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

   1 //#include "sysdeps.h"
   2 #include "quicksync_encoder.h"
   3
   4 #include <movit/util.h>
   5 #include <EGL/eglplatform.h>
   6 #include <X11/X.h>
   7 #include <X11/Xlib.h>
   8 #include <assert.h>
   9 #include <epoxy/egl.h>
  10 #include <libdrm/drm_fourcc.h>
  11 #include <stdio.h>
  12 #include <stdlib.h>
  13 #include <string.h>
  14 #include <fcntl.h>
  15 #include <va/va.h>
  16 #include <va/va_drm.h>
  17 #include <va/va_drmcommon.h>
  18 #include <va/va_enc_h264.h>
  19 #include <va/va_x11.h>
  20 #include <algorithm>
  21 #include <condition_variable>
  22 #include <cstdint>
  23 #include <map>
  24 #include <memory>
  25 #include <mutex>
  26 #include <queue>
  27 #include <string>
  28 #include <thread>
  29 #include <utility>
  30
  31 #include "audio_encoder.h"
  32 #include "context.h"
  33 #include "defs.h"
  34 #include "flags.h"
  35 #include "mux.h"
  36 #include "timebase.h"
  37 #include "x264_encoder.h"
  38
  39 using namespace std;
  40
  41 class QOpenGLContext;
  42 class QSurface;
  43
  44 #define CHECK_VASTATUS(va_status, func)                                 \
  45     if (va_status != VA_STATUS_SUCCESS) {                               \
  46         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  47         exit(1);                                                        \
  48     }
  49
  50 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  51
  52 //#include "loadsurface.h"
  53
  54 #define NAL_REF_IDC_NONE        0
  55 #define NAL_REF_IDC_LOW         1
  56 #define NAL_REF_IDC_MEDIUM      2
  57 #define NAL_REF_IDC_HIGH        3
  58
  59 #define NAL_NON_IDR             1
  60 #define NAL_IDR                 5
  61 #define NAL_SPS                 7
  62 #define NAL_PPS                 8
  63 #define NAL_SEI                 6
  64
  65 #define SLICE_TYPE_P            0
  66 #define SLICE_TYPE_B            1
  67 #define SLICE_TYPE_I            2
  68 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  69 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  70 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  71
  72
  73 #define ENTROPY_MODE_CAVLC      0
  74 #define ENTROPY_MODE_CABAC      1
  75
  76 #define PROFILE_IDC_BASELINE    66
  77 #define PROFILE_IDC_MAIN        77
  78 #define PROFILE_IDC_HIGH        100
  79
  80 #define BITSTREAM_ALLOCATE_STEPPING     4096
  81 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  82 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  83 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  84
  85 static constexpr unsigned int MaxFrameNum = (2<<16);
  86 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  87 static constexpr unsigned int Log2MaxFrameNum = 16;
  88 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  89 static constexpr int rc_default_modes[] = {  // Priority list of modes.
  90     VA_RC_VBR,
  91     VA_RC_CQP,
  92     VA_RC_VBR_CONSTRAINED,
  93     VA_RC_CBR,
  94     VA_RC_VCM,
  95     VA_RC_NONE,
  96 };
  97
  98 /* thread to save coded data */
  99 #define SRC_SURFACE_FREE        0
 100 #define SRC_SURFACE_IN_ENCODING 1
 101
 102 struct __bitstream {
 103     unsigned int *buffer;
 104     int bit_offset;
 105     int max_size_in_dword;
 106 };
 107 typedef struct __bitstream bitstream;
 108
 109 using namespace std;
 110
 111 // H.264 video comes out in encoding order (e.g. with two B-frames:
 112 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 113 // come in the right order. Since we do everything, including waiting
 114 // for the frames to come out of OpenGL, in encoding order, we need
 115 // a reordering buffer for uncompressed frames so that they come out
 116 // correctly. We go the super-lazy way of not making it understand
 117 // anything about the true order (which introduces some extra latency,
 118 // though); we know that for N B-frames we need at most (N-1) frames
 119 // in the reorder buffer, and can just sort on that.
 120 //
 121 // The class also deals with keeping a freelist as needed.
 122 class FrameReorderer {
 123 public:
 124         FrameReorderer(unsigned queue_length, int width, int height);
 125
 126         struct Frame {
 127                 int64_t pts, duration;
 128                 uint8_t *data;
 129
 130                 // Invert to get the smallest pts first.
 131                 bool operator< (const Frame &other) const { return pts > other.pts; }
 132         };
 133
 134         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 135         // Does _not_ take ownership of data; a copy is taken if needed.
 136         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 137         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 138         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data);
 139
 140         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 141         Frame get_first_frame();
 142
 143         bool empty() const { return frames.empty(); }
 144
 145 private:
 146         unsigned queue_length;
 147         int width, height;
 148
 149         priority_queue<Frame> frames;
 150         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 151
 152         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 153         // except priority_queue doesn't work well with movable-only types.
 154         vector<unique_ptr<uint8_t[]>> owner;
 155 };
 156
 157 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 158     : queue_length(queue_length), width(width), height(height)
 159 {
 160         for (unsigned i = 0; i < queue_length; ++i) {
 161                 owner.emplace_back(new uint8_t[width * height * 2]);
 162                 freelist.push(owner.back().get());
 163         }
 164 }
 165
 166 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data)
 167 {
 168         if (queue_length == 0) {
 169                 return Frame{pts, duration, data};
 170         }
 171
 172         assert(!freelist.empty());
 173         uint8_t *storage = freelist.top();
 174         freelist.pop();
 175         memcpy(storage, data, width * height * 2);
 176         frames.push(Frame{pts, duration, storage});
 177
 178         if (frames.size() >= queue_length) {
 179                 return get_first_frame();
 180         } else {
 181                 return Frame{-1, -1, nullptr};
 182         }
 183 }
 184
 185 FrameReorderer::Frame FrameReorderer::get_first_frame()
 186 {
 187         assert(!frames.empty());
 188         Frame storage = frames.top();
 189         frames.pop();
 190         freelist.push(storage.data);
 191         return storage;
 192 }
 193
 194 class QuickSyncEncoderImpl {
 195 public:
 196         QuickSyncEncoderImpl(QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux);
 197         ~QuickSyncEncoderImpl();
 198         void add_audio(int64_t pts, vector<float> audio);
 199         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 200         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
 201         void shutdown();
 202         void open_output_file(const std::string &filename);
 203         void close_output_file();
 204
 205 private:
 206         struct storage_task {
 207                 unsigned long long display_order;
 208                 int frame_type;
 209                 vector<float> audio;
 210                 int64_t pts, dts, duration;
 211         };
 212         struct PendingFrame {
 213                 RefCountedGLsync fence;
 214                 vector<RefCountedFrame> input_frames;
 215                 int64_t pts, duration;
 216         };
 217
 218         // So we never get negative dts.
 219         int64_t global_delay() const {
 220                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 221         }
 222
 223         void encode_thread_func();
 224         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 225         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 226         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 227                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 228         void storage_task_thread();
 229         void encode_remaining_audio();
 230         void storage_task_enqueue(storage_task task);
 231         void save_codeddata(storage_task task);
 232         int render_packedsequence();
 233         int render_packedpicture();
 234         void render_packedslice();
 235         int render_sequence();
 236         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 237         void sps_rbsp(bitstream *bs);
 238         void pps_rbsp(bitstream *bs);
 239         int build_packed_pic_buffer(unsigned char **header_buffer);
 240         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 241         void slice_header(bitstream *bs);
 242         int build_packed_seq_buffer(unsigned char **header_buffer);
 243         int build_packed_slice_buffer(unsigned char **header_buffer);
 244         int init_va(const string &va_display);
 245         int deinit_va();
 246         void enable_zerocopy_if_possible();
 247         VADisplay va_open_display(const string &va_display);
 248         void va_close_display(VADisplay va_dpy);
 249         int setup_encode();
 250         int release_encode();
 251         void update_ReferenceFrames(int frame_type);
 252         int update_RefPicList(int frame_type);
 253
 254         bool is_shutdown = false;
 255         bool use_zerocopy;
 256         int drm_fd = -1;
 257
 258         thread encode_thread, storage_thread;
 259
 260         mutex storage_task_queue_mutex;
 261         condition_variable storage_task_queue_changed;
 262         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 263         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 264         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 265
 266         mutex frame_queue_mutex;
 267         condition_variable frame_queue_nonempty;
 268         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 269
 270         int current_storage_frame;
 271
 272         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 273         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 274         QSurface *surface;
 275
 276         unique_ptr<AudioEncoder> file_audio_encoder;
 277         unique_ptr<AudioEncoder> stream_audio_encoder;
 278
 279         Mux* stream_mux;  // To HTTP.
 280         unique_ptr<Mux> file_mux;  // To local disk.
 281
 282         unique_ptr<FrameReorderer> reorderer;
 283         unique_ptr<X264Encoder> x264_encoder;  // nullptr if not using x264.
 284
 285         Display *x11_display = nullptr;
 286
 287         // Encoder parameters
 288         VADisplay va_dpy;
 289         VAProfile h264_profile = (VAProfile)~0;
 290         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 291         int config_attrib_num = 0, enc_packed_header_idx;
 292
 293         struct GLSurface {
 294                 VASurfaceID src_surface, ref_surface;
 295                 VABufferID coded_buf;
 296
 297                 VAImage surface_image;
 298                 GLuint y_tex, cbcr_tex;
 299
 300                 // Only if use_zerocopy == true.
 301                 EGLImage y_egl_image, cbcr_egl_image;
 302
 303                 // Only if use_zerocopy == false.
 304                 GLuint pbo;
 305                 uint8_t *y_ptr, *cbcr_ptr;
 306                 size_t y_offset, cbcr_offset;
 307         };
 308         GLSurface gl_surfaces[SURFACE_NUM];
 309
 310         VAConfigID config_id;
 311         VAContextID context_id;
 312         VAEncSequenceParameterBufferH264 seq_param;
 313         VAEncPictureParameterBufferH264 pic_param;
 314         VAEncSliceParameterBufferH264 slice_param;
 315         VAPictureH264 CurrentCurrPic;
 316         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 317
 318         // Static quality settings.
 319         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 320         static constexpr unsigned int num_ref_frames = 2;
 321         static constexpr int initial_qp = 15;
 322         static constexpr int minimal_qp = 0;
 323         static constexpr int intra_period = 30;
 324         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 325
 326         // Quality settings that are meant to be static, but might be overridden
 327         // by the profile.
 328         int constraint_set_flag = 0;
 329         int h264_packedheader = 0; /* support pack header? */
 330         int h264_maxref = (1<<16|1);
 331         int h264_entropy_mode = 1; /* cabac */
 332         int ip_period = 3;
 333
 334         int rc_mode = -1;
 335         unsigned int current_frame_num = 0;
 336         unsigned int numShortTerm = 0;
 337
 338         int frame_width;
 339         int frame_height;
 340         int frame_width_mbaligned;
 341         int frame_height_mbaligned;
 342 };
 343
 344 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 345 // but if we don't delete it here, we get leaks. The GStreamer implementation
 346 // does the same.
 347 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 348 {
 349     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 350     CHECK_VASTATUS(va_status, "vaRenderPicture");
 351
 352     for (int i = 0; i < num_buffers; ++i) {
 353         va_status = vaDestroyBuffer(dpy, buffers[i]);
 354         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 355     }
 356 }
 357
 358 static unsigned int
 359 va_swap32(unsigned int val)
 360 {
 361     unsigned char *pval = (unsigned char *)&val;
 362
 363     return ((pval[0] << 24)     |
 364             (pval[1] << 16)     |
 365             (pval[2] << 8)      |
 366             (pval[3] << 0));
 367 }
 368
 369 static void
 370 bitstream_start(bitstream *bs)
 371 {
 372     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 373     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 374     bs->bit_offset = 0;
 375 }
 376
 377 static void
 378 bitstream_end(bitstream *bs)
 379 {
 380     int pos = (bs->bit_offset >> 5);
 381     int bit_offset = (bs->bit_offset & 0x1f);
 382     int bit_left = 32 - bit_offset;
 383
 384     if (bit_offset) {
 385         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 386     }
 387 }
 388
 389 static void
 390 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 391 {
 392     int pos = (bs->bit_offset >> 5);
 393     int bit_offset = (bs->bit_offset & 0x1f);
 394     int bit_left = 32 - bit_offset;
 395
 396     if (!size_in_bits)
 397         return;
 398
 399     bs->bit_offset += size_in_bits;
 400
 401     if (bit_left > size_in_bits) {
 402         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 403     } else {
 404         size_in_bits -= bit_left;
 405         if (bit_left >= 32) {
 406             bs->buffer[pos] = (val >> size_in_bits);
 407         } else {
 408             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 409         }
 410         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 411
 412         if (pos + 1 == bs->max_size_in_dword) {
 413             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 414             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 415         }
 416
 417         bs->buffer[pos + 1] = val;
 418     }
 419 }
 420
 421 static void
 422 bitstream_put_ue(bitstream *bs, unsigned int val)
 423 {
 424     int size_in_bits = 0;
 425     int tmp_val = ++val;
 426
 427     while (tmp_val) {
 428         tmp_val >>= 1;
 429         size_in_bits++;
 430     }
 431
 432     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 433     bitstream_put_ui(bs, val, size_in_bits);
 434 }
 435
 436 static void
 437 bitstream_put_se(bitstream *bs, int val)
 438 {
 439     unsigned int new_val;
 440
 441     if (val <= 0)
 442         new_val = -2 * val;
 443     else
 444         new_val = 2 * val - 1;
 445
 446     bitstream_put_ue(bs, new_val);
 447 }
 448
 449 static void
 450 bitstream_byte_aligning(bitstream *bs, int bit)
 451 {
 452     int bit_offset = (bs->bit_offset & 0x7);
 453     int bit_left = 8 - bit_offset;
 454     int new_val;
 455
 456     if (!bit_offset)
 457         return;
 458
 459     assert(bit == 0 || bit == 1);
 460
 461     if (bit)
 462         new_val = (1 << bit_left) - 1;
 463     else
 464         new_val = 0;
 465
 466     bitstream_put_ui(bs, new_val, bit_left);
 467 }
 468
 469 static void
 470 rbsp_trailing_bits(bitstream *bs)
 471 {
 472     bitstream_put_ui(bs, 1, 1);
 473     bitstream_byte_aligning(bs, 0);
 474 }
 475
 476 static void nal_start_code_prefix(bitstream *bs)
 477 {
 478     bitstream_put_ui(bs, 0x00000001, 32);
 479 }
 480
 481 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 482 {
 483     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 484     bitstream_put_ui(bs, nal_ref_idc, 2);
 485     bitstream_put_ui(bs, nal_unit_type, 5);
 486 }
 487
 488 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 489 {
 490     int profile_idc = PROFILE_IDC_BASELINE;
 491
 492     if (h264_profile  == VAProfileH264High)
 493         profile_idc = PROFILE_IDC_HIGH;
 494     else if (h264_profile  == VAProfileH264Main)
 495         profile_idc = PROFILE_IDC_MAIN;
 496
 497     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 498     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 499     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 500     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 501     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 502     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 503     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 504     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 505
 506     if ( profile_idc == PROFILE_IDC_HIGH) {
 507         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 508         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 509         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 510         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 511         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 512     }
 513
 514     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 515     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 516
 517     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 518         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 519     else {
 520         assert(0);
 521     }
 522
 523     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 524     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 525
 526     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 527     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 528     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 529
 530     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 531         assert(0);
 532     }
 533
 534     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 535     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 536
 537     if (seq_param.frame_cropping_flag) {
 538         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 539         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 540         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 541         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 542     }
 543
 544     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 545     if ( false ) {
 546         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 547     } else {
 548         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 549         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 550         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 551         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 552         {
 553             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 554             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 555             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 556             {
 557                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 558                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 559                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 560             }
 561         }
 562         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 563         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 564         {
 565             bitstream_put_ui(bs, 1, 32);  // FPS
 566             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 567             bitstream_put_ui(bs, 1, 1);
 568         }
 569         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 570         {
 571             // hrd_parameters
 572             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 573             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 574             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 575
 576             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 577             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 578             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 579
 580             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 581             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 582             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 583             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 584         }
 585         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 586         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 587
 588         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 589         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 590     }
 591
 592     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 593 }
 594
 595
 596 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 597 {
 598     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 599     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 600
 601     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 602
 603     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 604
 605     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 606
 607     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 608     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 609
 610     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 611     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 612
 613     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 614     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 615     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 616
 617     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 618     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 619     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 620
 621     /* more_rbsp_data */
 622     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 623     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 624     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 625
 626     rbsp_trailing_bits(bs);
 627 }
 628
 629 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 630 {
 631     int first_mb_in_slice = slice_param.macroblock_address;
 632
 633     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 634     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 635     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 636     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 637
 638     /* frame_mbs_only_flag == 1 */
 639     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 640         /* FIXME: */
 641         assert(0);
 642     }
 643
 644     if (pic_param.pic_fields.bits.idr_pic_flag)
 645         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 646
 647     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 648         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 649         /* pic_order_present_flag == 0 */
 650     } else {
 651         /* FIXME: */
 652         assert(0);
 653     }
 654
 655     /* redundant_pic_cnt_present_flag == 0 */
 656     /* slice type */
 657     if (IS_P_SLICE(slice_param.slice_type)) {
 658         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 659
 660         if (slice_param.num_ref_idx_active_override_flag)
 661             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 662
 663         /* ref_pic_list_reordering */
 664         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 665     } else if (IS_B_SLICE(slice_param.slice_type)) {
 666         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 667
 668         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 669
 670         if (slice_param.num_ref_idx_active_override_flag) {
 671             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 672             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 673         }
 674
 675         /* ref_pic_list_reordering */
 676         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 677         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 678     }
 679
 680     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 681          IS_P_SLICE(slice_param.slice_type)) ||
 682         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 683          IS_B_SLICE(slice_param.slice_type))) {
 684         /* FIXME: fill weight/offset table */
 685         assert(0);
 686     }
 687
 688     /* dec_ref_pic_marking */
 689     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 690         unsigned char no_output_of_prior_pics_flag = 0;
 691         unsigned char long_term_reference_flag = 0;
 692         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 693
 694         if (pic_param.pic_fields.bits.idr_pic_flag) {
 695             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 696             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 697         } else {
 698             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 699         }
 700     }
 701
 702     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 703         !IS_I_SLICE(slice_param.slice_type))
 704         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 705
 706     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 707
 708     /* ignore for SP/SI */
 709
 710     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 711         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 712
 713         if (slice_param.disable_deblocking_filter_idc != 1) {
 714             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 715             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 716         }
 717     }
 718
 719     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 720         bitstream_byte_aligning(bs, 1);
 721     }
 722 }
 723
 724 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 725 {
 726     bitstream bs;
 727
 728     bitstream_start(&bs);
 729     nal_start_code_prefix(&bs);
 730     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 731     pps_rbsp(&bs);
 732     bitstream_end(&bs);
 733
 734     *header_buffer = (unsigned char *)bs.buffer;
 735     return bs.bit_offset;
 736 }
 737
 738 int
 739 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 740 {
 741     bitstream bs;
 742
 743     bitstream_start(&bs);
 744     nal_start_code_prefix(&bs);
 745     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 746     sps_rbsp(&bs);
 747     bitstream_end(&bs);
 748
 749     *header_buffer = (unsigned char *)bs.buffer;
 750     return bs.bit_offset;
 751 }
 752
 753 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 754 {
 755     bitstream bs;
 756     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 757     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 758
 759     bitstream_start(&bs);
 760     nal_start_code_prefix(&bs);
 761
 762     if (IS_I_SLICE(slice_param.slice_type)) {
 763         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 764     } else if (IS_P_SLICE(slice_param.slice_type)) {
 765         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 766     } else {
 767         assert(IS_B_SLICE(slice_param.slice_type));
 768         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 769     }
 770
 771     slice_header(&bs);
 772     bitstream_end(&bs);
 773
 774     *header_buffer = (unsigned char *)bs.buffer;
 775     return bs.bit_offset;
 776 }
 777
 778
 779 /*
 780   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 781   1) period between Frame #X and Frame #N = #X - #N
 782   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 783   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 784   4) intra_period and intra_idr_period take precedence over ip_period
 785   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 786      of I/IDR frames, see bellow examples
 787   -------------------------------------------------------------------
 788   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 789   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 790   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 791   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 792   1            1                ignored    IDR IDR IDR IDR...
 793   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 794   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 795   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 796                                               (PBB)(IBB)(PBB)(IBB)...
 797   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 798                                            IDRPPPPP IPPPPP IPPPPP...
 799   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 800                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 801                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 802                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 803                                            {IDR(PBB)(PBB)}                     (6/6/3)
 804                                            {IDR(PBB)(PBB)}.
 805 */
 806
 807 // General pts/dts strategy:
 808 //
 809 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 810 // bit tricky. We assume first of all that the frame rate never goes _above_
 811 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 812 // in at least this speed, as long at dts <= pts (the frame is not attempted
 813 // presented before it is decoded). Furthermore, we never have longer chains of
 814 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 815 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 816 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 817 // frame, with an order O <= C.)
 818 //
 819 // Many strategies are possible, but we establish these rules:
 820 //
 821 //  - Tip frames have dts = pts - (C-O)*N.
 822 //  - Non-tip frames have dts = dts_last + N.
 823 //
 824 // An example, with C=2 and N=10 and the data flow showed with arrows:
 825 //
 826 //        I  B  P  B  B  P
 827 //   pts: 30 40 50 60 70 80
 828 //        ↓  ↓     ↓
 829 //   dts: 10 30 20 60 50←40
 830 //         |  |  ↑        ↑
 831 //         `--|--'        |
 832 //             `----------'
 833 //
 834 // To show that this works fine also with irregular spacings, let's say that
 835 // the third frame is delayed a bit (something earlier was dropped). Now the
 836 // situation looks like this:
 837 //
 838 //        I  B  P  B  B   P
 839 //   pts: 30 40 80 90 100 110
 840 //        ↓  ↓     ↓
 841 //   dts: 10 30 20 90 50←40
 842 //         |  |  ↑        ↑
 843 //         `--|--'        |
 844 //             `----------'
 845 //
 846 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 847 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 848 //
 849 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 850 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 851 // a tip frame and should be given a dts based on the previous one.
 852 #define FRAME_P 0
 853 #define FRAME_B 1
 854 #define FRAME_I 2
 855 #define FRAME_IDR 7
 856 void encoding2display_order(
 857     int encoding_order, int intra_period,
 858     int intra_idr_period, int ip_period,
 859     int *displaying_order,
 860     int *frame_type, int *pts_lag)
 861 {
 862     int encoding_order_gop = 0;
 863
 864     *pts_lag = 0;
 865
 866     if (intra_period == 1) { /* all are I/IDR frames */
 867         *displaying_order = encoding_order;
 868         if (intra_idr_period == 0)
 869             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 870         else
 871             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 872         return;
 873     }
 874
 875     if (intra_period == 0)
 876         intra_idr_period = 0;
 877
 878     if (ip_period == 1) {
 879         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 880         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 881         *displaying_order = encoding_order;
 882
 883         if (encoding_order_gop == 0) { /* the first frame */
 884             *frame_type = FRAME_IDR;
 885         } else if (intra_period != 0 && /* have I frames */
 886                    encoding_order_gop >= 2 &&
 887                    (encoding_order_gop % intra_period == 0)) {
 888             *frame_type = FRAME_I;
 889         } else {
 890             *frame_type = FRAME_P;
 891         }
 892         return;
 893     }
 894
 895     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 896     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 897     *pts_lag = -1;  // Most frames are not tip frames.
 898
 899     if (encoding_order_gop == 0) { /* the first frame */
 900         *frame_type = FRAME_IDR;
 901         *displaying_order = encoding_order;
 902         // IDR frames are a special case; I honestly can't find the logic behind
 903         // why this is the right thing, but it seems to line up nicely in practice :-)
 904         *pts_lag = TIMEBASE / MAX_FPS;
 905     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 906         *frame_type = FRAME_B;
 907         *displaying_order = encoding_order - 1;
 908         if ((encoding_order_gop % ip_period) == 0) {
 909             *pts_lag = 0;  // Last B-frame.
 910         }
 911     } else if (intra_period != 0 && /* have I frames */
 912                encoding_order_gop >= 2 &&
 913                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 914         *frame_type = FRAME_I;
 915         *displaying_order = encoding_order + ip_period - 1;
 916     } else {
 917         *frame_type = FRAME_P;
 918         *displaying_order = encoding_order + ip_period - 1;
 919     }
 920 }
 921
 922
 923 static const char *rc_to_string(int rc_mode)
 924 {
 925     switch (rc_mode) {
 926     case VA_RC_NONE:
 927         return "NONE";
 928     case VA_RC_CBR:
 929         return "CBR";
 930     case VA_RC_VBR:
 931         return "VBR";
 932     case VA_RC_VCM:
 933         return "VCM";
 934     case VA_RC_CQP:
 935         return "CQP";
 936     case VA_RC_VBR_CONSTRAINED:
 937         return "VBR_CONSTRAINED";
 938     default:
 939         return "Unknown";
 940     }
 941 }
 942
 943 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 944 {
 945         if (global_flags.uncompressed_video_to_http) {
 946                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 947                 use_zerocopy = false;
 948         } else if (global_flags.x264_video_to_http) {
 949                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 950                 use_zerocopy = false;
 951         } else {
 952                 use_zerocopy = true;
 953         }
 954 }
 955
 956 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 957 {
 958         if (va_display.empty()) {
 959                 x11_display = XOpenDisplay(NULL);
 960                 if (!x11_display) {
 961                         fprintf(stderr, "error: can't connect to X server!\n");
 962                         return NULL;
 963                 }
 964                 enable_zerocopy_if_possible();
 965                 return vaGetDisplay(x11_display);
 966         } else if (va_display[0] != '/') {
 967                 x11_display = XOpenDisplay(va_display.c_str());
 968                 if (!x11_display) {
 969                         fprintf(stderr, "error: can't connect to X server!\n");
 970                         return NULL;
 971                 }
 972                 enable_zerocopy_if_possible();
 973                 return vaGetDisplay(x11_display);
 974         } else {
 975                 drm_fd = open(va_display.c_str(), O_RDWR);
 976                 if (drm_fd == -1) {
 977                         perror(va_display.c_str());
 978                         return NULL;
 979                 }
 980                 use_zerocopy = false;
 981                 return vaGetDisplayDRM(drm_fd);
 982         }
 983 }
 984
 985 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
 986 {
 987         if (x11_display) {
 988                 XCloseDisplay(x11_display);
 989                 x11_display = nullptr;
 990         }
 991         if (drm_fd != -1) {
 992                 close(drm_fd);
 993         }
 994 }
 995
 996 int QuickSyncEncoderImpl::init_va(const string &va_display)
 997 {
 998     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 999     VAEntrypoint *entrypoints;
1000     int num_entrypoints, slice_entrypoint;
1001     int support_encode = 0;
1002     int major_ver, minor_ver;
1003     VAStatus va_status;
1004     unsigned int i;
1005
1006     va_dpy = va_open_display(va_display);
1007     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1008     CHECK_VASTATUS(va_status, "vaInitialize");
1009
1010     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1011     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1012     if (!entrypoints) {
1013         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1014         exit(1);
1015     }
1016
1017     /* use the highest profile */
1018     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1019         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1020             continue;
1021
1022         h264_profile = profile_list[i];
1023         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1024         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1025             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1026                 support_encode = 1;
1027                 break;
1028             }
1029         }
1030         if (support_encode == 1)
1031             break;
1032     }
1033
1034     if (support_encode == 0) {
1035         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1036         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1037         printf("to use VA-API against DRM instead of X11.\n");
1038         exit(1);
1039     } else {
1040         switch (h264_profile) {
1041             case VAProfileH264Baseline:
1042                 ip_period = 1;
1043                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1044                 h264_entropy_mode = 0;
1045                 break;
1046             case VAProfileH264ConstrainedBaseline:
1047                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1048                 ip_period = 1;
1049                 break;
1050
1051             case VAProfileH264Main:
1052                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1053                 break;
1054
1055             case VAProfileH264High:
1056                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1057                 break;
1058             default:
1059                 h264_profile = VAProfileH264Baseline;
1060                 ip_period = 1;
1061                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1062                 break;
1063         }
1064     }
1065
1066     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1067
1068     /* find out the format for the render target, and rate control mode */
1069     for (i = 0; i < VAConfigAttribTypeMax; i++)
1070         attrib[i].type = (VAConfigAttribType)i;
1071
1072     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1073                                       &attrib[0], VAConfigAttribTypeMax);
1074     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1075     /* check the interested configattrib */
1076     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1077         printf("Not find desired YUV420 RT format\n");
1078         exit(1);
1079     } else {
1080         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1081         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1082         config_attrib_num++;
1083     }
1084
1085     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1086         int tmp = attrib[VAConfigAttribRateControl].value;
1087
1088         if (rc_mode == -1 || !(rc_mode & tmp))  {
1089             if (rc_mode != -1) {
1090                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1091             }
1092
1093             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1094                 if (rc_default_modes[i] & tmp) {
1095                     rc_mode = rc_default_modes[i];
1096                     break;
1097                 }
1098             }
1099         }
1100
1101         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1102         config_attrib[config_attrib_num].value = rc_mode;
1103         config_attrib_num++;
1104     }
1105
1106
1107     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1108         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1109
1110         h264_packedheader = 1;
1111         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1112         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1113
1114         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1115             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1116         }
1117
1118         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1119             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1120         }
1121
1122         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1123             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1124         }
1125
1126         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1127             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1128         }
1129
1130         enc_packed_header_idx = config_attrib_num;
1131         config_attrib_num++;
1132     }
1133
1134     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1135         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1136         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1137         config_attrib_num++;
1138     }
1139
1140     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1141         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1142     }
1143
1144     free(entrypoints);
1145     return 0;
1146 }
1147
1148 int QuickSyncEncoderImpl::setup_encode()
1149 {
1150     VAStatus va_status;
1151     VASurfaceID *tmp_surfaceid;
1152     int codedbuf_size, i;
1153     static VASurfaceID src_surface[SURFACE_NUM];
1154     static VASurfaceID ref_surface[SURFACE_NUM];
1155
1156     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1157             &config_attrib[0], config_attrib_num, &config_id);
1158     CHECK_VASTATUS(va_status, "vaCreateConfig");
1159
1160     /* create source surfaces */
1161     va_status = vaCreateSurfaces(va_dpy,
1162                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1163                                  &src_surface[0], SURFACE_NUM,
1164                                  NULL, 0);
1165     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1166
1167     /* create reference surfaces */
1168     va_status = vaCreateSurfaces(va_dpy,
1169                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1170                                  &ref_surface[0], SURFACE_NUM,
1171                                  NULL, 0);
1172     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1173
1174     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1175     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1176     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1177
1178     /* Create a context for this encode pipe */
1179     va_status = vaCreateContext(va_dpy, config_id,
1180                                 frame_width_mbaligned, frame_height_mbaligned,
1181                                 VA_PROGRESSIVE,
1182                                 tmp_surfaceid, 2 * SURFACE_NUM,
1183                                 &context_id);
1184     CHECK_VASTATUS(va_status, "vaCreateContext");
1185     free(tmp_surfaceid);
1186
1187     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1188
1189     for (i = 0; i < SURFACE_NUM; i++) {
1190         /* create coded buffer once for all
1191          * other VA buffers which won't be used again after vaRenderPicture.
1192          * so APP can always vaCreateBuffer for every frame
1193          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1194          * so VA won't maintain the coded buffer
1195          */
1196         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1197                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1198         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1199     }
1200
1201     /* create OpenGL objects */
1202     //glGenFramebuffers(SURFACE_NUM, fbos);
1203
1204     for (i = 0; i < SURFACE_NUM; i++) {
1205         glGenTextures(1, &gl_surfaces[i].y_tex);
1206         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1207
1208         if (!use_zerocopy) {
1209             // Create Y image.
1210             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1211             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1212
1213             // Create CbCr image.
1214             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1215             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1216
1217             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1218             // buffers, due to potentially differing pitch.
1219             glGenBuffers(1, &gl_surfaces[i].pbo);
1220             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1221             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1222             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1223             gl_surfaces[i].y_offset = 0;
1224             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1225             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1226             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1227             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1228         }
1229     }
1230
1231     for (i = 0; i < SURFACE_NUM; i++) {
1232         gl_surfaces[i].src_surface = src_surface[i];
1233         gl_surfaces[i].ref_surface = ref_surface[i];
1234     }
1235
1236     return 0;
1237 }
1238
1239 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1240 //
1241 //   2 1 0 [3] 4 8 9
1242 template<class T, class C>
1243 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1244 {
1245         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1246         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1247         sort(middle, end, less_than);
1248 }
1249
1250 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1251 {
1252     int i;
1253
1254     if (frame_type == FRAME_B)
1255         return;
1256
1257     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1258     numShortTerm++;
1259     if (numShortTerm > num_ref_frames)
1260         numShortTerm = num_ref_frames;
1261     for (i=numShortTerm-1; i>0; i--)
1262         ReferenceFrames[i] = ReferenceFrames[i-1];
1263     ReferenceFrames[0] = CurrentCurrPic;
1264
1265     current_frame_num++;
1266     if (current_frame_num > MaxFrameNum)
1267         current_frame_num = 0;
1268 }
1269
1270
1271 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1272 {
1273     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1274         return a.frame_idx > b.frame_idx;
1275     };
1276     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1277         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1278     };
1279     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1280         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1281     };
1282
1283     if (frame_type == FRAME_P) {
1284         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1285         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1286     } else if (frame_type == FRAME_B) {
1287         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1288         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1289
1290         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1291         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1292     }
1293
1294     return 0;
1295 }
1296
1297
1298 int QuickSyncEncoderImpl::render_sequence()
1299 {
1300     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1301     VAStatus va_status;
1302     VAEncMiscParameterBuffer *misc_param;
1303     VAEncMiscParameterRateControl *misc_rate_ctrl;
1304
1305     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1306     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1307     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1308     seq_param.bits_per_second = frame_bitrate;
1309
1310     seq_param.intra_period = intra_period;
1311     seq_param.intra_idr_period = intra_idr_period;
1312     seq_param.ip_period = ip_period;
1313
1314     seq_param.max_num_ref_frames = num_ref_frames;
1315     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1316     seq_param.time_scale = TIMEBASE * 2;
1317     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1318     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1319     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1320     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1321     seq_param.seq_fields.bits.chroma_format_idc = 1;
1322     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1323
1324     if (frame_width != frame_width_mbaligned ||
1325         frame_height != frame_height_mbaligned) {
1326         seq_param.frame_cropping_flag = 1;
1327         seq_param.frame_crop_left_offset = 0;
1328         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1329         seq_param.frame_crop_top_offset = 0;
1330         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1331     }
1332
1333     va_status = vaCreateBuffer(va_dpy, context_id,
1334                                VAEncSequenceParameterBufferType,
1335                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1336     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1337
1338     va_status = vaCreateBuffer(va_dpy, context_id,
1339                                VAEncMiscParameterBufferType,
1340                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1341                                1, NULL, &rc_param_buf);
1342     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1343
1344     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1345     misc_param->type = VAEncMiscParameterTypeRateControl;
1346     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1347     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1348     misc_rate_ctrl->bits_per_second = frame_bitrate;
1349     misc_rate_ctrl->target_percentage = 66;
1350     misc_rate_ctrl->window_size = 1000;
1351     misc_rate_ctrl->initial_qp = initial_qp;
1352     misc_rate_ctrl->min_qp = minimal_qp;
1353     misc_rate_ctrl->basic_unit_size = 0;
1354     vaUnmapBuffer(va_dpy, rc_param_buf);
1355
1356     render_id[0] = seq_param_buf;
1357     render_id[1] = rc_param_buf;
1358
1359     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1360
1361     return 0;
1362 }
1363
1364 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1365 {
1366     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1367     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1368     int PicOrderCntMsb, TopFieldOrderCnt;
1369
1370     if (frame_type == FRAME_IDR)
1371         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1372     else {
1373         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1374         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1375     }
1376
1377     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1378         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1379         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1380     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1381              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1382         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1383     else
1384         PicOrderCntMsb = prevPicOrderCntMsb;
1385
1386     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1387
1388     if (frame_type != FRAME_B) {
1389         PicOrderCntMsb_ref = PicOrderCntMsb;
1390         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1391     }
1392
1393     return TopFieldOrderCnt;
1394 }
1395
1396 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1397 {
1398     VABufferID pic_param_buf;
1399     VAStatus va_status;
1400     int i = 0;
1401
1402     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1403     pic_param.CurrPic.frame_idx = current_frame_num;
1404     pic_param.CurrPic.flags = 0;
1405     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1406     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1407     CurrentCurrPic = pic_param.CurrPic;
1408
1409     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1410     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1411         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1412         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1413     }
1414
1415     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1416     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1417     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1418     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1419     pic_param.frame_num = current_frame_num;
1420     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1421     pic_param.last_picture = false;  // FIXME
1422     pic_param.pic_init_qp = initial_qp;
1423
1424     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1425                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1426     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1427
1428     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1429
1430     return 0;
1431 }
1432
1433 int QuickSyncEncoderImpl::render_packedsequence()
1434 {
1435     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1436     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1437     unsigned int length_in_bits;
1438     unsigned char *packedseq_buffer = NULL;
1439     VAStatus va_status;
1440
1441     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1442
1443     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1444
1445     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1446     packedheader_param_buffer.has_emulation_bytes = 0;
1447     va_status = vaCreateBuffer(va_dpy,
1448                                context_id,
1449                                VAEncPackedHeaderParameterBufferType,
1450                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1451                                &packedseq_para_bufid);
1452     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1453
1454     va_status = vaCreateBuffer(va_dpy,
1455                                context_id,
1456                                VAEncPackedHeaderDataBufferType,
1457                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1458                                &packedseq_data_bufid);
1459     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1460
1461     render_id[0] = packedseq_para_bufid;
1462     render_id[1] = packedseq_data_bufid;
1463     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1464
1465     free(packedseq_buffer);
1466
1467     return 0;
1468 }
1469
1470
1471 int QuickSyncEncoderImpl::render_packedpicture()
1472 {
1473     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1474     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1475     unsigned int length_in_bits;
1476     unsigned char *packedpic_buffer = NULL;
1477     VAStatus va_status;
1478
1479     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1480     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1481     packedheader_param_buffer.bit_length = length_in_bits;
1482     packedheader_param_buffer.has_emulation_bytes = 0;
1483
1484     va_status = vaCreateBuffer(va_dpy,
1485                                context_id,
1486                                VAEncPackedHeaderParameterBufferType,
1487                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1488                                &packedpic_para_bufid);
1489     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1490
1491     va_status = vaCreateBuffer(va_dpy,
1492                                context_id,
1493                                VAEncPackedHeaderDataBufferType,
1494                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1495                                &packedpic_data_bufid);
1496     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1497
1498     render_id[0] = packedpic_para_bufid;
1499     render_id[1] = packedpic_data_bufid;
1500     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1501
1502     free(packedpic_buffer);
1503
1504     return 0;
1505 }
1506
1507 void QuickSyncEncoderImpl::render_packedslice()
1508 {
1509     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1510     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1511     unsigned int length_in_bits;
1512     unsigned char *packedslice_buffer = NULL;
1513     VAStatus va_status;
1514
1515     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1516     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1517     packedheader_param_buffer.bit_length = length_in_bits;
1518     packedheader_param_buffer.has_emulation_bytes = 0;
1519
1520     va_status = vaCreateBuffer(va_dpy,
1521                                context_id,
1522                                VAEncPackedHeaderParameterBufferType,
1523                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1524                                &packedslice_para_bufid);
1525     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1526
1527     va_status = vaCreateBuffer(va_dpy,
1528                                context_id,
1529                                VAEncPackedHeaderDataBufferType,
1530                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1531                                &packedslice_data_bufid);
1532     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1533
1534     render_id[0] = packedslice_para_bufid;
1535     render_id[1] = packedslice_data_bufid;
1536     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1537
1538     free(packedslice_buffer);
1539 }
1540
1541 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1542 {
1543     VABufferID slice_param_buf;
1544     VAStatus va_status;
1545     int i;
1546
1547     update_RefPicList(frame_type);
1548
1549     /* one frame, one slice */
1550     slice_param.macroblock_address = 0;
1551     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1552     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1553     if (frame_type == FRAME_IDR) {
1554         if (encoding_frame_num != 0)
1555             ++slice_param.idr_pic_id;
1556     } else if (frame_type == FRAME_P) {
1557         int refpiclist0_max = h264_maxref & 0xffff;
1558         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1559
1560         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1561             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1562             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1563         }
1564     } else if (frame_type == FRAME_B) {
1565         int refpiclist0_max = h264_maxref & 0xffff;
1566         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1567
1568         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1569         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1570             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1571             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1572         }
1573
1574         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1575         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1576             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1577             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1578         }
1579     }
1580
1581     slice_param.slice_alpha_c0_offset_div2 = 0;
1582     slice_param.slice_beta_offset_div2 = 0;
1583     slice_param.direct_spatial_mv_pred_flag = 1;
1584     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1585
1586
1587     if (h264_packedheader &&
1588         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1589         render_packedslice();
1590
1591     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1592                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1593     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1594
1595     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1596
1597     return 0;
1598 }
1599
1600
1601
1602 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1603 {
1604         VACodedBufferSegment *buf_list = NULL;
1605         VAStatus va_status;
1606
1607         string data;
1608
1609         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1610         CHECK_VASTATUS(va_status, "vaMapBuffer");
1611         while (buf_list != NULL) {
1612                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1613                 buf_list = (VACodedBufferSegment *) buf_list->next;
1614         }
1615         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1616
1617         {
1618                 // Add video.
1619                 AVPacket pkt;
1620                 memset(&pkt, 0, sizeof(pkt));
1621                 pkt.buf = nullptr;
1622                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1623                 pkt.size = data.size();
1624                 pkt.stream_index = 0;
1625                 if (task.frame_type == FRAME_IDR) {
1626                         pkt.flags = AV_PKT_FLAG_KEY;
1627                 } else {
1628                         pkt.flags = 0;
1629                 }
1630                 pkt.duration = task.duration;
1631                 if (file_mux) {
1632                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1633                 }
1634                 if (!global_flags.uncompressed_video_to_http &&
1635                     !global_flags.x264_video_to_http) {
1636                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1637                 }
1638         }
1639         // Encode and add all audio frames up to and including the pts of this video frame.
1640         for ( ;; ) {
1641                 int64_t audio_pts;
1642                 vector<float> audio;
1643                 {
1644                         unique_lock<mutex> lock(frame_queue_mutex);
1645                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1646                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1647                         auto it = pending_audio_frames.begin();
1648                         if (it->first > task.pts) break;
1649                         audio_pts = it->first;
1650                         audio = move(it->second);
1651                         pending_audio_frames.erase(it);
1652                 }
1653
1654                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1655                 if (stream_audio_encoder) {
1656                         stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1657                 }
1658
1659                 if (audio_pts == task.pts) break;
1660         }
1661 }
1662
1663
1664 // this is weird. but it seems to put a new frame onto the queue
1665 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1666 {
1667         unique_lock<mutex> lock(storage_task_queue_mutex);
1668         storage_task_queue.push(move(task));
1669         storage_task_queue_changed.notify_all();
1670 }
1671
1672 void QuickSyncEncoderImpl::storage_task_thread()
1673 {
1674         for ( ;; ) {
1675                 storage_task current;
1676                 {
1677                         // wait until there's an encoded frame
1678                         unique_lock<mutex> lock(storage_task_queue_mutex);
1679                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1680                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1681                         current = move(storage_task_queue.front());
1682                         storage_task_queue.pop();
1683                 }
1684
1685                 VAStatus va_status;
1686
1687                 // waits for data, then saves it to disk.
1688                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1689                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1690                 save_codeddata(move(current));
1691
1692                 {
1693                         unique_lock<mutex> lock(storage_task_queue_mutex);
1694                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1695                         storage_task_queue_changed.notify_all();
1696                 }
1697         }
1698 }
1699
1700 int QuickSyncEncoderImpl::release_encode()
1701 {
1702         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1703                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1704                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1705                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1706
1707                 if (!use_zerocopy) {
1708                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1709                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1710                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1711                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1712                 }
1713                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1714                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1715         }
1716
1717         vaDestroyContext(va_dpy, context_id);
1718         vaDestroyConfig(va_dpy, config_id);
1719
1720         return 0;
1721 }
1722
1723 int QuickSyncEncoderImpl::deinit_va()
1724 {
1725     vaTerminate(va_dpy);
1726
1727     va_close_display(va_dpy);
1728
1729     return 0;
1730 }
1731
1732 namespace {
1733
1734 }  // namespace
1735
1736 QuickSyncEncoderImpl::QuickSyncEncoderImpl(QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux)
1737         : current_storage_frame(0), surface(surface), stream_mux(stream_mux), frame_width(width), frame_height(height)
1738 {
1739         if (global_flags.stream_audio_codec_name.empty()) {
1740                 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, { file_mux.get(), stream_mux }));
1741         } else {
1742                 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, { file_mux.get() }));
1743                 stream_audio_encoder.reset(new AudioEncoder(global_flags.stream_audio_codec_name, global_flags.stream_audio_codec_bitrate, { stream_mux }));
1744         }
1745
1746         frame_width_mbaligned = (frame_width + 15) & (~15);
1747         frame_height_mbaligned = (frame_height + 15) & (~15);
1748
1749         //print_input();
1750
1751         if (global_flags.uncompressed_video_to_http ||
1752             global_flags.x264_video_to_http) {
1753                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1754         }
1755         if (global_flags.x264_video_to_http) {
1756                 x264_encoder.reset(new X264Encoder(stream_mux));
1757         }
1758
1759         init_va(va_display);
1760         setup_encode();
1761
1762         // No frames are ready yet.
1763         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1764
1765         memset(&seq_param, 0, sizeof(seq_param));
1766         memset(&pic_param, 0, sizeof(pic_param));
1767         memset(&slice_param, 0, sizeof(slice_param));
1768
1769         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1770
1771         encode_thread = thread([this]{
1772                 //SDL_GL_MakeCurrent(window, context);
1773                 QOpenGLContext *context = create_context(this->surface);
1774                 eglBindAPI(EGL_OPENGL_API);
1775                 if (!make_current(context, this->surface)) {
1776                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1777                                 eglGetError());
1778                         exit(1);
1779                 }
1780                 encode_thread_func();
1781         });
1782 }
1783
1784 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1785 {
1786         shutdown();
1787 }
1788
1789 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1790 {
1791         assert(!is_shutdown);
1792         {
1793                 // Wait until this frame slot is done encoding.
1794                 unique_lock<mutex> lock(storage_task_queue_mutex);
1795                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1796                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1797                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1798                 }
1799                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1800                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1801                 if (storage_thread_should_quit) return false;
1802         }
1803
1804         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1805         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1806         *y_tex = surf->y_tex;
1807         *cbcr_tex = surf->cbcr_tex;
1808
1809         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1810         CHECK_VASTATUS(va_status, "vaDeriveImage");
1811
1812         if (use_zerocopy) {
1813                 VABufferInfo buf_info;
1814                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1815                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1816                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1817
1818                 // Create Y image.
1819                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1820                 EGLint y_attribs[] = {
1821                         EGL_WIDTH, frame_width,
1822                         EGL_HEIGHT, frame_height,
1823                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1824                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1825                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1826                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1827                         EGL_NONE
1828                 };
1829
1830                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1831                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1832
1833                 // Associate Y image to a texture.
1834                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1835                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1836
1837                 // Create CbCr image.
1838                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1839                 EGLint cbcr_attribs[] = {
1840                         EGL_WIDTH, frame_width,
1841                         EGL_HEIGHT, frame_height,
1842                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1843                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1844                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1845                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1846                         EGL_NONE
1847                 };
1848
1849                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1850                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1851
1852                 // Associate CbCr image to a texture.
1853                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1854                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1855         }
1856
1857         return true;
1858 }
1859
1860 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1861 {
1862         assert(!is_shutdown);
1863         {
1864                 unique_lock<mutex> lock(frame_queue_mutex);
1865                 pending_audio_frames[pts] = move(audio);
1866         }
1867         frame_queue_nonempty.notify_all();
1868 }
1869
1870 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1871 {
1872         assert(!is_shutdown);
1873
1874         if (!use_zerocopy) {
1875                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1876
1877                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1878                 check_error();
1879
1880                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1881                 check_error();
1882
1883                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1884                 check_error();
1885                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1886                 check_error();
1887
1888                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1889                 check_error();
1890                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1891                 check_error();
1892
1893                 glBindTexture(GL_TEXTURE_2D, 0);
1894                 check_error();
1895                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1896                 check_error();
1897
1898                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1899                 check_error();
1900         }
1901
1902         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1903         check_error();
1904         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1905         check_error();
1906
1907         {
1908                 unique_lock<mutex> lock(frame_queue_mutex);
1909                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1910                 ++current_storage_frame;
1911         }
1912         frame_queue_nonempty.notify_all();
1913         return fence;
1914 }
1915
1916 void QuickSyncEncoderImpl::shutdown()
1917 {
1918         if (is_shutdown) {
1919                 return;
1920         }
1921
1922         {
1923                 unique_lock<mutex> lock(frame_queue_mutex);
1924                 encode_thread_should_quit = true;
1925                 frame_queue_nonempty.notify_all();
1926         }
1927         encode_thread.join();
1928         x264_encoder.reset();
1929         {
1930                 unique_lock<mutex> lock(storage_task_queue_mutex);
1931                 storage_thread_should_quit = true;
1932                 frame_queue_nonempty.notify_all();
1933                 storage_task_queue_changed.notify_all();
1934         }
1935         storage_thread.join();
1936         encode_remaining_audio();
1937
1938         release_encode();
1939         deinit_va();
1940         is_shutdown = true;
1941 }
1942
1943 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1944 {
1945         AVFormatContext *avctx = avformat_alloc_context();
1946         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1947         assert(filename.size() < sizeof(avctx->filename) - 1);
1948         strcpy(avctx->filename, filename.c_str());
1949
1950         string url = "file:" + filename;
1951         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1952         if (ret < 0) {
1953                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1954                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1955                 exit(1);
1956         }
1957
1958         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, file_audio_encoder->get_codec(), TIMEBASE, DEFAULT_AUDIO_OUTPUT_BIT_RATE, nullptr));
1959 }
1960
1961 void QuickSyncEncoderImpl::close_output_file()
1962 {
1963         file_mux.reset();
1964 }
1965
1966 void QuickSyncEncoderImpl::encode_thread_func()
1967 {
1968         int64_t last_dts = -1;
1969         int gop_start_display_frame_num = 0;
1970         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1971                 PendingFrame frame;
1972                 int pts_lag;
1973                 int frame_type, display_frame_num;
1974                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1975                                        &display_frame_num, &frame_type, &pts_lag);
1976                 if (frame_type == FRAME_IDR) {
1977                         numShortTerm = 0;
1978                         current_frame_num = 0;
1979                         gop_start_display_frame_num = display_frame_num;
1980                 }
1981
1982                 {
1983                         unique_lock<mutex> lock(frame_queue_mutex);
1984                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1985                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1986                         });
1987                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1988                                 // We have queued frames that were supposed to be B-frames,
1989                                 // but will be no P-frame to encode them against. Encode them all
1990                                 // as P-frames instead. Note that this happens under the mutex,
1991                                 // but nobody else uses it at this point, since we're shutting down,
1992                                 // so there's no contention.
1993                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1994                                 return;
1995                         } else {
1996                                 frame = move(pending_video_frames[display_frame_num]);
1997                                 pending_video_frames.erase(display_frame_num);
1998                         }
1999                 }
2000
2001                 // Determine the dts of this frame.
2002                 int64_t dts;
2003                 if (pts_lag == -1) {
2004                         assert(last_dts != -1);
2005                         dts = last_dts + (TIMEBASE / MAX_FPS);
2006                 } else {
2007                         dts = frame.pts - pts_lag;
2008                 }
2009                 last_dts = dts;
2010
2011                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2012         }
2013 }
2014
2015 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2016 {
2017         if (pending_video_frames.empty()) {
2018                 return;
2019         }
2020
2021         for (auto &pending_frame : pending_video_frames) {
2022                 int display_frame_num = pending_frame.first;
2023                 assert(display_frame_num > 0);
2024                 PendingFrame frame = move(pending_frame.second);
2025                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2026                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2027                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2028                 last_dts = dts;
2029         }
2030
2031         if (global_flags.uncompressed_video_to_http ||
2032             global_flags.x264_video_to_http) {
2033                 // Add frames left in reorderer.
2034                 while (!reorderer->empty()) {
2035                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2036                         if (global_flags.uncompressed_video_to_http) {
2037                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2038                         } else {
2039                                 assert(global_flags.x264_video_to_http);
2040                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2041                         }
2042                 }
2043         }
2044 }
2045
2046 void QuickSyncEncoderImpl::encode_remaining_audio()
2047 {
2048         // This really ought to be empty by now, but just to be sure...
2049         for (auto &pending_frame : pending_audio_frames) {
2050                 int64_t audio_pts = pending_frame.first;
2051                 vector<float> audio = move(pending_frame.second);
2052
2053                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2054                 if (stream_audio_encoder) {
2055                         stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2056                 }
2057         }
2058         pending_audio_frames.clear();
2059
2060         // Encode any leftover audio in the queues, and also any delayed frames.
2061         file_audio_encoder->encode_last_audio();
2062         if (stream_audio_encoder) {
2063                 stream_audio_encoder->encode_last_audio();
2064         }
2065 }
2066
2067 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2068 {
2069         AVPacket pkt;
2070         memset(&pkt, 0, sizeof(pkt));
2071         pkt.buf = nullptr;
2072         pkt.data = const_cast<uint8_t *>(data);
2073         pkt.size = frame_width * frame_height * 2;
2074         pkt.stream_index = 0;
2075         pkt.flags = AV_PKT_FLAG_KEY;
2076         pkt.duration = duration;
2077         stream_mux->add_packet(pkt, pts, pts);
2078 }
2079
2080 namespace {
2081
2082 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2083 {
2084         if (src_width == dst_pitch) {
2085                 memcpy(dst, src, src_width * height);
2086         } else {
2087                 for (size_t y = 0; y < height; ++y) {
2088                         const uint8_t *sptr = src + y * src_width;
2089                         uint8_t *dptr = dst + y * dst_pitch;
2090                         memcpy(dptr, sptr, src_width);
2091                 }
2092         }
2093 }
2094
2095 }  // namespace
2096
2097 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2098                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2099 {
2100         // Wait for the GPU to be done with the frame.
2101         GLenum sync_status;
2102         do {
2103                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2104                 check_error();
2105         } while (sync_status == GL_TIMEOUT_EXPIRED);
2106         assert(sync_status != GL_WAIT_FAILED);
2107
2108         // Release back any input frames we needed to render this frame.
2109         frame.input_frames.clear();
2110
2111         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2112         VAStatus va_status;
2113
2114         if (use_zerocopy) {
2115                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2116                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2117                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2118                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2119         } else {
2120                 unsigned char *surface_p = nullptr;
2121                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2122
2123                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2124                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2125
2126                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2127                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2128
2129                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2130                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2131
2132                 if (global_flags.uncompressed_video_to_http ||
2133                     global_flags.x264_video_to_http) {
2134                         // Add uncompressed video. (Note that pts == dts here.)
2135                         // Delay needs to match audio.
2136                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2137                         if (output_frame.data != nullptr) {
2138                                 if (global_flags.uncompressed_video_to_http) {
2139                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2140                                 } else {
2141                                         assert(global_flags.x264_video_to_http);
2142                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2143                                 }
2144                         }
2145                 }
2146         }
2147
2148         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2149         CHECK_VASTATUS(va_status, "vaDestroyImage");
2150
2151         // Schedule the frame for encoding.
2152         VASurfaceID va_surface = surf->src_surface;
2153         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2154         CHECK_VASTATUS(va_status, "vaBeginPicture");
2155
2156         if (frame_type == FRAME_IDR) {
2157                 render_sequence();
2158                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2159                 if (h264_packedheader) {
2160                         render_packedsequence();
2161                         render_packedpicture();
2162                 }
2163         } else {
2164                 //render_sequence();
2165                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2166         }
2167         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2168
2169         va_status = vaEndPicture(va_dpy, context_id);
2170         CHECK_VASTATUS(va_status, "vaEndPicture");
2171
2172         // so now the data is done encoding (well, async job kicked off)...
2173         // we send that to the storage thread
2174         storage_task tmp;
2175         tmp.display_order = display_frame_num;
2176         tmp.frame_type = frame_type;
2177         tmp.pts = pts;
2178         tmp.dts = dts;
2179         tmp.duration = duration;
2180         storage_task_enqueue(move(tmp));
2181
2182         update_ReferenceFrames(frame_type);
2183 }
2184
2185 // Proxy object.
2186 QuickSyncEncoder::QuickSyncEncoder(QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux)
2187         : impl(new QuickSyncEncoderImpl(surface, va_display, width, height, stream_mux)) {}
2188
2189 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2190 QuickSyncEncoder::~QuickSyncEncoder() {}
2191
2192 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2193 {
2194         impl->add_audio(pts, audio);
2195 }
2196
2197 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2198 {
2199         return impl->begin_frame(y_tex, cbcr_tex);
2200 }
2201
2202 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2203 {
2204         return impl->end_frame(pts, duration, input_frames);
2205 }
2206
2207 void QuickSyncEncoder::shutdown()
2208 {
2209         impl->shutdown();
2210 }
2211
2212 void QuickSyncEncoder::open_output_file(const std::string &filename)
2213 {
2214         impl->open_output_file(filename);
2215 }
2216
2217 void QuickSyncEncoder::close_output_file()
2218 {
2219         impl->close_output_file();
2220 }