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

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