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 release_gl_resources();
 204         void set_stream_mux(Mux *mux)
 205         {
 206                 stream_mux = mux;
 207         }
 208
 209 private:
 210         struct storage_task {
 211                 unsigned long long display_order;
 212                 int frame_type;
 213                 vector<float> audio;
 214                 int64_t pts, dts, duration;
 215         };
 216         struct PendingFrame {
 217                 RefCountedGLsync fence;
 218                 vector<RefCountedFrame> input_frames;
 219                 int64_t pts, duration;
 220         };
 221
 222         // So we never get negative dts.
 223         int64_t global_delay() const {
 224                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 225         }
 226
 227         void open_output_file(const std::string &filename);
 228         void encode_thread_func();
 229         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 230         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 231         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 232                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 233         void storage_task_thread();
 234         void encode_remaining_audio();
 235         void storage_task_enqueue(storage_task task);
 236         void save_codeddata(storage_task task);
 237         int render_packedsequence();
 238         int render_packedpicture();
 239         void render_packedslice();
 240         int render_sequence();
 241         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 242         void sps_rbsp(bitstream *bs);
 243         void pps_rbsp(bitstream *bs);
 244         int build_packed_pic_buffer(unsigned char **header_buffer);
 245         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 246         void slice_header(bitstream *bs);
 247         int build_packed_seq_buffer(unsigned char **header_buffer);
 248         int build_packed_slice_buffer(unsigned char **header_buffer);
 249         int init_va(const string &va_display);
 250         int deinit_va();
 251         void enable_zerocopy_if_possible();
 252         VADisplay va_open_display(const string &va_display);
 253         void va_close_display(VADisplay va_dpy);
 254         int setup_encode();
 255         void release_encode();
 256         void update_ReferenceFrames(int frame_type);
 257         int update_RefPicList(int frame_type);
 258
 259         bool is_shutdown = false;
 260         bool has_released_gl_resources = false;
 261         bool use_zerocopy;
 262         int drm_fd = -1;
 263
 264         thread encode_thread, storage_thread;
 265
 266         mutex storage_task_queue_mutex;
 267         condition_variable storage_task_queue_changed;
 268         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 269         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 270         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 271
 272         mutex frame_queue_mutex;
 273         condition_variable frame_queue_nonempty;
 274         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 275
 276         int current_storage_frame;
 277
 278         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 279         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 280         movit::ResourcePool *resource_pool;
 281         QSurface *surface;
 282
 283         unique_ptr<AudioEncoder> file_audio_encoder;
 284         AudioEncoder *stream_audio_encoder;
 285
 286         unique_ptr<FrameReorderer> reorderer;
 287         X264Encoder *x264_encoder;  // nullptr if not using x264.
 288
 289         Mux* stream_mux = nullptr;  // To HTTP.
 290         unique_ptr<Mux> file_mux;  // To local disk.
 291
 292         Display *x11_display = nullptr;
 293
 294         // Encoder parameters
 295         VADisplay va_dpy;
 296         VAProfile h264_profile = (VAProfile)~0;
 297         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 298         int config_attrib_num = 0, enc_packed_header_idx;
 299
 300         struct GLSurface {
 301                 VASurfaceID src_surface, ref_surface;
 302                 VABufferID coded_buf;
 303
 304                 VAImage surface_image;
 305                 GLuint y_tex, cbcr_tex;
 306
 307                 // Only if use_zerocopy == true.
 308                 EGLImage y_egl_image, cbcr_egl_image;
 309
 310                 // Only if use_zerocopy == false.
 311                 GLuint pbo;
 312                 uint8_t *y_ptr, *cbcr_ptr;
 313                 size_t y_offset, cbcr_offset;
 314         };
 315         GLSurface gl_surfaces[SURFACE_NUM];
 316
 317         VAConfigID config_id;
 318         VAContextID context_id;
 319         VAEncSequenceParameterBufferH264 seq_param;
 320         VAEncPictureParameterBufferH264 pic_param;
 321         VAEncSliceParameterBufferH264 slice_param;
 322         VAPictureH264 CurrentCurrPic;
 323         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 324
 325         // Static quality settings.
 326         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 327         static constexpr unsigned int num_ref_frames = 2;
 328         static constexpr int initial_qp = 15;
 329         static constexpr int minimal_qp = 0;
 330         static constexpr int intra_period = 30;
 331         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 332
 333         // Quality settings that are meant to be static, but might be overridden
 334         // by the profile.
 335         int constraint_set_flag = 0;
 336         int h264_packedheader = 0; /* support pack header? */
 337         int h264_maxref = (1<<16|1);
 338         int h264_entropy_mode = 1; /* cabac */
 339         int ip_period = 3;
 340
 341         int rc_mode = -1;
 342         unsigned int current_frame_num = 0;
 343         unsigned int numShortTerm = 0;
 344
 345         int frame_width;
 346         int frame_height;
 347         int frame_width_mbaligned;
 348         int frame_height_mbaligned;
 349 };
 350
 351 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 352 // but if we don't delete it here, we get leaks. The GStreamer implementation
 353 // does the same.
 354 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 355 {
 356     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 357     CHECK_VASTATUS(va_status, "vaRenderPicture");
 358
 359     for (int i = 0; i < num_buffers; ++i) {
 360         va_status = vaDestroyBuffer(dpy, buffers[i]);
 361         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 362     }
 363 }
 364
 365 static unsigned int
 366 va_swap32(unsigned int val)
 367 {
 368     unsigned char *pval = (unsigned char *)&val;
 369
 370     return ((pval[0] << 24)     |
 371             (pval[1] << 16)     |
 372             (pval[2] << 8)      |
 373             (pval[3] << 0));
 374 }
 375
 376 static void
 377 bitstream_start(bitstream *bs)
 378 {
 379     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 380     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 381     bs->bit_offset = 0;
 382 }
 383
 384 static void
 385 bitstream_end(bitstream *bs)
 386 {
 387     int pos = (bs->bit_offset >> 5);
 388     int bit_offset = (bs->bit_offset & 0x1f);
 389     int bit_left = 32 - bit_offset;
 390
 391     if (bit_offset) {
 392         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 393     }
 394 }
 395
 396 static void
 397 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 398 {
 399     int pos = (bs->bit_offset >> 5);
 400     int bit_offset = (bs->bit_offset & 0x1f);
 401     int bit_left = 32 - bit_offset;
 402
 403     if (!size_in_bits)
 404         return;
 405
 406     bs->bit_offset += size_in_bits;
 407
 408     if (bit_left > size_in_bits) {
 409         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 410     } else {
 411         size_in_bits -= bit_left;
 412         if (bit_left >= 32) {
 413             bs->buffer[pos] = (val >> size_in_bits);
 414         } else {
 415             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 416         }
 417         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 418
 419         if (pos + 1 == bs->max_size_in_dword) {
 420             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 421             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 422         }
 423
 424         bs->buffer[pos + 1] = val;
 425     }
 426 }
 427
 428 static void
 429 bitstream_put_ue(bitstream *bs, unsigned int val)
 430 {
 431     int size_in_bits = 0;
 432     int tmp_val = ++val;
 433
 434     while (tmp_val) {
 435         tmp_val >>= 1;
 436         size_in_bits++;
 437     }
 438
 439     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 440     bitstream_put_ui(bs, val, size_in_bits);
 441 }
 442
 443 static void
 444 bitstream_put_se(bitstream *bs, int val)
 445 {
 446     unsigned int new_val;
 447
 448     if (val <= 0)
 449         new_val = -2 * val;
 450     else
 451         new_val = 2 * val - 1;
 452
 453     bitstream_put_ue(bs, new_val);
 454 }
 455
 456 static void
 457 bitstream_byte_aligning(bitstream *bs, int bit)
 458 {
 459     int bit_offset = (bs->bit_offset & 0x7);
 460     int bit_left = 8 - bit_offset;
 461     int new_val;
 462
 463     if (!bit_offset)
 464         return;
 465
 466     assert(bit == 0 || bit == 1);
 467
 468     if (bit)
 469         new_val = (1 << bit_left) - 1;
 470     else
 471         new_val = 0;
 472
 473     bitstream_put_ui(bs, new_val, bit_left);
 474 }
 475
 476 static void
 477 rbsp_trailing_bits(bitstream *bs)
 478 {
 479     bitstream_put_ui(bs, 1, 1);
 480     bitstream_byte_aligning(bs, 0);
 481 }
 482
 483 static void nal_start_code_prefix(bitstream *bs)
 484 {
 485     bitstream_put_ui(bs, 0x00000001, 32);
 486 }
 487
 488 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 489 {
 490     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 491     bitstream_put_ui(bs, nal_ref_idc, 2);
 492     bitstream_put_ui(bs, nal_unit_type, 5);
 493 }
 494
 495 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 496 {
 497     int profile_idc = PROFILE_IDC_BASELINE;
 498
 499     if (h264_profile  == VAProfileH264High)
 500         profile_idc = PROFILE_IDC_HIGH;
 501     else if (h264_profile  == VAProfileH264Main)
 502         profile_idc = PROFILE_IDC_MAIN;
 503
 504     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 505     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 506     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 507     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 508     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 509     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 510     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 511     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 512
 513     if ( profile_idc == PROFILE_IDC_HIGH) {
 514         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 515         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 516         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 517         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 518         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 519     }
 520
 521     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 522     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 523
 524     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 525         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 526     else {
 527         assert(0);
 528     }
 529
 530     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 531     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 532
 533     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 534     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 535     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 536
 537     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 538         assert(0);
 539     }
 540
 541     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 542     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 543
 544     if (seq_param.frame_cropping_flag) {
 545         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 546         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 547         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 548         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 549     }
 550
 551     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 552     if ( false ) {
 553         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 554     } else {
 555         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 556         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 557         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 558         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 559         {
 560             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 561             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 562             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 563             {
 564                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 565                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 566                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 567             }
 568         }
 569         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 570         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 571         {
 572             bitstream_put_ui(bs, 1, 32);  // FPS
 573             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 574             bitstream_put_ui(bs, 1, 1);
 575         }
 576         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 577         {
 578             // hrd_parameters
 579             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 580             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 581             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 582
 583             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 584             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 585             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 586
 587             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 588             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 589             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 590             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 591         }
 592         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 593         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 594
 595         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 596         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 597     }
 598
 599     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 600 }
 601
 602
 603 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 604 {
 605     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 606     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 607
 608     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 609
 610     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 611
 612     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 613
 614     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 615     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 616
 617     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 618     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 619
 620     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 621     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 622     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 623
 624     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 625     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 626     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 627
 628     /* more_rbsp_data */
 629     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 630     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 631     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 632
 633     rbsp_trailing_bits(bs);
 634 }
 635
 636 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 637 {
 638     int first_mb_in_slice = slice_param.macroblock_address;
 639
 640     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 641     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 642     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 643     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 644
 645     /* frame_mbs_only_flag == 1 */
 646     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 647         /* FIXME: */
 648         assert(0);
 649     }
 650
 651     if (pic_param.pic_fields.bits.idr_pic_flag)
 652         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 653
 654     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 655         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 656         /* pic_order_present_flag == 0 */
 657     } else {
 658         /* FIXME: */
 659         assert(0);
 660     }
 661
 662     /* redundant_pic_cnt_present_flag == 0 */
 663     /* slice type */
 664     if (IS_P_SLICE(slice_param.slice_type)) {
 665         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 666
 667         if (slice_param.num_ref_idx_active_override_flag)
 668             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 669
 670         /* ref_pic_list_reordering */
 671         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 672     } else if (IS_B_SLICE(slice_param.slice_type)) {
 673         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 674
 675         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 676
 677         if (slice_param.num_ref_idx_active_override_flag) {
 678             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 679             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 680         }
 681
 682         /* ref_pic_list_reordering */
 683         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 684         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 685     }
 686
 687     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 688          IS_P_SLICE(slice_param.slice_type)) ||
 689         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 690          IS_B_SLICE(slice_param.slice_type))) {
 691         /* FIXME: fill weight/offset table */
 692         assert(0);
 693     }
 694
 695     /* dec_ref_pic_marking */
 696     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 697         unsigned char no_output_of_prior_pics_flag = 0;
 698         unsigned char long_term_reference_flag = 0;
 699         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 700
 701         if (pic_param.pic_fields.bits.idr_pic_flag) {
 702             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 703             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 704         } else {
 705             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 706         }
 707     }
 708
 709     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 710         !IS_I_SLICE(slice_param.slice_type))
 711         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 712
 713     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 714
 715     /* ignore for SP/SI */
 716
 717     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 718         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 719
 720         if (slice_param.disable_deblocking_filter_idc != 1) {
 721             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 722             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 723         }
 724     }
 725
 726     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 727         bitstream_byte_aligning(bs, 1);
 728     }
 729 }
 730
 731 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 732 {
 733     bitstream bs;
 734
 735     bitstream_start(&bs);
 736     nal_start_code_prefix(&bs);
 737     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 738     pps_rbsp(&bs);
 739     bitstream_end(&bs);
 740
 741     *header_buffer = (unsigned char *)bs.buffer;
 742     return bs.bit_offset;
 743 }
 744
 745 int
 746 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 747 {
 748     bitstream bs;
 749
 750     bitstream_start(&bs);
 751     nal_start_code_prefix(&bs);
 752     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 753     sps_rbsp(&bs);
 754     bitstream_end(&bs);
 755
 756     *header_buffer = (unsigned char *)bs.buffer;
 757     return bs.bit_offset;
 758 }
 759
 760 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 761 {
 762     bitstream bs;
 763     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 764     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 765
 766     bitstream_start(&bs);
 767     nal_start_code_prefix(&bs);
 768
 769     if (IS_I_SLICE(slice_param.slice_type)) {
 770         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 771     } else if (IS_P_SLICE(slice_param.slice_type)) {
 772         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 773     } else {
 774         assert(IS_B_SLICE(slice_param.slice_type));
 775         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 776     }
 777
 778     slice_header(&bs);
 779     bitstream_end(&bs);
 780
 781     *header_buffer = (unsigned char *)bs.buffer;
 782     return bs.bit_offset;
 783 }
 784
 785
 786 /*
 787   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 788   1) period between Frame #X and Frame #N = #X - #N
 789   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 790   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 791   4) intra_period and intra_idr_period take precedence over ip_period
 792   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 793      of I/IDR frames, see bellow examples
 794   -------------------------------------------------------------------
 795   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 796   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 797   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 798   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 799   1            1                ignored    IDR IDR IDR IDR...
 800   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 801   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 802   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 803                                               (PBB)(IBB)(PBB)(IBB)...
 804   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 805                                            IDRPPPPP IPPPPP IPPPPP...
 806   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 807                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 808                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 809                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 810                                            {IDR(PBB)(PBB)}                     (6/6/3)
 811                                            {IDR(PBB)(PBB)}.
 812 */
 813
 814 // General pts/dts strategy:
 815 //
 816 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 817 // bit tricky. We assume first of all that the frame rate never goes _above_
 818 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 819 // in at least this speed, as long at dts <= pts (the frame is not attempted
 820 // presented before it is decoded). Furthermore, we never have longer chains of
 821 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 822 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 823 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 824 // frame, with an order O <= C.)
 825 //
 826 // Many strategies are possible, but we establish these rules:
 827 //
 828 //  - Tip frames have dts = pts - (C-O)*N.
 829 //  - Non-tip frames have dts = dts_last + N.
 830 //
 831 // An example, with C=2 and N=10 and the data flow showed with arrows:
 832 //
 833 //        I  B  P  B  B  P
 834 //   pts: 30 40 50 60 70 80
 835 //        ↓  ↓     ↓
 836 //   dts: 10 30 20 60 50←40
 837 //         |  |  ↑        ↑
 838 //         `--|--'        |
 839 //             `----------'
 840 //
 841 // To show that this works fine also with irregular spacings, let's say that
 842 // the third frame is delayed a bit (something earlier was dropped). Now the
 843 // situation looks like this:
 844 //
 845 //        I  B  P  B  B   P
 846 //   pts: 30 40 80 90 100 110
 847 //        ↓  ↓     ↓
 848 //   dts: 10 30 20 90 50←40
 849 //         |  |  ↑        ↑
 850 //         `--|--'        |
 851 //             `----------'
 852 //
 853 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 854 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 855 //
 856 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 857 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 858 // a tip frame and should be given a dts based on the previous one.
 859 #define FRAME_P 0
 860 #define FRAME_B 1
 861 #define FRAME_I 2
 862 #define FRAME_IDR 7
 863 void encoding2display_order(
 864     int encoding_order, int intra_period,
 865     int intra_idr_period, int ip_period,
 866     int *displaying_order,
 867     int *frame_type, int *pts_lag)
 868 {
 869     int encoding_order_gop = 0;
 870
 871     *pts_lag = 0;
 872
 873     if (intra_period == 1) { /* all are I/IDR frames */
 874         *displaying_order = encoding_order;
 875         if (intra_idr_period == 0)
 876             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 877         else
 878             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 879         return;
 880     }
 881
 882     if (intra_period == 0)
 883         intra_idr_period = 0;
 884
 885     if (ip_period == 1) {
 886         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 887         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 888         *displaying_order = encoding_order;
 889
 890         if (encoding_order_gop == 0) { /* the first frame */
 891             *frame_type = FRAME_IDR;
 892         } else if (intra_period != 0 && /* have I frames */
 893                    encoding_order_gop >= 2 &&
 894                    (encoding_order_gop % intra_period == 0)) {
 895             *frame_type = FRAME_I;
 896         } else {
 897             *frame_type = FRAME_P;
 898         }
 899         return;
 900     }
 901
 902     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 903     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 904     *pts_lag = -1;  // Most frames are not tip frames.
 905
 906     if (encoding_order_gop == 0) { /* the first frame */
 907         *frame_type = FRAME_IDR;
 908         *displaying_order = encoding_order;
 909         // IDR frames are a special case; I honestly can't find the logic behind
 910         // why this is the right thing, but it seems to line up nicely in practice :-)
 911         *pts_lag = TIMEBASE / MAX_FPS;
 912     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 913         *frame_type = FRAME_B;
 914         *displaying_order = encoding_order - 1;
 915         if ((encoding_order_gop % ip_period) == 0) {
 916             *pts_lag = 0;  // Last B-frame.
 917         }
 918     } else if (intra_period != 0 && /* have I frames */
 919                encoding_order_gop >= 2 &&
 920                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 921         *frame_type = FRAME_I;
 922         *displaying_order = encoding_order + ip_period - 1;
 923     } else {
 924         *frame_type = FRAME_P;
 925         *displaying_order = encoding_order + ip_period - 1;
 926     }
 927 }
 928
 929
 930 static const char *rc_to_string(int rc_mode)
 931 {
 932     switch (rc_mode) {
 933     case VA_RC_NONE:
 934         return "NONE";
 935     case VA_RC_CBR:
 936         return "CBR";
 937     case VA_RC_VBR:
 938         return "VBR";
 939     case VA_RC_VCM:
 940         return "VCM";
 941     case VA_RC_CQP:
 942         return "CQP";
 943     case VA_RC_VBR_CONSTRAINED:
 944         return "VBR_CONSTRAINED";
 945     default:
 946         return "Unknown";
 947     }
 948 }
 949
 950 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 951 {
 952         if (global_flags.uncompressed_video_to_http) {
 953                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 954                 use_zerocopy = false;
 955         } else if (global_flags.x264_video_to_http) {
 956                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 957                 use_zerocopy = false;
 958         } else {
 959                 use_zerocopy = true;
 960         }
 961 }
 962
 963 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 964 {
 965         if (va_display.empty()) {
 966                 x11_display = XOpenDisplay(NULL);
 967                 if (!x11_display) {
 968                         fprintf(stderr, "error: can't connect to X server!\n");
 969                         return NULL;
 970                 }
 971                 enable_zerocopy_if_possible();
 972                 return vaGetDisplay(x11_display);
 973         } else if (va_display[0] != '/') {
 974                 x11_display = XOpenDisplay(va_display.c_str());
 975                 if (!x11_display) {
 976                         fprintf(stderr, "error: can't connect to X server!\n");
 977                         return NULL;
 978                 }
 979                 enable_zerocopy_if_possible();
 980                 return vaGetDisplay(x11_display);
 981         } else {
 982                 drm_fd = open(va_display.c_str(), O_RDWR);
 983                 if (drm_fd == -1) {
 984                         perror(va_display.c_str());
 985                         return NULL;
 986                 }
 987                 use_zerocopy = false;
 988                 return vaGetDisplayDRM(drm_fd);
 989         }
 990 }
 991
 992 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
 993 {
 994         if (x11_display) {
 995                 XCloseDisplay(x11_display);
 996                 x11_display = nullptr;
 997         }
 998         if (drm_fd != -1) {
 999                 close(drm_fd);
1000         }
1001 }
1002
1003 int QuickSyncEncoderImpl::init_va(const string &va_display)
1004 {
1005     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1006     VAEntrypoint *entrypoints;
1007     int num_entrypoints, slice_entrypoint;
1008     int support_encode = 0;
1009     int major_ver, minor_ver;
1010     VAStatus va_status;
1011     unsigned int i;
1012
1013     va_dpy = va_open_display(va_display);
1014     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1015     CHECK_VASTATUS(va_status, "vaInitialize");
1016
1017     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1018     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1019     if (!entrypoints) {
1020         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1021         exit(1);
1022     }
1023
1024     /* use the highest profile */
1025     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1026         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1027             continue;
1028
1029         h264_profile = profile_list[i];
1030         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1031         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1032             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1033                 support_encode = 1;
1034                 break;
1035             }
1036         }
1037         if (support_encode == 1)
1038             break;
1039     }
1040
1041     if (support_encode == 0) {
1042         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1043         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1044         printf("to use VA-API against DRM instead of X11.\n");
1045         exit(1);
1046     } else {
1047         switch (h264_profile) {
1048             case VAProfileH264Baseline:
1049                 ip_period = 1;
1050                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1051                 h264_entropy_mode = 0;
1052                 break;
1053             case VAProfileH264ConstrainedBaseline:
1054                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1055                 ip_period = 1;
1056                 break;
1057
1058             case VAProfileH264Main:
1059                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1060                 break;
1061
1062             case VAProfileH264High:
1063                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1064                 break;
1065             default:
1066                 h264_profile = VAProfileH264Baseline;
1067                 ip_period = 1;
1068                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1069                 break;
1070         }
1071     }
1072
1073     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1074
1075     /* find out the format for the render target, and rate control mode */
1076     for (i = 0; i < VAConfigAttribTypeMax; i++)
1077         attrib[i].type = (VAConfigAttribType)i;
1078
1079     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1080                                       &attrib[0], VAConfigAttribTypeMax);
1081     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1082     /* check the interested configattrib */
1083     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1084         printf("Not find desired YUV420 RT format\n");
1085         exit(1);
1086     } else {
1087         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1088         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1089         config_attrib_num++;
1090     }
1091
1092     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1093         int tmp = attrib[VAConfigAttribRateControl].value;
1094
1095         if (rc_mode == -1 || !(rc_mode & tmp))  {
1096             if (rc_mode != -1) {
1097                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1098             }
1099
1100             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1101                 if (rc_default_modes[i] & tmp) {
1102                     rc_mode = rc_default_modes[i];
1103                     break;
1104                 }
1105             }
1106         }
1107
1108         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1109         config_attrib[config_attrib_num].value = rc_mode;
1110         config_attrib_num++;
1111     }
1112
1113
1114     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1115         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1116
1117         h264_packedheader = 1;
1118         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1119         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1120
1121         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1122             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1123         }
1124
1125         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1126             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1127         }
1128
1129         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1130             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1131         }
1132
1133         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1134             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1135         }
1136
1137         enc_packed_header_idx = config_attrib_num;
1138         config_attrib_num++;
1139     }
1140
1141     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1142         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1143         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1144         config_attrib_num++;
1145     }
1146
1147     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1148         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1149     }
1150
1151     free(entrypoints);
1152     return 0;
1153 }
1154
1155 int QuickSyncEncoderImpl::setup_encode()
1156 {
1157     VAStatus va_status;
1158     VASurfaceID *tmp_surfaceid;
1159     int codedbuf_size, i;
1160     VASurfaceID src_surface[SURFACE_NUM];
1161     VASurfaceID ref_surface[SURFACE_NUM];
1162
1163     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1164             &config_attrib[0], config_attrib_num, &config_id);
1165     CHECK_VASTATUS(va_status, "vaCreateConfig");
1166
1167     /* create source surfaces */
1168     va_status = vaCreateSurfaces(va_dpy,
1169                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1170                                  &src_surface[0], SURFACE_NUM,
1171                                  NULL, 0);
1172     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1173
1174     /* create reference surfaces */
1175     va_status = vaCreateSurfaces(va_dpy,
1176                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1177                                  &ref_surface[0], SURFACE_NUM,
1178                                  NULL, 0);
1179     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1180
1181     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1182     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1183     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1184
1185     /* Create a context for this encode pipe */
1186     va_status = vaCreateContext(va_dpy, config_id,
1187                                 frame_width_mbaligned, frame_height_mbaligned,
1188                                 VA_PROGRESSIVE,
1189                                 tmp_surfaceid, 2 * SURFACE_NUM,
1190                                 &context_id);
1191     CHECK_VASTATUS(va_status, "vaCreateContext");
1192     free(tmp_surfaceid);
1193
1194     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1195
1196     for (i = 0; i < SURFACE_NUM; i++) {
1197         /* create coded buffer once for all
1198          * other VA buffers which won't be used again after vaRenderPicture.
1199          * so APP can always vaCreateBuffer for every frame
1200          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1201          * so VA won't maintain the coded buffer
1202          */
1203         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1204                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1205         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1206     }
1207
1208     /* create OpenGL objects */
1209     //glGenFramebuffers(SURFACE_NUM, fbos);
1210
1211     for (i = 0; i < SURFACE_NUM; i++) {
1212         if (use_zerocopy) {
1213             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
1214             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
1215         } else {
1216             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
1217             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
1218
1219             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1220             // buffers, due to potentially differing pitch.
1221             glGenBuffers(1, &gl_surfaces[i].pbo);
1222             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1223             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1224             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1225             gl_surfaces[i].y_offset = 0;
1226             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1227             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1228             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1229             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1230         }
1231     }
1232
1233     for (i = 0; i < SURFACE_NUM; i++) {
1234         gl_surfaces[i].src_surface = src_surface[i];
1235         gl_surfaces[i].ref_surface = ref_surface[i];
1236     }
1237
1238     return 0;
1239 }
1240
1241 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1242 //
1243 //   2 1 0 [3] 4 8 9
1244 template<class T, class C>
1245 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1246 {
1247         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1248         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1249         sort(middle, end, less_than);
1250 }
1251
1252 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1253 {
1254     int i;
1255
1256     if (frame_type == FRAME_B)
1257         return;
1258
1259     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1260     numShortTerm++;
1261     if (numShortTerm > num_ref_frames)
1262         numShortTerm = num_ref_frames;
1263     for (i=numShortTerm-1; i>0; i--)
1264         ReferenceFrames[i] = ReferenceFrames[i-1];
1265     ReferenceFrames[0] = CurrentCurrPic;
1266
1267     current_frame_num++;
1268     if (current_frame_num > MaxFrameNum)
1269         current_frame_num = 0;
1270 }
1271
1272
1273 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1274 {
1275     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1276         return a.frame_idx > b.frame_idx;
1277     };
1278     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1279         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1280     };
1281     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1282         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1283     };
1284
1285     if (frame_type == FRAME_P) {
1286         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1287         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1288     } else if (frame_type == FRAME_B) {
1289         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1290         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1291
1292         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1293         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1294     }
1295
1296     return 0;
1297 }
1298
1299
1300 int QuickSyncEncoderImpl::render_sequence()
1301 {
1302     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1303     VAStatus va_status;
1304     VAEncMiscParameterBuffer *misc_param;
1305     VAEncMiscParameterRateControl *misc_rate_ctrl;
1306
1307     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1308     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1309     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1310     seq_param.bits_per_second = frame_bitrate;
1311
1312     seq_param.intra_period = intra_period;
1313     seq_param.intra_idr_period = intra_idr_period;
1314     seq_param.ip_period = ip_period;
1315
1316     seq_param.max_num_ref_frames = num_ref_frames;
1317     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1318     seq_param.time_scale = TIMEBASE * 2;
1319     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1320     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1321     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1322     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1323     seq_param.seq_fields.bits.chroma_format_idc = 1;
1324     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1325
1326     if (frame_width != frame_width_mbaligned ||
1327         frame_height != frame_height_mbaligned) {
1328         seq_param.frame_cropping_flag = 1;
1329         seq_param.frame_crop_left_offset = 0;
1330         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1331         seq_param.frame_crop_top_offset = 0;
1332         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1333     }
1334
1335     va_status = vaCreateBuffer(va_dpy, context_id,
1336                                VAEncSequenceParameterBufferType,
1337                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1338     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1339
1340     va_status = vaCreateBuffer(va_dpy, context_id,
1341                                VAEncMiscParameterBufferType,
1342                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1343                                1, NULL, &rc_param_buf);
1344     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1345
1346     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1347     misc_param->type = VAEncMiscParameterTypeRateControl;
1348     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1349     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1350     misc_rate_ctrl->bits_per_second = frame_bitrate;
1351     misc_rate_ctrl->target_percentage = 66;
1352     misc_rate_ctrl->window_size = 1000;
1353     misc_rate_ctrl->initial_qp = initial_qp;
1354     misc_rate_ctrl->min_qp = minimal_qp;
1355     misc_rate_ctrl->basic_unit_size = 0;
1356     vaUnmapBuffer(va_dpy, rc_param_buf);
1357
1358     render_id[0] = seq_param_buf;
1359     render_id[1] = rc_param_buf;
1360
1361     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1362
1363     return 0;
1364 }
1365
1366 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1367 {
1368     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1369     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1370     int PicOrderCntMsb, TopFieldOrderCnt;
1371
1372     if (frame_type == FRAME_IDR)
1373         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1374     else {
1375         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1376         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1377     }
1378
1379     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1380         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1381         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1382     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1383              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1384         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1385     else
1386         PicOrderCntMsb = prevPicOrderCntMsb;
1387
1388     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1389
1390     if (frame_type != FRAME_B) {
1391         PicOrderCntMsb_ref = PicOrderCntMsb;
1392         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1393     }
1394
1395     return TopFieldOrderCnt;
1396 }
1397
1398 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1399 {
1400     VABufferID pic_param_buf;
1401     VAStatus va_status;
1402     int i = 0;
1403
1404     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1405     pic_param.CurrPic.frame_idx = current_frame_num;
1406     pic_param.CurrPic.flags = 0;
1407     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1408     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1409     CurrentCurrPic = pic_param.CurrPic;
1410
1411     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1412     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1413         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1414         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1415     }
1416
1417     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1418     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1419     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1420     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1421     pic_param.frame_num = current_frame_num;
1422     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1423     pic_param.last_picture = false;  // FIXME
1424     pic_param.pic_init_qp = initial_qp;
1425
1426     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1427                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1428     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1429
1430     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1431
1432     return 0;
1433 }
1434
1435 int QuickSyncEncoderImpl::render_packedsequence()
1436 {
1437     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1438     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1439     unsigned int length_in_bits;
1440     unsigned char *packedseq_buffer = NULL;
1441     VAStatus va_status;
1442
1443     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1444
1445     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1446
1447     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1448     packedheader_param_buffer.has_emulation_bytes = 0;
1449     va_status = vaCreateBuffer(va_dpy,
1450                                context_id,
1451                                VAEncPackedHeaderParameterBufferType,
1452                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1453                                &packedseq_para_bufid);
1454     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1455
1456     va_status = vaCreateBuffer(va_dpy,
1457                                context_id,
1458                                VAEncPackedHeaderDataBufferType,
1459                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1460                                &packedseq_data_bufid);
1461     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1462
1463     render_id[0] = packedseq_para_bufid;
1464     render_id[1] = packedseq_data_bufid;
1465     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1466
1467     free(packedseq_buffer);
1468
1469     return 0;
1470 }
1471
1472
1473 int QuickSyncEncoderImpl::render_packedpicture()
1474 {
1475     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1476     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1477     unsigned int length_in_bits;
1478     unsigned char *packedpic_buffer = NULL;
1479     VAStatus va_status;
1480
1481     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1482     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1483     packedheader_param_buffer.bit_length = length_in_bits;
1484     packedheader_param_buffer.has_emulation_bytes = 0;
1485
1486     va_status = vaCreateBuffer(va_dpy,
1487                                context_id,
1488                                VAEncPackedHeaderParameterBufferType,
1489                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1490                                &packedpic_para_bufid);
1491     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1492
1493     va_status = vaCreateBuffer(va_dpy,
1494                                context_id,
1495                                VAEncPackedHeaderDataBufferType,
1496                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1497                                &packedpic_data_bufid);
1498     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1499
1500     render_id[0] = packedpic_para_bufid;
1501     render_id[1] = packedpic_data_bufid;
1502     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1503
1504     free(packedpic_buffer);
1505
1506     return 0;
1507 }
1508
1509 void QuickSyncEncoderImpl::render_packedslice()
1510 {
1511     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1512     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1513     unsigned int length_in_bits;
1514     unsigned char *packedslice_buffer = NULL;
1515     VAStatus va_status;
1516
1517     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1518     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1519     packedheader_param_buffer.bit_length = length_in_bits;
1520     packedheader_param_buffer.has_emulation_bytes = 0;
1521
1522     va_status = vaCreateBuffer(va_dpy,
1523                                context_id,
1524                                VAEncPackedHeaderParameterBufferType,
1525                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1526                                &packedslice_para_bufid);
1527     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1528
1529     va_status = vaCreateBuffer(va_dpy,
1530                                context_id,
1531                                VAEncPackedHeaderDataBufferType,
1532                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1533                                &packedslice_data_bufid);
1534     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1535
1536     render_id[0] = packedslice_para_bufid;
1537     render_id[1] = packedslice_data_bufid;
1538     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1539
1540     free(packedslice_buffer);
1541 }
1542
1543 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1544 {
1545     VABufferID slice_param_buf;
1546     VAStatus va_status;
1547     int i;
1548
1549     update_RefPicList(frame_type);
1550
1551     /* one frame, one slice */
1552     slice_param.macroblock_address = 0;
1553     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1554     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1555     if (frame_type == FRAME_IDR) {
1556         if (encoding_frame_num != 0)
1557             ++slice_param.idr_pic_id;
1558     } else if (frame_type == FRAME_P) {
1559         int refpiclist0_max = h264_maxref & 0xffff;
1560         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1561
1562         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1563             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1564             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1565         }
1566     } else if (frame_type == FRAME_B) {
1567         int refpiclist0_max = h264_maxref & 0xffff;
1568         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1569
1570         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1571         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1572             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1573             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1574         }
1575
1576         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1577         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1578             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1579             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1580         }
1581     }
1582
1583     slice_param.slice_alpha_c0_offset_div2 = 0;
1584     slice_param.slice_beta_offset_div2 = 0;
1585     slice_param.direct_spatial_mv_pred_flag = 1;
1586     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1587
1588
1589     if (h264_packedheader &&
1590         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1591         render_packedslice();
1592
1593     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1594                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1595     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1596
1597     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1598
1599     return 0;
1600 }
1601
1602
1603
1604 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1605 {
1606         VACodedBufferSegment *buf_list = NULL;
1607         VAStatus va_status;
1608
1609         string data;
1610
1611         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1612         CHECK_VASTATUS(va_status, "vaMapBuffer");
1613         while (buf_list != NULL) {
1614                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1615                 buf_list = (VACodedBufferSegment *) buf_list->next;
1616         }
1617         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1618
1619         {
1620                 // Add video.
1621                 AVPacket pkt;
1622                 memset(&pkt, 0, sizeof(pkt));
1623                 pkt.buf = nullptr;
1624                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1625                 pkt.size = data.size();
1626                 pkt.stream_index = 0;
1627                 if (task.frame_type == FRAME_IDR) {
1628                         pkt.flags = AV_PKT_FLAG_KEY;
1629                 } else {
1630                         pkt.flags = 0;
1631                 }
1632                 pkt.duration = task.duration;
1633                 if (file_mux) {
1634                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1635                 }
1636                 if (!global_flags.uncompressed_video_to_http &&
1637                     !global_flags.x264_video_to_http) {
1638                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1639                 }
1640         }
1641         // Encode and add all audio frames up to and including the pts of this video frame.
1642         for ( ;; ) {
1643                 int64_t audio_pts;
1644                 vector<float> audio;
1645                 {
1646                         unique_lock<mutex> lock(frame_queue_mutex);
1647                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1648                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1649                         auto it = pending_audio_frames.begin();
1650                         if (it->first > task.pts) break;
1651                         audio_pts = it->first;
1652                         audio = move(it->second);
1653                         pending_audio_frames.erase(it);
1654                 }
1655
1656                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1657                 stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1658
1659                 if (audio_pts == task.pts) break;
1660         }
1661 }
1662
1663
1664 // this is weird. but it seems to put a new frame onto the queue
1665 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1666 {
1667         unique_lock<mutex> lock(storage_task_queue_mutex);
1668         storage_task_queue.push(move(task));
1669         storage_task_queue_changed.notify_all();
1670 }
1671
1672 void QuickSyncEncoderImpl::storage_task_thread()
1673 {
1674         for ( ;; ) {
1675                 storage_task current;
1676                 {
1677                         // wait until there's an encoded frame
1678                         unique_lock<mutex> lock(storage_task_queue_mutex);
1679                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1680                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1681                         current = move(storage_task_queue.front());
1682                         storage_task_queue.pop();
1683                 }
1684
1685                 VAStatus va_status;
1686
1687                 // waits for data, then saves it to disk.
1688                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1689                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1690                 save_codeddata(move(current));
1691
1692                 {
1693                         unique_lock<mutex> lock(storage_task_queue_mutex);
1694                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1695                         storage_task_queue_changed.notify_all();
1696                 }
1697         }
1698 }
1699
1700 void QuickSyncEncoderImpl::release_encode()
1701 {
1702         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1703                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1704                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1705                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1706         }
1707
1708         vaDestroyContext(va_dpy, context_id);
1709         vaDestroyConfig(va_dpy, config_id);
1710 }
1711
1712 void QuickSyncEncoderImpl::release_gl_resources()
1713 {
1714         assert(is_shutdown);
1715         if (has_released_gl_resources) {
1716                 return;
1717         }
1718
1719         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1720                 if (!use_zerocopy) {
1721                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1722                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1723                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1724                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1725                 }
1726                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1727                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1728         }
1729
1730         has_released_gl_resources = true;
1731 }
1732
1733 int QuickSyncEncoderImpl::deinit_va()
1734 {
1735     vaTerminate(va_dpy);
1736
1737     va_close_display(va_dpy);
1738
1739     return 0;
1740 }
1741
1742 namespace {
1743
1744 }  // namespace
1745
1746 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)
1747         : 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)
1748 {
1749         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1750         open_output_file(filename);
1751         file_audio_encoder->add_mux(file_mux.get());
1752
1753         frame_width_mbaligned = (frame_width + 15) & (~15);
1754         frame_height_mbaligned = (frame_height + 15) & (~15);
1755
1756         //print_input();
1757
1758         if (global_flags.uncompressed_video_to_http ||
1759             global_flags.x264_video_to_http) {
1760                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1761         }
1762         if (global_flags.x264_video_to_http) {
1763                 assert(x264_encoder != nullptr);
1764         } else {
1765                 assert(x264_encoder == nullptr);
1766         }
1767
1768         init_va(va_display);
1769         setup_encode();
1770
1771         // No frames are ready yet.
1772         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1773
1774         memset(&seq_param, 0, sizeof(seq_param));
1775         memset(&pic_param, 0, sizeof(pic_param));
1776         memset(&slice_param, 0, sizeof(slice_param));
1777
1778         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1779
1780         encode_thread = thread([this]{
1781                 //SDL_GL_MakeCurrent(window, context);
1782                 QOpenGLContext *context = create_context(this->surface);
1783                 eglBindAPI(EGL_OPENGL_API);
1784                 if (!make_current(context, this->surface)) {
1785                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1786                                 eglGetError());
1787                         exit(1);
1788                 }
1789                 encode_thread_func();
1790                 delete_context(context);
1791         });
1792 }
1793
1794 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1795 {
1796         shutdown();
1797         release_gl_resources();
1798 }
1799
1800 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1801 {
1802         assert(!is_shutdown);
1803         {
1804                 // Wait until this frame slot is done encoding.
1805                 unique_lock<mutex> lock(storage_task_queue_mutex);
1806                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1807                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1808                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1809                 }
1810                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1811                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1812                 if (storage_thread_should_quit) return false;
1813         }
1814
1815         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1816         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1817         *y_tex = surf->y_tex;
1818         *cbcr_tex = surf->cbcr_tex;
1819
1820         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1821         CHECK_VASTATUS(va_status, "vaDeriveImage");
1822
1823         if (use_zerocopy) {
1824                 VABufferInfo buf_info;
1825                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1826                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1827                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1828
1829                 // Create Y image.
1830                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1831                 EGLint y_attribs[] = {
1832                         EGL_WIDTH, frame_width,
1833                         EGL_HEIGHT, frame_height,
1834                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1835                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1836                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1837                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1838                         EGL_NONE
1839                 };
1840
1841                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1842                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1843
1844                 // Associate Y image to a texture.
1845                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1846                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1847
1848                 // Create CbCr image.
1849                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1850                 EGLint cbcr_attribs[] = {
1851                         EGL_WIDTH, frame_width,
1852                         EGL_HEIGHT, frame_height,
1853                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1854                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1855                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1856                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1857                         EGL_NONE
1858                 };
1859
1860                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1861                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1862
1863                 // Associate CbCr image to a texture.
1864                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1865                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1866         }
1867
1868         return true;
1869 }
1870
1871 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1872 {
1873         assert(!is_shutdown);
1874         {
1875                 unique_lock<mutex> lock(frame_queue_mutex);
1876                 pending_audio_frames[pts] = move(audio);
1877         }
1878         frame_queue_nonempty.notify_all();
1879 }
1880
1881 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1882 {
1883         assert(!is_shutdown);
1884
1885         if (!use_zerocopy) {
1886                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1887
1888                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1889                 check_error();
1890
1891                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1892                 check_error();
1893
1894                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1895                 check_error();
1896                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1897                 check_error();
1898
1899                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1900                 check_error();
1901                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1902                 check_error();
1903
1904                 glBindTexture(GL_TEXTURE_2D, 0);
1905                 check_error();
1906                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1907                 check_error();
1908
1909                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1910                 check_error();
1911         }
1912
1913         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1914         check_error();
1915         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1916         check_error();
1917
1918         {
1919                 unique_lock<mutex> lock(frame_queue_mutex);
1920                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1921                 ++current_storage_frame;
1922         }
1923         frame_queue_nonempty.notify_all();
1924         return fence;
1925 }
1926
1927 void QuickSyncEncoderImpl::shutdown()
1928 {
1929         if (is_shutdown) {
1930                 return;
1931         }
1932
1933         {
1934                 unique_lock<mutex> lock(frame_queue_mutex);
1935                 encode_thread_should_quit = true;
1936                 frame_queue_nonempty.notify_all();
1937         }
1938         encode_thread.join();
1939         {
1940                 unique_lock<mutex> lock(storage_task_queue_mutex);
1941                 storage_thread_should_quit = true;
1942                 frame_queue_nonempty.notify_all();
1943                 storage_task_queue_changed.notify_all();
1944         }
1945         storage_thread.join();
1946         encode_remaining_audio();
1947
1948         release_encode();
1949         deinit_va();
1950         file_mux.reset();
1951         is_shutdown = true;
1952 }
1953
1954 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1955 {
1956         AVFormatContext *avctx = avformat_alloc_context();
1957         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1958         assert(filename.size() < sizeof(avctx->filename) - 1);
1959         strcpy(avctx->filename, filename.c_str());
1960
1961         string url = "file:" + filename;
1962         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1963         if (ret < 0) {
1964                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1965                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1966                 exit(1);
1967         }
1968
1969         string video_extradata = "";  // FIXME: See other comment about global headers.
1970         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, file_audio_encoder->get_ctx(), TIMEBASE, nullptr));
1971 }
1972
1973 void QuickSyncEncoderImpl::encode_thread_func()
1974 {
1975         int64_t last_dts = -1;
1976         int gop_start_display_frame_num = 0;
1977         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1978                 PendingFrame frame;
1979                 int pts_lag;
1980                 int frame_type, display_frame_num;
1981                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1982                                        &display_frame_num, &frame_type, &pts_lag);
1983                 if (frame_type == FRAME_IDR) {
1984                         numShortTerm = 0;
1985                         current_frame_num = 0;
1986                         gop_start_display_frame_num = display_frame_num;
1987                 }
1988
1989                 {
1990                         unique_lock<mutex> lock(frame_queue_mutex);
1991                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1992                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1993                         });
1994                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1995                                 // We have queued frames that were supposed to be B-frames,
1996                                 // but will be no P-frame to encode them against. Encode them all
1997                                 // as P-frames instead. Note that this happens under the mutex,
1998                                 // but nobody else uses it at this point, since we're shutting down,
1999                                 // so there's no contention.
2000                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
2001                                 return;
2002                         } else {
2003                                 frame = move(pending_video_frames[display_frame_num]);
2004                                 pending_video_frames.erase(display_frame_num);
2005                         }
2006                 }
2007
2008                 // Determine the dts of this frame.
2009                 int64_t dts;
2010                 if (pts_lag == -1) {
2011                         assert(last_dts != -1);
2012                         dts = last_dts + (TIMEBASE / MAX_FPS);
2013                 } else {
2014                         dts = frame.pts - pts_lag;
2015                 }
2016                 last_dts = dts;
2017
2018                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2019         }
2020 }
2021
2022 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2023 {
2024         if (pending_video_frames.empty()) {
2025                 return;
2026         }
2027
2028         for (auto &pending_frame : pending_video_frames) {
2029                 int display_frame_num = pending_frame.first;
2030                 assert(display_frame_num > 0);
2031                 PendingFrame frame = move(pending_frame.second);
2032                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2033                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2034                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2035                 last_dts = dts;
2036         }
2037
2038         if (global_flags.uncompressed_video_to_http ||
2039             global_flags.x264_video_to_http) {
2040                 // Add frames left in reorderer.
2041                 while (!reorderer->empty()) {
2042                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2043                         if (global_flags.uncompressed_video_to_http) {
2044                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2045                         } else {
2046                                 assert(global_flags.x264_video_to_http);
2047                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2048                         }
2049                 }
2050         }
2051 }
2052
2053 void QuickSyncEncoderImpl::encode_remaining_audio()
2054 {
2055         // This really ought to be empty by now, but just to be sure...
2056         for (auto &pending_frame : pending_audio_frames) {
2057                 int64_t audio_pts = pending_frame.first;
2058                 vector<float> audio = move(pending_frame.second);
2059
2060                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2061                 if (stream_audio_encoder) {
2062                         stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2063                 }
2064         }
2065         pending_audio_frames.clear();
2066
2067         // Encode any leftover audio in the queues, and also any delayed frames.
2068         // Note: stream_audio_encoder is not owned by us, so don't call encode_last_audio().
2069         file_audio_encoder->encode_last_audio();
2070 }
2071
2072 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2073 {
2074         AVPacket pkt;
2075         memset(&pkt, 0, sizeof(pkt));
2076         pkt.buf = nullptr;
2077         pkt.data = const_cast<uint8_t *>(data);
2078         pkt.size = frame_width * frame_height * 2;
2079         pkt.stream_index = 0;
2080         pkt.flags = AV_PKT_FLAG_KEY;
2081         pkt.duration = duration;
2082         stream_mux->add_packet(pkt, pts, pts);
2083 }
2084
2085 namespace {
2086
2087 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2088 {
2089         if (src_width == dst_pitch) {
2090                 memcpy(dst, src, src_width * height);
2091         } else {
2092                 for (size_t y = 0; y < height; ++y) {
2093                         const uint8_t *sptr = src + y * src_width;
2094                         uint8_t *dptr = dst + y * dst_pitch;
2095                         memcpy(dptr, sptr, src_width);
2096                 }
2097         }
2098 }
2099
2100 }  // namespace
2101
2102 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2103                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2104 {
2105         // Wait for the GPU to be done with the frame.
2106         GLenum sync_status;
2107         do {
2108                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2109                 check_error();
2110         } while (sync_status == GL_TIMEOUT_EXPIRED);
2111         assert(sync_status != GL_WAIT_FAILED);
2112
2113         // Release back any input frames we needed to render this frame.
2114         frame.input_frames.clear();
2115
2116         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2117         VAStatus va_status;
2118
2119         if (use_zerocopy) {
2120                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2121                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2122                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2123                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2124         } else {
2125                 unsigned char *surface_p = nullptr;
2126                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2127
2128                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2129                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2130
2131                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2132                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2133
2134                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2135                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2136
2137                 if (global_flags.uncompressed_video_to_http ||
2138                     global_flags.x264_video_to_http) {
2139                         // Add uncompressed video. (Note that pts == dts here.)
2140                         // Delay needs to match audio.
2141                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2142                         if (output_frame.data != nullptr) {
2143                                 if (global_flags.uncompressed_video_to_http) {
2144                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2145                                 } else {
2146                                         assert(global_flags.x264_video_to_http);
2147                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2148                                 }
2149                         }
2150                 }
2151         }
2152
2153         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2154         CHECK_VASTATUS(va_status, "vaDestroyImage");
2155
2156         // Schedule the frame for encoding.
2157         VASurfaceID va_surface = surf->src_surface;
2158         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2159         CHECK_VASTATUS(va_status, "vaBeginPicture");
2160
2161         if (frame_type == FRAME_IDR) {
2162                 // FIXME: If the mux wants global headers, we should not put the
2163                 // SPS/PPS before each IDR frame, but rather put it into the
2164                 // codec extradata (formatted differently?).
2165                 render_sequence();
2166                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2167                 if (h264_packedheader) {
2168                         render_packedsequence();
2169                         render_packedpicture();
2170                 }
2171         } else {
2172                 //render_sequence();
2173                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2174         }
2175         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2176
2177         va_status = vaEndPicture(va_dpy, context_id);
2178         CHECK_VASTATUS(va_status, "vaEndPicture");
2179
2180         // so now the data is done encoding (well, async job kicked off)...
2181         // we send that to the storage thread
2182         storage_task tmp;
2183         tmp.display_order = display_frame_num;
2184         tmp.frame_type = frame_type;
2185         tmp.pts = pts;
2186         tmp.dts = dts;
2187         tmp.duration = duration;
2188         storage_task_enqueue(move(tmp));
2189
2190         update_ReferenceFrames(frame_type);
2191 }
2192
2193 // Proxy object.
2194 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)
2195         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, stream_audio_encoder, x264_encoder)) {}
2196
2197 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2198 QuickSyncEncoder::~QuickSyncEncoder() {}
2199
2200 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2201 {
2202         impl->add_audio(pts, audio);
2203 }
2204
2205 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2206 {
2207         return impl->begin_frame(y_tex, cbcr_tex);
2208 }
2209
2210 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2211 {
2212         return impl->end_frame(pts, duration, input_frames);
2213 }
2214
2215 void QuickSyncEncoder::shutdown()
2216 {
2217         impl->shutdown();
2218 }
2219
2220 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2221 {
2222         impl->set_stream_mux(mux);
2223 }
2224