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

   1 //#include "sysdeps.h"
   2 #include "quicksync_encoder.h"
   3
   4 #include <movit/resource_pool.h>
   5 #include <movit/util.h>
   6 #include <EGL/eglplatform.h>
   7 #include <X11/X.h>
   8 #include <X11/Xlib.h>
   9 #include <assert.h>
  10 #include <epoxy/egl.h>
  11 #include <libdrm/drm_fourcc.h>
  12 #include <stdio.h>
  13 #include <stdlib.h>
  14 #include <string.h>
  15 #include <fcntl.h>
  16 #include <va/va.h>
  17 #include <va/va_drm.h>
  18 #include <va/va_drmcommon.h>
  19 #include <va/va_enc_h264.h>
  20 #include <va/va_x11.h>
  21 #include <algorithm>
  22 #include <condition_variable>
  23 #include <cstdint>
  24 #include <map>
  25 #include <memory>
  26 #include <mutex>
  27 #include <queue>
  28 #include <string>
  29 #include <thread>
  30 #include <utility>
  31
  32 #include "audio_encoder.h"
  33 #include "context.h"
  34 #include "defs.h"
  35 #include "flags.h"
  36 #include "mux.h"
  37 #include "timebase.h"
  38 #include "x264_encoder.h"
  39
  40 using namespace std;
  41
  42 class QOpenGLContext;
  43 class QSurface;
  44
  45 #define CHECK_VASTATUS(va_status, func)                                 \
  46     if (va_status != VA_STATUS_SUCCESS) {                               \
  47         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  48         exit(1);                                                        \
  49     }
  50
  51 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  52
  53 //#include "loadsurface.h"
  54
  55 #define NAL_REF_IDC_NONE        0
  56 #define NAL_REF_IDC_LOW         1
  57 #define NAL_REF_IDC_MEDIUM      2
  58 #define NAL_REF_IDC_HIGH        3
  59
  60 #define NAL_NON_IDR             1
  61 #define NAL_IDR                 5
  62 #define NAL_SPS                 7
  63 #define NAL_PPS                 8
  64 #define NAL_SEI                 6
  65
  66 #define SLICE_TYPE_P            0
  67 #define SLICE_TYPE_B            1
  68 #define SLICE_TYPE_I            2
  69 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  70 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  71 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  72
  73
  74 #define ENTROPY_MODE_CAVLC      0
  75 #define ENTROPY_MODE_CABAC      1
  76
  77 #define PROFILE_IDC_BASELINE    66
  78 #define PROFILE_IDC_MAIN        77
  79 #define PROFILE_IDC_HIGH        100
  80
  81 #define BITSTREAM_ALLOCATE_STEPPING     4096
  82 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  83 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  84 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  85
  86 static constexpr unsigned int MaxFrameNum = (2<<16);
  87 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  88 static constexpr unsigned int Log2MaxFrameNum = 16;
  89 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  90 static constexpr int rc_default_modes[] = {  // Priority list of modes.
  91     VA_RC_VBR,
  92     VA_RC_CQP,
  93     VA_RC_VBR_CONSTRAINED,
  94     VA_RC_CBR,
  95     VA_RC_VCM,
  96     VA_RC_NONE,
  97 };
  98
  99 /* thread to save coded data */
 100 #define SRC_SURFACE_FREE        0
 101 #define SRC_SURFACE_IN_ENCODING 1
 102
 103 struct __bitstream {
 104     unsigned int *buffer;
 105     int bit_offset;
 106     int max_size_in_dword;
 107 };
 108 typedef struct __bitstream bitstream;
 109
 110 using namespace std;
 111
 112 // H.264 video comes out in encoding order (e.g. with two B-frames:
 113 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 114 // come in the right order. Since we do everything, including waiting
 115 // for the frames to come out of OpenGL, in encoding order, we need
 116 // a reordering buffer for uncompressed frames so that they come out
 117 // correctly. We go the super-lazy way of not making it understand
 118 // anything about the true order (which introduces some extra latency,
 119 // though); we know that for N B-frames we need at most (N-1) frames
 120 // in the reorder buffer, and can just sort on that.
 121 //
 122 // The class also deals with keeping a freelist as needed.
 123 class FrameReorderer {
 124 public:
 125         FrameReorderer(unsigned queue_length, int width, int height);
 126
 127         struct Frame {
 128                 int64_t pts, duration;
 129                 uint8_t *data;
 130
 131                 // Invert to get the smallest pts first.
 132                 bool operator< (const Frame &other) const { return pts > other.pts; }
 133         };
 134
 135         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 136         // Does _not_ take ownership of data; a copy is taken if needed.
 137         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 138         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 139         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data);
 140
 141         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 142         Frame get_first_frame();
 143
 144         bool empty() const { return frames.empty(); }
 145
 146 private:
 147         unsigned queue_length;
 148         int width, height;
 149
 150         priority_queue<Frame> frames;
 151         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 152
 153         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 154         // except priority_queue doesn't work well with movable-only types.
 155         vector<unique_ptr<uint8_t[]>> owner;
 156 };
 157
 158 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 159     : queue_length(queue_length), width(width), height(height)
 160 {
 161         for (unsigned i = 0; i < queue_length; ++i) {
 162                 owner.emplace_back(new uint8_t[width * height * 2]);
 163                 freelist.push(owner.back().get());
 164         }
 165 }
 166
 167 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data)
 168 {
 169         if (queue_length == 0) {
 170                 return Frame{pts, duration, data};
 171         }
 172
 173         assert(!freelist.empty());
 174         uint8_t *storage = freelist.top();
 175         freelist.pop();
 176         memcpy(storage, data, width * height * 2);
 177         frames.push(Frame{pts, duration, storage});
 178
 179         if (frames.size() >= queue_length) {
 180                 return get_first_frame();
 181         } else {
 182                 return Frame{-1, -1, nullptr};
 183         }
 184 }
 185
 186 FrameReorderer::Frame FrameReorderer::get_first_frame()
 187 {
 188         assert(!frames.empty());
 189         Frame storage = frames.top();
 190         frames.pop();
 191         freelist.push(storage.data);
 192         return storage;
 193 }
 194
 195 class QuickSyncEncoderImpl {
 196 public:
 197         QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder);
 198         ~QuickSyncEncoderImpl();
 199         void add_audio(int64_t pts, vector<float> audio);
 200         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 201         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
 202         void shutdown();
 203
 204 private:
 205         struct storage_task {
 206                 unsigned long long display_order;
 207                 int frame_type;
 208                 vector<float> audio;
 209                 int64_t pts, dts, duration;
 210         };
 211         struct PendingFrame {
 212                 RefCountedGLsync fence;
 213                 vector<RefCountedFrame> input_frames;
 214                 int64_t pts, duration;
 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         void open_output_file(const std::string &filename);
 223         void encode_thread_func();
 224         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 225         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 226         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 227                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 228         void storage_task_thread();
 229         void encode_remaining_audio();
 230         void storage_task_enqueue(storage_task task);
 231         void save_codeddata(storage_task task);
 232         int render_packedsequence();
 233         int render_packedpicture();
 234         void render_packedslice();
 235         int render_sequence();
 236         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 237         void sps_rbsp(bitstream *bs);
 238         void pps_rbsp(bitstream *bs);
 239         int build_packed_pic_buffer(unsigned char **header_buffer);
 240         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 241         void slice_header(bitstream *bs);
 242         int build_packed_seq_buffer(unsigned char **header_buffer);
 243         int build_packed_slice_buffer(unsigned char **header_buffer);
 244         int init_va(const string &va_display);
 245         int deinit_va();
 246         void enable_zerocopy_if_possible();
 247         VADisplay va_open_display(const string &va_display);
 248         void va_close_display(VADisplay va_dpy);
 249         int setup_encode();
 250         int release_encode();
 251         void update_ReferenceFrames(int frame_type);
 252         int update_RefPicList(int frame_type);
 253
 254         bool is_shutdown = false;
 255         bool use_zerocopy;
 256         int drm_fd = -1;
 257
 258         thread encode_thread, storage_thread;
 259
 260         mutex storage_task_queue_mutex;
 261         condition_variable storage_task_queue_changed;
 262         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 263         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 264         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 265
 266         mutex frame_queue_mutex;
 267         condition_variable frame_queue_nonempty;
 268         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 269
 270         int current_storage_frame;
 271
 272         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 273         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 274         movit::ResourcePool *resource_pool;
 275         QSurface *surface;
 276
 277         unique_ptr<AudioEncoder> file_audio_encoder;
 278         AudioEncoder *stream_audio_encoder;
 279
 280         unique_ptr<FrameReorderer> reorderer;
 281         X264Encoder *x264_encoder;  // nullptr if not using x264.
 282
 283         Mux* stream_mux;  // To HTTP.
 284         unique_ptr<Mux> file_mux;  // To local disk.
 285
 286         Display *x11_display = nullptr;
 287
 288         // Encoder parameters
 289         VADisplay va_dpy;
 290         VAProfile h264_profile = (VAProfile)~0;
 291         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 292         int config_attrib_num = 0, enc_packed_header_idx;
 293
 294         struct GLSurface {
 295                 VASurfaceID src_surface, ref_surface;
 296                 VABufferID coded_buf;
 297
 298                 VAImage surface_image;
 299                 GLuint y_tex, cbcr_tex;
 300
 301                 // Only if use_zerocopy == true.
 302                 EGLImage y_egl_image, cbcr_egl_image;
 303
 304                 // Only if use_zerocopy == false.
 305                 GLuint pbo;
 306                 uint8_t *y_ptr, *cbcr_ptr;
 307                 size_t y_offset, cbcr_offset;
 308         };
 309         GLSurface gl_surfaces[SURFACE_NUM];
 310
 311         VAConfigID config_id;
 312         VAContextID context_id;
 313         VAEncSequenceParameterBufferH264 seq_param;
 314         VAEncPictureParameterBufferH264 pic_param;
 315         VAEncSliceParameterBufferH264 slice_param;
 316         VAPictureH264 CurrentCurrPic;
 317         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 318
 319         // Static quality settings.
 320         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 321         static constexpr unsigned int num_ref_frames = 2;
 322         static constexpr int initial_qp = 15;
 323         static constexpr int minimal_qp = 0;
 324         static constexpr int intra_period = 30;
 325         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 326
 327         // Quality settings that are meant to be static, but might be overridden
 328         // by the profile.
 329         int constraint_set_flag = 0;
 330         int h264_packedheader = 0; /* support pack header? */
 331         int h264_maxref = (1<<16|1);
 332         int h264_entropy_mode = 1; /* cabac */
 333         int ip_period = 3;
 334
 335         int rc_mode = -1;
 336         unsigned int current_frame_num = 0;
 337         unsigned int numShortTerm = 0;
 338
 339         int frame_width;
 340         int frame_height;
 341         int frame_width_mbaligned;
 342         int frame_height_mbaligned;
 343 };
 344
 345 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 346 // but if we don't delete it here, we get leaks. The GStreamer implementation
 347 // does the same.
 348 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 349 {
 350     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 351     CHECK_VASTATUS(va_status, "vaRenderPicture");
 352
 353     for (int i = 0; i < num_buffers; ++i) {
 354         va_status = vaDestroyBuffer(dpy, buffers[i]);
 355         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 356     }
 357 }
 358
 359 static unsigned int
 360 va_swap32(unsigned int val)
 361 {
 362     unsigned char *pval = (unsigned char *)&val;
 363
 364     return ((pval[0] << 24)     |
 365             (pval[1] << 16)     |
 366             (pval[2] << 8)      |
 367             (pval[3] << 0));
 368 }
 369
 370 static void
 371 bitstream_start(bitstream *bs)
 372 {
 373     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 374     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 375     bs->bit_offset = 0;
 376 }
 377
 378 static void
 379 bitstream_end(bitstream *bs)
 380 {
 381     int pos = (bs->bit_offset >> 5);
 382     int bit_offset = (bs->bit_offset & 0x1f);
 383     int bit_left = 32 - bit_offset;
 384
 385     if (bit_offset) {
 386         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 387     }
 388 }
 389
 390 static void
 391 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 392 {
 393     int pos = (bs->bit_offset >> 5);
 394     int bit_offset = (bs->bit_offset & 0x1f);
 395     int bit_left = 32 - bit_offset;
 396
 397     if (!size_in_bits)
 398         return;
 399
 400     bs->bit_offset += size_in_bits;
 401
 402     if (bit_left > size_in_bits) {
 403         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 404     } else {
 405         size_in_bits -= bit_left;
 406         if (bit_left >= 32) {
 407             bs->buffer[pos] = (val >> size_in_bits);
 408         } else {
 409             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 410         }
 411         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 412
 413         if (pos + 1 == bs->max_size_in_dword) {
 414             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 415             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 416         }
 417
 418         bs->buffer[pos + 1] = val;
 419     }
 420 }
 421
 422 static void
 423 bitstream_put_ue(bitstream *bs, unsigned int val)
 424 {
 425     int size_in_bits = 0;
 426     int tmp_val = ++val;
 427
 428     while (tmp_val) {
 429         tmp_val >>= 1;
 430         size_in_bits++;
 431     }
 432
 433     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 434     bitstream_put_ui(bs, val, size_in_bits);
 435 }
 436
 437 static void
 438 bitstream_put_se(bitstream *bs, int val)
 439 {
 440     unsigned int new_val;
 441
 442     if (val <= 0)
 443         new_val = -2 * val;
 444     else
 445         new_val = 2 * val - 1;
 446
 447     bitstream_put_ue(bs, new_val);
 448 }
 449
 450 static void
 451 bitstream_byte_aligning(bitstream *bs, int bit)
 452 {
 453     int bit_offset = (bs->bit_offset & 0x7);
 454     int bit_left = 8 - bit_offset;
 455     int new_val;
 456
 457     if (!bit_offset)
 458         return;
 459
 460     assert(bit == 0 || bit == 1);
 461
 462     if (bit)
 463         new_val = (1 << bit_left) - 1;
 464     else
 465         new_val = 0;
 466
 467     bitstream_put_ui(bs, new_val, bit_left);
 468 }
 469
 470 static void
 471 rbsp_trailing_bits(bitstream *bs)
 472 {
 473     bitstream_put_ui(bs, 1, 1);
 474     bitstream_byte_aligning(bs, 0);
 475 }
 476
 477 static void nal_start_code_prefix(bitstream *bs)
 478 {
 479     bitstream_put_ui(bs, 0x00000001, 32);
 480 }
 481
 482 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 483 {
 484     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 485     bitstream_put_ui(bs, nal_ref_idc, 2);
 486     bitstream_put_ui(bs, nal_unit_type, 5);
 487 }
 488
 489 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 490 {
 491     int profile_idc = PROFILE_IDC_BASELINE;
 492
 493     if (h264_profile  == VAProfileH264High)
 494         profile_idc = PROFILE_IDC_HIGH;
 495     else if (h264_profile  == VAProfileH264Main)
 496         profile_idc = PROFILE_IDC_MAIN;
 497
 498     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 499     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 500     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 501     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 502     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 503     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 504     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 505     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 506
 507     if ( profile_idc == PROFILE_IDC_HIGH) {
 508         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 509         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 510         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 511         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 512         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 513     }
 514
 515     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 516     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 517
 518     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 519         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 520     else {
 521         assert(0);
 522     }
 523
 524     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 525     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 526
 527     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 528     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 529     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 530
 531     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 532         assert(0);
 533     }
 534
 535     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 536     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 537
 538     if (seq_param.frame_cropping_flag) {
 539         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 540         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 541         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 542         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 543     }
 544
 545     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 546     if ( false ) {
 547         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 548     } else {
 549         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 550         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 551         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 552         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 553         {
 554             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 555             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 556             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 557             {
 558                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 559                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 560                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 561             }
 562         }
 563         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 564         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 565         {
 566             bitstream_put_ui(bs, 1, 32);  // FPS
 567             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 568             bitstream_put_ui(bs, 1, 1);
 569         }
 570         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 571         {
 572             // hrd_parameters
 573             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 574             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 575             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 576
 577             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 578             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 579             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 580
 581             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 582             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 583             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 584             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 585         }
 586         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 587         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 588
 589         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 590         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 591     }
 592
 593     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 594 }
 595
 596
 597 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 598 {
 599     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 600     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 601
 602     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 603
 604     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 605
 606     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 607
 608     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 609     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 610
 611     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 612     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 613
 614     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 615     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 616     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 617
 618     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 619     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 620     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 621
 622     /* more_rbsp_data */
 623     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 624     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 625     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 626
 627     rbsp_trailing_bits(bs);
 628 }
 629
 630 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 631 {
 632     int first_mb_in_slice = slice_param.macroblock_address;
 633
 634     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 635     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 636     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 637     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 638
 639     /* frame_mbs_only_flag == 1 */
 640     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 641         /* FIXME: */
 642         assert(0);
 643     }
 644
 645     if (pic_param.pic_fields.bits.idr_pic_flag)
 646         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 647
 648     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 649         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 650         /* pic_order_present_flag == 0 */
 651     } else {
 652         /* FIXME: */
 653         assert(0);
 654     }
 655
 656     /* redundant_pic_cnt_present_flag == 0 */
 657     /* slice type */
 658     if (IS_P_SLICE(slice_param.slice_type)) {
 659         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 660
 661         if (slice_param.num_ref_idx_active_override_flag)
 662             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 663
 664         /* ref_pic_list_reordering */
 665         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 666     } else if (IS_B_SLICE(slice_param.slice_type)) {
 667         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 668
 669         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 670
 671         if (slice_param.num_ref_idx_active_override_flag) {
 672             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 673             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 674         }
 675
 676         /* ref_pic_list_reordering */
 677         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 678         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 679     }
 680
 681     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 682          IS_P_SLICE(slice_param.slice_type)) ||
 683         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 684          IS_B_SLICE(slice_param.slice_type))) {
 685         /* FIXME: fill weight/offset table */
 686         assert(0);
 687     }
 688
 689     /* dec_ref_pic_marking */
 690     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 691         unsigned char no_output_of_prior_pics_flag = 0;
 692         unsigned char long_term_reference_flag = 0;
 693         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 694
 695         if (pic_param.pic_fields.bits.idr_pic_flag) {
 696             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 697             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 698         } else {
 699             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 700         }
 701     }
 702
 703     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 704         !IS_I_SLICE(slice_param.slice_type))
 705         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 706
 707     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 708
 709     /* ignore for SP/SI */
 710
 711     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 712         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 713
 714         if (slice_param.disable_deblocking_filter_idc != 1) {
 715             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 716             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 717         }
 718     }
 719
 720     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 721         bitstream_byte_aligning(bs, 1);
 722     }
 723 }
 724
 725 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 726 {
 727     bitstream bs;
 728
 729     bitstream_start(&bs);
 730     nal_start_code_prefix(&bs);
 731     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 732     pps_rbsp(&bs);
 733     bitstream_end(&bs);
 734
 735     *header_buffer = (unsigned char *)bs.buffer;
 736     return bs.bit_offset;
 737 }
 738
 739 int
 740 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 741 {
 742     bitstream bs;
 743
 744     bitstream_start(&bs);
 745     nal_start_code_prefix(&bs);
 746     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 747     sps_rbsp(&bs);
 748     bitstream_end(&bs);
 749
 750     *header_buffer = (unsigned char *)bs.buffer;
 751     return bs.bit_offset;
 752 }
 753
 754 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 755 {
 756     bitstream bs;
 757     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 758     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 759
 760     bitstream_start(&bs);
 761     nal_start_code_prefix(&bs);
 762
 763     if (IS_I_SLICE(slice_param.slice_type)) {
 764         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 765     } else if (IS_P_SLICE(slice_param.slice_type)) {
 766         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 767     } else {
 768         assert(IS_B_SLICE(slice_param.slice_type));
 769         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 770     }
 771
 772     slice_header(&bs);
 773     bitstream_end(&bs);
 774
 775     *header_buffer = (unsigned char *)bs.buffer;
 776     return bs.bit_offset;
 777 }
 778
 779
 780 /*
 781   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 782   1) period between Frame #X and Frame #N = #X - #N
 783   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 784   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 785   4) intra_period and intra_idr_period take precedence over ip_period
 786   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 787      of I/IDR frames, see bellow examples
 788   -------------------------------------------------------------------
 789   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 790   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 791   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 792   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 793   1            1                ignored    IDR IDR IDR IDR...
 794   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 795   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 796   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 797                                               (PBB)(IBB)(PBB)(IBB)...
 798   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 799                                            IDRPPPPP IPPPPP IPPPPP...
 800   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 801                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 802                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 803                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 804                                            {IDR(PBB)(PBB)}                     (6/6/3)
 805                                            {IDR(PBB)(PBB)}.
 806 */
 807
 808 // General pts/dts strategy:
 809 //
 810 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 811 // bit tricky. We assume first of all that the frame rate never goes _above_
 812 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 813 // in at least this speed, as long at dts <= pts (the frame is not attempted
 814 // presented before it is decoded). Furthermore, we never have longer chains of
 815 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 816 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 817 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 818 // frame, with an order O <= C.)
 819 //
 820 // Many strategies are possible, but we establish these rules:
 821 //
 822 //  - Tip frames have dts = pts - (C-O)*N.
 823 //  - Non-tip frames have dts = dts_last + N.
 824 //
 825 // An example, with C=2 and N=10 and the data flow showed with arrows:
 826 //
 827 //        I  B  P  B  B  P
 828 //   pts: 30 40 50 60 70 80
 829 //        ↓  ↓     ↓
 830 //   dts: 10 30 20 60 50←40
 831 //         |  |  ↑        ↑
 832 //         `--|--'        |
 833 //             `----------'
 834 //
 835 // To show that this works fine also with irregular spacings, let's say that
 836 // the third frame is delayed a bit (something earlier was dropped). Now the
 837 // situation looks like this:
 838 //
 839 //        I  B  P  B  B   P
 840 //   pts: 30 40 80 90 100 110
 841 //        ↓  ↓     ↓
 842 //   dts: 10 30 20 90 50←40
 843 //         |  |  ↑        ↑
 844 //         `--|--'        |
 845 //             `----------'
 846 //
 847 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 848 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 849 //
 850 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 851 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 852 // a tip frame and should be given a dts based on the previous one.
 853 #define FRAME_P 0
 854 #define FRAME_B 1
 855 #define FRAME_I 2
 856 #define FRAME_IDR 7
 857 void encoding2display_order(
 858     int encoding_order, int intra_period,
 859     int intra_idr_period, int ip_period,
 860     int *displaying_order,
 861     int *frame_type, int *pts_lag)
 862 {
 863     int encoding_order_gop = 0;
 864
 865     *pts_lag = 0;
 866
 867     if (intra_period == 1) { /* all are I/IDR frames */
 868         *displaying_order = encoding_order;
 869         if (intra_idr_period == 0)
 870             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 871         else
 872             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 873         return;
 874     }
 875
 876     if (intra_period == 0)
 877         intra_idr_period = 0;
 878
 879     if (ip_period == 1) {
 880         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 881         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 882         *displaying_order = encoding_order;
 883
 884         if (encoding_order_gop == 0) { /* the first frame */
 885             *frame_type = FRAME_IDR;
 886         } else if (intra_period != 0 && /* have I frames */
 887                    encoding_order_gop >= 2 &&
 888                    (encoding_order_gop % intra_period == 0)) {
 889             *frame_type = FRAME_I;
 890         } else {
 891             *frame_type = FRAME_P;
 892         }
 893         return;
 894     }
 895
 896     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 897     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 898     *pts_lag = -1;  // Most frames are not tip frames.
 899
 900     if (encoding_order_gop == 0) { /* the first frame */
 901         *frame_type = FRAME_IDR;
 902         *displaying_order = encoding_order;
 903         // IDR frames are a special case; I honestly can't find the logic behind
 904         // why this is the right thing, but it seems to line up nicely in practice :-)
 905         *pts_lag = TIMEBASE / MAX_FPS;
 906     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 907         *frame_type = FRAME_B;
 908         *displaying_order = encoding_order - 1;
 909         if ((encoding_order_gop % ip_period) == 0) {
 910             *pts_lag = 0;  // Last B-frame.
 911         }
 912     } else if (intra_period != 0 && /* have I frames */
 913                encoding_order_gop >= 2 &&
 914                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 915         *frame_type = FRAME_I;
 916         *displaying_order = encoding_order + ip_period - 1;
 917     } else {
 918         *frame_type = FRAME_P;
 919         *displaying_order = encoding_order + ip_period - 1;
 920     }
 921 }
 922
 923
 924 static const char *rc_to_string(int rc_mode)
 925 {
 926     switch (rc_mode) {
 927     case VA_RC_NONE:
 928         return "NONE";
 929     case VA_RC_CBR:
 930         return "CBR";
 931     case VA_RC_VBR:
 932         return "VBR";
 933     case VA_RC_VCM:
 934         return "VCM";
 935     case VA_RC_CQP:
 936         return "CQP";
 937     case VA_RC_VBR_CONSTRAINED:
 938         return "VBR_CONSTRAINED";
 939     default:
 940         return "Unknown";
 941     }
 942 }
 943
 944 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 945 {
 946         if (global_flags.uncompressed_video_to_http) {
 947                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 948                 use_zerocopy = false;
 949         } else if (global_flags.x264_video_to_http) {
 950                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 951                 use_zerocopy = false;
 952         } else {
 953                 use_zerocopy = true;
 954         }
 955 }
 956
 957 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 958 {
 959         if (va_display.empty()) {
 960                 x11_display = XOpenDisplay(NULL);
 961                 if (!x11_display) {
 962                         fprintf(stderr, "error: can't connect to X server!\n");
 963                         return NULL;
 964                 }
 965                 enable_zerocopy_if_possible();
 966                 return vaGetDisplay(x11_display);
 967         } else if (va_display[0] != '/') {
 968                 x11_display = XOpenDisplay(va_display.c_str());
 969                 if (!x11_display) {
 970                         fprintf(stderr, "error: can't connect to X server!\n");
 971                         return NULL;
 972                 }
 973                 enable_zerocopy_if_possible();
 974                 return vaGetDisplay(x11_display);
 975         } else {
 976                 drm_fd = open(va_display.c_str(), O_RDWR);
 977                 if (drm_fd == -1) {
 978                         perror(va_display.c_str());
 979                         return NULL;
 980                 }
 981                 use_zerocopy = false;
 982                 return vaGetDisplayDRM(drm_fd);
 983         }
 984 }
 985
 986 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
 987 {
 988         if (x11_display) {
 989                 XCloseDisplay(x11_display);
 990                 x11_display = nullptr;
 991         }
 992         if (drm_fd != -1) {
 993                 close(drm_fd);
 994         }
 995 }
 996
 997 int QuickSyncEncoderImpl::init_va(const string &va_display)
 998 {
 999     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1000     VAEntrypoint *entrypoints;
1001     int num_entrypoints, slice_entrypoint;
1002     int support_encode = 0;
1003     int major_ver, minor_ver;
1004     VAStatus va_status;
1005     unsigned int i;
1006
1007     va_dpy = va_open_display(va_display);
1008     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1009     CHECK_VASTATUS(va_status, "vaInitialize");
1010
1011     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1012     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1013     if (!entrypoints) {
1014         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1015         exit(1);
1016     }
1017
1018     /* use the highest profile */
1019     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1020         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1021             continue;
1022
1023         h264_profile = profile_list[i];
1024         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1025         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1026             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1027                 support_encode = 1;
1028                 break;
1029             }
1030         }
1031         if (support_encode == 1)
1032             break;
1033     }
1034
1035     if (support_encode == 0) {
1036         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1037         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1038         printf("to use VA-API against DRM instead of X11.\n");
1039         exit(1);
1040     } else {
1041         switch (h264_profile) {
1042             case VAProfileH264Baseline:
1043                 ip_period = 1;
1044                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1045                 h264_entropy_mode = 0;
1046                 break;
1047             case VAProfileH264ConstrainedBaseline:
1048                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1049                 ip_period = 1;
1050                 break;
1051
1052             case VAProfileH264Main:
1053                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1054                 break;
1055
1056             case VAProfileH264High:
1057                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1058                 break;
1059             default:
1060                 h264_profile = VAProfileH264Baseline;
1061                 ip_period = 1;
1062                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1063                 break;
1064         }
1065     }
1066
1067     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1068
1069     /* find out the format for the render target, and rate control mode */
1070     for (i = 0; i < VAConfigAttribTypeMax; i++)
1071         attrib[i].type = (VAConfigAttribType)i;
1072
1073     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1074                                       &attrib[0], VAConfigAttribTypeMax);
1075     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1076     /* check the interested configattrib */
1077     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1078         printf("Not find desired YUV420 RT format\n");
1079         exit(1);
1080     } else {
1081         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1082         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1083         config_attrib_num++;
1084     }
1085
1086     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1087         int tmp = attrib[VAConfigAttribRateControl].value;
1088
1089         if (rc_mode == -1 || !(rc_mode & tmp))  {
1090             if (rc_mode != -1) {
1091                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1092             }
1093
1094             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1095                 if (rc_default_modes[i] & tmp) {
1096                     rc_mode = rc_default_modes[i];
1097                     break;
1098                 }
1099             }
1100         }
1101
1102         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1103         config_attrib[config_attrib_num].value = rc_mode;
1104         config_attrib_num++;
1105     }
1106
1107
1108     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1109         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1110
1111         h264_packedheader = 1;
1112         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1113         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1114
1115         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1116             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1117         }
1118
1119         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1120             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1121         }
1122
1123         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1124             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1125         }
1126
1127         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1128             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1129         }
1130
1131         enc_packed_header_idx = config_attrib_num;
1132         config_attrib_num++;
1133     }
1134
1135     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1136         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1137         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1138         config_attrib_num++;
1139     }
1140
1141     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1142         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1143     }
1144
1145     free(entrypoints);
1146     return 0;
1147 }
1148
1149 int QuickSyncEncoderImpl::setup_encode()
1150 {
1151     VAStatus va_status;
1152     VASurfaceID *tmp_surfaceid;
1153     int codedbuf_size, i;
1154     VASurfaceID src_surface[SURFACE_NUM];
1155     VASurfaceID ref_surface[SURFACE_NUM];
1156
1157     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1158             &config_attrib[0], config_attrib_num, &config_id);
1159     CHECK_VASTATUS(va_status, "vaCreateConfig");
1160
1161     /* create source surfaces */
1162     va_status = vaCreateSurfaces(va_dpy,
1163                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1164                                  &src_surface[0], SURFACE_NUM,
1165                                  NULL, 0);
1166     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1167
1168     /* create reference surfaces */
1169     va_status = vaCreateSurfaces(va_dpy,
1170                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1171                                  &ref_surface[0], SURFACE_NUM,
1172                                  NULL, 0);
1173     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1174
1175     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1176     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1177     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1178
1179     /* Create a context for this encode pipe */
1180     va_status = vaCreateContext(va_dpy, config_id,
1181                                 frame_width_mbaligned, frame_height_mbaligned,
1182                                 VA_PROGRESSIVE,
1183                                 tmp_surfaceid, 2 * SURFACE_NUM,
1184                                 &context_id);
1185     CHECK_VASTATUS(va_status, "vaCreateContext");
1186     free(tmp_surfaceid);
1187
1188     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1189
1190     for (i = 0; i < SURFACE_NUM; i++) {
1191         /* create coded buffer once for all
1192          * other VA buffers which won't be used again after vaRenderPicture.
1193          * so APP can always vaCreateBuffer for every frame
1194          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1195          * so VA won't maintain the coded buffer
1196          */
1197         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1198                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1199         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1200     }
1201
1202     /* create OpenGL objects */
1203     //glGenFramebuffers(SURFACE_NUM, fbos);
1204
1205     for (i = 0; i < SURFACE_NUM; i++) {
1206         if (use_zerocopy) {
1207             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
1208             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
1209         } else {
1210             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
1211             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
1212
1213             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1214             // buffers, due to potentially differing pitch.
1215             glGenBuffers(1, &gl_surfaces[i].pbo);
1216             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1217             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1218             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1219             gl_surfaces[i].y_offset = 0;
1220             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1221             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1222             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1223             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1224         }
1225     }
1226
1227     for (i = 0; i < SURFACE_NUM; i++) {
1228         gl_surfaces[i].src_surface = src_surface[i];
1229         gl_surfaces[i].ref_surface = ref_surface[i];
1230     }
1231
1232     return 0;
1233 }
1234
1235 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1236 //
1237 //   2 1 0 [3] 4 8 9
1238 template<class T, class C>
1239 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1240 {
1241         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1242         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1243         sort(middle, end, less_than);
1244 }
1245
1246 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1247 {
1248     int i;
1249
1250     if (frame_type == FRAME_B)
1251         return;
1252
1253     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1254     numShortTerm++;
1255     if (numShortTerm > num_ref_frames)
1256         numShortTerm = num_ref_frames;
1257     for (i=numShortTerm-1; i>0; i--)
1258         ReferenceFrames[i] = ReferenceFrames[i-1];
1259     ReferenceFrames[0] = CurrentCurrPic;
1260
1261     current_frame_num++;
1262     if (current_frame_num > MaxFrameNum)
1263         current_frame_num = 0;
1264 }
1265
1266
1267 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1268 {
1269     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1270         return a.frame_idx > b.frame_idx;
1271     };
1272     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1273         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1274     };
1275     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1276         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1277     };
1278
1279     if (frame_type == FRAME_P) {
1280         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1281         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1282     } else if (frame_type == FRAME_B) {
1283         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1284         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1285
1286         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1287         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1288     }
1289
1290     return 0;
1291 }
1292
1293
1294 int QuickSyncEncoderImpl::render_sequence()
1295 {
1296     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1297     VAStatus va_status;
1298     VAEncMiscParameterBuffer *misc_param;
1299     VAEncMiscParameterRateControl *misc_rate_ctrl;
1300
1301     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1302     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1303     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1304     seq_param.bits_per_second = frame_bitrate;
1305
1306     seq_param.intra_period = intra_period;
1307     seq_param.intra_idr_period = intra_idr_period;
1308     seq_param.ip_period = ip_period;
1309
1310     seq_param.max_num_ref_frames = num_ref_frames;
1311     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1312     seq_param.time_scale = TIMEBASE * 2;
1313     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1314     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1315     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1316     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1317     seq_param.seq_fields.bits.chroma_format_idc = 1;
1318     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1319
1320     if (frame_width != frame_width_mbaligned ||
1321         frame_height != frame_height_mbaligned) {
1322         seq_param.frame_cropping_flag = 1;
1323         seq_param.frame_crop_left_offset = 0;
1324         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1325         seq_param.frame_crop_top_offset = 0;
1326         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1327     }
1328
1329     va_status = vaCreateBuffer(va_dpy, context_id,
1330                                VAEncSequenceParameterBufferType,
1331                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1332     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1333
1334     va_status = vaCreateBuffer(va_dpy, context_id,
1335                                VAEncMiscParameterBufferType,
1336                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1337                                1, NULL, &rc_param_buf);
1338     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1339
1340     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1341     misc_param->type = VAEncMiscParameterTypeRateControl;
1342     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1343     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1344     misc_rate_ctrl->bits_per_second = frame_bitrate;
1345     misc_rate_ctrl->target_percentage = 66;
1346     misc_rate_ctrl->window_size = 1000;
1347     misc_rate_ctrl->initial_qp = initial_qp;
1348     misc_rate_ctrl->min_qp = minimal_qp;
1349     misc_rate_ctrl->basic_unit_size = 0;
1350     vaUnmapBuffer(va_dpy, rc_param_buf);
1351
1352     render_id[0] = seq_param_buf;
1353     render_id[1] = rc_param_buf;
1354
1355     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1356
1357     return 0;
1358 }
1359
1360 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1361 {
1362     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1363     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1364     int PicOrderCntMsb, TopFieldOrderCnt;
1365
1366     if (frame_type == FRAME_IDR)
1367         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1368     else {
1369         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1370         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1371     }
1372
1373     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1374         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1375         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1376     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1377              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1378         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1379     else
1380         PicOrderCntMsb = prevPicOrderCntMsb;
1381
1382     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1383
1384     if (frame_type != FRAME_B) {
1385         PicOrderCntMsb_ref = PicOrderCntMsb;
1386         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1387     }
1388
1389     return TopFieldOrderCnt;
1390 }
1391
1392 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1393 {
1394     VABufferID pic_param_buf;
1395     VAStatus va_status;
1396     int i = 0;
1397
1398     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1399     pic_param.CurrPic.frame_idx = current_frame_num;
1400     pic_param.CurrPic.flags = 0;
1401     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1402     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1403     CurrentCurrPic = pic_param.CurrPic;
1404
1405     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1406     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1407         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1408         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1409     }
1410
1411     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1412     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1413     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1414     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1415     pic_param.frame_num = current_frame_num;
1416     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1417     pic_param.last_picture = false;  // FIXME
1418     pic_param.pic_init_qp = initial_qp;
1419
1420     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1421                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1422     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1423
1424     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1425
1426     return 0;
1427 }
1428
1429 int QuickSyncEncoderImpl::render_packedsequence()
1430 {
1431     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1432     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1433     unsigned int length_in_bits;
1434     unsigned char *packedseq_buffer = NULL;
1435     VAStatus va_status;
1436
1437     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1438
1439     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1440
1441     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1442     packedheader_param_buffer.has_emulation_bytes = 0;
1443     va_status = vaCreateBuffer(va_dpy,
1444                                context_id,
1445                                VAEncPackedHeaderParameterBufferType,
1446                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1447                                &packedseq_para_bufid);
1448     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1449
1450     va_status = vaCreateBuffer(va_dpy,
1451                                context_id,
1452                                VAEncPackedHeaderDataBufferType,
1453                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1454                                &packedseq_data_bufid);
1455     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1456
1457     render_id[0] = packedseq_para_bufid;
1458     render_id[1] = packedseq_data_bufid;
1459     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1460
1461     free(packedseq_buffer);
1462
1463     return 0;
1464 }
1465
1466
1467 int QuickSyncEncoderImpl::render_packedpicture()
1468 {
1469     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1470     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1471     unsigned int length_in_bits;
1472     unsigned char *packedpic_buffer = NULL;
1473     VAStatus va_status;
1474
1475     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1476     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1477     packedheader_param_buffer.bit_length = length_in_bits;
1478     packedheader_param_buffer.has_emulation_bytes = 0;
1479
1480     va_status = vaCreateBuffer(va_dpy,
1481                                context_id,
1482                                VAEncPackedHeaderParameterBufferType,
1483                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1484                                &packedpic_para_bufid);
1485     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1486
1487     va_status = vaCreateBuffer(va_dpy,
1488                                context_id,
1489                                VAEncPackedHeaderDataBufferType,
1490                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1491                                &packedpic_data_bufid);
1492     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1493
1494     render_id[0] = packedpic_para_bufid;
1495     render_id[1] = packedpic_data_bufid;
1496     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1497
1498     free(packedpic_buffer);
1499
1500     return 0;
1501 }
1502
1503 void QuickSyncEncoderImpl::render_packedslice()
1504 {
1505     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1506     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1507     unsigned int length_in_bits;
1508     unsigned char *packedslice_buffer = NULL;
1509     VAStatus va_status;
1510
1511     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1512     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1513     packedheader_param_buffer.bit_length = length_in_bits;
1514     packedheader_param_buffer.has_emulation_bytes = 0;
1515
1516     va_status = vaCreateBuffer(va_dpy,
1517                                context_id,
1518                                VAEncPackedHeaderParameterBufferType,
1519                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1520                                &packedslice_para_bufid);
1521     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1522
1523     va_status = vaCreateBuffer(va_dpy,
1524                                context_id,
1525                                VAEncPackedHeaderDataBufferType,
1526                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1527                                &packedslice_data_bufid);
1528     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1529
1530     render_id[0] = packedslice_para_bufid;
1531     render_id[1] = packedslice_data_bufid;
1532     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1533
1534     free(packedslice_buffer);
1535 }
1536
1537 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1538 {
1539     VABufferID slice_param_buf;
1540     VAStatus va_status;
1541     int i;
1542
1543     update_RefPicList(frame_type);
1544
1545     /* one frame, one slice */
1546     slice_param.macroblock_address = 0;
1547     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1548     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1549     if (frame_type == FRAME_IDR) {
1550         if (encoding_frame_num != 0)
1551             ++slice_param.idr_pic_id;
1552     } else if (frame_type == FRAME_P) {
1553         int refpiclist0_max = h264_maxref & 0xffff;
1554         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1555
1556         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1557             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1558             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1559         }
1560     } else if (frame_type == FRAME_B) {
1561         int refpiclist0_max = h264_maxref & 0xffff;
1562         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1563
1564         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1565         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1566             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1567             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1568         }
1569
1570         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1571         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1572             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1573             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1574         }
1575     }
1576
1577     slice_param.slice_alpha_c0_offset_div2 = 0;
1578     slice_param.slice_beta_offset_div2 = 0;
1579     slice_param.direct_spatial_mv_pred_flag = 1;
1580     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1581
1582
1583     if (h264_packedheader &&
1584         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1585         render_packedslice();
1586
1587     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1588                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1589     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1590
1591     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1592
1593     return 0;
1594 }
1595
1596
1597
1598 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1599 {
1600         VACodedBufferSegment *buf_list = NULL;
1601         VAStatus va_status;
1602
1603         string data;
1604
1605         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1606         CHECK_VASTATUS(va_status, "vaMapBuffer");
1607         while (buf_list != NULL) {
1608                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1609                 buf_list = (VACodedBufferSegment *) buf_list->next;
1610         }
1611         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1612
1613         {
1614                 // Add video.
1615                 AVPacket pkt;
1616                 memset(&pkt, 0, sizeof(pkt));
1617                 pkt.buf = nullptr;
1618                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1619                 pkt.size = data.size();
1620                 pkt.stream_index = 0;
1621                 if (task.frame_type == FRAME_IDR) {
1622                         pkt.flags = AV_PKT_FLAG_KEY;
1623                 } else {
1624                         pkt.flags = 0;
1625                 }
1626                 pkt.duration = task.duration;
1627                 if (file_mux) {
1628                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1629                 }
1630                 if (!global_flags.uncompressed_video_to_http &&
1631                     !global_flags.x264_video_to_http) {
1632                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1633                 }
1634         }
1635         // Encode and add all audio frames up to and including the pts of this video frame.
1636         for ( ;; ) {
1637                 int64_t audio_pts;
1638                 vector<float> audio;
1639                 {
1640                         unique_lock<mutex> lock(frame_queue_mutex);
1641                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1642                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1643                         auto it = pending_audio_frames.begin();
1644                         if (it->first > task.pts) break;
1645                         audio_pts = it->first;
1646                         audio = move(it->second);
1647                         pending_audio_frames.erase(it);
1648                 }
1649
1650                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1651                 stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1652
1653                 if (audio_pts == task.pts) break;
1654         }
1655 }
1656
1657
1658 // this is weird. but it seems to put a new frame onto the queue
1659 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1660 {
1661         unique_lock<mutex> lock(storage_task_queue_mutex);
1662         storage_task_queue.push(move(task));
1663         storage_task_queue_changed.notify_all();
1664 }
1665
1666 void QuickSyncEncoderImpl::storage_task_thread()
1667 {
1668         for ( ;; ) {
1669                 storage_task current;
1670                 {
1671                         // wait until there's an encoded frame
1672                         unique_lock<mutex> lock(storage_task_queue_mutex);
1673                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1674                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1675                         current = move(storage_task_queue.front());
1676                         storage_task_queue.pop();
1677                 }
1678
1679                 VAStatus va_status;
1680
1681                 // waits for data, then saves it to disk.
1682                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1683                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1684                 save_codeddata(move(current));
1685
1686                 {
1687                         unique_lock<mutex> lock(storage_task_queue_mutex);
1688                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1689                         storage_task_queue_changed.notify_all();
1690                 }
1691         }
1692 }
1693
1694 int QuickSyncEncoderImpl::release_encode()
1695 {
1696         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1697                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1698                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1699                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1700
1701                 if (!use_zerocopy) {
1702                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1703                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1704                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1705                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1706                 }
1707                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1708                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1709         }
1710
1711         vaDestroyContext(va_dpy, context_id);
1712         vaDestroyConfig(va_dpy, config_id);
1713
1714         return 0;
1715 }
1716
1717 int QuickSyncEncoderImpl::deinit_va()
1718 {
1719     vaTerminate(va_dpy);
1720
1721     va_close_display(va_dpy);
1722
1723     return 0;
1724 }
1725
1726 namespace {
1727
1728 }  // namespace
1729
1730 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder)
1731         : current_storage_frame(0), resource_pool(resource_pool), surface(surface), stream_audio_encoder(stream_audio_encoder), x264_encoder(x264_encoder), stream_mux(stream_mux), frame_width(width), frame_height(height)
1732 {
1733         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE));
1734         open_output_file(filename);
1735         file_audio_encoder->add_mux(file_mux.get());
1736
1737         frame_width_mbaligned = (frame_width + 15) & (~15);
1738         frame_height_mbaligned = (frame_height + 15) & (~15);
1739
1740         //print_input();
1741
1742         if (global_flags.uncompressed_video_to_http ||
1743             global_flags.x264_video_to_http) {
1744                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1745         }
1746         if (global_flags.x264_video_to_http) {
1747                 assert(x264_encoder != nullptr);
1748         } else {
1749                 assert(x264_encoder == nullptr);
1750         }
1751
1752         init_va(va_display);
1753         setup_encode();
1754
1755         // No frames are ready yet.
1756         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1757
1758         memset(&seq_param, 0, sizeof(seq_param));
1759         memset(&pic_param, 0, sizeof(pic_param));
1760         memset(&slice_param, 0, sizeof(slice_param));
1761
1762         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1763
1764         encode_thread = thread([this]{
1765                 //SDL_GL_MakeCurrent(window, context);
1766                 QOpenGLContext *context = create_context(this->surface);
1767                 eglBindAPI(EGL_OPENGL_API);
1768                 if (!make_current(context, this->surface)) {
1769                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1770                                 eglGetError());
1771                         exit(1);
1772                 }
1773                 encode_thread_func();
1774                 delete_context(context);
1775         });
1776 }
1777
1778 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1779 {
1780         shutdown();
1781 }
1782
1783 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1784 {
1785         assert(!is_shutdown);
1786         {
1787                 // Wait until this frame slot is done encoding.
1788                 unique_lock<mutex> lock(storage_task_queue_mutex);
1789                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1790                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1791                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1792                 }
1793                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1794                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1795                 if (storage_thread_should_quit) return false;
1796         }
1797
1798         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1799         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1800         *y_tex = surf->y_tex;
1801         *cbcr_tex = surf->cbcr_tex;
1802
1803         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1804         CHECK_VASTATUS(va_status, "vaDeriveImage");
1805
1806         if (use_zerocopy) {
1807                 VABufferInfo buf_info;
1808                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1809                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1810                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1811
1812                 // Create Y image.
1813                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1814                 EGLint y_attribs[] = {
1815                         EGL_WIDTH, frame_width,
1816                         EGL_HEIGHT, frame_height,
1817                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1818                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1819                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1820                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1821                         EGL_NONE
1822                 };
1823
1824                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1825                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1826
1827                 // Associate Y image to a texture.
1828                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1829                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1830
1831                 // Create CbCr image.
1832                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1833                 EGLint cbcr_attribs[] = {
1834                         EGL_WIDTH, frame_width,
1835                         EGL_HEIGHT, frame_height,
1836                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1837                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1838                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1839                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1840                         EGL_NONE
1841                 };
1842
1843                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1844                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1845
1846                 // Associate CbCr image to a texture.
1847                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1848                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1849         }
1850
1851         return true;
1852 }
1853
1854 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1855 {
1856         assert(!is_shutdown);
1857         {
1858                 unique_lock<mutex> lock(frame_queue_mutex);
1859                 pending_audio_frames[pts] = move(audio);
1860         }
1861         frame_queue_nonempty.notify_all();
1862 }
1863
1864 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1865 {
1866         assert(!is_shutdown);
1867
1868         if (!use_zerocopy) {
1869                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1870
1871                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1872                 check_error();
1873
1874                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1875                 check_error();
1876
1877                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1878                 check_error();
1879                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1880                 check_error();
1881
1882                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1883                 check_error();
1884                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1885                 check_error();
1886
1887                 glBindTexture(GL_TEXTURE_2D, 0);
1888                 check_error();
1889                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1890                 check_error();
1891
1892                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1893                 check_error();
1894         }
1895
1896         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1897         check_error();
1898         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1899         check_error();
1900
1901         {
1902                 unique_lock<mutex> lock(frame_queue_mutex);
1903                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1904                 ++current_storage_frame;
1905         }
1906         frame_queue_nonempty.notify_all();
1907         return fence;
1908 }
1909
1910 void QuickSyncEncoderImpl::shutdown()
1911 {
1912         if (is_shutdown) {
1913                 return;
1914         }
1915
1916         {
1917                 unique_lock<mutex> lock(frame_queue_mutex);
1918                 encode_thread_should_quit = true;
1919                 frame_queue_nonempty.notify_all();
1920         }
1921         encode_thread.join();
1922         {
1923                 unique_lock<mutex> lock(storage_task_queue_mutex);
1924                 storage_thread_should_quit = true;
1925                 frame_queue_nonempty.notify_all();
1926                 storage_task_queue_changed.notify_all();
1927         }
1928         storage_thread.join();
1929         encode_remaining_audio();
1930
1931         release_encode();
1932         deinit_va();
1933         file_mux.reset();
1934         is_shutdown = true;
1935 }
1936
1937 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1938 {
1939         AVFormatContext *avctx = avformat_alloc_context();
1940         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1941         assert(filename.size() < sizeof(avctx->filename) - 1);
1942         strcpy(avctx->filename, filename.c_str());
1943
1944         string url = "file:" + filename;
1945         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1946         if (ret < 0) {
1947                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1948                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1949                 exit(1);
1950         }
1951
1952         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, file_audio_encoder->get_codec(), TIMEBASE, DEFAULT_AUDIO_OUTPUT_BIT_RATE, nullptr));
1953 }
1954
1955 void QuickSyncEncoderImpl::encode_thread_func()
1956 {
1957         int64_t last_dts = -1;
1958         int gop_start_display_frame_num = 0;
1959         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1960                 PendingFrame frame;
1961                 int pts_lag;
1962                 int frame_type, display_frame_num;
1963                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1964                                        &display_frame_num, &frame_type, &pts_lag);
1965                 if (frame_type == FRAME_IDR) {
1966                         numShortTerm = 0;
1967                         current_frame_num = 0;
1968                         gop_start_display_frame_num = display_frame_num;
1969                 }
1970
1971                 {
1972                         unique_lock<mutex> lock(frame_queue_mutex);
1973                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1974                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1975                         });
1976                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1977                                 // We have queued frames that were supposed to be B-frames,
1978                                 // but will be no P-frame to encode them against. Encode them all
1979                                 // as P-frames instead. Note that this happens under the mutex,
1980                                 // but nobody else uses it at this point, since we're shutting down,
1981                                 // so there's no contention.
1982                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1983                                 return;
1984                         } else {
1985                                 frame = move(pending_video_frames[display_frame_num]);
1986                                 pending_video_frames.erase(display_frame_num);
1987                         }
1988                 }
1989
1990                 // Determine the dts of this frame.
1991                 int64_t dts;
1992                 if (pts_lag == -1) {
1993                         assert(last_dts != -1);
1994                         dts = last_dts + (TIMEBASE / MAX_FPS);
1995                 } else {
1996                         dts = frame.pts - pts_lag;
1997                 }
1998                 last_dts = dts;
1999
2000                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2001         }
2002 }
2003
2004 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2005 {
2006         if (pending_video_frames.empty()) {
2007                 return;
2008         }
2009
2010         for (auto &pending_frame : pending_video_frames) {
2011                 int display_frame_num = pending_frame.first;
2012                 assert(display_frame_num > 0);
2013                 PendingFrame frame = move(pending_frame.second);
2014                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2015                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2016                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2017                 last_dts = dts;
2018         }
2019
2020         if (global_flags.uncompressed_video_to_http ||
2021             global_flags.x264_video_to_http) {
2022                 // Add frames left in reorderer.
2023                 while (!reorderer->empty()) {
2024                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2025                         if (global_flags.uncompressed_video_to_http) {
2026                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2027                         } else {
2028                                 assert(global_flags.x264_video_to_http);
2029                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2030                         }
2031                 }
2032         }
2033 }
2034
2035 void QuickSyncEncoderImpl::encode_remaining_audio()
2036 {
2037         // This really ought to be empty by now, but just to be sure...
2038         for (auto &pending_frame : pending_audio_frames) {
2039                 int64_t audio_pts = pending_frame.first;
2040                 vector<float> audio = move(pending_frame.second);
2041
2042                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2043                 if (stream_audio_encoder) {
2044                         stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2045                 }
2046         }
2047         pending_audio_frames.clear();
2048
2049         // Encode any leftover audio in the queues, and also any delayed frames.
2050         // Note: stream_audio_encoder is not owned by us, so don't call encode_last_audio().
2051         file_audio_encoder->encode_last_audio();
2052 }
2053
2054 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2055 {
2056         AVPacket pkt;
2057         memset(&pkt, 0, sizeof(pkt));
2058         pkt.buf = nullptr;
2059         pkt.data = const_cast<uint8_t *>(data);
2060         pkt.size = frame_width * frame_height * 2;
2061         pkt.stream_index = 0;
2062         pkt.flags = AV_PKT_FLAG_KEY;
2063         pkt.duration = duration;
2064         stream_mux->add_packet(pkt, pts, pts);
2065 }
2066
2067 namespace {
2068
2069 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2070 {
2071         if (src_width == dst_pitch) {
2072                 memcpy(dst, src, src_width * height);
2073         } else {
2074                 for (size_t y = 0; y < height; ++y) {
2075                         const uint8_t *sptr = src + y * src_width;
2076                         uint8_t *dptr = dst + y * dst_pitch;
2077                         memcpy(dptr, sptr, src_width);
2078                 }
2079         }
2080 }
2081
2082 }  // namespace
2083
2084 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2085                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2086 {
2087         // Wait for the GPU to be done with the frame.
2088         GLenum sync_status;
2089         do {
2090                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2091                 check_error();
2092         } while (sync_status == GL_TIMEOUT_EXPIRED);
2093         assert(sync_status != GL_WAIT_FAILED);
2094
2095         // Release back any input frames we needed to render this frame.
2096         frame.input_frames.clear();
2097
2098         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2099         VAStatus va_status;
2100
2101         if (use_zerocopy) {
2102                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2103                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2104                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2105                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2106         } else {
2107                 unsigned char *surface_p = nullptr;
2108                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2109
2110                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2111                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2112
2113                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2114                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2115
2116                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2117                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2118
2119                 if (global_flags.uncompressed_video_to_http ||
2120                     global_flags.x264_video_to_http) {
2121                         // Add uncompressed video. (Note that pts == dts here.)
2122                         // Delay needs to match audio.
2123                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2124                         if (output_frame.data != nullptr) {
2125                                 if (global_flags.uncompressed_video_to_http) {
2126                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2127                                 } else {
2128                                         assert(global_flags.x264_video_to_http);
2129                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2130                                 }
2131                         }
2132                 }
2133         }
2134
2135         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2136         CHECK_VASTATUS(va_status, "vaDestroyImage");
2137
2138         // Schedule the frame for encoding.
2139         VASurfaceID va_surface = surf->src_surface;
2140         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2141         CHECK_VASTATUS(va_status, "vaBeginPicture");
2142
2143         if (frame_type == FRAME_IDR) {
2144                 render_sequence();
2145                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2146                 if (h264_packedheader) {
2147                         render_packedsequence();
2148                         render_packedpicture();
2149                 }
2150         } else {
2151                 //render_sequence();
2152                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2153         }
2154         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2155
2156         va_status = vaEndPicture(va_dpy, context_id);
2157         CHECK_VASTATUS(va_status, "vaEndPicture");
2158
2159         // so now the data is done encoding (well, async job kicked off)...
2160         // we send that to the storage thread
2161         storage_task tmp;
2162         tmp.display_order = display_frame_num;
2163         tmp.frame_type = frame_type;
2164         tmp.pts = pts;
2165         tmp.dts = dts;
2166         tmp.duration = duration;
2167         storage_task_enqueue(move(tmp));
2168
2169         update_ReferenceFrames(frame_type);
2170 }
2171
2172 // Proxy object.
2173 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder)
2174         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, stream_mux, stream_audio_encoder, x264_encoder)) {}
2175
2176 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2177 QuickSyncEncoder::~QuickSyncEncoder() {}
2178
2179 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2180 {
2181         impl->add_audio(pts, audio);
2182 }
2183
2184 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2185 {
2186         return impl->begin_frame(y_tex, cbcr_tex);
2187 }
2188
2189 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2190 {
2191         return impl->end_frame(pts, duration, input_frames);
2192 }
2193
2194 void QuickSyncEncoder::shutdown()
2195 {
2196         impl->shutdown();
2197 }