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