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 extern "C" {
  11 #include <libavcodec/avcodec.h>
  12 #include <libavformat/avformat.h>
  13 #include <libavresample/avresample.h>
  14 #include <libavutil/channel_layout.h>
  15 #include <libavutil/frame.h>
  16 #include <libavutil/rational.h>
  17 #include <libavutil/samplefmt.h>
  18 #include <libavutil/opt.h>
  19 }
  20 #include <libdrm/drm_fourcc.h>
  21 #include <stdio.h>
  22 #include <stdlib.h>
  23 #include <string.h>
  24 #include <fcntl.h>
  25 #include <va/va.h>
  26 #include <va/va_drm.h>
  27 #include <va/va_drmcommon.h>
  28 #include <va/va_enc_h264.h>
  29 #include <va/va_x11.h>
  30 #include <algorithm>
  31 #include <condition_variable>
  32 #include <cstdint>
  33 #include <map>
  34 #include <memory>
  35 #include <mutex>
  36 #include <queue>
  37 #include <string>
  38 #include <thread>
  39 #include <utility>
  40
  41 #include "context.h"
  42 #include "defs.h"
  43 #include "flags.h"
  44 #include "httpd.h"
  45 #include "mux.h"
  46 #include "timebase.h"
  47 #include "x264_encoder.h"
  48
  49 using namespace std;
  50
  51 class QOpenGLContext;
  52 class QSurface;
  53
  54 #define CHECK_VASTATUS(va_status, func)                                 \
  55     if (va_status != VA_STATUS_SUCCESS) {                               \
  56         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  57         exit(1);                                                        \
  58     }
  59
  60 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  61
  62 //#include "loadsurface.h"
  63
  64 #define NAL_REF_IDC_NONE        0
  65 #define NAL_REF_IDC_LOW         1
  66 #define NAL_REF_IDC_MEDIUM      2
  67 #define NAL_REF_IDC_HIGH        3
  68
  69 #define NAL_NON_IDR             1
  70 #define NAL_IDR                 5
  71 #define NAL_SPS                 7
  72 #define NAL_PPS                 8
  73 #define NAL_SEI                 6
  74
  75 #define SLICE_TYPE_P            0
  76 #define SLICE_TYPE_B            1
  77 #define SLICE_TYPE_I            2
  78 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  79 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  80 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  81
  82
  83 #define ENTROPY_MODE_CAVLC      0
  84 #define ENTROPY_MODE_CABAC      1
  85
  86 #define PROFILE_IDC_BASELINE    66
  87 #define PROFILE_IDC_MAIN        77
  88 #define PROFILE_IDC_HIGH        100
  89
  90 #define BITSTREAM_ALLOCATE_STEPPING     4096
  91 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  92 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  93 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  94
  95 static constexpr unsigned int MaxFrameNum = (2<<16);
  96 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  97 static constexpr unsigned int Log2MaxFrameNum = 16;
  98 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  99 static constexpr int rc_default_modes[] = {  // Priority list of modes.
 100     VA_RC_VBR,
 101     VA_RC_CQP,
 102     VA_RC_VBR_CONSTRAINED,
 103     VA_RC_CBR,
 104     VA_RC_VCM,
 105     VA_RC_NONE,
 106 };
 107
 108 /* thread to save coded data */
 109 #define SRC_SURFACE_FREE        0
 110 #define SRC_SURFACE_IN_ENCODING 1
 111
 112 struct __bitstream {
 113     unsigned int *buffer;
 114     int bit_offset;
 115     int max_size_in_dword;
 116 };
 117 typedef struct __bitstream bitstream;
 118
 119 using namespace std;
 120
 121 // H.264 video comes out in encoding order (e.g. with two B-frames:
 122 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 123 // come in the right order. Since we do everything, including waiting
 124 // for the frames to come out of OpenGL, in encoding order, we need
 125 // a reordering buffer for uncompressed frames so that they come out
 126 // correctly. We go the super-lazy way of not making it understand
 127 // anything about the true order (which introduces some extra latency,
 128 // though); we know that for N B-frames we need at most (N-1) frames
 129 // in the reorder buffer, and can just sort on that.
 130 //
 131 // The class also deals with keeping a freelist as needed.
 132 class FrameReorderer {
 133 public:
 134         FrameReorderer(unsigned queue_length, int width, int height);
 135
 136         struct Frame {
 137                 int64_t pts, duration;
 138                 uint8_t *data;
 139
 140                 // Invert to get the smallest pts first.
 141                 bool operator< (const Frame &other) const { return pts > other.pts; }
 142         };
 143
 144         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 145         // Does _not_ take ownership of data; a copy is taken if needed.
 146         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 147         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 148         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data);
 149
 150         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 151         Frame get_first_frame();
 152
 153         bool empty() const { return frames.empty(); }
 154
 155 private:
 156         unsigned queue_length;
 157         int width, height;
 158
 159         priority_queue<Frame> frames;
 160         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 161
 162         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 163         // except priority_queue doesn't work well with movable-only types.
 164         vector<unique_ptr<uint8_t[]>> owner;
 165 };
 166
 167 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 168     : queue_length(queue_length), width(width), height(height)
 169 {
 170         for (unsigned i = 0; i < queue_length; ++i) {
 171                 owner.emplace_back(new uint8_t[width * height * 2]);
 172                 freelist.push(owner.back().get());
 173         }
 174 }
 175
 176 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data)
 177 {
 178         if (queue_length == 0) {
 179                 return Frame{pts, duration, data};
 180         }
 181
 182         assert(!freelist.empty());
 183         uint8_t *storage = freelist.top();
 184         freelist.pop();
 185         memcpy(storage, data, width * height * 2);
 186         frames.push(Frame{pts, duration, storage});
 187
 188         if (frames.size() >= queue_length) {
 189                 return get_first_frame();
 190         } else {
 191                 return Frame{-1, -1, nullptr};
 192         }
 193 }
 194
 195 FrameReorderer::Frame FrameReorderer::get_first_frame()
 196 {
 197         assert(!frames.empty());
 198         Frame storage = frames.top();
 199         frames.pop();
 200         freelist.push(storage.data);
 201         return storage;
 202 }
 203
 204 class QuickSyncEncoderImpl : public KeyFrameSignalReceiver {
 205 public:
 206         QuickSyncEncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
 207         ~QuickSyncEncoderImpl();
 208         void add_audio(int64_t pts, vector<float> audio);
 209         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 210         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
 211         void shutdown();
 212         void open_output_file(const std::string &filename);
 213         void close_output_file();
 214
 215         virtual void signal_keyframe() override {
 216                 stream_mux_writing_keyframes = true;
 217         }
 218
 219 private:
 220         struct storage_task {
 221                 unsigned long long display_order;
 222                 int frame_type;
 223                 vector<float> audio;
 224                 int64_t pts, dts, duration;
 225         };
 226         struct PendingFrame {
 227                 RefCountedGLsync fence;
 228                 vector<RefCountedFrame> input_frames;
 229                 int64_t pts, duration;
 230         };
 231
 232         // So we never get negative dts.
 233         int64_t global_delay() const {
 234                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 235         }
 236
 237         void encode_thread_func();
 238         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 239         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 240         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 241                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 242         void storage_task_thread();
 243         void encode_audio(const vector<float> &audio,
 244                           vector<float> *audio_queue,
 245                           int64_t audio_pts,
 246                           AVCodecContext *ctx,
 247                           AVAudioResampleContext *resampler,
 248                           const vector<Mux *> &muxes);
 249         void encode_audio_one_frame(const float *audio,
 250                                     size_t num_samples,  // In each channel.
 251                                     int64_t audio_pts,
 252                                     AVCodecContext *ctx,
 253                                     AVAudioResampleContext *resampler,
 254                                     const vector<Mux *> &muxes);
 255         void encode_last_audio(vector<float> *audio_queue,
 256                                int64_t audio_pts,
 257                                AVCodecContext *ctx,
 258                                AVAudioResampleContext *resampler,
 259                                const vector<Mux *> &muxes);
 260         void encode_remaining_audio();
 261         void storage_task_enqueue(storage_task task);
 262         void save_codeddata(storage_task task);
 263         int render_packedsequence();
 264         int render_packedpicture();
 265         void render_packedslice();
 266         int render_sequence();
 267         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 268         void sps_rbsp(bitstream *bs);
 269         void pps_rbsp(bitstream *bs);
 270         int build_packed_pic_buffer(unsigned char **header_buffer);
 271         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 272         void slice_header(bitstream *bs);
 273         int build_packed_seq_buffer(unsigned char **header_buffer);
 274         int build_packed_slice_buffer(unsigned char **header_buffer);
 275         int init_va(const string &va_display);
 276         int deinit_va();
 277         void enable_zerocopy_if_possible();
 278         VADisplay va_open_display(const string &va_display);
 279         void va_close_display(VADisplay va_dpy);
 280         int setup_encode();
 281         int release_encode();
 282         void update_ReferenceFrames(int frame_type);
 283         int update_RefPicList(int frame_type);
 284         void open_output_stream();
 285         void close_output_stream();
 286         static int write_packet_thunk(void *opaque, uint8_t *buf, int buf_size);
 287         int write_packet(uint8_t *buf, int buf_size);
 288
 289         bool is_shutdown = false;
 290         bool use_zerocopy;
 291         int drm_fd = -1;
 292
 293         thread encode_thread, storage_thread;
 294
 295         mutex storage_task_queue_mutex;
 296         condition_variable storage_task_queue_changed;
 297         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 298         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 299         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 300
 301         mutex frame_queue_mutex;
 302         condition_variable frame_queue_nonempty;
 303         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 304
 305         int current_storage_frame;
 306
 307         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 308         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 309         int64_t last_audio_pts = 0;  // The first pts after all audio we've encoded.
 310         QSurface *surface;
 311
 312         AVCodecContext *context_audio_file;
 313         AVCodecContext *context_audio_stream = nullptr;  // nullptr = don't code separate audio for stream.
 314
 315         AVAudioResampleContext *resampler_audio_file = nullptr;
 316         AVAudioResampleContext *resampler_audio_stream = nullptr;
 317
 318         vector<float> audio_queue_file;
 319         vector<float> audio_queue_stream;
 320
 321         unique_ptr<Mux> stream_mux;  // To HTTP.
 322         unique_ptr<Mux> file_mux;  // To local disk.
 323
 324         // While Mux object is constructing, <stream_mux_writing_header> is true,
 325         // and the header is being collected into stream_mux_header.
 326         bool stream_mux_writing_header;
 327         string stream_mux_header;
 328
 329         bool stream_mux_writing_keyframes = false;
 330
 331         AVFrame *audio_frame = nullptr;
 332         HTTPD *httpd;
 333         unique_ptr<FrameReorderer> reorderer;
 334         unique_ptr<X264Encoder> x264_encoder;  // nullptr if not using x264.
 335
 336         Display *x11_display = nullptr;
 337
 338         // Encoder parameters
 339         VADisplay va_dpy;
 340         VAProfile h264_profile = (VAProfile)~0;
 341         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 342         int config_attrib_num = 0, enc_packed_header_idx;
 343
 344         struct GLSurface {
 345                 VASurfaceID src_surface, ref_surface;
 346                 VABufferID coded_buf;
 347
 348                 VAImage surface_image;
 349                 GLuint y_tex, cbcr_tex;
 350
 351                 // Only if use_zerocopy == true.
 352                 EGLImage y_egl_image, cbcr_egl_image;
 353
 354                 // Only if use_zerocopy == false.
 355                 GLuint pbo;
 356                 uint8_t *y_ptr, *cbcr_ptr;
 357                 size_t y_offset, cbcr_offset;
 358         };
 359         GLSurface gl_surfaces[SURFACE_NUM];
 360
 361         VAConfigID config_id;
 362         VAContextID context_id;
 363         VAEncSequenceParameterBufferH264 seq_param;
 364         VAEncPictureParameterBufferH264 pic_param;
 365         VAEncSliceParameterBufferH264 slice_param;
 366         VAPictureH264 CurrentCurrPic;
 367         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 368
 369         // Static quality settings.
 370         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 371         static constexpr unsigned int num_ref_frames = 2;
 372         static constexpr int initial_qp = 15;
 373         static constexpr int minimal_qp = 0;
 374         static constexpr int intra_period = 30;
 375         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 376
 377         // Quality settings that are meant to be static, but might be overridden
 378         // by the profile.
 379         int constraint_set_flag = 0;
 380         int h264_packedheader = 0; /* support pack header? */
 381         int h264_maxref = (1<<16|1);
 382         int h264_entropy_mode = 1; /* cabac */
 383         int ip_period = 3;
 384
 385         int rc_mode = -1;
 386         unsigned int current_frame_num = 0;
 387         unsigned int numShortTerm = 0;
 388
 389         int frame_width;
 390         int frame_height;
 391         int frame_width_mbaligned;
 392         int frame_height_mbaligned;
 393 };
 394
 395 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 396 // but if we don't delete it here, we get leaks. The GStreamer implementation
 397 // does the same.
 398 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 399 {
 400     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 401     CHECK_VASTATUS(va_status, "vaRenderPicture");
 402
 403     for (int i = 0; i < num_buffers; ++i) {
 404         va_status = vaDestroyBuffer(dpy, buffers[i]);
 405         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 406     }
 407 }
 408
 409 static unsigned int
 410 va_swap32(unsigned int val)
 411 {
 412     unsigned char *pval = (unsigned char *)&val;
 413
 414     return ((pval[0] << 24)     |
 415             (pval[1] << 16)     |
 416             (pval[2] << 8)      |
 417             (pval[3] << 0));
 418 }
 419
 420 static void
 421 bitstream_start(bitstream *bs)
 422 {
 423     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 424     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 425     bs->bit_offset = 0;
 426 }
 427
 428 static void
 429 bitstream_end(bitstream *bs)
 430 {
 431     int pos = (bs->bit_offset >> 5);
 432     int bit_offset = (bs->bit_offset & 0x1f);
 433     int bit_left = 32 - bit_offset;
 434
 435     if (bit_offset) {
 436         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 437     }
 438 }
 439
 440 static void
 441 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 442 {
 443     int pos = (bs->bit_offset >> 5);
 444     int bit_offset = (bs->bit_offset & 0x1f);
 445     int bit_left = 32 - bit_offset;
 446
 447     if (!size_in_bits)
 448         return;
 449
 450     bs->bit_offset += size_in_bits;
 451
 452     if (bit_left > size_in_bits) {
 453         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 454     } else {
 455         size_in_bits -= bit_left;
 456         if (bit_left >= 32) {
 457             bs->buffer[pos] = (val >> size_in_bits);
 458         } else {
 459             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 460         }
 461         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 462
 463         if (pos + 1 == bs->max_size_in_dword) {
 464             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 465             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 466         }
 467
 468         bs->buffer[pos + 1] = val;
 469     }
 470 }
 471
 472 static void
 473 bitstream_put_ue(bitstream *bs, unsigned int val)
 474 {
 475     int size_in_bits = 0;
 476     int tmp_val = ++val;
 477
 478     while (tmp_val) {
 479         tmp_val >>= 1;
 480         size_in_bits++;
 481     }
 482
 483     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 484     bitstream_put_ui(bs, val, size_in_bits);
 485 }
 486
 487 static void
 488 bitstream_put_se(bitstream *bs, int val)
 489 {
 490     unsigned int new_val;
 491
 492     if (val <= 0)
 493         new_val = -2 * val;
 494     else
 495         new_val = 2 * val - 1;
 496
 497     bitstream_put_ue(bs, new_val);
 498 }
 499
 500 static void
 501 bitstream_byte_aligning(bitstream *bs, int bit)
 502 {
 503     int bit_offset = (bs->bit_offset & 0x7);
 504     int bit_left = 8 - bit_offset;
 505     int new_val;
 506
 507     if (!bit_offset)
 508         return;
 509
 510     assert(bit == 0 || bit == 1);
 511
 512     if (bit)
 513         new_val = (1 << bit_left) - 1;
 514     else
 515         new_val = 0;
 516
 517     bitstream_put_ui(bs, new_val, bit_left);
 518 }
 519
 520 static void
 521 rbsp_trailing_bits(bitstream *bs)
 522 {
 523     bitstream_put_ui(bs, 1, 1);
 524     bitstream_byte_aligning(bs, 0);
 525 }
 526
 527 static void nal_start_code_prefix(bitstream *bs)
 528 {
 529     bitstream_put_ui(bs, 0x00000001, 32);
 530 }
 531
 532 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 533 {
 534     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 535     bitstream_put_ui(bs, nal_ref_idc, 2);
 536     bitstream_put_ui(bs, nal_unit_type, 5);
 537 }
 538
 539 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 540 {
 541     int profile_idc = PROFILE_IDC_BASELINE;
 542
 543     if (h264_profile  == VAProfileH264High)
 544         profile_idc = PROFILE_IDC_HIGH;
 545     else if (h264_profile  == VAProfileH264Main)
 546         profile_idc = PROFILE_IDC_MAIN;
 547
 548     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 549     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 550     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 551     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 552     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 553     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 554     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 555     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 556
 557     if ( profile_idc == PROFILE_IDC_HIGH) {
 558         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 559         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 560         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 561         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 562         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 563     }
 564
 565     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 566     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 567
 568     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 569         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 570     else {
 571         assert(0);
 572     }
 573
 574     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 575     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 576
 577     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 578     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 579     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 580
 581     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 582         assert(0);
 583     }
 584
 585     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 586     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 587
 588     if (seq_param.frame_cropping_flag) {
 589         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 590         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 591         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 592         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 593     }
 594
 595     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 596     if ( false ) {
 597         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 598     } else {
 599         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 600         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 601         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 602         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 603         {
 604             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 605             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 606             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 607             {
 608                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 609                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 610                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 611             }
 612         }
 613         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 614         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 615         {
 616             bitstream_put_ui(bs, 1, 32);  // FPS
 617             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 618             bitstream_put_ui(bs, 1, 1);
 619         }
 620         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 621         {
 622             // hrd_parameters
 623             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 624             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 625             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 626
 627             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 628             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 629             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 630
 631             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 632             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 633             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 634             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 635         }
 636         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 637         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 638
 639         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 640         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 641     }
 642
 643     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 644 }
 645
 646
 647 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 648 {
 649     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 650     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 651
 652     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 653
 654     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 655
 656     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 657
 658     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 659     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 660
 661     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 662     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 663
 664     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 665     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 666     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 667
 668     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 669     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 670     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 671
 672     /* more_rbsp_data */
 673     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 674     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 675     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 676
 677     rbsp_trailing_bits(bs);
 678 }
 679
 680 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 681 {
 682     int first_mb_in_slice = slice_param.macroblock_address;
 683
 684     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 685     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 686     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 687     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 688
 689     /* frame_mbs_only_flag == 1 */
 690     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 691         /* FIXME: */
 692         assert(0);
 693     }
 694
 695     if (pic_param.pic_fields.bits.idr_pic_flag)
 696         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 697
 698     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 699         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 700         /* pic_order_present_flag == 0 */
 701     } else {
 702         /* FIXME: */
 703         assert(0);
 704     }
 705
 706     /* redundant_pic_cnt_present_flag == 0 */
 707     /* slice type */
 708     if (IS_P_SLICE(slice_param.slice_type)) {
 709         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 710
 711         if (slice_param.num_ref_idx_active_override_flag)
 712             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 713
 714         /* ref_pic_list_reordering */
 715         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 716     } else if (IS_B_SLICE(slice_param.slice_type)) {
 717         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 718
 719         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 720
 721         if (slice_param.num_ref_idx_active_override_flag) {
 722             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 723             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 724         }
 725
 726         /* ref_pic_list_reordering */
 727         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 728         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 729     }
 730
 731     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 732          IS_P_SLICE(slice_param.slice_type)) ||
 733         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 734          IS_B_SLICE(slice_param.slice_type))) {
 735         /* FIXME: fill weight/offset table */
 736         assert(0);
 737     }
 738
 739     /* dec_ref_pic_marking */
 740     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 741         unsigned char no_output_of_prior_pics_flag = 0;
 742         unsigned char long_term_reference_flag = 0;
 743         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 744
 745         if (pic_param.pic_fields.bits.idr_pic_flag) {
 746             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 747             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 748         } else {
 749             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 750         }
 751     }
 752
 753     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 754         !IS_I_SLICE(slice_param.slice_type))
 755         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 756
 757     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 758
 759     /* ignore for SP/SI */
 760
 761     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 762         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 763
 764         if (slice_param.disable_deblocking_filter_idc != 1) {
 765             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 766             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 767         }
 768     }
 769
 770     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 771         bitstream_byte_aligning(bs, 1);
 772     }
 773 }
 774
 775 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 776 {
 777     bitstream bs;
 778
 779     bitstream_start(&bs);
 780     nal_start_code_prefix(&bs);
 781     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 782     pps_rbsp(&bs);
 783     bitstream_end(&bs);
 784
 785     *header_buffer = (unsigned char *)bs.buffer;
 786     return bs.bit_offset;
 787 }
 788
 789 int
 790 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 791 {
 792     bitstream bs;
 793
 794     bitstream_start(&bs);
 795     nal_start_code_prefix(&bs);
 796     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 797     sps_rbsp(&bs);
 798     bitstream_end(&bs);
 799
 800     *header_buffer = (unsigned char *)bs.buffer;
 801     return bs.bit_offset;
 802 }
 803
 804 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 805 {
 806     bitstream bs;
 807     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 808     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 809
 810     bitstream_start(&bs);
 811     nal_start_code_prefix(&bs);
 812
 813     if (IS_I_SLICE(slice_param.slice_type)) {
 814         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 815     } else if (IS_P_SLICE(slice_param.slice_type)) {
 816         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 817     } else {
 818         assert(IS_B_SLICE(slice_param.slice_type));
 819         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 820     }
 821
 822     slice_header(&bs);
 823     bitstream_end(&bs);
 824
 825     *header_buffer = (unsigned char *)bs.buffer;
 826     return bs.bit_offset;
 827 }
 828
 829
 830 /*
 831   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 832   1) period between Frame #X and Frame #N = #X - #N
 833   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 834   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 835   4) intra_period and intra_idr_period take precedence over ip_period
 836   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 837      of I/IDR frames, see bellow examples
 838   -------------------------------------------------------------------
 839   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 840   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 841   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 842   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 843   1            1                ignored    IDR IDR IDR IDR...
 844   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 845   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 846   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 847                                               (PBB)(IBB)(PBB)(IBB)...
 848   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 849                                            IDRPPPPP IPPPPP IPPPPP...
 850   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 851                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 852                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 853                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 854                                            {IDR(PBB)(PBB)}                     (6/6/3)
 855                                            {IDR(PBB)(PBB)}.
 856 */
 857
 858 // General pts/dts strategy:
 859 //
 860 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 861 // bit tricky. We assume first of all that the frame rate never goes _above_
 862 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 863 // in at least this speed, as long at dts <= pts (the frame is not attempted
 864 // presented before it is decoded). Furthermore, we never have longer chains of
 865 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 866 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 867 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 868 // frame, with an order O <= C.)
 869 //
 870 // Many strategies are possible, but we establish these rules:
 871 //
 872 //  - Tip frames have dts = pts - (C-O)*N.
 873 //  - Non-tip frames have dts = dts_last + N.
 874 //
 875 // An example, with C=2 and N=10 and the data flow showed with arrows:
 876 //
 877 //        I  B  P  B  B  P
 878 //   pts: 30 40 50 60 70 80
 879 //        ↓  ↓     ↓
 880 //   dts: 10 30 20 60 50←40
 881 //         |  |  ↑        ↑
 882 //         `--|--'        |
 883 //             `----------'
 884 //
 885 // To show that this works fine also with irregular spacings, let's say that
 886 // the third frame is delayed a bit (something earlier was dropped). Now the
 887 // situation looks like this:
 888 //
 889 //        I  B  P  B  B   P
 890 //   pts: 30 40 80 90 100 110
 891 //        ↓  ↓     ↓
 892 //   dts: 10 30 20 90 50←40
 893 //         |  |  ↑        ↑
 894 //         `--|--'        |
 895 //             `----------'
 896 //
 897 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 898 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 899 //
 900 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 901 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 902 // a tip frame and should be given a dts based on the previous one.
 903 #define FRAME_P 0
 904 #define FRAME_B 1
 905 #define FRAME_I 2
 906 #define FRAME_IDR 7
 907 void encoding2display_order(
 908     int encoding_order, int intra_period,
 909     int intra_idr_period, int ip_period,
 910     int *displaying_order,
 911     int *frame_type, int *pts_lag)
 912 {
 913     int encoding_order_gop = 0;
 914
 915     *pts_lag = 0;
 916
 917     if (intra_period == 1) { /* all are I/IDR frames */
 918         *displaying_order = encoding_order;
 919         if (intra_idr_period == 0)
 920             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 921         else
 922             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 923         return;
 924     }
 925
 926     if (intra_period == 0)
 927         intra_idr_period = 0;
 928
 929     if (ip_period == 1) {
 930         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 931         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 932         *displaying_order = encoding_order;
 933
 934         if (encoding_order_gop == 0) { /* the first frame */
 935             *frame_type = FRAME_IDR;
 936         } else if (intra_period != 0 && /* have I frames */
 937                    encoding_order_gop >= 2 &&
 938                    (encoding_order_gop % intra_period == 0)) {
 939             *frame_type = FRAME_I;
 940         } else {
 941             *frame_type = FRAME_P;
 942         }
 943         return;
 944     }
 945
 946     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 947     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 948     *pts_lag = -1;  // Most frames are not tip frames.
 949
 950     if (encoding_order_gop == 0) { /* the first frame */
 951         *frame_type = FRAME_IDR;
 952         *displaying_order = encoding_order;
 953         // IDR frames are a special case; I honestly can't find the logic behind
 954         // why this is the right thing, but it seems to line up nicely in practice :-)
 955         *pts_lag = TIMEBASE / MAX_FPS;
 956     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 957         *frame_type = FRAME_B;
 958         *displaying_order = encoding_order - 1;
 959         if ((encoding_order_gop % ip_period) == 0) {
 960             *pts_lag = 0;  // Last B-frame.
 961         }
 962     } else if (intra_period != 0 && /* have I frames */
 963                encoding_order_gop >= 2 &&
 964                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 965         *frame_type = FRAME_I;
 966         *displaying_order = encoding_order + ip_period - 1;
 967     } else {
 968         *frame_type = FRAME_P;
 969         *displaying_order = encoding_order + ip_period - 1;
 970     }
 971 }
 972
 973
 974 static const char *rc_to_string(int rc_mode)
 975 {
 976     switch (rc_mode) {
 977     case VA_RC_NONE:
 978         return "NONE";
 979     case VA_RC_CBR:
 980         return "CBR";
 981     case VA_RC_VBR:
 982         return "VBR";
 983     case VA_RC_VCM:
 984         return "VCM";
 985     case VA_RC_CQP:
 986         return "CQP";
 987     case VA_RC_VBR_CONSTRAINED:
 988         return "VBR_CONSTRAINED";
 989     default:
 990         return "Unknown";
 991     }
 992 }
 993
 994 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 995 {
 996         if (global_flags.uncompressed_video_to_http) {
 997                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 998                 use_zerocopy = false;
 999         } else if (global_flags.x264_video_to_http) {
1000                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
1001                 use_zerocopy = false;
1002         } else {
1003                 use_zerocopy = true;
1004         }
1005 }
1006
1007 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
1008 {
1009         if (va_display.empty()) {
1010                 x11_display = XOpenDisplay(NULL);
1011                 if (!x11_display) {
1012                         fprintf(stderr, "error: can't connect to X server!\n");
1013                         return NULL;
1014                 }
1015                 enable_zerocopy_if_possible();
1016                 return vaGetDisplay(x11_display);
1017         } else if (va_display[0] != '/') {
1018                 x11_display = XOpenDisplay(va_display.c_str());
1019                 if (!x11_display) {
1020                         fprintf(stderr, "error: can't connect to X server!\n");
1021                         return NULL;
1022                 }
1023                 enable_zerocopy_if_possible();
1024                 return vaGetDisplay(x11_display);
1025         } else {
1026                 drm_fd = open(va_display.c_str(), O_RDWR);
1027                 if (drm_fd == -1) {
1028                         perror(va_display.c_str());
1029                         return NULL;
1030                 }
1031                 use_zerocopy = false;
1032                 return vaGetDisplayDRM(drm_fd);
1033         }
1034 }
1035
1036 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
1037 {
1038         if (x11_display) {
1039                 XCloseDisplay(x11_display);
1040                 x11_display = nullptr;
1041         }
1042         if (drm_fd != -1) {
1043                 close(drm_fd);
1044         }
1045 }
1046
1047 int QuickSyncEncoderImpl::init_va(const string &va_display)
1048 {
1049     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1050     VAEntrypoint *entrypoints;
1051     int num_entrypoints, slice_entrypoint;
1052     int support_encode = 0;
1053     int major_ver, minor_ver;
1054     VAStatus va_status;
1055     unsigned int i;
1056
1057     va_dpy = va_open_display(va_display);
1058     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1059     CHECK_VASTATUS(va_status, "vaInitialize");
1060
1061     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1062     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1063     if (!entrypoints) {
1064         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1065         exit(1);
1066     }
1067
1068     /* use the highest profile */
1069     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1070         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1071             continue;
1072
1073         h264_profile = profile_list[i];
1074         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1075         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1076             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1077                 support_encode = 1;
1078                 break;
1079             }
1080         }
1081         if (support_encode == 1)
1082             break;
1083     }
1084
1085     if (support_encode == 0) {
1086         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1087         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1088         printf("to use VA-API against DRM instead of X11.\n");
1089         exit(1);
1090     } else {
1091         switch (h264_profile) {
1092             case VAProfileH264Baseline:
1093                 ip_period = 1;
1094                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1095                 h264_entropy_mode = 0;
1096                 break;
1097             case VAProfileH264ConstrainedBaseline:
1098                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1099                 ip_period = 1;
1100                 break;
1101
1102             case VAProfileH264Main:
1103                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1104                 break;
1105
1106             case VAProfileH264High:
1107                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1108                 break;
1109             default:
1110                 h264_profile = VAProfileH264Baseline;
1111                 ip_period = 1;
1112                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1113                 break;
1114         }
1115     }
1116
1117     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1118
1119     /* find out the format for the render target, and rate control mode */
1120     for (i = 0; i < VAConfigAttribTypeMax; i++)
1121         attrib[i].type = (VAConfigAttribType)i;
1122
1123     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1124                                       &attrib[0], VAConfigAttribTypeMax);
1125     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1126     /* check the interested configattrib */
1127     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1128         printf("Not find desired YUV420 RT format\n");
1129         exit(1);
1130     } else {
1131         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1132         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1133         config_attrib_num++;
1134     }
1135
1136     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1137         int tmp = attrib[VAConfigAttribRateControl].value;
1138
1139         if (rc_mode == -1 || !(rc_mode & tmp))  {
1140             if (rc_mode != -1) {
1141                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1142             }
1143
1144             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1145                 if (rc_default_modes[i] & tmp) {
1146                     rc_mode = rc_default_modes[i];
1147                     break;
1148                 }
1149             }
1150         }
1151
1152         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1153         config_attrib[config_attrib_num].value = rc_mode;
1154         config_attrib_num++;
1155     }
1156
1157
1158     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1159         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1160
1161         h264_packedheader = 1;
1162         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1163         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1164
1165         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1166             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1167         }
1168
1169         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1170             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1171         }
1172
1173         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1174             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1175         }
1176
1177         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1178             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1179         }
1180
1181         enc_packed_header_idx = config_attrib_num;
1182         config_attrib_num++;
1183     }
1184
1185     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1186         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1187         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1188         config_attrib_num++;
1189     }
1190
1191     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1192         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1193     }
1194
1195     free(entrypoints);
1196     return 0;
1197 }
1198
1199 int QuickSyncEncoderImpl::setup_encode()
1200 {
1201     VAStatus va_status;
1202     VASurfaceID *tmp_surfaceid;
1203     int codedbuf_size, i;
1204     static VASurfaceID src_surface[SURFACE_NUM];
1205     static VASurfaceID ref_surface[SURFACE_NUM];
1206
1207     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1208             &config_attrib[0], config_attrib_num, &config_id);
1209     CHECK_VASTATUS(va_status, "vaCreateConfig");
1210
1211     /* create source surfaces */
1212     va_status = vaCreateSurfaces(va_dpy,
1213                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1214                                  &src_surface[0], SURFACE_NUM,
1215                                  NULL, 0);
1216     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1217
1218     /* create reference surfaces */
1219     va_status = vaCreateSurfaces(va_dpy,
1220                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1221                                  &ref_surface[0], SURFACE_NUM,
1222                                  NULL, 0);
1223     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1224
1225     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1226     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1227     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1228
1229     /* Create a context for this encode pipe */
1230     va_status = vaCreateContext(va_dpy, config_id,
1231                                 frame_width_mbaligned, frame_height_mbaligned,
1232                                 VA_PROGRESSIVE,
1233                                 tmp_surfaceid, 2 * SURFACE_NUM,
1234                                 &context_id);
1235     CHECK_VASTATUS(va_status, "vaCreateContext");
1236     free(tmp_surfaceid);
1237
1238     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1239
1240     for (i = 0; i < SURFACE_NUM; i++) {
1241         /* create coded buffer once for all
1242          * other VA buffers which won't be used again after vaRenderPicture.
1243          * so APP can always vaCreateBuffer for every frame
1244          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1245          * so VA won't maintain the coded buffer
1246          */
1247         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1248                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1249         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1250     }
1251
1252     /* create OpenGL objects */
1253     //glGenFramebuffers(SURFACE_NUM, fbos);
1254
1255     for (i = 0; i < SURFACE_NUM; i++) {
1256         glGenTextures(1, &gl_surfaces[i].y_tex);
1257         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1258
1259         if (!use_zerocopy) {
1260             // Create Y image.
1261             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1262             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1263
1264             // Create CbCr image.
1265             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1266             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1267
1268             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1269             // buffers, due to potentially differing pitch.
1270             glGenBuffers(1, &gl_surfaces[i].pbo);
1271             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1272             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1273             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1274             gl_surfaces[i].y_offset = 0;
1275             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1276             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1277             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1278             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1279         }
1280     }
1281
1282     for (i = 0; i < SURFACE_NUM; i++) {
1283         gl_surfaces[i].src_surface = src_surface[i];
1284         gl_surfaces[i].ref_surface = ref_surface[i];
1285     }
1286
1287     return 0;
1288 }
1289
1290 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1291 //
1292 //   2 1 0 [3] 4 8 9
1293 template<class T, class C>
1294 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1295 {
1296         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1297         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1298         sort(middle, end, less_than);
1299 }
1300
1301 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1302 {
1303     int i;
1304
1305     if (frame_type == FRAME_B)
1306         return;
1307
1308     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1309     numShortTerm++;
1310     if (numShortTerm > num_ref_frames)
1311         numShortTerm = num_ref_frames;
1312     for (i=numShortTerm-1; i>0; i--)
1313         ReferenceFrames[i] = ReferenceFrames[i-1];
1314     ReferenceFrames[0] = CurrentCurrPic;
1315
1316     current_frame_num++;
1317     if (current_frame_num > MaxFrameNum)
1318         current_frame_num = 0;
1319 }
1320
1321
1322 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1323 {
1324     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1325         return a.frame_idx > b.frame_idx;
1326     };
1327     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1328         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1329     };
1330     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1331         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1332     };
1333
1334     if (frame_type == FRAME_P) {
1335         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1336         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1337     } else if (frame_type == FRAME_B) {
1338         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1339         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1340
1341         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1342         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1343     }
1344
1345     return 0;
1346 }
1347
1348
1349 int QuickSyncEncoderImpl::render_sequence()
1350 {
1351     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1352     VAStatus va_status;
1353     VAEncMiscParameterBuffer *misc_param;
1354     VAEncMiscParameterRateControl *misc_rate_ctrl;
1355
1356     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1357     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1358     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1359     seq_param.bits_per_second = frame_bitrate;
1360
1361     seq_param.intra_period = intra_period;
1362     seq_param.intra_idr_period = intra_idr_period;
1363     seq_param.ip_period = ip_period;
1364
1365     seq_param.max_num_ref_frames = num_ref_frames;
1366     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1367     seq_param.time_scale = TIMEBASE * 2;
1368     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1369     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1370     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1371     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1372     seq_param.seq_fields.bits.chroma_format_idc = 1;
1373     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1374
1375     if (frame_width != frame_width_mbaligned ||
1376         frame_height != frame_height_mbaligned) {
1377         seq_param.frame_cropping_flag = 1;
1378         seq_param.frame_crop_left_offset = 0;
1379         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1380         seq_param.frame_crop_top_offset = 0;
1381         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1382     }
1383
1384     va_status = vaCreateBuffer(va_dpy, context_id,
1385                                VAEncSequenceParameterBufferType,
1386                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1387     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1388
1389     va_status = vaCreateBuffer(va_dpy, context_id,
1390                                VAEncMiscParameterBufferType,
1391                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1392                                1, NULL, &rc_param_buf);
1393     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1394
1395     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1396     misc_param->type = VAEncMiscParameterTypeRateControl;
1397     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1398     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1399     misc_rate_ctrl->bits_per_second = frame_bitrate;
1400     misc_rate_ctrl->target_percentage = 66;
1401     misc_rate_ctrl->window_size = 1000;
1402     misc_rate_ctrl->initial_qp = initial_qp;
1403     misc_rate_ctrl->min_qp = minimal_qp;
1404     misc_rate_ctrl->basic_unit_size = 0;
1405     vaUnmapBuffer(va_dpy, rc_param_buf);
1406
1407     render_id[0] = seq_param_buf;
1408     render_id[1] = rc_param_buf;
1409
1410     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1411
1412     return 0;
1413 }
1414
1415 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1416 {
1417     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1418     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1419     int PicOrderCntMsb, TopFieldOrderCnt;
1420
1421     if (frame_type == FRAME_IDR)
1422         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1423     else {
1424         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1425         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1426     }
1427
1428     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1429         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1430         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1431     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1432              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1433         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1434     else
1435         PicOrderCntMsb = prevPicOrderCntMsb;
1436
1437     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1438
1439     if (frame_type != FRAME_B) {
1440         PicOrderCntMsb_ref = PicOrderCntMsb;
1441         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1442     }
1443
1444     return TopFieldOrderCnt;
1445 }
1446
1447 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1448 {
1449     VABufferID pic_param_buf;
1450     VAStatus va_status;
1451     int i = 0;
1452
1453     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1454     pic_param.CurrPic.frame_idx = current_frame_num;
1455     pic_param.CurrPic.flags = 0;
1456     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1457     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1458     CurrentCurrPic = pic_param.CurrPic;
1459
1460     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1461     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1462         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1463         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1464     }
1465
1466     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1467     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1468     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1469     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1470     pic_param.frame_num = current_frame_num;
1471     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1472     pic_param.last_picture = false;  // FIXME
1473     pic_param.pic_init_qp = initial_qp;
1474
1475     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1476                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1477     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1478
1479     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1480
1481     return 0;
1482 }
1483
1484 int QuickSyncEncoderImpl::render_packedsequence()
1485 {
1486     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1487     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1488     unsigned int length_in_bits;
1489     unsigned char *packedseq_buffer = NULL;
1490     VAStatus va_status;
1491
1492     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1493
1494     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1495
1496     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1497     packedheader_param_buffer.has_emulation_bytes = 0;
1498     va_status = vaCreateBuffer(va_dpy,
1499                                context_id,
1500                                VAEncPackedHeaderParameterBufferType,
1501                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1502                                &packedseq_para_bufid);
1503     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1504
1505     va_status = vaCreateBuffer(va_dpy,
1506                                context_id,
1507                                VAEncPackedHeaderDataBufferType,
1508                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1509                                &packedseq_data_bufid);
1510     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1511
1512     render_id[0] = packedseq_para_bufid;
1513     render_id[1] = packedseq_data_bufid;
1514     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1515
1516     free(packedseq_buffer);
1517
1518     return 0;
1519 }
1520
1521
1522 int QuickSyncEncoderImpl::render_packedpicture()
1523 {
1524     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1525     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1526     unsigned int length_in_bits;
1527     unsigned char *packedpic_buffer = NULL;
1528     VAStatus va_status;
1529
1530     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1531     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1532     packedheader_param_buffer.bit_length = length_in_bits;
1533     packedheader_param_buffer.has_emulation_bytes = 0;
1534
1535     va_status = vaCreateBuffer(va_dpy,
1536                                context_id,
1537                                VAEncPackedHeaderParameterBufferType,
1538                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1539                                &packedpic_para_bufid);
1540     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1541
1542     va_status = vaCreateBuffer(va_dpy,
1543                                context_id,
1544                                VAEncPackedHeaderDataBufferType,
1545                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1546                                &packedpic_data_bufid);
1547     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1548
1549     render_id[0] = packedpic_para_bufid;
1550     render_id[1] = packedpic_data_bufid;
1551     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1552
1553     free(packedpic_buffer);
1554
1555     return 0;
1556 }
1557
1558 void QuickSyncEncoderImpl::render_packedslice()
1559 {
1560     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1561     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1562     unsigned int length_in_bits;
1563     unsigned char *packedslice_buffer = NULL;
1564     VAStatus va_status;
1565
1566     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1567     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1568     packedheader_param_buffer.bit_length = length_in_bits;
1569     packedheader_param_buffer.has_emulation_bytes = 0;
1570
1571     va_status = vaCreateBuffer(va_dpy,
1572                                context_id,
1573                                VAEncPackedHeaderParameterBufferType,
1574                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1575                                &packedslice_para_bufid);
1576     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1577
1578     va_status = vaCreateBuffer(va_dpy,
1579                                context_id,
1580                                VAEncPackedHeaderDataBufferType,
1581                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1582                                &packedslice_data_bufid);
1583     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1584
1585     render_id[0] = packedslice_para_bufid;
1586     render_id[1] = packedslice_data_bufid;
1587     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1588
1589     free(packedslice_buffer);
1590 }
1591
1592 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1593 {
1594     VABufferID slice_param_buf;
1595     VAStatus va_status;
1596     int i;
1597
1598     update_RefPicList(frame_type);
1599
1600     /* one frame, one slice */
1601     slice_param.macroblock_address = 0;
1602     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1603     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1604     if (frame_type == FRAME_IDR) {
1605         if (encoding_frame_num != 0)
1606             ++slice_param.idr_pic_id;
1607     } else if (frame_type == FRAME_P) {
1608         int refpiclist0_max = h264_maxref & 0xffff;
1609         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1610
1611         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1612             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1613             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1614         }
1615     } else if (frame_type == FRAME_B) {
1616         int refpiclist0_max = h264_maxref & 0xffff;
1617         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1618
1619         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1620         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1621             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1622             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1623         }
1624
1625         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1626         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1627             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1628             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1629         }
1630     }
1631
1632     slice_param.slice_alpha_c0_offset_div2 = 0;
1633     slice_param.slice_beta_offset_div2 = 0;
1634     slice_param.direct_spatial_mv_pred_flag = 1;
1635     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1636
1637
1638     if (h264_packedheader &&
1639         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1640         render_packedslice();
1641
1642     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1643                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1644     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1645
1646     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1647
1648     return 0;
1649 }
1650
1651
1652
1653 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1654 {
1655         VACodedBufferSegment *buf_list = NULL;
1656         VAStatus va_status;
1657
1658         string data;
1659
1660         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1661         CHECK_VASTATUS(va_status, "vaMapBuffer");
1662         while (buf_list != NULL) {
1663                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1664                 buf_list = (VACodedBufferSegment *) buf_list->next;
1665         }
1666         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1667
1668         {
1669                 // Add video.
1670                 AVPacket pkt;
1671                 memset(&pkt, 0, sizeof(pkt));
1672                 pkt.buf = nullptr;
1673                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1674                 pkt.size = data.size();
1675                 pkt.stream_index = 0;
1676                 if (task.frame_type == FRAME_IDR) {
1677                         pkt.flags = AV_PKT_FLAG_KEY;
1678                 } else {
1679                         pkt.flags = 0;
1680                 }
1681                 pkt.duration = task.duration;
1682                 if (file_mux) {
1683                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1684                 }
1685                 if (!global_flags.uncompressed_video_to_http &&
1686                     !global_flags.x264_video_to_http) {
1687                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1688                 }
1689         }
1690         // Encode and add all audio frames up to and including the pts of this video frame.
1691         for ( ;; ) {
1692                 int64_t audio_pts;
1693                 vector<float> audio;
1694                 {
1695                         unique_lock<mutex> lock(frame_queue_mutex);
1696                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1697                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1698                         auto it = pending_audio_frames.begin();
1699                         if (it->first > task.pts) break;
1700                         audio_pts = it->first;
1701                         audio = move(it->second);
1702                         pending_audio_frames.erase(it);
1703                 }
1704
1705                 if (context_audio_stream) {
1706                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { file_mux.get() });
1707                         encode_audio(audio, &audio_queue_stream, audio_pts, context_audio_stream, resampler_audio_stream, { stream_mux.get() });
1708                 } else {
1709                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { stream_mux.get(), file_mux.get() });
1710                 }
1711                 last_audio_pts = audio_pts + audio.size() * TIMEBASE / (OUTPUT_FREQUENCY * 2);
1712
1713                 if (audio_pts == task.pts) break;
1714         }
1715 }
1716
1717 void QuickSyncEncoderImpl::encode_audio(
1718         const vector<float> &audio,
1719         vector<float> *audio_queue,
1720         int64_t audio_pts,
1721         AVCodecContext *ctx,
1722         AVAudioResampleContext *resampler,
1723         const vector<Mux *> &muxes)
1724 {
1725         if (ctx->frame_size == 0) {
1726                 // No queueing needed.
1727                 assert(audio_queue->empty());
1728                 assert(audio.size() % 2 == 0);
1729                 encode_audio_one_frame(&audio[0], audio.size() / 2, audio_pts, ctx, resampler, muxes);
1730                 return;
1731         }
1732
1733         int64_t sample_offset = audio_queue->size();
1734
1735         audio_queue->insert(audio_queue->end(), audio.begin(), audio.end());
1736         size_t sample_num;
1737         for (sample_num = 0;
1738              sample_num + ctx->frame_size * 2 <= audio_queue->size();
1739              sample_num += ctx->frame_size * 2) {
1740                 int64_t adjusted_audio_pts = audio_pts + (int64_t(sample_num) - sample_offset) * TIMEBASE / (OUTPUT_FREQUENCY * 2);
1741                 encode_audio_one_frame(&(*audio_queue)[sample_num],
1742                                        ctx->frame_size,
1743                                        adjusted_audio_pts,
1744                                        ctx,
1745                                        resampler,
1746                                        muxes);
1747         }
1748         audio_queue->erase(audio_queue->begin(), audio_queue->begin() + sample_num);
1749 }
1750
1751 void QuickSyncEncoderImpl::encode_audio_one_frame(
1752         const float *audio,
1753         size_t num_samples,
1754         int64_t audio_pts,
1755         AVCodecContext *ctx,
1756         AVAudioResampleContext *resampler,
1757         const vector<Mux *> &muxes)
1758 {
1759         audio_frame->pts = audio_pts + global_delay();
1760         audio_frame->nb_samples = num_samples;
1761         audio_frame->channel_layout = AV_CH_LAYOUT_STEREO;
1762         audio_frame->format = ctx->sample_fmt;
1763         audio_frame->sample_rate = OUTPUT_FREQUENCY;
1764
1765         if (av_samples_alloc(audio_frame->data, nullptr, 2, num_samples, ctx->sample_fmt, 0) < 0) {
1766                 fprintf(stderr, "Could not allocate %ld samples.\n", num_samples);
1767                 exit(1);
1768         }
1769
1770         if (avresample_convert(resampler, audio_frame->data, 0, num_samples,
1771                                (uint8_t **)&audio, 0, num_samples) < 0) {
1772                 fprintf(stderr, "Audio conversion failed.\n");
1773                 exit(1);
1774         }
1775
1776         AVPacket pkt;
1777         av_init_packet(&pkt);
1778         pkt.data = nullptr;
1779         pkt.size = 0;
1780         int got_output = 0;
1781         avcodec_encode_audio2(ctx, &pkt, audio_frame, &got_output);
1782         if (got_output) {
1783                 pkt.stream_index = 1;
1784                 pkt.flags = 0;
1785                 for (Mux *mux : muxes) {
1786                         mux->add_packet(pkt, pkt.pts, pkt.dts);
1787                 }
1788         }
1789
1790         av_freep(&audio_frame->data[0]);
1791
1792         av_frame_unref(audio_frame);
1793         av_free_packet(&pkt);
1794 }
1795
1796 void QuickSyncEncoderImpl::encode_last_audio(
1797         vector<float> *audio_queue,
1798         int64_t audio_pts,
1799         AVCodecContext *ctx,
1800         AVAudioResampleContext *resampler,
1801         const vector<Mux *> &muxes)
1802 {
1803         if (!audio_queue->empty()) {
1804                 // Last frame can be whatever size we want.
1805                 assert(audio_queue->size() % 2 == 0);
1806                 encode_audio_one_frame(&(*audio_queue)[0], audio_queue->size() / 2, audio_pts, ctx, resampler, muxes);
1807                 audio_queue->clear();
1808         }
1809
1810         if (ctx->codec->capabilities & AV_CODEC_CAP_DELAY) {
1811                 // Collect any delayed frames.
1812                 for ( ;; ) {
1813                         int got_output = 0;
1814                         AVPacket pkt;
1815                         av_init_packet(&pkt);
1816                         pkt.data = nullptr;
1817                         pkt.size = 0;
1818                         avcodec_encode_audio2(ctx, &pkt, nullptr, &got_output);
1819                         if (!got_output) break;
1820
1821                         pkt.stream_index = 1;
1822                         pkt.flags = 0;
1823                         for (Mux *mux : muxes) {
1824                                 mux->add_packet(pkt, pkt.pts, pkt.dts);
1825                         }
1826                         av_free_packet(&pkt);
1827                 }
1828         }
1829 }
1830
1831 // this is weird. but it seems to put a new frame onto the queue
1832 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1833 {
1834         unique_lock<mutex> lock(storage_task_queue_mutex);
1835         storage_task_queue.push(move(task));
1836         storage_task_queue_changed.notify_all();
1837 }
1838
1839 void QuickSyncEncoderImpl::storage_task_thread()
1840 {
1841         for ( ;; ) {
1842                 storage_task current;
1843                 {
1844                         // wait until there's an encoded frame
1845                         unique_lock<mutex> lock(storage_task_queue_mutex);
1846                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1847                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1848                         current = move(storage_task_queue.front());
1849                         storage_task_queue.pop();
1850                 }
1851
1852                 VAStatus va_status;
1853
1854                 // waits for data, then saves it to disk.
1855                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1856                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1857                 save_codeddata(move(current));
1858
1859                 {
1860                         unique_lock<mutex> lock(storage_task_queue_mutex);
1861                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1862                         storage_task_queue_changed.notify_all();
1863                 }
1864         }
1865 }
1866
1867 int QuickSyncEncoderImpl::release_encode()
1868 {
1869         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1870                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1871                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1872                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1873
1874                 if (!use_zerocopy) {
1875                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1876                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1877                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1878                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1879                 }
1880                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1881                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1882         }
1883
1884         vaDestroyContext(va_dpy, context_id);
1885         vaDestroyConfig(va_dpy, config_id);
1886
1887         return 0;
1888 }
1889
1890 int QuickSyncEncoderImpl::deinit_va()
1891 {
1892     vaTerminate(va_dpy);
1893
1894     va_close_display(va_dpy);
1895
1896     return 0;
1897 }
1898
1899 namespace {
1900
1901 void init_audio_encoder(const string &codec_name, int bit_rate, AVCodecContext **ctx, AVAudioResampleContext **resampler)
1902 {
1903         AVCodec *codec_audio = avcodec_find_encoder_by_name(codec_name.c_str());
1904         if (codec_audio == nullptr) {
1905                 fprintf(stderr, "ERROR: Could not find codec '%s'\n", codec_name.c_str());
1906                 exit(1);
1907         }
1908
1909         AVCodecContext *context_audio = avcodec_alloc_context3(codec_audio);
1910         context_audio->bit_rate = bit_rate;
1911         context_audio->sample_rate = OUTPUT_FREQUENCY;
1912         context_audio->sample_fmt = codec_audio->sample_fmts[0];
1913         context_audio->channels = 2;
1914         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1915         context_audio->time_base = AVRational{1, TIMEBASE};
1916         context_audio->flags |= CODEC_FLAG_GLOBAL_HEADER;
1917         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1918                 fprintf(stderr, "Could not open codec '%s'\n", codec_name.c_str());
1919                 exit(1);
1920         }
1921
1922         *ctx = context_audio;
1923
1924         *resampler = avresample_alloc_context();
1925         if (*resampler == nullptr) {
1926                 fprintf(stderr, "Allocating resampler failed.\n");
1927                 exit(1);
1928         }
1929
1930         av_opt_set_int(*resampler, "in_channel_layout",  AV_CH_LAYOUT_STEREO,       0);
1931         av_opt_set_int(*resampler, "out_channel_layout", AV_CH_LAYOUT_STEREO,       0);
1932         av_opt_set_int(*resampler, "in_sample_rate",     OUTPUT_FREQUENCY,          0);
1933         av_opt_set_int(*resampler, "out_sample_rate",    OUTPUT_FREQUENCY,          0);
1934         av_opt_set_int(*resampler, "in_sample_fmt",      AV_SAMPLE_FMT_FLT,         0);
1935         av_opt_set_int(*resampler, "out_sample_fmt",     context_audio->sample_fmt, 0);
1936
1937         if (avresample_open(*resampler) < 0) {
1938                 fprintf(stderr, "Could not open resample context.\n");
1939                 exit(1);
1940         }
1941 }
1942
1943 }  // namespace
1944
1945 QuickSyncEncoderImpl::QuickSyncEncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1946         : current_storage_frame(0), surface(surface), httpd(httpd), frame_width(width), frame_height(height)
1947 {
1948         init_audio_encoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, &context_audio_file, &resampler_audio_file);
1949
1950         if (!global_flags.stream_audio_codec_name.empty()) {
1951                 init_audio_encoder(global_flags.stream_audio_codec_name,
1952                         global_flags.stream_audio_codec_bitrate, &context_audio_stream, &resampler_audio_stream);
1953         }
1954
1955         frame_width_mbaligned = (frame_width + 15) & (~15);
1956         frame_height_mbaligned = (frame_height + 15) & (~15);
1957
1958         open_output_stream();
1959
1960         audio_frame = av_frame_alloc();
1961
1962         //print_input();
1963
1964         if (global_flags.uncompressed_video_to_http ||
1965             global_flags.x264_video_to_http) {
1966                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1967         }
1968         if (global_flags.x264_video_to_http) {
1969                 x264_encoder.reset(new X264Encoder(stream_mux.get()));
1970         }
1971
1972         init_va(va_display);
1973         setup_encode();
1974
1975         // No frames are ready yet.
1976         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1977
1978         memset(&seq_param, 0, sizeof(seq_param));
1979         memset(&pic_param, 0, sizeof(pic_param));
1980         memset(&slice_param, 0, sizeof(slice_param));
1981
1982         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1983
1984         encode_thread = thread([this]{
1985                 //SDL_GL_MakeCurrent(window, context);
1986                 QOpenGLContext *context = create_context(this->surface);
1987                 eglBindAPI(EGL_OPENGL_API);
1988                 if (!make_current(context, this->surface)) {
1989                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1990                                 eglGetError());
1991                         exit(1);
1992                 }
1993                 encode_thread_func();
1994         });
1995 }
1996
1997 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1998 {
1999         shutdown();
2000         av_frame_free(&audio_frame);
2001         avresample_free(&resampler_audio_file);
2002         avresample_free(&resampler_audio_stream);
2003         avcodec_free_context(&context_audio_file);
2004         avcodec_free_context(&context_audio_stream);
2005         close_output_stream();
2006 }
2007
2008 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2009 {
2010         assert(!is_shutdown);
2011         {
2012                 // Wait until this frame slot is done encoding.
2013                 unique_lock<mutex> lock(storage_task_queue_mutex);
2014                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
2015                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
2016                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
2017                 }
2018                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
2019                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
2020                 if (storage_thread_should_quit) return false;
2021         }
2022
2023         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
2024         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
2025         *y_tex = surf->y_tex;
2026         *cbcr_tex = surf->cbcr_tex;
2027
2028         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
2029         CHECK_VASTATUS(va_status, "vaDeriveImage");
2030
2031         if (use_zerocopy) {
2032                 VABufferInfo buf_info;
2033                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
2034                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
2035                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
2036
2037                 // Create Y image.
2038                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
2039                 EGLint y_attribs[] = {
2040                         EGL_WIDTH, frame_width,
2041                         EGL_HEIGHT, frame_height,
2042                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
2043                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
2044                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
2045                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
2046                         EGL_NONE
2047                 };
2048
2049                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
2050                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
2051
2052                 // Associate Y image to a texture.
2053                 glBindTexture(GL_TEXTURE_2D, *y_tex);
2054                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
2055
2056                 // Create CbCr image.
2057                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
2058                 EGLint cbcr_attribs[] = {
2059                         EGL_WIDTH, frame_width,
2060                         EGL_HEIGHT, frame_height,
2061                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
2062                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
2063                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
2064                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
2065                         EGL_NONE
2066                 };
2067
2068                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
2069                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
2070
2071                 // Associate CbCr image to a texture.
2072                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
2073                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
2074         }
2075
2076         return true;
2077 }
2078
2079 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
2080 {
2081         assert(!is_shutdown);
2082         {
2083                 unique_lock<mutex> lock(frame_queue_mutex);
2084                 pending_audio_frames[pts] = move(audio);
2085         }
2086         frame_queue_nonempty.notify_all();
2087 }
2088
2089 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2090 {
2091         assert(!is_shutdown);
2092
2093         if (!use_zerocopy) {
2094                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
2095
2096                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
2097                 check_error();
2098
2099                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
2100                 check_error();
2101
2102                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
2103                 check_error();
2104                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
2105                 check_error();
2106
2107                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
2108                 check_error();
2109                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
2110                 check_error();
2111
2112                 glBindTexture(GL_TEXTURE_2D, 0);
2113                 check_error();
2114                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
2115                 check_error();
2116
2117                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
2118                 check_error();
2119         }
2120
2121         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
2122         check_error();
2123         glFlush();  // Make the H.264 thread see the fence as soon as possible.
2124         check_error();
2125
2126         {
2127                 unique_lock<mutex> lock(frame_queue_mutex);
2128                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
2129                 ++current_storage_frame;
2130         }
2131         frame_queue_nonempty.notify_all();
2132         return fence;
2133 }
2134
2135 void QuickSyncEncoderImpl::shutdown()
2136 {
2137         if (is_shutdown) {
2138                 return;
2139         }
2140
2141         {
2142                 unique_lock<mutex> lock(frame_queue_mutex);
2143                 encode_thread_should_quit = true;
2144                 frame_queue_nonempty.notify_all();
2145         }
2146         encode_thread.join();
2147         x264_encoder.reset();
2148         {
2149                 unique_lock<mutex> lock(storage_task_queue_mutex);
2150                 storage_thread_should_quit = true;
2151                 frame_queue_nonempty.notify_all();
2152                 storage_task_queue_changed.notify_all();
2153         }
2154         storage_thread.join();
2155         encode_remaining_audio();
2156
2157         release_encode();
2158         deinit_va();
2159         is_shutdown = true;
2160 }
2161
2162 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
2163 {
2164         AVFormatContext *avctx = avformat_alloc_context();
2165         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
2166         assert(filename.size() < sizeof(avctx->filename) - 1);
2167         strcpy(avctx->filename, filename.c_str());
2168
2169         string url = "file:" + filename;
2170         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
2171         if (ret < 0) {
2172                 char tmp[AV_ERROR_MAX_STRING_SIZE];
2173                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
2174                 exit(1);
2175         }
2176
2177         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, context_audio_file->codec, TIMEBASE, DEFAULT_AUDIO_OUTPUT_BIT_RATE, nullptr));
2178 }
2179
2180 void QuickSyncEncoderImpl::close_output_file()
2181 {
2182         file_mux.reset();
2183 }
2184
2185 void QuickSyncEncoderImpl::open_output_stream()
2186 {
2187         AVFormatContext *avctx = avformat_alloc_context();
2188         AVOutputFormat *oformat = av_guess_format(global_flags.stream_mux_name.c_str(), nullptr, nullptr);
2189         assert(oformat != nullptr);
2190         avctx->oformat = oformat;
2191
2192         string codec_name;
2193         int bit_rate;
2194
2195         if (global_flags.stream_audio_codec_name.empty()) {
2196                 codec_name = AUDIO_OUTPUT_CODEC_NAME;
2197                 bit_rate = DEFAULT_AUDIO_OUTPUT_BIT_RATE;
2198         } else {
2199                 codec_name = global_flags.stream_audio_codec_name;
2200                 bit_rate = global_flags.stream_audio_codec_bitrate;
2201         }
2202
2203         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
2204         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, &QuickSyncEncoderImpl::write_packet_thunk, nullptr);
2205
2206         Mux::Codec video_codec;
2207         if (global_flags.uncompressed_video_to_http) {
2208                 video_codec = Mux::CODEC_NV12;
2209         } else {
2210                 video_codec = Mux::CODEC_H264;
2211         }
2212
2213         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
2214         AVCodec *codec_audio = avcodec_find_encoder_by_name(codec_name.c_str());
2215         if (codec_audio == nullptr) {
2216                 fprintf(stderr, "ERROR: Could not find codec '%s'\n", codec_name.c_str());
2217                 exit(1);
2218         }
2219
2220         int time_base = global_flags.stream_coarse_timebase ? COARSE_TIMEBASE : TIMEBASE;
2221         stream_mux_writing_header = true;
2222         stream_mux.reset(new Mux(avctx, frame_width, frame_height, video_codec, codec_audio, time_base, bit_rate, this));
2223         stream_mux_writing_header = false;
2224         httpd->set_header(stream_mux_header);
2225         stream_mux_header.clear();
2226 }
2227
2228 void QuickSyncEncoderImpl::close_output_stream()
2229 {
2230         stream_mux.reset();
2231 }
2232
2233 int QuickSyncEncoderImpl::write_packet_thunk(void *opaque, uint8_t *buf, int buf_size)
2234 {
2235         QuickSyncEncoderImpl *h264_encoder = (QuickSyncEncoderImpl *)opaque;
2236         return h264_encoder->write_packet(buf, buf_size);
2237 }
2238
2239 int QuickSyncEncoderImpl::write_packet(uint8_t *buf, int buf_size)
2240 {
2241         if (stream_mux_writing_header) {
2242                 stream_mux_header.append((char *)buf, buf_size);
2243         } else {
2244                 httpd->add_data((char *)buf, buf_size, stream_mux_writing_keyframes);
2245                 stream_mux_writing_keyframes = false;
2246         }
2247         return buf_size;
2248 }
2249
2250 void QuickSyncEncoderImpl::encode_thread_func()
2251 {
2252         int64_t last_dts = -1;
2253         int gop_start_display_frame_num = 0;
2254         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
2255                 PendingFrame frame;
2256                 int pts_lag;
2257                 int frame_type, display_frame_num;
2258                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
2259                                        &display_frame_num, &frame_type, &pts_lag);
2260                 if (frame_type == FRAME_IDR) {
2261                         numShortTerm = 0;
2262                         current_frame_num = 0;
2263                         gop_start_display_frame_num = display_frame_num;
2264                 }
2265
2266                 {
2267                         unique_lock<mutex> lock(frame_queue_mutex);
2268                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
2269                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
2270                         });
2271                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
2272                                 // We have queued frames that were supposed to be B-frames,
2273                                 // but will be no P-frame to encode them against. Encode them all
2274                                 // as P-frames instead. Note that this happens under the mutex,
2275                                 // but nobody else uses it at this point, since we're shutting down,
2276                                 // so there's no contention.
2277                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
2278                                 return;
2279                         } else {
2280                                 frame = move(pending_video_frames[display_frame_num]);
2281                                 pending_video_frames.erase(display_frame_num);
2282                         }
2283                 }
2284
2285                 // Determine the dts of this frame.
2286                 int64_t dts;
2287                 if (pts_lag == -1) {
2288                         assert(last_dts != -1);
2289                         dts = last_dts + (TIMEBASE / MAX_FPS);
2290                 } else {
2291                         dts = frame.pts - pts_lag;
2292                 }
2293                 last_dts = dts;
2294
2295                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2296         }
2297 }
2298
2299 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2300 {
2301         if (pending_video_frames.empty()) {
2302                 return;
2303         }
2304
2305         for (auto &pending_frame : pending_video_frames) {
2306                 int display_frame_num = pending_frame.first;
2307                 assert(display_frame_num > 0);
2308                 PendingFrame frame = move(pending_frame.second);
2309                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2310                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2311                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2312                 last_dts = dts;
2313         }
2314
2315         if (global_flags.uncompressed_video_to_http ||
2316             global_flags.x264_video_to_http) {
2317                 // Add frames left in reorderer.
2318                 while (!reorderer->empty()) {
2319                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2320                         if (global_flags.uncompressed_video_to_http) {
2321                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2322                         } else {
2323                                 assert(global_flags.x264_video_to_http);
2324                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2325                         }
2326                 }
2327         }
2328 }
2329
2330 void QuickSyncEncoderImpl::encode_remaining_audio()
2331 {
2332         // This really ought to be empty by now, but just to be sure...
2333         for (auto &pending_frame : pending_audio_frames) {
2334                 int64_t audio_pts = pending_frame.first;
2335                 vector<float> audio = move(pending_frame.second);
2336
2337                 if (context_audio_stream) {
2338                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { file_mux.get() });
2339                         encode_audio(audio, &audio_queue_stream, audio_pts, context_audio_stream, resampler_audio_stream, { stream_mux.get() });
2340                 } else {
2341                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { stream_mux.get(), file_mux.get() });
2342                 }
2343                 last_audio_pts = audio_pts + audio.size() * TIMEBASE / (OUTPUT_FREQUENCY * 2);
2344         }
2345         pending_audio_frames.clear();
2346
2347         // Encode any leftover audio in the queues, and also any delayed frames.
2348         if (context_audio_stream) {
2349                 encode_last_audio(&audio_queue_file, last_audio_pts, context_audio_file, resampler_audio_file, { file_mux.get() });
2350                 encode_last_audio(&audio_queue_stream, last_audio_pts, context_audio_stream, resampler_audio_stream, { stream_mux.get() });
2351         } else {
2352                 encode_last_audio(&audio_queue_file, last_audio_pts, context_audio_file, resampler_audio_file, { stream_mux.get(), file_mux.get() });
2353         }
2354 }
2355
2356 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2357 {
2358         AVPacket pkt;
2359         memset(&pkt, 0, sizeof(pkt));
2360         pkt.buf = nullptr;
2361         pkt.data = const_cast<uint8_t *>(data);
2362         pkt.size = frame_width * frame_height * 2;
2363         pkt.stream_index = 0;
2364         pkt.flags = AV_PKT_FLAG_KEY;
2365         pkt.duration = duration;
2366         stream_mux->add_packet(pkt, pts, pts);
2367 }
2368
2369 namespace {
2370
2371 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2372 {
2373         if (src_width == dst_pitch) {
2374                 memcpy(dst, src, src_width * height);
2375         } else {
2376                 for (size_t y = 0; y < height; ++y) {
2377                         const uint8_t *sptr = src + y * src_width;
2378                         uint8_t *dptr = dst + y * dst_pitch;
2379                         memcpy(dptr, sptr, src_width);
2380                 }
2381         }
2382 }
2383
2384 }  // namespace
2385
2386 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2387                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2388 {
2389         // Wait for the GPU to be done with the frame.
2390         GLenum sync_status;
2391         do {
2392                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2393                 check_error();
2394         } while (sync_status == GL_TIMEOUT_EXPIRED);
2395         assert(sync_status != GL_WAIT_FAILED);
2396
2397         // Release back any input frames we needed to render this frame.
2398         frame.input_frames.clear();
2399
2400         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2401         VAStatus va_status;
2402
2403         if (use_zerocopy) {
2404                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2405                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2406                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2407                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2408         } else {
2409                 unsigned char *surface_p = nullptr;
2410                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2411
2412                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2413                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2414
2415                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2416                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2417
2418                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2419                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2420
2421                 if (global_flags.uncompressed_video_to_http ||
2422                     global_flags.x264_video_to_http) {
2423                         // Add uncompressed video. (Note that pts == dts here.)
2424                         // Delay needs to match audio.
2425                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2426                         if (output_frame.data != nullptr) {
2427                                 if (global_flags.uncompressed_video_to_http) {
2428                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2429                                 } else {
2430                                         assert(global_flags.x264_video_to_http);
2431                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2432                                 }
2433                         }
2434                 }
2435         }
2436
2437         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2438         CHECK_VASTATUS(va_status, "vaDestroyImage");
2439
2440         // Schedule the frame for encoding.
2441         VASurfaceID va_surface = surf->src_surface;
2442         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2443         CHECK_VASTATUS(va_status, "vaBeginPicture");
2444
2445         if (frame_type == FRAME_IDR) {
2446                 render_sequence();
2447                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2448                 if (h264_packedheader) {
2449                         render_packedsequence();
2450                         render_packedpicture();
2451                 }
2452         } else {
2453                 //render_sequence();
2454                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2455         }
2456         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2457
2458         va_status = vaEndPicture(va_dpy, context_id);
2459         CHECK_VASTATUS(va_status, "vaEndPicture");
2460
2461         // so now the data is done encoding (well, async job kicked off)...
2462         // we send that to the storage thread
2463         storage_task tmp;
2464         tmp.display_order = display_frame_num;
2465         tmp.frame_type = frame_type;
2466         tmp.pts = pts;
2467         tmp.dts = dts;
2468         tmp.duration = duration;
2469         storage_task_enqueue(move(tmp));
2470
2471         update_ReferenceFrames(frame_type);
2472 }
2473
2474 // Proxy object.
2475 QuickSyncEncoder::QuickSyncEncoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
2476         : impl(new QuickSyncEncoderImpl(surface, va_display, width, height, httpd)) {}
2477
2478 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2479 QuickSyncEncoder::~QuickSyncEncoder() {}
2480
2481 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2482 {
2483         impl->add_audio(pts, audio);
2484 }
2485
2486 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2487 {
2488         return impl->begin_frame(y_tex, cbcr_tex);
2489 }
2490
2491 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2492 {
2493         return impl->end_frame(pts, duration, input_frames);
2494 }
2495
2496 void QuickSyncEncoder::shutdown()
2497 {
2498         impl->shutdown();
2499 }
2500
2501 void QuickSyncEncoder::open_output_file(const std::string &filename)
2502 {
2503         impl->open_output_file(filename);
2504 }
2505
2506 void QuickSyncEncoder::close_output_file()
2507 {
2508         impl->close_output_file();
2509 }