrPic;
1408
1409     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1410     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1411         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1412         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1413     }
1414
1415     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1416     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1417     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1418     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1419     pic_param.frame_num = current_frame_num;
1420     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1421     pic_param.last_picture = false;  // FIXME
1422     pic_param.pic_init_qp = initial_qp;
1423
1424     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1425                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1426     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1427
1428     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1429
1430     return 0;
1431 }
1432
1433 int QuickSyncEncoderImpl::render_packedsequence()
1434 {
1435     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1436     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1437     unsigned int length_in_bits;
1438     unsigned char *packedseq_buffer = NULL;
1439     VAStatus va_status;
1440
1441     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1442
1443     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1444
1445     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1446     packedheader_param_buffer.has_emulation_bytes = 0;
1447     va_status = vaCreateBuffer(va_dpy,
1448                                context_id,
1449                                VAEncPackedHeaderParameterBufferType,
1450                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1451                                &packedseq_para_bufid);
1452     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1453
1454     va_status = vaCreateBuffer(va_dpy,
1455                                context_id,
1456                                VAEncPackedHeaderDataBufferType,
1457                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1458                                &packedseq_data_bufid);
1459     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1460
1461     render_id[0] = packedseq_para_bufid;
1462     render_id[1] = packedseq_data_bufid;
1463     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1464
1465     free(packedseq_buffer);
1466
1467     return 0;
1468 }
1469
1470
1471 int QuickSyncEncoderImpl::render_packedpicture()
1472 {
1473     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1474     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1475     unsigned int length_in_bits;
1476     unsigned char *packedpic_buffer = NULL;
1477     VAStatus va_status;
1478
1479     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1480     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1481     packedheader_param_buffer.bit_length = length_in_bits;
1482     packedheader_param_buffer.has_emulation_bytes = 0;
1483
1484     va_status = vaCreateBuffer(va_dpy,
1485                                context_id,
1486                                VAEncPackedHeaderParameterBufferType,
1487                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1488                                &packedpic_para_bufid);
1489     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1490
1491     va_status = vaCreateBuffer(va_dpy,
1492                                context_id,
1493                                VAEncPackedHeaderDataBufferType,
1494                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1495                                &packedpic_data_bufid);
1496     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1497
1498     render_id[0] = packedpic_para_bufid;
1499     render_id[1] = packedpic_data_bufid;
1500     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1501
1502     free(packedpic_buffer);
1503
1504     return 0;
1505 }
1506
1507 void QuickSyncEncoderImpl::render_packedslice()
1508 {
1509     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1510     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1511     unsigned int length_in_bits;
1512     unsigned char *packedslice_buffer = NULL;
1513     VAStatus va_status;
1514
1515     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1516     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1517     packedheader_param_buffer.bit_length = length_in_bits;
1518     packedheader_param_buffer.has_emulation_bytes = 0;
1519
1520     va_status = vaCreateBuffer(va_dpy,
1521                                context_id,
1522                                VAEncPackedHeaderParameterBufferType,
1523                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1524                                &packedslice_para_bufid);
1525     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1526
1527     va_status = vaCreateBuffer(va_dpy,
1528                                context_id,
1529                                VAEncPackedHeaderDataBufferType,
1530                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1531                                &packedslice_data_bufid);
1532     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1533
1534     render_id[0] = packedslice_para_bufid;
1535     render_id[1] = packedslice_data_bufid;
1536     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1537
1538     free(packedslice_buffer);
1539 }
1540
1541 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1542 {
1543     VABufferID slice_param_buf;
1544     VAStatus va_status;
1545     int i;
1546
1547     update_RefPicList(frame_type);
1548
1549     /* one frame, one slice */
1550     slice_param.macroblock_address = 0;
1551     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1552     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1553     if (frame_type == FRAME_IDR) {
1554         if (encoding_frame_num != 0)
1555             ++slice_param.idr_pic_id;
1556     } else if (frame_type == FRAME_P) {
1557         int refpiclist0_max = h264_maxref & 0xffff;
1558         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1559
1560         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1561             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1562             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1563         }
1564     } else if (frame_type == FRAME_B) {
1565         int refpiclist0_max = h264_maxref & 0xffff;
1566         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1567
1568         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1569         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1570             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1571             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1572         }
1573
1574         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1575         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1576             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1577             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1578         }
1579     }
1580
1581     slice_param.slice_alpha_c0_offset_div2 = 0;
1582     slice_param.slice_beta_offset_div2 = 0;
1583     slice_param.direct_spatial_mv_pred_flag = 1;
1584     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1585
1586
1587     if (h264_packedheader &&
1588         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1589         render_packedslice();
1590
1591     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1592                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1593     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1594
1595     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1596
1597     return 0;
1598 }
1599
1600
1601
1602 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1603 {
1604         VACodedBufferSegment *buf_list = NULL;
1605         VAStatus va_status;
1606
1607         string data;
1608
1609         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1610         CHECK_VASTATUS(va_status, "vaMapBuffer");
1611         while (buf_list != NULL) {
1612                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1613                 buf_list = (VACodedBufferSegment *) buf_list->next;
1614         }
1615         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1616
1617         {
1618                 // Add video.
1619                 AVPacket pkt;
1620                 memset(&pkt, 0, sizeof(pkt));
1621                 pkt.buf = nullptr;
1622                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1623                 pkt.size = data.size();
1624                 pkt.stream_index = 0;
1625                 if (task.frame_type == FRAME_IDR) {
1626                         pkt.flags = AV_PKT_FLAG_KEY;
1627                 } else {
1628                         pkt.flags = 0;
1629                 }
1630                 pkt.duration = task.duration;
1631                 if (file_mux) {
1632                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1633                 }
1634                 if (!global_flags.uncompressed_video_to_http &&
1635                     !global_flags.x264_video_to_http) {
1636                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1637                 }
1638         }
1639         // Encode and add all audio frames up to and including the pts of this video frame.
1640         for ( ;; ) {
1641                 int64_t audio_pts;
1642                 vector<float> audio;
1643                 {
1644                         unique_lock<mutex> lock(frame_queue_mutex);
1645                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1646                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1647                         auto it = pending_audio_frames.begin();
1648                         if (it->first > task.pts) break;
1649                         audio_pts = it->first;
1650                         audio = move(it->second);
1651                         pending_audio_frames.erase(it);
1652                 }
1653
1654                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1655                 stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1656
1657                 if (audio_pts == task.pts) break;
1658         }
1659 }
1660
1661
1662 // this is weird. but it seems to put a new frame onto the queue
1663 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1664 {
1665         unique_lock<mutex> lock(storage_task_queue_mutex);
1666         storage_task_queue.push(move(task));
1667         storage_task_queue_changed.notify_all();
1668 }
1669
1670 void QuickSyncEncoderImpl::storage_task_thread()
1671 {
1672         for ( ;; ) {
1673                 storage_task current;
1674                 {
1675                         // wait until there's an encoded frame
1676                         unique_lock<mutex> lock(storage_task_queue_mutex);
1677                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1678                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1679                         current = move(storage_task_queue.front());
1680                         storage_task_queue.pop();
1681                 }
1682
1683                 VAStatus va_status;
1684
1685                 // waits for data, then saves it to disk.
1686                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1687                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1688                 save_codeddata(move(current));
1689
1690                 {
1691                         unique_lock<mutex> lock(storage_task_queue_mutex);
1692                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1693                         storage_task_queue_changed.notify_all();
1694                 }
1695         }
1696 }
1697
1698 int QuickSyncEncoderImpl::release_encode()
1699 {
1700         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1701                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1702                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1703                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1704
1705                 if (!use_zerocopy) {
1706                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1707                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1708                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1709                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1710                 }
1711                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1712                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1713         }
1714
1715         vaDestroyContext(va_dpy, context_id);
1716         vaDestroyConfig(va_dpy, config_id);
1717
1718         return 0;
1719 }
1720
1721 int QuickSyncEncoderImpl::deinit_va()
1722 {
1723     vaTerminate(va_dpy);
1724
1725     va_close_display(va_dpy);
1726
1727     return 0;
1728 }
1729
1730 namespace {
1731
1732 }  // namespace
1733
1734 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder)
1735         : current_storage_frame(0), resource_pool(resource_pool), surface(surface), stream_audio_encoder(stream_audio_encoder), x264_encoder(x264_encoder), frame_width(width), frame_height(height)
1736 {
1737         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1738         open_output_file(filename);
1739         file_audio_encoder->add_mux(file_mux.get());
1740
1741         frame_width_mbaligned = (frame_width + 15) & (~15);
1742         frame_height_mbaligned = (frame_height + 15) & (~15);
1743
1744         //print_input();
1745
1746         if (global_flags.uncompressed_video_to_http ||
1747             global_flags.x264_video_to_http) {
1748                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1749         }
1750         if (global_flags.x264_video_to_http) {
1751                 assert(x264_encoder != nullptr);
1752         } else {
1753                 assert(x264_encoder == nullptr);
1754         }
1755
1756         init_va(va_display);
1757         setup_encode();
1758
1759         // No frames are ready yet.
1760         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1761
1762         memset(&seq_param, 0, sizeof(seq_param));
1763         memset(&pic_param, 0, sizeof(pic_param));
1764         memset(&slice_param, 0, sizeof(slice_param));
1765
1766         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1767
1768         encode_thread = thread([this]{
1769                 //SDL_GL_MakeCurrent(window, context);
1770                 QOpenGLContext *context = create_context(this->surface);
1771                 eglBindAPI(EGL_OPENGL_API);
1772                 if (!make_current(context, this->surface)) {
1773                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1774                                 eglGetError());
1775                         exit(1);
1776                 }
1777                 encode_thread_func();
1778                 delete_context(context);
1779         });
1780 }
1781
1782 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1783 {
1784         shutdown();
1785 }
1786
1787 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1788 {
1789         assert(!is_shutdown);
1790         {
1791                 // Wait until this frame slot is done encoding.
1792                 unique_lock<mutex> lock(storage_task_queue_mutex);
1793                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1794                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1795                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1796                 }
1797                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1798                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1799                 if (storage_thread_should_quit) return false;
1800         }
1801
1802         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1803         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1804         *y_tex = surf->y_tex;
1805         *cbcr_tex = surf->cbcr_tex;
1806
1807         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1808         CHECK_VASTATUS(va_status, "vaDeriveImage");
1809
1810         if (use_zerocopy) {
1811                 VABufferInfo buf_info;
1812                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1813                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1814                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1815
1816                 // Create Y image.
1817                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1818                 EGLint y_attribs[] = {
1819                         EGL_WIDTH, frame_width,
1820                         EGL_HEIGHT, frame_height,
1821                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1822                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1823                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1824                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1825                         EGL_NONE
1826                 };
1827
1828                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1829                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1830
1831                 // Associate Y image to a texture.
1832                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1833                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1834
1835                 // Create CbCr image.
1836                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1837                 EGLint cbcr_attribs[] = {
1838                         EGL_WIDTH, frame_width,
1839                         EGL_HEIGHT, frame_height,
1840                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1841                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1842                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1843                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1844                         EGL_NONE
1845                 };
1846
1847                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1848                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1849
1850                 // Associate CbCr image to a texture.
1851                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1852                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1853         }
1854
1855         return true;
1856 }
1857
1858 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1859 {
1860         assert(!is_shutdown);
1861         {
1862                 unique_lock<mutex> lock(frame_queue_mutex);
1863                 pending_audio_frames[pts] = move(audio);
1864         }
1865         frame_queue_nonempty.notify_all();
1866 }
1867
1868 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1869 {
1870         assert(!is_shutdown);
1871
1872         if (!use_zerocopy) {
1873                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1874
1875                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1876                 check_error();
1877
1878                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1879                 check_error();
1880
1881                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1882                 check_error();
1883                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1884                 check_error();
1885
1886                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1887                 check_error();
1888                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1889                 check_error();
1890
1891                 glBindTexture(GL_TEXTURE_2D, 0);
1892                 check_error();
1893                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1894                 check_error();
1895
1896                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1897                 check_error();
1898         }
1899
1900         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1901         check_error();
1902         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1903         check_error();
1904
1905         {
1906                 unique_lock<mutex> lock(frame_queue_mutex);
1907                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1908                 ++current_storage_frame;
1909         }
1910         frame_queue_nonempty.notify_all();
1911         return fence;
1912 }
1913
1914 void QuickSyncEncoderImpl::shutdown()
1915 {
1916         if (is_shutdown) {
1917                 return;
1918         }
1919
1920         {
1921                 unique_lock<mutex> lock(frame_queue_mutex);
1922                 encode_thread_should_quit = true;
1923                 frame_queue_nonempty.notify_all();
1924         }
1925         encode_thread.join();
1926         {
1927                 unique_lock<mutex> lock(storage_task_queue_mutex);
1928                 storage_thread_should_quit = true;
1929                 frame_queue_nonempty.notify_all();
1930                 storage_task_queue_changed.notify_all();
1931         }
1932         storage_thread.join();
1933         encode_remaining_audio();
1934
1935         release_encode();
1936         deinit_va();
1937         file_mux.reset();
1938         is_shutdown = true;
1939 }
1940
1941 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1942 {
1943         AVFormatContext *avctx = avformat_alloc_context();
1944         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1945         assert(filename.size() < sizeof(avctx->filename) - 1);
1946         strcpy(avctx->filename, filename.c_str());
1947
1948         string url = "file:" + filename;
1949         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1950         if (ret < 0) {
1951                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1952                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1953                 exit(1);
1954         }
1955
1956         string video_extradata = "";  // FIXME: See other comment about global headers.
1957         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, file_audio_encoder->get_ctx(), TIMEBASE, nullptr));
1958 }
1959
1960 void QuickSyncEncoderImpl::encode_thread_func()
1961 {
1962         int64_t last_dts = -1;
1963         int gop_start_display_frame_num = 0;
1964         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1965                 PendingFrame frame;
1966                 int pts_lag;
1967                 int frame_type, display_frame_num;
1968                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1969                                        &display_frame_num, &frame_type, &pts_lag);
1970                 if (frame_type == FRAME_IDR) {
1971                         numShortTerm = 0;
1972                         current_frame_num = 0;
1973                         gop_start_display_frame_num = display_frame_num;
1974                 }
1975
1976                 {
1977                         unique_lock<mutex> lock(frame_queue_mutex);
1978                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1979                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1980                         });
1981                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1982                                 // We have queued frames that were supposed to be B-frames,
1983                                 // but will be no P-frame to encode them against. Encode them all
1984                                 // as P-frames instead. Note that this happens under the mutex,
1985                                 // but nobody else uses it at this point, since we're shutting down,
1986                                 // so there's no contention.
1987                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1988                                 return;
1989                         } else {
1990                                 frame = move(pending_video_frames[display_frame_num]);
1991                                 pending_video_frames.erase(display_frame_num);
1992                         }
1993                 }
1994
1995                 // Determine the dts of this frame.
1996                 int64_t dts;
1997                 if (pts_lag == -1) {
1998                         assert(last_dts != -1);
1999                         dts = last_dts + (TIMEBASE / MAX_FPS);
2000                 } else {
2001                         dts = frame.pts - pts_lag;
2002                 }
2003                 last_dts = dts;
2004
2005                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2006         }
2007 }
2008
2009 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2010 {
2011         if (pending_video_frames.empty()) {
2012                 return;
2013         }
2014
2015         for (auto &pending_frame : pending_video_frames) {
2016                 int display_frame_num = pending_frame.first;
2017                 assert(display_frame_num > 0);
2018                 PendingFrame frame = move(pending_frame.second);
2019                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2020                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2021                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2022                 last_dts = dts;
2023         }
2024
2025         if (global_flags.uncompressed_video_to_http ||
2026             global_flags.x264_video_to_http) {
2027                 // Add frames left in reorderer.
2028                 while (!reorderer->empty()) {
2029                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2030                         if (global_flags.uncompressed_video_to_http) {
2031                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2032                         } else {
2033                                 assert(global_flags.x264_video_to_http);
2034                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2035                         }
2036                 }
2037         }
2038 }
2039
2040 void QuickSyncEncoderImpl::encode_remaining_audio()
2041 {
2042         // This really ought to be empty by now, but just to be sure...
2043         for (auto &pending_frame : pending_audio_frames) {
2044                 int64_t audio_pts = pending_frame.first;
2045                 vector<float> audio = move(pending_frame.second);
2046
2047                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2048                 if (stream_audio_encoder) {
2049                         stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2050                 }
2051         }
2052         pending_audio_frames.clear();
2053
2054         // Encode any leftover audio in the queues, and also any delayed frames.
2055         // Note: stream_audio_encoder is not owned by us, so don't call encode_last_audio().
2056         file_audio_encoder->encode_last_audio();
2057 }
2058
2059 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2060 {
2061         AVPacket pkt;
2062         memset(&pkt, 0, sizeof(pkt));
2063         pkt.buf = nullptr;
2064         pkt.data = const_cast<uint8_t *>(data);
2065         pkt.size = frame_width * frame_height * 2;
2066         pkt.stream_index = 0;
2067         pkt.flags = AV_PKT_FLAG_KEY;
2068         pkt.duration = duration;
2069         stream_mux->add_packet(pkt, pts, pts);
2070 }
2071
2072 namespace {
2073
2074 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2075 {
2076         if (src_width == dst_pitch) {
2077                 memcpy(dst, src, src_width * height);
2078         } else {
2079                 for (size_t y = 0; y < height; ++y) {
2080                         const uint8_t *sptr = src + y * src_width;
2081                         uint8_t *dptr = dst + y * dst_pitch;
2082                         memcpy(dptr, sptr, src_width);
2083                 }
2084         }
2085 }
2086
2087 }  // namespace
2088
2089 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2090                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2091 {
2092         // Wait for the GPU to be done with the frame.
2093         GLenum sync_status;
2094         do {
2095                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2096                 check_error();
2097         } while (sync_status == GL_TIMEOUT_EXPIRED);
2098         assert(sync_status != GL_WAIT_FAILED);
2099
2100         // Release back any input frames we needed to render this frame.
2101         frame.input_frames.clear();
2102
2103         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2104         VAStatus va_status;
2105
2106         if (use_zerocopy) {
2107                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2108                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2109                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2110                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2111         } else {
2112                 unsigned char *surface_p = nullptr;
2113                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2114
2115                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2116                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2117
2118                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2119                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2120
2121                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2122                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2123
2124                 if (global_flags.uncompressed_video_to_http ||
2125                     global_flags.x264_video_to_http) {
2126                         // Add uncompressed video. (Note that pts == dts here.)
2127                         // Delay needs to match audio.
2128                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2129                         if (output_frame.data != nullptr) {
2130                                 if (global_flags.uncompressed_video_to_http) {
2131                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2132                                 } else {
2133                                         assert(global_flags.x264_video_to_http);
2134                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2135                                 }
2136                         }
2137                 }
2138         }
2139
2140         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2141         CHECK_VASTATUS(va_status, "vaDestroyImage");
2142
2143         // Schedule the frame for encoding.
2144         VASurfaceID va_surface = surf->src_surface;
2145         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2146         CHECK_VASTATUS(va_status, "vaBeginPicture");
2147
2148         if (frame_type == FRAME_IDR) {
2149                 // FIXME: If the mux wants global headers, we should not put the
2150                 // SPS/PPS before each IDR frame, but rather put it into the
2151                 // codec extradata (formatted differently?).
2152                 render_sequence();
2153                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2154                 if (h264_packedheader) {
2155                         render_packedsequence();
2156                         render_packedpicture();
2157                 }
2158         } else {
2159                 //render_sequence();
2160                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2161         }
2162         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2163
2164         va_status = vaEndPicture(va_dpy, context_id);
2165         CHECK_VASTATUS(va_status, "vaEndPicture");
2166
2167         // so now the data is done encoding (well, async job kicked off)...
2168         // we send that to the storage thread
2169         storage_task tmp;
2170         tmp.display_order = display_frame_num;
2171         tmp.frame_type = frame_type;
2172         tmp.pts = pts;
2173         tmp.dts = dts;
2174         tmp.duration = duration;
2175         storage_task_enqueue(move(tmp));
2176
2177         update_ReferenceFrames(frame_type);
2178 }
2179
2180 // Proxy object.
2181 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder)
2182         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, stream_audio_encoder, x264_encoder)) {}
2183
2184 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2185 QuickSyncEncoder::~QuickSyncEncoder() {}
2186
2187 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2188 {
2189         impl->add_audio(pts, audio);
2190 }
2191
2192 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2193 {
2194         return impl->begin_frame(y_tex, cbcr_tex);
2195 }
2196
2197 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2198 {
2199         return impl->end_frame(pts, duration, input_frames);
2200 }
2201
2202 void QuickSyncEncoder::shutdown()
2203 {
2204         impl->shutdown();
2205 }
2206
2207 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2208 {
2209         impl->set_stream_mux(mux);
2210 }
2211