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

   1 #include "quicksync_encoder.h"
   2
   3 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
   4 #include <movit/util.h>
   5
   6 #include <EGL/eglplatform.h>
   7 #include <X11/Xlib.h>
   8 #include <assert.h>
   9 #include <epoxy/egl.h>
  10 #include <fcntl.h>
  11 #include <stdio.h>
  12 #include <stdlib.h>
  13 #include <string.h>
  14 #include <unistd.h>
  15 #include <va/va.h>
  16 #include <va/va_drm.h>
  17 #include <va/va_drmcommon.h>
  18 #include <va/va_enc_h264.h>
  19 #include <va/va_x11.h>
  20 #include <algorithm>
  21 #include <condition_variable>
  22 #include <cstddef>
  23 #include <cstdint>
  24 #include <functional>
  25 #include <map>
  26 #include <memory>
  27 #include <mutex>
  28 #include <queue>
  29 #include <stack>
  30 #include <string>
  31 #include <thread>
  32 #include <utility>
  33
  34 extern "C" {
  35
  36 #include <libavcodec/avcodec.h>
  37 #include <libavformat/avio.h>
  38 #include <libavutil/error.h>
  39 #include <libdrm/drm_fourcc.h>
  40
  41 }  // namespace
  42
  43 #include "audio_encoder.h"
  44 #include "context.h"
  45 #include "defs.h"
  46 #include "disk_space_estimator.h"
  47 #include "ffmpeg_raii.h"
  48 #include "flags.h"
  49 #include "mux.h"
  50 #include "ref_counted_frame.h"
  51 #include "timebase.h"
  52 #include "x264_encoder.h"
  53
  54 using namespace std;
  55 using namespace std::placeholders;
  56
  57 class QOpenGLContext;
  58 class QSurface;
  59
  60 #define CHECK_VASTATUS(va_status, func)                                 \
  61     if (va_status != VA_STATUS_SUCCESS) {                               \
  62         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  63         exit(1);                                                        \
  64     }
  65
  66 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  67
  68 //#include "loadsurface.h"
  69
  70 #define NAL_REF_IDC_NONE        0
  71 #define NAL_REF_IDC_LOW         1
  72 #define NAL_REF_IDC_MEDIUM      2
  73 #define NAL_REF_IDC_HIGH        3
  74
  75 #define NAL_NON_IDR             1
  76 #define NAL_IDR                 5
  77 #define NAL_SPS                 7
  78 #define NAL_PPS                 8
  79 #define NAL_SEI                 6
  80
  81 #define SLICE_TYPE_P            0
  82 #define SLICE_TYPE_B            1
  83 #define SLICE_TYPE_I            2
  84 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  85 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  86 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  87
  88
  89 #define ENTROPY_MODE_CAVLC      0
  90 #define ENTROPY_MODE_CABAC      1
  91
  92 #define PROFILE_IDC_BASELINE    66
  93 #define PROFILE_IDC_MAIN        77
  94 #define PROFILE_IDC_HIGH        100
  95
  96 #define BITSTREAM_ALLOCATE_STEPPING     4096
  97 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  98 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  99 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
 100
 101 static constexpr unsigned int MaxFrameNum = (2<<16);
 102 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
 103 static constexpr unsigned int Log2MaxFrameNum = 16;
 104 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
 105 static constexpr int rc_default_modes[] = {  // Priority list of modes.
 106     VA_RC_VBR,
 107     VA_RC_CQP,
 108     VA_RC_VBR_CONSTRAINED,
 109     VA_RC_CBR,
 110     VA_RC_VCM,
 111     VA_RC_NONE,
 112 };
 113
 114 /* thread to save coded data */
 115 #define SRC_SURFACE_FREE        0
 116 #define SRC_SURFACE_IN_ENCODING 1
 117
 118 struct __bitstream {
 119     unsigned int *buffer;
 120     int bit_offset;
 121     int max_size_in_dword;
 122 };
 123 typedef struct __bitstream bitstream;
 124
 125 using namespace std;
 126
 127 // H.264 video comes out in encoding order (e.g. with two B-frames:
 128 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 129 // come in the right order. Since we do everything, including waiting
 130 // for the frames to come out of OpenGL, in encoding order, we need
 131 // a reordering buffer for uncompressed frames so that they come out
 132 // correctly. We go the super-lazy way of not making it understand
 133 // anything about the true order (which introduces some extra latency,
 134 // though); we know that for N B-frames we need at most (N-1) frames
 135 // in the reorder buffer, and can just sort on that.
 136 //
 137 // The class also deals with keeping a freelist as needed.
 138 class FrameReorderer {
 139 public:
 140         FrameReorderer(unsigned queue_length, int width, int height);
 141
 142         struct Frame {
 143                 int64_t pts, duration;
 144                 uint8_t *data;
 145
 146                 // Invert to get the smallest pts first.
 147                 bool operator< (const Frame &other) const { return pts > other.pts; }
 148         };
 149
 150         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 151         // Does _not_ take ownership of data; a copy is taken if needed.
 152         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 153         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 154         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data);
 155
 156         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 157         Frame get_first_frame();
 158
 159         bool empty() const { return frames.empty(); }
 160
 161 private:
 162         unsigned queue_length;
 163         int width, height;
 164
 165         priority_queue<Frame> frames;
 166         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 167
 168         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 169         // except priority_queue doesn't work well with movable-only types.
 170         vector<unique_ptr<uint8_t[]>> owner;
 171 };
 172
 173 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 174     : queue_length(queue_length), width(width), height(height)
 175 {
 176         for (unsigned i = 0; i < queue_length; ++i) {
 177                 owner.emplace_back(new uint8_t[width * height * 2]);
 178                 freelist.push(owner.back().get());
 179         }
 180 }
 181
 182 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data)
 183 {
 184         if (queue_length == 0) {
 185                 return Frame{pts, duration, data};
 186         }
 187
 188         assert(!freelist.empty());
 189         uint8_t *storage = freelist.top();
 190         freelist.pop();
 191         memcpy(storage, data, width * height * 2);
 192         frames.push(Frame{pts, duration, storage});
 193
 194         if (frames.size() >= queue_length) {
 195                 return get_first_frame();
 196         } else {
 197                 return Frame{-1, -1, nullptr};
 198         }
 199 }
 200
 201 FrameReorderer::Frame FrameReorderer::get_first_frame()
 202 {
 203         assert(!frames.empty());
 204         Frame storage = frames.top();
 205         frames.pop();
 206         freelist.push(storage.data);
 207         return storage;
 208 }
 209
 210 class QuickSyncEncoderImpl {
 211 public:
 212         QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator);
 213         ~QuickSyncEncoderImpl();
 214         void add_audio(int64_t pts, vector<float> audio);
 215         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 216         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
 217         void shutdown();
 218         void release_gl_resources();
 219         void set_stream_mux(Mux *mux)
 220         {
 221                 stream_mux = mux;
 222         }
 223
 224         // So we never get negative dts.
 225         int64_t global_delay() const {
 226                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 227         }
 228
 229 private:
 230         struct storage_task {
 231                 unsigned long long display_order;
 232                 int frame_type;
 233                 vector<float> audio;
 234                 int64_t pts, dts, duration;
 235         };
 236         struct PendingFrame {
 237                 RefCountedGLsync fence;
 238                 vector<RefCountedFrame> input_frames;
 239                 int64_t pts, duration;
 240         };
 241
 242         void open_output_file(const std::string &filename);
 243         void encode_thread_func();
 244         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 245         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 246         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 247                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 248         void storage_task_thread();
 249         void storage_task_enqueue(storage_task task);
 250         void save_codeddata(storage_task task);
 251         int render_packedsequence();
 252         int render_packedpicture();
 253         void render_packedslice();
 254         int render_sequence();
 255         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 256         void sps_rbsp(bitstream *bs);
 257         void pps_rbsp(bitstream *bs);
 258         int build_packed_pic_buffer(unsigned char **header_buffer);
 259         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 260         void slice_header(bitstream *bs);
 261         int build_packed_seq_buffer(unsigned char **header_buffer);
 262         int build_packed_slice_buffer(unsigned char **header_buffer);
 263         int init_va(const string &va_display);
 264         int deinit_va();
 265         void enable_zerocopy_if_possible();
 266         VADisplay va_open_display(const string &va_display);
 267         void va_close_display(VADisplay va_dpy);
 268         int setup_encode();
 269         void release_encode();
 270         void update_ReferenceFrames(int frame_type);
 271         int update_RefPicList(int frame_type);
 272
 273         bool is_shutdown = false;
 274         bool has_released_gl_resources = false;
 275         bool use_zerocopy;
 276         int drm_fd = -1;
 277
 278         thread encode_thread, storage_thread;
 279
 280         mutex storage_task_queue_mutex;
 281         condition_variable storage_task_queue_changed;
 282         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 283         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 284         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 285
 286         mutex frame_queue_mutex;
 287         condition_variable frame_queue_nonempty;
 288         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 289
 290         int current_storage_frame;
 291
 292         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 293         movit::ResourcePool *resource_pool;
 294         QSurface *surface;
 295
 296         unique_ptr<AudioEncoder> file_audio_encoder;
 297
 298         unique_ptr<FrameReorderer> reorderer;
 299         X264Encoder *x264_encoder;  // nullptr if not using x264.
 300
 301         Mux* stream_mux = nullptr;  // To HTTP.
 302         unique_ptr<Mux> file_mux;  // To local disk.
 303
 304         Display *x11_display = nullptr;
 305
 306         // Encoder parameters
 307         VADisplay va_dpy;
 308         VAProfile h264_profile = (VAProfile)~0;
 309         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 310         int config_attrib_num = 0, enc_packed_header_idx;
 311
 312         struct GLSurface {
 313                 VASurfaceID src_surface, ref_surface;
 314                 VABufferID coded_buf;
 315
 316                 VAImage surface_image;
 317                 GLuint y_tex, cbcr_tex;
 318
 319                 // Only if use_zerocopy == true.
 320                 EGLImage y_egl_image, cbcr_egl_image;
 321
 322                 // Only if use_zerocopy == false.
 323                 GLuint pbo;
 324                 uint8_t *y_ptr, *cbcr_ptr;
 325                 size_t y_offset, cbcr_offset;
 326         };
 327         GLSurface gl_surfaces[SURFACE_NUM];
 328
 329         VAConfigID config_id;
 330         VAContextID context_id;
 331         VAEncSequenceParameterBufferH264 seq_param;
 332         VAEncPictureParameterBufferH264 pic_param;
 333         VAEncSliceParameterBufferH264 slice_param;
 334         VAPictureH264 CurrentCurrPic;
 335         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 336
 337         // Static quality settings.
 338         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 339         static constexpr unsigned int num_ref_frames = 2;
 340         static constexpr int initial_qp = 15;
 341         static constexpr int minimal_qp = 0;
 342         static constexpr int intra_period = 30;
 343         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 344
 345         // Quality settings that are meant to be static, but might be overridden
 346         // by the profile.
 347         int constraint_set_flag = 0;
 348         int h264_packedheader = 0; /* support pack header? */
 349         int h264_maxref = (1<<16|1);
 350         int h264_entropy_mode = 1; /* cabac */
 351         int ip_period = 3;
 352
 353         int rc_mode = -1;
 354         unsigned int current_frame_num = 0;
 355         unsigned int numShortTerm = 0;
 356
 357         int frame_width;
 358         int frame_height;
 359         int frame_width_mbaligned;
 360         int frame_height_mbaligned;
 361
 362         DiskSpaceEstimator *disk_space_estimator;
 363 };
 364
 365 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 366 // but if we don't delete it here, we get leaks. The GStreamer implementation
 367 // does the same.
 368 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 369 {
 370     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 371     CHECK_VASTATUS(va_status, "vaRenderPicture");
 372
 373     for (int i = 0; i < num_buffers; ++i) {
 374         va_status = vaDestroyBuffer(dpy, buffers[i]);
 375         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 376     }
 377 }
 378
 379 static unsigned int
 380 va_swap32(unsigned int val)
 381 {
 382     unsigned char *pval = (unsigned char *)&val;
 383
 384     return ((pval[0] << 24)     |
 385             (pval[1] << 16)     |
 386             (pval[2] << 8)      |
 387             (pval[3] << 0));
 388 }
 389
 390 static void
 391 bitstream_start(bitstream *bs)
 392 {
 393     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 394     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 395     bs->bit_offset = 0;
 396 }
 397
 398 static void
 399 bitstream_end(bitstream *bs)
 400 {
 401     int pos = (bs->bit_offset >> 5);
 402     int bit_offset = (bs->bit_offset & 0x1f);
 403     int bit_left = 32 - bit_offset;
 404
 405     if (bit_offset) {
 406         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 407     }
 408 }
 409
 410 static void
 411 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 412 {
 413     int pos = (bs->bit_offset >> 5);
 414     int bit_offset = (bs->bit_offset & 0x1f);
 415     int bit_left = 32 - bit_offset;
 416
 417     if (!size_in_bits)
 418         return;
 419
 420     bs->bit_offset += size_in_bits;
 421
 422     if (bit_left > size_in_bits) {
 423         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 424     } else {
 425         size_in_bits -= bit_left;
 426         if (bit_left >= 32) {
 427             bs->buffer[pos] = (val >> size_in_bits);
 428         } else {
 429             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 430         }
 431         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 432
 433         if (pos + 1 == bs->max_size_in_dword) {
 434             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 435             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 436         }
 437
 438         bs->buffer[pos + 1] = val;
 439     }
 440 }
 441
 442 static void
 443 bitstream_put_ue(bitstream *bs, unsigned int val)
 444 {
 445     int size_in_bits = 0;
 446     int tmp_val = ++val;
 447
 448     while (tmp_val) {
 449         tmp_val >>= 1;
 450         size_in_bits++;
 451     }
 452
 453     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 454     bitstream_put_ui(bs, val, size_in_bits);
 455 }
 456
 457 static void
 458 bitstream_put_se(bitstream *bs, int val)
 459 {
 460     unsigned int new_val;
 461
 462     if (val <= 0)
 463         new_val = -2 * val;
 464     else
 465         new_val = 2 * val - 1;
 466
 467     bitstream_put_ue(bs, new_val);
 468 }
 469
 470 static void
 471 bitstream_byte_aligning(bitstream *bs, int bit)
 472 {
 473     int bit_offset = (bs->bit_offset & 0x7);
 474     int bit_left = 8 - bit_offset;
 475     int new_val;
 476
 477     if (!bit_offset)
 478         return;
 479
 480     assert(bit == 0 || bit == 1);
 481
 482     if (bit)
 483         new_val = (1 << bit_left) - 1;
 484     else
 485         new_val = 0;
 486
 487     bitstream_put_ui(bs, new_val, bit_left);
 488 }
 489
 490 static void
 491 rbsp_trailing_bits(bitstream *bs)
 492 {
 493     bitstream_put_ui(bs, 1, 1);
 494     bitstream_byte_aligning(bs, 0);
 495 }
 496
 497 static void nal_start_code_prefix(bitstream *bs)
 498 {
 499     bitstream_put_ui(bs, 0x00000001, 32);
 500 }
 501
 502 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 503 {
 504     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 505     bitstream_put_ui(bs, nal_ref_idc, 2);
 506     bitstream_put_ui(bs, nal_unit_type, 5);
 507 }
 508
 509 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 510 {
 511     int profile_idc = PROFILE_IDC_BASELINE;
 512
 513     if (h264_profile  == VAProfileH264High)
 514         profile_idc = PROFILE_IDC_HIGH;
 515     else if (h264_profile  == VAProfileH264Main)
 516         profile_idc = PROFILE_IDC_MAIN;
 517
 518     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 519     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 520     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 521     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 522     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 523     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 524     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 525     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 526
 527     if ( profile_idc == PROFILE_IDC_HIGH) {
 528         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 529         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 530         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 531         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 532         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 533     }
 534
 535     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 536     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 537
 538     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 539         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 540     else {
 541         assert(0);
 542     }
 543
 544     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 545     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 546
 547     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 548     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 549     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 550
 551     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 552         assert(0);
 553     }
 554
 555     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 556     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 557
 558     if (seq_param.frame_cropping_flag) {
 559         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 560         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 561         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 562         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 563     }
 564
 565     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 566     if ( false ) {
 567         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 568     } else {
 569         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 570         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 571         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 572         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 573         {
 574             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 575             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 576             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 577             {
 578                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 579                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 580                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 581             }
 582         }
 583         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 584         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 585         {
 586             bitstream_put_ui(bs, 1, 32);  // FPS
 587             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 588             bitstream_put_ui(bs, 1, 1);
 589         }
 590         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 591         {
 592             // hrd_parameters
 593             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 594             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 595             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 596
 597             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 598             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 599             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 600
 601             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 602             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 603             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 604             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 605         }
 606         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 607         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 608
 609         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 610         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 611     }
 612
 613     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 614 }
 615
 616
 617 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 618 {
 619     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 620     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 621
 622     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 623
 624     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 625
 626     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 627
 628     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 629     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 630
 631     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 632     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 633
 634     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 635     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 636     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 637
 638     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 639     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 640     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 641
 642     /* more_rbsp_data */
 643     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 644     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 645     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 646
 647     rbsp_trailing_bits(bs);
 648 }
 649
 650 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 651 {
 652     int first_mb_in_slice = slice_param.macroblock_address;
 653
 654     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 655     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 656     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 657     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 658
 659     /* frame_mbs_only_flag == 1 */
 660     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 661         /* FIXME: */
 662         assert(0);
 663     }
 664
 665     if (pic_param.pic_fields.bits.idr_pic_flag)
 666         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 667
 668     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 669         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 670         /* pic_order_present_flag == 0 */
 671     } else {
 672         /* FIXME: */
 673         assert(0);
 674     }
 675
 676     /* redundant_pic_cnt_present_flag == 0 */
 677     /* slice type */
 678     if (IS_P_SLICE(slice_param.slice_type)) {
 679         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 680
 681         if (slice_param.num_ref_idx_active_override_flag)
 682             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 683
 684         /* ref_pic_list_reordering */
 685         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 686     } else if (IS_B_SLICE(slice_param.slice_type)) {
 687         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 688
 689         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 690
 691         if (slice_param.num_ref_idx_active_override_flag) {
 692             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 693             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 694         }
 695
 696         /* ref_pic_list_reordering */
 697         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 698         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 699     }
 700
 701     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 702          IS_P_SLICE(slice_param.slice_type)) ||
 703         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 704          IS_B_SLICE(slice_param.slice_type))) {
 705         /* FIXME: fill weight/offset table */
 706         assert(0);
 707     }
 708
 709     /* dec_ref_pic_marking */
 710     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 711         unsigned char no_output_of_prior_pics_flag = 0;
 712         unsigned char long_term_reference_flag = 0;
 713         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 714
 715         if (pic_param.pic_fields.bits.idr_pic_flag) {
 716             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 717             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 718         } else {
 719             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 720         }
 721     }
 722
 723     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 724         !IS_I_SLICE(slice_param.slice_type))
 725         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 726
 727     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 728
 729     /* ignore for SP/SI */
 730
 731     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 732         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 733
 734         if (slice_param.disable_deblocking_filter_idc != 1) {
 735             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 736             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 737         }
 738     }
 739
 740     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 741         bitstream_byte_aligning(bs, 1);
 742     }
 743 }
 744
 745 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 746 {
 747     bitstream bs;
 748
 749     bitstream_start(&bs);
 750     nal_start_code_prefix(&bs);
 751     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 752     pps_rbsp(&bs);
 753     bitstream_end(&bs);
 754
 755     *header_buffer = (unsigned char *)bs.buffer;
 756     return bs.bit_offset;
 757 }
 758
 759 int
 760 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 761 {
 762     bitstream bs;
 763
 764     bitstream_start(&bs);
 765     nal_start_code_prefix(&bs);
 766     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 767     sps_rbsp(&bs);
 768     bitstream_end(&bs);
 769
 770     *header_buffer = (unsigned char *)bs.buffer;
 771     return bs.bit_offset;
 772 }
 773
 774 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 775 {
 776     bitstream bs;
 777     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 778     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 779
 780     bitstream_start(&bs);
 781     nal_start_code_prefix(&bs);
 782
 783     if (IS_I_SLICE(slice_param.slice_type)) {
 784         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 785     } else if (IS_P_SLICE(slice_param.slice_type)) {
 786         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 787     } else {
 788         assert(IS_B_SLICE(slice_param.slice_type));
 789         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 790     }
 791
 792     slice_header(&bs);
 793     bitstream_end(&bs);
 794
 795     *header_buffer = (unsigned char *)bs.buffer;
 796     return bs.bit_offset;
 797 }
 798
 799
 800 /*
 801   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 802   1) period between Frame #X and Frame #N = #X - #N
 803   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 804   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 805   4) intra_period and intra_idr_period take precedence over ip_period
 806   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 807      of I/IDR frames, see bellow examples
 808   -------------------------------------------------------------------
 809   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 810   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 811   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 812   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 813   1            1                ignored    IDR IDR IDR IDR...
 814   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 815   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 816   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 817                                               (PBB)(IBB)(PBB)(IBB)...
 818   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 819                                            IDRPPPPP IPPPPP IPPPPP...
 820   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 821                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 822                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 823                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 824                                            {IDR(PBB)(PBB)}                     (6/6/3)
 825                                            {IDR(PBB)(PBB)}.
 826 */
 827
 828 // General pts/dts strategy:
 829 //
 830 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 831 // bit tricky. We assume first of all that the frame rate never goes _above_
 832 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 833 // in at least this speed, as long at dts <= pts (the frame is not attempted
 834 // presented before it is decoded). Furthermore, we never have longer chains of
 835 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 836 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 837 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 838 // frame, with an order O <= C.)
 839 //
 840 // Many strategies are possible, but we establish these rules:
 841 //
 842 //  - Tip frames have dts = pts - (C-O)*N.
 843 //  - Non-tip frames have dts = dts_last + N.
 844 //
 845 // An example, with C=2 and N=10 and the data flow showed with arrows:
 846 //
 847 //        I  B  P  B  B  P
 848 //   pts: 30 40 50 60 70 80
 849 //        ↓  ↓     ↓
 850 //   dts: 10 30 20 60 50←40
 851 //         |  |  ↑        ↑
 852 //         `--|--'        |
 853 //             `----------'
 854 //
 855 // To show that this works fine also with irregular spacings, let's say that
 856 // the third frame is delayed a bit (something earlier was dropped). Now the
 857 // situation looks like this:
 858 //
 859 //        I  B  P  B  B   P
 860 //   pts: 30 40 80 90 100 110
 861 //        ↓  ↓     ↓
 862 //   dts: 10 30 20 90 50←40
 863 //         |  |  ↑        ↑
 864 //         `--|--'        |
 865 //             `----------'
 866 //
 867 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 868 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 869 //
 870 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 871 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 872 // a tip frame and should be given a dts based on the previous one.
 873 #define FRAME_P 0
 874 #define FRAME_B 1
 875 #define FRAME_I 2
 876 #define FRAME_IDR 7
 877 void encoding2display_order(
 878     int encoding_order, int intra_period,
 879     int intra_idr_period, int ip_period,
 880     int *displaying_order,
 881     int *frame_type, int *pts_lag)
 882 {
 883     int encoding_order_gop = 0;
 884
 885     *pts_lag = 0;
 886
 887     if (intra_period == 1) { /* all are I/IDR frames */
 888         *displaying_order = encoding_order;
 889         if (intra_idr_period == 0)
 890             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 891         else
 892             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 893         return;
 894     }
 895
 896     if (intra_period == 0)
 897         intra_idr_period = 0;
 898
 899     if (ip_period == 1) {
 900         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 901         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 902         *displaying_order = encoding_order;
 903
 904         if (encoding_order_gop == 0) { /* the first frame */
 905             *frame_type = FRAME_IDR;
 906         } else if (intra_period != 0 && /* have I frames */
 907                    encoding_order_gop >= 2 &&
 908                    (encoding_order_gop % intra_period == 0)) {
 909             *frame_type = FRAME_I;
 910         } else {
 911             *frame_type = FRAME_P;
 912         }
 913         return;
 914     }
 915
 916     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 917     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 918     *pts_lag = -1;  // Most frames are not tip frames.
 919
 920     if (encoding_order_gop == 0) { /* the first frame */
 921         *frame_type = FRAME_IDR;
 922         *displaying_order = encoding_order;
 923         // IDR frames are a special case; I honestly can't find the logic behind
 924         // why this is the right thing, but it seems to line up nicely in practice :-)
 925         *pts_lag = TIMEBASE / MAX_FPS;
 926     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 927         *frame_type = FRAME_B;
 928         *displaying_order = encoding_order - 1;
 929         if ((encoding_order_gop % ip_period) == 0) {
 930             *pts_lag = 0;  // Last B-frame.
 931         }
 932     } else if (intra_period != 0 && /* have I frames */
 933                encoding_order_gop >= 2 &&
 934                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 935         *frame_type = FRAME_I;
 936         *displaying_order = encoding_order + ip_period - 1;
 937     } else {
 938         *frame_type = FRAME_P;
 939         *displaying_order = encoding_order + ip_period - 1;
 940     }
 941 }
 942
 943
 944 static const char *rc_to_string(int rc_mode)
 945 {
 946     switch (rc_mode) {
 947     case VA_RC_NONE:
 948         return "NONE";
 949     case VA_RC_CBR:
 950         return "CBR";
 951     case VA_RC_VBR:
 952         return "VBR";
 953     case VA_RC_VCM:
 954         return "VCM";
 955     case VA_RC_CQP:
 956         return "CQP";
 957     case VA_RC_VBR_CONSTRAINED:
 958         return "VBR_CONSTRAINED";
 959     default:
 960         return "Unknown";
 961     }
 962 }
 963
 964 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 965 {
 966         if (global_flags.uncompressed_video_to_http) {
 967                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 968                 use_zerocopy = false;
 969         } else if (global_flags.x264_video_to_http) {
 970                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 971                 use_zerocopy = false;
 972         } else {
 973                 use_zerocopy = true;
 974         }
 975 }
 976
 977 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 978 {
 979         if (va_display.empty()) {
 980                 x11_display = XOpenDisplay(NULL);
 981                 if (!x11_display) {
 982                         fprintf(stderr, "error: can't connect to X server!\n");
 983                         return NULL;
 984                 }
 985                 enable_zerocopy_if_possible();
 986                 return vaGetDisplay(x11_display);
 987         } else if (va_display[0] != '/') {
 988                 x11_display = XOpenDisplay(va_display.c_str());
 989                 if (!x11_display) {
 990                         fprintf(stderr, "error: can't connect to X server!\n");
 991                         return NULL;
 992                 }
 993                 enable_zerocopy_if_possible();
 994                 return vaGetDisplay(x11_display);
 995         } else {
 996                 drm_fd = open(va_display.c_str(), O_RDWR);
 997                 if (drm_fd == -1) {
 998                         perror(va_display.c_str());
 999                         return NULL;
1000                 }
1001                 use_zerocopy = false;
1002                 return vaGetDisplayDRM(drm_fd);
1003         }
1004 }
1005
1006 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
1007 {
1008         if (x11_display) {
1009                 XCloseDisplay(x11_display);
1010                 x11_display = nullptr;
1011         }
1012         if (drm_fd != -1) {
1013                 close(drm_fd);
1014         }
1015 }
1016
1017 int QuickSyncEncoderImpl::init_va(const string &va_display)
1018 {
1019     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1020     VAEntrypoint *entrypoints;
1021     int num_entrypoints, slice_entrypoint;
1022     int support_encode = 0;
1023     int major_ver, minor_ver;
1024     VAStatus va_status;
1025     unsigned int i;
1026
1027     va_dpy = va_open_display(va_display);
1028     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1029     CHECK_VASTATUS(va_status, "vaInitialize");
1030
1031     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1032     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1033     if (!entrypoints) {
1034         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1035         exit(1);
1036     }
1037
1038     /* use the highest profile */
1039     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1040         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1041             continue;
1042
1043         h264_profile = profile_list[i];
1044         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1045         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1046             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1047                 support_encode = 1;
1048                 break;
1049             }
1050         }
1051         if (support_encode == 1)
1052             break;
1053     }
1054
1055     if (support_encode == 0) {
1056         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1057         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1058         printf("to use VA-API against DRM instead of X11.\n");
1059         exit(1);
1060     } else {
1061         switch (h264_profile) {
1062             case VAProfileH264Baseline:
1063                 ip_period = 1;
1064                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1065                 h264_entropy_mode = 0;
1066                 break;
1067             case VAProfileH264ConstrainedBaseline:
1068                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1069                 ip_period = 1;
1070                 break;
1071
1072             case VAProfileH264Main:
1073                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1074                 break;
1075
1076             case VAProfileH264High:
1077                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1078                 break;
1079             default:
1080                 h264_profile = VAProfileH264Baseline;
1081                 ip_period = 1;
1082                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1083                 break;
1084         }
1085     }
1086
1087     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1088
1089     /* find out the format for the render target, and rate control mode */
1090     for (i = 0; i < VAConfigAttribTypeMax; i++)
1091         attrib[i].type = (VAConfigAttribType)i;
1092
1093     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1094                                       &attrib[0], VAConfigAttribTypeMax);
1095     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1096     /* check the interested configattrib */
1097     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1098         printf("Not find desired YUV420 RT format\n");
1099         exit(1);
1100     } else {
1101         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1102         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1103         config_attrib_num++;
1104     }
1105
1106     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1107         int tmp = attrib[VAConfigAttribRateControl].value;
1108
1109         if (rc_mode == -1 || !(rc_mode & tmp))  {
1110             if (rc_mode != -1) {
1111                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1112             }
1113
1114             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1115                 if (rc_default_modes[i] & tmp) {
1116                     rc_mode = rc_default_modes[i];
1117                     break;
1118                 }
1119             }
1120         }
1121
1122         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1123         config_attrib[config_attrib_num].value = rc_mode;
1124         config_attrib_num++;
1125     }
1126
1127
1128     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1129         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1130
1131         h264_packedheader = 1;
1132         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1133         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1134
1135         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1136             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1137         }
1138
1139         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1140             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1141         }
1142
1143         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1144             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1145         }
1146
1147         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1148             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1149         }
1150
1151         enc_packed_header_idx = config_attrib_num;
1152         config_attrib_num++;
1153     }
1154
1155     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1156         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1157         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1158         config_attrib_num++;
1159     }
1160
1161     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1162         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1163     }
1164
1165     free(entrypoints);
1166     return 0;
1167 }
1168
1169 int QuickSyncEncoderImpl::setup_encode()
1170 {
1171     VAStatus va_status;
1172     VASurfaceID *tmp_surfaceid;
1173     int codedbuf_size, i;
1174     VASurfaceID src_surface[SURFACE_NUM];
1175     VASurfaceID ref_surface[SURFACE_NUM];
1176
1177     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1178             &config_attrib[0], config_attrib_num, &config_id);
1179     CHECK_VASTATUS(va_status, "vaCreateConfig");
1180
1181     /* create source surfaces */
1182     va_status = vaCreateSurfaces(va_dpy,
1183                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1184                                  &src_surface[0], SURFACE_NUM,
1185                                  NULL, 0);
1186     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1187
1188     /* create reference surfaces */
1189     va_status = vaCreateSurfaces(va_dpy,
1190                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1191                                  &ref_surface[0], SURFACE_NUM,
1192                                  NULL, 0);
1193     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1194
1195     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1196     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1197     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1198
1199     /* Create a context for this encode pipe */
1200     va_status = vaCreateContext(va_dpy, config_id,
1201                                 frame_width_mbaligned, frame_height_mbaligned,
1202                                 VA_PROGRESSIVE,
1203                                 tmp_surfaceid, 2 * SURFACE_NUM,
1204                                 &context_id);
1205     CHECK_VASTATUS(va_status, "vaCreateContext");
1206     free(tmp_surfaceid);
1207
1208     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1209
1210     for (i = 0; i < SURFACE_NUM; i++) {
1211         /* create coded buffer once for all
1212          * other VA buffers which won't be used again after vaRenderPicture.
1213          * so APP can always vaCreateBuffer for every frame
1214          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1215          * so VA won't maintain the coded buffer
1216          */
1217         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1218                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1219         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1220     }
1221
1222     /* create OpenGL objects */
1223     //glGenFramebuffers(SURFACE_NUM, fbos);
1224
1225     for (i = 0; i < SURFACE_NUM; i++) {
1226         if (use_zerocopy) {
1227             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
1228             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
1229         } else {
1230             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
1231             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
1232
1233             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1234             // buffers, due to potentially differing pitch.
1235             glGenBuffers(1, &gl_surfaces[i].pbo);
1236             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1237             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1238             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1239             gl_surfaces[i].y_offset = 0;
1240             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1241             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1242             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1243             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1244         }
1245     }
1246
1247     for (i = 0; i < SURFACE_NUM; i++) {
1248         gl_surfaces[i].src_surface = src_surface[i];
1249         gl_surfaces[i].ref_surface = ref_surface[i];
1250     }
1251
1252     return 0;
1253 }
1254
1255 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1256 //
1257 //   2 1 0 [3] 4 8 9
1258 template<class T, class C>
1259 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1260 {
1261         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1262         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1263         sort(middle, end, less_than);
1264 }
1265
1266 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1267 {
1268     int i;
1269
1270     if (frame_type == FRAME_B)
1271         return;
1272
1273     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1274     numShortTerm++;
1275     if (numShortTerm > num_ref_frames)
1276         numShortTerm = num_ref_frames;
1277     for (i=numShortTerm-1; i>0; i--)
1278         ReferenceFrames[i] = ReferenceFrames[i-1];
1279     ReferenceFrames[0] = CurrentCurrPic;
1280
1281     current_frame_num++;
1282     if (current_frame_num > MaxFrameNum)
1283         current_frame_num = 0;
1284 }
1285
1286
1287 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1288 {
1289     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1290         return a.frame_idx > b.frame_idx;
1291     };
1292     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1293         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1294     };
1295     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1296         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1297     };
1298
1299     if (frame_type == FRAME_P) {
1300         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1301         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1302     } else if (frame_type == FRAME_B) {
1303         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1304         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1305
1306         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1307         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1308     }
1309
1310     return 0;
1311 }
1312
1313
1314 int QuickSyncEncoderImpl::render_sequence()
1315 {
1316     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1317     VAStatus va_status;
1318     VAEncMiscParameterBuffer *misc_param;
1319     VAEncMiscParameterRateControl *misc_rate_ctrl;
1320
1321     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1322     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1323     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1324     seq_param.bits_per_second = frame_bitrate;
1325
1326     seq_param.intra_period = intra_period;
1327     seq_param.intra_idr_period = intra_idr_period;
1328     seq_param.ip_period = ip_period;
1329
1330     seq_param.max_num_ref_frames = num_ref_frames;
1331     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1332     seq_param.time_scale = TIMEBASE * 2;
1333     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1334     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1335     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1336     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1337     seq_param.seq_fields.bits.chroma_format_idc = 1;
1338     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1339
1340     if (frame_width != frame_width_mbaligned ||
1341         frame_height != frame_height_mbaligned) {
1342         seq_param.frame_cropping_flag = 1;
1343         seq_param.frame_crop_left_offset = 0;
1344         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1345         seq_param.frame_crop_top_offset = 0;
1346         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1347     }
1348
1349     va_status = vaCreateBuffer(va_dpy, context_id,
1350                                VAEncSequenceParameterBufferType,
1351                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1352     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1353
1354     va_status = vaCreateBuffer(va_dpy, context_id,
1355                                VAEncMiscParameterBufferType,
1356                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1357                                1, NULL, &rc_param_buf);
1358     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1359
1360     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1361     misc_param->type = VAEncMiscParameterTypeRateControl;
1362     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1363     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1364     misc_rate_ctrl->bits_per_second = frame_bitrate;
1365     misc_rate_ctrl->target_percentage = 66;
1366     misc_rate_ctrl->window_size = 1000;
1367     misc_rate_ctrl->initial_qp = initial_qp;
1368     misc_rate_ctrl->min_qp = minimal_qp;
1369     misc_rate_ctrl->basic_unit_size = 0;
1370     vaUnmapBuffer(va_dpy, rc_param_buf);
1371
1372     render_id[0] = seq_param_buf;
1373     render_id[1] = rc_param_buf;
1374
1375     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1376
1377     return 0;
1378 }
1379
1380 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1381 {
1382     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1383     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1384     int PicOrderCntMsb, TopFieldOrderCnt;
1385
1386     if (frame_type == FRAME_IDR)
1387         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1388     else {
1389         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1390         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1391     }
1392
1393     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1394         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1395         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1396     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1397              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1398         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1399     else
1400         PicOrderCntMsb = prevPicOrderCntMsb;
1401
1402     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1403
1404     if (frame_type != FRAME_B) {
1405         PicOrderCntMsb_ref = PicOrderCntMsb;
1406         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1407     }
1408
1409     return TopFieldOrderCnt;
1410 }
1411
1412 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1413 {
1414     VABufferID pic_param_buf;
1415     VAStatus va_status;
1416     int i = 0;
1417
1418     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1419     pic_param.CurrPic.frame_idx = current_frame_num;
1420     pic_param.CurrPic.flags = 0;
1421     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1422     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1423     CurrentCurrPic = pic_param.CurrPic;
1424
1425     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1426     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1427         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1428         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1429     }
1430
1431     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1432     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1433     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1434     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1435     pic_param.frame_num = current_frame_num;
1436     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1437     pic_param.last_picture = false;  // FIXME
1438     pic_param.pic_init_qp = initial_qp;
1439
1440     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1441                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1442     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1443
1444     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1445
1446     return 0;
1447 }
1448
1449 int QuickSyncEncoderImpl::render_packedsequence()
1450 {
1451     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1452     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1453     unsigned int length_in_bits;
1454     unsigned char *packedseq_buffer = NULL;
1455     VAStatus va_status;
1456
1457     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1458
1459     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1460
1461     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1462     packedheader_param_buffer.has_emulation_bytes = 0;
1463     va_status = vaCreateBuffer(va_dpy,
1464                                context_id,
1465                                VAEncPackedHeaderParameterBufferType,
1466                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1467                                &packedseq_para_bufid);
1468     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1469
1470     va_status = vaCreateBuffer(va_dpy,
1471                                context_id,
1472                                VAEncPackedHeaderDataBufferType,
1473                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1474                                &packedseq_data_bufid);
1475     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1476
1477     render_id[0] = packedseq_para_bufid;
1478     render_id[1] = packedseq_data_bufid;
1479     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1480
1481     free(packedseq_buffer);
1482
1483     return 0;
1484 }
1485
1486
1487 int QuickSyncEncoderImpl::render_packedpicture()
1488 {
1489     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1490     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1491     unsigned int length_in_bits;
1492     unsigned char *packedpic_buffer = NULL;
1493     VAStatus va_status;
1494
1495     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1496     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1497     packedheader_param_buffer.bit_length = length_in_bits;
1498     packedheader_param_buffer.has_emulation_bytes = 0;
1499
1500     va_status = vaCreateBuffer(va_dpy,
1501                                context_id,
1502                                VAEncPackedHeaderParameterBufferType,
1503                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1504                                &packedpic_para_bufid);
1505     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1506
1507     va_status = vaCreateBuffer(va_dpy,
1508                                context_id,
1509                                VAEncPackedHeaderDataBufferType,
1510                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1511                                &packedpic_data_bufid);
1512     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1513
1514     render_id[0] = packedpic_para_bufid;
1515     render_id[1] = packedpic_data_bufid;
1516     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1517
1518     free(packedpic_buffer);
1519
1520     return 0;
1521 }
1522
1523 void QuickSyncEncoderImpl::render_packedslice()
1524 {
1525     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1526     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1527     unsigned int length_in_bits;
1528     unsigned char *packedslice_buffer = NULL;
1529     VAStatus va_status;
1530
1531     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1532     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1533     packedheader_param_buffer.bit_length = length_in_bits;
1534     packedheader_param_buffer.has_emulation_bytes = 0;
1535
1536     va_status = vaCreateBuffer(va_dpy,
1537                                context_id,
1538                                VAEncPackedHeaderParameterBufferType,
1539                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1540                                &packedslice_para_bufid);
1541     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1542
1543     va_status = vaCreateBuffer(va_dpy,
1544                                context_id,
1545                                VAEncPackedHeaderDataBufferType,
1546                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1547                                &packedslice_data_bufid);
1548     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1549
1550     render_id[0] = packedslice_para_bufid;
1551     render_id[1] = packedslice_data_bufid;
1552     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1553
1554     free(packedslice_buffer);
1555 }
1556
1557 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1558 {
1559     VABufferID slice_param_buf;
1560     VAStatus va_status;
1561     int i;
1562
1563     update_RefPicList(frame_type);
1564
1565     /* one frame, one slice */
1566     slice_param.macroblock_address = 0;
1567     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1568     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1569     if (frame_type == FRAME_IDR) {
1570         if (encoding_frame_num != 0)
1571             ++slice_param.idr_pic_id;
1572     } else if (frame_type == FRAME_P) {
1573         int refpiclist0_max = h264_maxref & 0xffff;
1574         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1575
1576         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1577             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1578             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1579         }
1580     } else if (frame_type == FRAME_B) {
1581         int refpiclist0_max = h264_maxref & 0xffff;
1582         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1583
1584         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1585         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1586             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1587             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1588         }
1589
1590         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1591         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1592             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1593             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1594         }
1595     }
1596
1597     slice_param.slice_alpha_c0_offset_div2 = 0;
1598     slice_param.slice_beta_offset_div2 = 0;
1599     slice_param.direct_spatial_mv_pred_flag = 1;
1600     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1601
1602
1603     if (h264_packedheader &&
1604         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1605         render_packedslice();
1606
1607     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1608                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1609     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1610
1611     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1612
1613     return 0;
1614 }
1615
1616
1617
1618 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1619 {
1620         VACodedBufferSegment *buf_list = NULL;
1621         VAStatus va_status;
1622
1623         string data;
1624
1625         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1626         CHECK_VASTATUS(va_status, "vaMapBuffer");
1627         while (buf_list != NULL) {
1628                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1629                 buf_list = (VACodedBufferSegment *) buf_list->next;
1630         }
1631         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1632
1633         {
1634                 // Add video.
1635                 AVPacket pkt;
1636                 memset(&pkt, 0, sizeof(pkt));
1637                 pkt.buf = nullptr;
1638                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1639                 pkt.size = data.size();
1640                 pkt.stream_index = 0;
1641                 if (task.frame_type == FRAME_IDR) {
1642                         pkt.flags = AV_PKT_FLAG_KEY;
1643                 } else {
1644                         pkt.flags = 0;
1645                 }
1646                 pkt.duration = task.duration;
1647                 if (file_mux) {
1648                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1649                 }
1650                 if (!global_flags.uncompressed_video_to_http &&
1651                     !global_flags.x264_video_to_http) {
1652                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1653                 }
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 void 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
1702         vaDestroyContext(va_dpy, context_id);
1703         vaDestroyConfig(va_dpy, config_id);
1704 }
1705
1706 void QuickSyncEncoderImpl::release_gl_resources()
1707 {
1708         assert(is_shutdown);
1709         if (has_released_gl_resources) {
1710                 return;
1711         }
1712
1713         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1714                 if (!use_zerocopy) {
1715                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1716                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1717                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1718                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1719                 }
1720                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1721                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1722         }
1723
1724         has_released_gl_resources = true;
1725 }
1726
1727 int QuickSyncEncoderImpl::deinit_va()
1728 {
1729     vaTerminate(va_dpy);
1730
1731     va_close_display(va_dpy);
1732
1733     return 0;
1734 }
1735
1736 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1737         : current_storage_frame(0), resource_pool(resource_pool), surface(surface), x264_encoder(x264_encoder), frame_width(width), frame_height(height), disk_space_estimator(disk_space_estimator)
1738 {
1739         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1740         open_output_file(filename);
1741         file_audio_encoder->add_mux(file_mux.get());
1742
1743         frame_width_mbaligned = (frame_width + 15) & (~15);
1744         frame_height_mbaligned = (frame_height + 15) & (~15);
1745
1746         //print_input();
1747
1748         if (global_flags.uncompressed_video_to_http ||
1749             global_flags.x264_video_to_http) {
1750                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1751         }
1752         if (global_flags.x264_video_to_http) {
1753                 assert(x264_encoder != nullptr);
1754         } else {
1755                 assert(x264_encoder == nullptr);
1756         }
1757
1758         init_va(va_display);
1759         setup_encode();
1760
1761         // No frames are ready yet.
1762         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1763
1764         memset(&seq_param, 0, sizeof(seq_param));
1765         memset(&pic_param, 0, sizeof(pic_param));
1766         memset(&slice_param, 0, sizeof(slice_param));
1767
1768         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1769
1770         encode_thread = thread([this]{
1771                 //SDL_GL_MakeCurrent(window, context);
1772                 QOpenGLContext *context = create_context(this->surface);
1773                 eglBindAPI(EGL_OPENGL_API);
1774                 if (!make_current(context, this->surface)) {
1775                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1776                                 eglGetError());
1777                         exit(1);
1778                 }
1779                 encode_thread_func();
1780                 delete_context(context);
1781         });
1782 }
1783
1784 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1785 {
1786         shutdown();
1787         release_gl_resources();
1788 }
1789
1790 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1791 {
1792         assert(!is_shutdown);
1793         {
1794                 // Wait until this frame slot is done encoding.
1795                 unique_lock<mutex> lock(storage_task_queue_mutex);
1796                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1797                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1798                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1799                 }
1800                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1801                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1802                 if (storage_thread_should_quit) return false;
1803         }
1804
1805         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1806         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1807         *y_tex = surf->y_tex;
1808         *cbcr_tex = surf->cbcr_tex;
1809
1810         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1811         CHECK_VASTATUS(va_status, "vaDeriveImage");
1812
1813         if (use_zerocopy) {
1814                 VABufferInfo buf_info;
1815                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1816                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1817                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1818
1819                 // Create Y image.
1820                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1821                 EGLint y_attribs[] = {
1822                         EGL_WIDTH, frame_width,
1823                         EGL_HEIGHT, frame_height,
1824                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1825                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1826                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1827                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1828                         EGL_NONE
1829                 };
1830
1831                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1832                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1833
1834                 // Associate Y image to a texture.
1835                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1836                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1837
1838                 // Create CbCr image.
1839                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1840                 EGLint cbcr_attribs[] = {
1841                         EGL_WIDTH, frame_width,
1842                         EGL_HEIGHT, frame_height,
1843                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1844                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1845                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1846                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1847                         EGL_NONE
1848                 };
1849
1850                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1851                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1852
1853                 // Associate CbCr image to a texture.
1854                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1855                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1856         }
1857
1858         return true;
1859 }
1860
1861 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1862 {
1863         assert(!is_shutdown);
1864         file_audio_encoder->encode_audio(audio, pts + global_delay());
1865 }
1866
1867 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1868 {
1869         assert(!is_shutdown);
1870
1871         if (!use_zerocopy) {
1872                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1873
1874                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1875                 check_error();
1876
1877                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1878                 check_error();
1879
1880                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1881                 check_error();
1882                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1883                 check_error();
1884
1885                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1886                 check_error();
1887                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1888                 check_error();
1889
1890                 glBindTexture(GL_TEXTURE_2D, 0);
1891                 check_error();
1892                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1893                 check_error();
1894
1895                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1896                 check_error();
1897         }
1898
1899         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1900         check_error();
1901         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1902         check_error();
1903
1904         {
1905                 unique_lock<mutex> lock(frame_queue_mutex);
1906                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1907                 ++current_storage_frame;
1908         }
1909         frame_queue_nonempty.notify_all();
1910         return fence;
1911 }
1912
1913 void QuickSyncEncoderImpl::shutdown()
1914 {
1915         if (is_shutdown) {
1916                 return;
1917         }
1918
1919         {
1920                 unique_lock<mutex> lock(frame_queue_mutex);
1921                 encode_thread_should_quit = true;
1922                 frame_queue_nonempty.notify_all();
1923         }
1924         encode_thread.join();
1925         {
1926                 unique_lock<mutex> lock(storage_task_queue_mutex);
1927                 storage_thread_should_quit = true;
1928                 frame_queue_nonempty.notify_all();
1929                 storage_task_queue_changed.notify_all();
1930         }
1931         storage_thread.join();
1932
1933         // Encode any leftover audio in the queues, and also any delayed frames.
1934         file_audio_encoder->encode_last_audio();
1935
1936         release_encode();
1937         deinit_va();
1938         file_mux.reset();
1939         is_shutdown = true;
1940 }
1941
1942 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1943 {
1944         AVFormatContext *avctx = avformat_alloc_context();
1945         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1946         assert(filename.size() < sizeof(avctx->filename) - 1);
1947         strcpy(avctx->filename, filename.c_str());
1948
1949         string url = "file:" + filename;
1950         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1951         if (ret < 0) {
1952                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1953                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1954                 exit(1);
1955         }
1956
1957         string video_extradata = "";  // FIXME: See other comment about global headers.
1958         AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1959         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1960                 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1)));
1961 }
1962
1963 void QuickSyncEncoderImpl::encode_thread_func()
1964 {
1965         int64_t last_dts = -1;
1966         int gop_start_display_frame_num = 0;
1967         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1968                 PendingFrame frame;
1969                 int pts_lag;
1970                 int frame_type, display_frame_num;
1971                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1972                                        &display_frame_num, &frame_type, &pts_lag);
1973                 if (frame_type == FRAME_IDR) {
1974                         numShortTerm = 0;
1975                         current_frame_num = 0;
1976                         gop_start_display_frame_num = display_frame_num;
1977                 }
1978
1979                 {
1980                         unique_lock<mutex> lock(frame_queue_mutex);
1981                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1982                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1983                         });
1984                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1985                                 // We have queued frames that were supposed to be B-frames,
1986                                 // but will be no P-frame to encode them against. Encode them all
1987                                 // as P-frames instead. Note that this happens under the mutex,
1988                                 // but nobody else uses it at this point, since we're shutting down,
1989                                 // so there's no contention.
1990                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1991                                 return;
1992                         } else {
1993                                 frame = move(pending_video_frames[display_frame_num]);
1994                                 pending_video_frames.erase(display_frame_num);
1995                         }
1996                 }
1997
1998                 // Determine the dts of this frame.
1999                 int64_t dts;
2000                 if (pts_lag == -1) {
2001                         assert(last_dts != -1);
2002                         dts = last_dts + (TIMEBASE / MAX_FPS);
2003                 } else {
2004                         dts = frame.pts - pts_lag;
2005                 }
2006                 last_dts = dts;
2007
2008                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2009         }
2010 }
2011
2012 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2013 {
2014         if (pending_video_frames.empty()) {
2015                 return;
2016         }
2017
2018         for (auto &pending_frame : pending_video_frames) {
2019                 int display_frame_num = pending_frame.first;
2020                 assert(display_frame_num > 0);
2021                 PendingFrame frame = move(pending_frame.second);
2022                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2023                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2024                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2025                 last_dts = dts;
2026         }
2027
2028         if (global_flags.uncompressed_video_to_http ||
2029             global_flags.x264_video_to_http) {
2030                 // Add frames left in reorderer.
2031                 while (!reorderer->empty()) {
2032                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2033                         if (global_flags.uncompressed_video_to_http) {
2034                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2035                         } else {
2036                                 assert(global_flags.x264_video_to_http);
2037                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2038                         }
2039                 }
2040         }
2041 }
2042
2043 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2044 {
2045         AVPacket pkt;
2046         memset(&pkt, 0, sizeof(pkt));
2047         pkt.buf = nullptr;
2048         pkt.data = const_cast<uint8_t *>(data);
2049         pkt.size = frame_width * frame_height * 2;
2050         pkt.stream_index = 0;
2051         pkt.flags = AV_PKT_FLAG_KEY;
2052         pkt.duration = duration;
2053         stream_mux->add_packet(pkt, pts, pts);
2054 }
2055
2056 namespace {
2057
2058 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2059 {
2060         if (src_width == dst_pitch) {
2061                 memcpy(dst, src, src_width * height);
2062         } else {
2063                 for (size_t y = 0; y < height; ++y) {
2064                         const uint8_t *sptr = src + y * src_width;
2065                         uint8_t *dptr = dst + y * dst_pitch;
2066                         memcpy(dptr, sptr, src_width);
2067                 }
2068         }
2069 }
2070
2071 }  // namespace
2072
2073 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2074                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2075 {
2076         // Wait for the GPU to be done with the frame.
2077         GLenum sync_status;
2078         do {
2079                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2080                 check_error();
2081         } while (sync_status == GL_TIMEOUT_EXPIRED);
2082         assert(sync_status != GL_WAIT_FAILED);
2083
2084         // Release back any input frames we needed to render this frame.
2085         frame.input_frames.clear();
2086
2087         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2088         VAStatus va_status;
2089
2090         if (use_zerocopy) {
2091                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2092                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2093                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2094                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2095         } else {
2096                 unsigned char *surface_p = nullptr;
2097                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2098
2099                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2100                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2101
2102                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2103                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2104
2105                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2106                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2107
2108                 if (global_flags.uncompressed_video_to_http ||
2109                     global_flags.x264_video_to_http) {
2110                         // Add uncompressed video. (Note that pts == dts here.)
2111                         // Delay needs to match audio.
2112                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2113                         if (output_frame.data != nullptr) {
2114                                 if (global_flags.uncompressed_video_to_http) {
2115                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2116                                 } else {
2117                                         assert(global_flags.x264_video_to_http);
2118                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2119                                 }
2120                         }
2121                 }
2122         }
2123
2124         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2125         CHECK_VASTATUS(va_status, "vaDestroyImage");
2126
2127         // Schedule the frame for encoding.
2128         VASurfaceID va_surface = surf->src_surface;
2129         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2130         CHECK_VASTATUS(va_status, "vaBeginPicture");
2131
2132         if (frame_type == FRAME_IDR) {
2133                 // FIXME: If the mux wants global headers, we should not put the
2134                 // SPS/PPS before each IDR frame, but rather put it into the
2135                 // codec extradata (formatted differently?).
2136                 render_sequence();
2137                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2138                 if (h264_packedheader) {
2139                         render_packedsequence();
2140                         render_packedpicture();
2141                 }
2142         } else {
2143                 //render_sequence();
2144                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2145         }
2146         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2147
2148         va_status = vaEndPicture(va_dpy, context_id);
2149         CHECK_VASTATUS(va_status, "vaEndPicture");
2150
2151         // so now the data is done encoding (well, async job kicked off)...
2152         // we send that to the storage thread
2153         storage_task tmp;
2154         tmp.display_order = display_frame_num;
2155         tmp.frame_type = frame_type;
2156         tmp.pts = pts;
2157         tmp.dts = dts;
2158         tmp.duration = duration;
2159         storage_task_enqueue(move(tmp));
2160
2161         update_ReferenceFrames(frame_type);
2162 }
2163
2164 // Proxy object.
2165 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
2166         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2167
2168 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2169 QuickSyncEncoder::~QuickSyncEncoder() {}
2170
2171 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2172 {
2173         impl->add_audio(pts, audio);
2174 }
2175
2176 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2177 {
2178         return impl->begin_frame(y_tex, cbcr_tex);
2179 }
2180
2181 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2182 {
2183         return impl->end_frame(pts, duration, input_frames);
2184 }
2185
2186 void QuickSyncEncoder::shutdown()
2187 {
2188         impl->shutdown();
2189 }
2190
2191 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2192 {
2193         impl->set_stream_mux(mux);
2194 }
2195
2196 int64_t QuickSyncEncoder::global_delay() const {
2197         return impl->global_delay();
2198 }