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 <map>
  25 #include <memory>
  26 #include <mutex>
  27 #include <queue>
  28 #include <stack>
  29 #include <string>
  30 #include <thread>
  31 #include <utility>
  32
  33 extern "C" {
  34
  35 #include <libavcodec/avcodec.h>
  36 #include <libavformat/avio.h>
  37 #include <libavutil/error.h>
  38 #include <libdrm/drm_fourcc.h>
  39
  40 }  // namespace
  41
  42 #include "audio_encoder.h"
  43 #include "context.h"
  44 #include "defs.h"
  45 #include "ffmpeg_raii.h"
  46 #include "flags.h"
  47 #include "mux.h"
  48 #include "ref_counted_frame.h"
  49 #include "timebase.h"
  50 #include "x264_encoder.h"
  51
  52 using namespace std;
  53
  54 class QOpenGLContext;
  55 class QSurface;
  56
  57 #define CHECK_VASTATUS(va_status, func)                                 \
  58     if (va_status != VA_STATUS_SUCCESS) {                               \
  59         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  60         exit(1);                                                        \
  61     }
  62
  63 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  64
  65 //#include "loadsurface.h"
  66
  67 #define NAL_REF_IDC_NONE        0
  68 #define NAL_REF_IDC_LOW         1
  69 #define NAL_REF_IDC_MEDIUM      2
  70 #define NAL_REF_IDC_HIGH        3
  71
  72 #define NAL_NON_IDR             1
  73 #define NAL_IDR                 5
  74 #define NAL_SPS                 7
  75 #define NAL_PPS                 8
  76 #define NAL_SEI                 6
  77
  78 #define SLICE_TYPE_P            0
  79 #define SLICE_TYPE_B            1
  80 #define SLICE_TYPE_I            2
  81 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  82 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  83 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  84
  85
  86 #define ENTROPY_MODE_CAVLC      0
  87 #define ENTROPY_MODE_CABAC      1
  88
  89 #define PROFILE_IDC_BASELINE    66
  90 #define PROFILE_IDC_MAIN        77
  91 #define PROFILE_IDC_HIGH        100
  92
  93 #define BITSTREAM_ALLOCATE_STEPPING     4096
  94 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  95 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  96 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  97
  98 static constexpr unsigned int MaxFrameNum = (2<<16);
  99 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
 100 static constexpr unsigned int Log2MaxFrameNum = 16;
 101 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
 102 static constexpr int rc_default_modes[] = {  // Priority list of modes.
 103     VA_RC_VBR,
 104     VA_RC_CQP,
 105     VA_RC_VBR_CONSTRAINED,
 106     VA_RC_CBR,
 107     VA_RC_VCM,
 108     VA_RC_NONE,
 109 };
 110
 111 /* thread to save coded data */
 112 #define SRC_SURFACE_FREE        0
 113 #define SRC_SURFACE_IN_ENCODING 1
 114
 115 struct __bitstream {
 116     unsigned int *buffer;
 117     int bit_offset;
 118     int max_size_in_dword;
 119 };
 120 typedef struct __bitstream bitstream;
 121
 122 using namespace std;
 123
 124 // H.264 video comes out in encoding order (e.g. with two B-frames:
 125 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 126 // come in the right order. Since we do everything, including waiting
 127 // for the frames to come out of OpenGL, in encoding order, we need
 128 // a reordering buffer for uncompressed frames so that they come out
 129 // correctly. We go the super-lazy way of not making it understand
 130 // anything about the true order (which introduces some extra latency,
 131 // though); we know that for N B-frames we need at most (N-1) frames
 132 // in the reorder buffer, and can just sort on that.
 133 //
 134 // The class also deals with keeping a freelist as needed.
 135 class FrameReorderer {
 136 public:
 137         FrameReorderer(unsigned queue_length, int width, int height);
 138
 139         struct Frame {
 140                 int64_t pts, duration;
 141                 uint8_t *data;
 142
 143                 // Invert to get the smallest pts first.
 144                 bool operator< (const Frame &other) const { return pts > other.pts; }
 145         };
 146
 147         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 148         // Does _not_ take ownership of data; a copy is taken if needed.
 149         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 150         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 151         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data);
 152
 153         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 154         Frame get_first_frame();
 155
 156         bool empty() const { return frames.empty(); }
 157
 158 private:
 159         unsigned queue_length;
 160         int width, height;
 161
 162         priority_queue<Frame> frames;
 163         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 164
 165         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 166         // except priority_queue doesn't work well with movable-only types.
 167         vector<unique_ptr<uint8_t[]>> owner;
 168 };
 169
 170 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 171     : queue_length(queue_length), width(width), height(height)
 172 {
 173         for (unsigned i = 0; i < queue_length; ++i) {
 174                 owner.emplace_back(new uint8_t[width * height * 2]);
 175                 freelist.push(owner.back().get());
 176         }
 177 }
 178
 179 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data)
 180 {
 181         if (queue_length == 0) {
 182                 return Frame{pts, duration, data};
 183         }
 184
 185         assert(!freelist.empty());
 186         uint8_t *storage = freelist.top();
 187         freelist.pop();
 188         memcpy(storage, data, width * height * 2);
 189         frames.push(Frame{pts, duration, storage});
 190
 191         if (frames.size() >= queue_length) {
 192                 return get_first_frame();
 193         } else {
 194                 return Frame{-1, -1, nullptr};
 195         }
 196 }
 197
 198 FrameReorderer::Frame FrameReorderer::get_first_frame()
 199 {
 200         assert(!frames.empty());
 201         Frame storage = frames.top();
 202         frames.pop();
 203         freelist.push(storage.data);
 204         return storage;
 205 }
 206
 207 class QuickSyncEncoderImpl {
 208 public:
 209         QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder);
 210         ~QuickSyncEncoderImpl();
 211         void add_audio(int64_t pts, vector<float> audio);
 212         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 213         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
 214         void shutdown();
 215         void release_gl_resources();
 216         void set_stream_mux(Mux *mux)
 217         {
 218                 stream_mux = mux;
 219         }
 220
 221         // So we never get negative dts.
 222         int64_t global_delay() const {
 223                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 224         }
 225
 226 private:
 227         struct storage_task {
 228                 unsigned long long display_order;
 229                 int frame_type;
 230                 vector<float> audio;
 231                 int64_t pts, dts, duration;
 232         };
 233         struct PendingFrame {
 234                 RefCountedGLsync fence;
 235                 vector<RefCountedFrame> input_frames;
 236                 int64_t pts, duration;
 237         };
 238
 239         void open_output_file(const std::string &filename);
 240         void encode_thread_func();
 241         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 242         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 243         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 244                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 245         void storage_task_thread();
 246         void storage_task_enqueue(storage_task task);
 247         void save_codeddata(storage_task task);
 248         int render_packedsequence();
 249         int render_packedpicture();
 250         void render_packedslice();
 251         int render_sequence();
 252         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 253         void sps_rbsp(bitstream *bs);
 254         void pps_rbsp(bitstream *bs);
 255         int build_packed_pic_buffer(unsigned char **header_buffer);
 256         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 257         void slice_header(bitstream *bs);
 258         int build_packed_seq_buffer(unsigned char **header_buffer);
 259         int build_packed_slice_buffer(unsigned char **header_buffer);
 260         int init_va(const string &va_display);
 261         int deinit_va();
 262         void enable_zerocopy_if_possible();
 263         VADisplay va_open_display(const string &va_display);
 264         void va_close_display(VADisplay va_dpy);
 265         int setup_encode();
 266         void release_encode();
 267         void update_ReferenceFrames(int frame_type);
 268         int update_RefPicList(int frame_type);
 269
 270         bool is_shutdown = false;
 271         bool has_released_gl_resources = false;
 272         bool use_zerocopy;
 273         int drm_fd = -1;
 274
 275         thread encode_thread, storage_thread;
 276
 277         mutex storage_task_queue_mutex;
 278         condition_variable storage_task_queue_changed;
 279         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 280         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 281         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 282
 283         mutex frame_queue_mutex;
 284         condition_variable frame_queue_nonempty;
 285         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 286
 287         int current_storage_frame;
 288
 289         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 290         movit::ResourcePool *resource_pool;
 291         QSurface *surface;
 292
 293         unique_ptr<AudioEncoder> file_audio_encoder;
 294
 295         unique_ptr<FrameReorderer> reorderer;
 296         X264Encoder *x264_encoder;  // nullptr if not using x264.
 297
 298         Mux* stream_mux = nullptr;  // To HTTP.
 299         unique_ptr<Mux> file_mux;  // To local disk.
 300
 301         Display *x11_display = nullptr;
 302
 303         // Encoder parameters
 304         VADisplay va_dpy;
 305         VAProfile h264_profile = (VAProfile)~0;
 306         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 307         int config_attrib_num = 0, enc_packed_header_idx;
 308
 309         struct GLSurface {
 310                 VASurfaceID src_surface, ref_surface;
 311                 VABufferID coded_buf;
 312
 313                 VAImage surface_image;
 314                 GLuint y_tex, cbcr_tex;
 315
 316                 // Only if use_zerocopy == true.
 317                 EGLImage y_egl_image, cbcr_egl_image;
 318
 319                 // Only if use_zerocopy == false.
 320                 GLuint pbo;
 321                 uint8_t *y_ptr, *cbcr_ptr;
 322                 size_t y_offset, cbcr_offset;
 323         };
 324         GLSurface gl_surfaces[SURFACE_NUM];
 325
 326         VAConfigID config_id;
 327         VAContextID context_id;
 328         VAEncSequenceParameterBufferH264 seq_param;
 329         VAEncPictureParameterBufferH264 pic_param;
 330         VAEncSliceParameterBufferH264 slice_param;
 331         VAPictureH264 CurrentCurrPic;
 332         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 333
 334         // Static quality settings.
 335         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 336         static constexpr unsigned int num_ref_frames = 2;
 337         static constexpr int initial_qp = 15;
 338         static constexpr int minimal_qp = 0;
 339         static constexpr int intra_period = 30;
 340         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 341
 342         // Quality settings that are meant to be static, but might be overridden
 343         // by the profile.
 344         int constraint_set_flag = 0;
 345         int h264_packedheader = 0; /* support pack header? */
 346         int h264_maxref = (1<<16|1);
 347         int h264_entropy_mode = 1; /* cabac */
 348         int ip_period = 3;
 349
 350         int rc_mode = -1;
 351         unsigned int current_frame_num = 0;
 352         unsigned int numShortTerm = 0;
 353
 354         int frame_width;
 355         int frame_height;
 356         int frame_width_mbaligned;
 357         int frame_height_mbaligned;
 358 };
 359
 360 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 361 // but if we don't delete it here, we get leaks. The GStreamer implementation
 362 // does the same.
 363 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 364 {
 365     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 366     CHECK_VASTATUS(va_status, "vaRenderPicture");
 367
 368     for (int i = 0; i < num_buffers; ++i) {
 369         va_status = vaDestroyBuffer(dpy, buffers[i]);
 370         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 371     }
 372 }
 373
 374 static unsigned int
 375 va_swap32(unsigned int val)
 376 {
 377     unsigned char *pval = (unsigned char *)&val;
 378
 379     return ((pval[0] << 24)     |
 380             (pval[1] << 16)     |
 381             (pval[2] << 8)      |
 382             (pval[3] << 0));
 383 }
 384
 385 static void
 386 bitstream_start(bitstream *bs)
 387 {
 388     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 389     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 390     bs->bit_offset = 0;
 391 }
 392
 393 static void
 394 bitstream_end(bitstream *bs)
 395 {
 396     int pos = (bs->bit_offset >> 5);
 397     int bit_offset = (bs->bit_offset & 0x1f);
 398     int bit_left = 32 - bit_offset;
 399
 400     if (bit_offset) {
 401         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 402     }
 403 }
 404
 405 static void
 406 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 407 {
 408     int pos = (bs->bit_offset >> 5);
 409     int bit_offset = (bs->bit_offset & 0x1f);
 410     int bit_left = 32 - bit_offset;
 411
 412     if (!size_in_bits)
 413         return;
 414
 415     bs->bit_offset += size_in_bits;
 416
 417     if (bit_left > size_in_bits) {
 418         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 419     } else {
 420         size_in_bits -= bit_left;
 421         if (bit_left >= 32) {
 422             bs->buffer[pos] = (val >> size_in_bits);
 423         } else {
 424             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 425         }
 426         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 427
 428         if (pos + 1 == bs->max_size_in_dword) {
 429             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 430             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 431         }
 432
 433         bs->buffer[pos + 1] = val;
 434     }
 435 }
 436
 437 static void
 438 bitstream_put_ue(bitstream *bs, unsigned int val)
 439 {
 440     int size_in_bits = 0;
 441     int tmp_val = ++val;
 442
 443     while (tmp_val) {
 444         tmp_val >>= 1;
 445         size_in_bits++;
 446     }
 447
 448     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 449     bitstream_put_ui(bs, val, size_in_bits);
 450 }
 451
 452 static void
 453 bitstream_put_se(bitstream *bs, int val)
 454 {
 455     unsigned int new_val;
 456
 457     if (val <= 0)
 458         new_val = -2 * val;
 459     else
 460         new_val = 2 * val - 1;
 461
 462     bitstream_put_ue(bs, new_val);
 463 }
 464
 465 static void
 466 bitstream_byte_aligning(bitstream *bs, int bit)
 467 {
 468     int bit_offset = (bs->bit_offset & 0x7);
 469     int bit_left = 8 - bit_offset;
 470     int new_val;
 471
 472     if (!bit_offset)
 473         return;
 474
 475     assert(bit == 0 || bit == 1);
 476
 477     if (bit)
 478         new_val = (1 << bit_left) - 1;
 479     else
 480         new_val = 0;
 481
 482     bitstream_put_ui(bs, new_val, bit_left);
 483 }
 484
 485 static void
 486 rbsp_trailing_bits(bitstream *bs)
 487 {
 488     bitstream_put_ui(bs, 1, 1);
 489     bitstream_byte_aligning(bs, 0);
 490 }
 491
 492 static void nal_start_code_prefix(bitstream *bs)
 493 {
 494     bitstream_put_ui(bs, 0x00000001, 32);
 495 }
 496
 497 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 498 {
 499     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 500     bitstream_put_ui(bs, nal_ref_idc, 2);
 501     bitstream_put_ui(bs, nal_unit_type, 5);
 502 }
 503
 504 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
 505 {
 506     int profile_idc = PROFILE_IDC_BASELINE;
 507
 508     if (h264_profile  == VAProfileH264High)
 509         profile_idc = PROFILE_IDC_HIGH;
 510     else if (h264_profile  == VAProfileH264Main)
 511         profile_idc = PROFILE_IDC_MAIN;
 512
 513     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 514     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 515     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 516     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 517     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 518     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 519     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 520     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 521
 522     if ( profile_idc == PROFILE_IDC_HIGH) {
 523         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 524         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 525         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 526         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 527         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 528     }
 529
 530     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 531     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 532
 533     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 534         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 535     else {
 536         assert(0);
 537     }
 538
 539     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 540     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 541
 542     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 543     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 544     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 545
 546     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 547         assert(0);
 548     }
 549
 550     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 551     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 552
 553     if (seq_param.frame_cropping_flag) {
 554         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 555         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 556         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 557         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 558     }
 559
 560     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 561     if ( false ) {
 562         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 563     } else {
 564         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 565         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 566         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 567         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 568         {
 569             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 570             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 571             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 572             {
 573                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 574                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 575                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 576             }
 577         }
 578         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 579         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 580         {
 581             bitstream_put_ui(bs, 1, 32);  // FPS
 582             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 583             bitstream_put_ui(bs, 1, 1);
 584         }
 585         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 586         {
 587             // hrd_parameters
 588             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 589             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 590             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 591
 592             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 593             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 594             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 595
 596             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 597             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 598             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 599             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 600         }
 601         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 602         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 603
 604         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 605         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 606     }
 607
 608     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 609 }
 610
 611
 612 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
 613 {
 614     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 615     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 616
 617     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 618
 619     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 620
 621     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 622
 623     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 624     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 625
 626     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 627     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 628
 629     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 630     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 631     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 632
 633     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 634     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 635     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 636
 637     /* more_rbsp_data */
 638     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 639     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 640     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 641
 642     rbsp_trailing_bits(bs);
 643 }
 644
 645 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
 646 {
 647     int first_mb_in_slice = slice_param.macroblock_address;
 648
 649     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 650     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 651     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 652     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 653
 654     /* frame_mbs_only_flag == 1 */
 655     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 656         /* FIXME: */
 657         assert(0);
 658     }
 659
 660     if (pic_param.pic_fields.bits.idr_pic_flag)
 661         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 662
 663     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 664         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 665         /* pic_order_present_flag == 0 */
 666     } else {
 667         /* FIXME: */
 668         assert(0);
 669     }
 670
 671     /* redundant_pic_cnt_present_flag == 0 */
 672     /* slice type */
 673     if (IS_P_SLICE(slice_param.slice_type)) {
 674         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 675
 676         if (slice_param.num_ref_idx_active_override_flag)
 677             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 678
 679         /* ref_pic_list_reordering */
 680         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 681     } else if (IS_B_SLICE(slice_param.slice_type)) {
 682         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 683
 684         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 685
 686         if (slice_param.num_ref_idx_active_override_flag) {
 687             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 688             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 689         }
 690
 691         /* ref_pic_list_reordering */
 692         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 693         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 694     }
 695
 696     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 697          IS_P_SLICE(slice_param.slice_type)) ||
 698         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 699          IS_B_SLICE(slice_param.slice_type))) {
 700         /* FIXME: fill weight/offset table */
 701         assert(0);
 702     }
 703
 704     /* dec_ref_pic_marking */
 705     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 706         unsigned char no_output_of_prior_pics_flag = 0;
 707         unsigned char long_term_reference_flag = 0;
 708         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 709
 710         if (pic_param.pic_fields.bits.idr_pic_flag) {
 711             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 712             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 713         } else {
 714             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 715         }
 716     }
 717
 718     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 719         !IS_I_SLICE(slice_param.slice_type))
 720         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 721
 722     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 723
 724     /* ignore for SP/SI */
 725
 726     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 727         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 728
 729         if (slice_param.disable_deblocking_filter_idc != 1) {
 730             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 731             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 732         }
 733     }
 734
 735     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 736         bitstream_byte_aligning(bs, 1);
 737     }
 738 }
 739
 740 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 741 {
 742     bitstream bs;
 743
 744     bitstream_start(&bs);
 745     nal_start_code_prefix(&bs);
 746     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 747     pps_rbsp(&bs);
 748     bitstream_end(&bs);
 749
 750     *header_buffer = (unsigned char *)bs.buffer;
 751     return bs.bit_offset;
 752 }
 753
 754 int
 755 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 756 {
 757     bitstream bs;
 758
 759     bitstream_start(&bs);
 760     nal_start_code_prefix(&bs);
 761     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 762     sps_rbsp(&bs);
 763     bitstream_end(&bs);
 764
 765     *header_buffer = (unsigned char *)bs.buffer;
 766     return bs.bit_offset;
 767 }
 768
 769 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 770 {
 771     bitstream bs;
 772     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 773     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 774
 775     bitstream_start(&bs);
 776     nal_start_code_prefix(&bs);
 777
 778     if (IS_I_SLICE(slice_param.slice_type)) {
 779         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 780     } else if (IS_P_SLICE(slice_param.slice_type)) {
 781         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 782     } else {
 783         assert(IS_B_SLICE(slice_param.slice_type));
 784         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 785     }
 786
 787     slice_header(&bs);
 788     bitstream_end(&bs);
 789
 790     *header_buffer = (unsigned char *)bs.buffer;
 791     return bs.bit_offset;
 792 }
 793
 794
 795 /*
 796   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 797   1) period between Frame #X and Frame #N = #X - #N
 798   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 799   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 800   4) intra_period and intra_idr_period take precedence over ip_period
 801   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 802      of I/IDR frames, see bellow examples
 803   -------------------------------------------------------------------
 804   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 805   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 806   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 807   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 808   1            1                ignored    IDR IDR IDR IDR...
 809   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 810   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 811   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 812                                               (PBB)(IBB)(PBB)(IBB)...
 813   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 814                                            IDRPPPPP IPPPPP IPPPPP...
 815   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 816                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 817                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 818                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 819                                            {IDR(PBB)(PBB)}                     (6/6/3)
 820                                            {IDR(PBB)(PBB)}.
 821 */
 822
 823 // General pts/dts strategy:
 824 //
 825 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 826 // bit tricky. We assume first of all that the frame rate never goes _above_
 827 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 828 // in at least this speed, as long at dts <= pts (the frame is not attempted
 829 // presented before it is decoded). Furthermore, we never have longer chains of
 830 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 831 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 832 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 833 // frame, with an order O <= C.)
 834 //
 835 // Many strategies are possible, but we establish these rules:
 836 //
 837 //  - Tip frames have dts = pts - (C-O)*N.
 838 //  - Non-tip frames have dts = dts_last + N.
 839 //
 840 // An example, with C=2 and N=10 and the data flow showed with arrows:
 841 //
 842 //        I  B  P  B  B  P
 843 //   pts: 30 40 50 60 70 80
 844 //        ↓  ↓     ↓
 845 //   dts: 10 30 20 60 50←40
 846 //         |  |  ↑        ↑
 847 //         `--|--'        |
 848 //             `----------'
 849 //
 850 // To show that this works fine also with irregular spacings, let's say that
 851 // the third frame is delayed a bit (something earlier was dropped). Now the
 852 // situation looks like this:
 853 //
 854 //        I  B  P  B  B   P
 855 //   pts: 30 40 80 90 100 110
 856 //        ↓  ↓     ↓
 857 //   dts: 10 30 20 90 50←40
 858 //         |  |  ↑        ↑
 859 //         `--|--'        |
 860 //             `----------'
 861 //
 862 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 863 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 864 //
 865 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 866 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 867 // a tip frame and should be given a dts based on the previous one.
 868 #define FRAME_P 0
 869 #define FRAME_B 1
 870 #define FRAME_I 2
 871 #define FRAME_IDR 7
 872 void encoding2display_order(
 873     int encoding_order, int intra_period,
 874     int intra_idr_period, int ip_period,
 875     int *displaying_order,
 876     int *frame_type, int *pts_lag)
 877 {
 878     int encoding_order_gop = 0;
 879
 880     *pts_lag = 0;
 881
 882     if (intra_period == 1) { /* all are I/IDR frames */
 883         *displaying_order = encoding_order;
 884         if (intra_idr_period == 0)
 885             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 886         else
 887             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 888         return;
 889     }
 890
 891     if (intra_period == 0)
 892         intra_idr_period = 0;
 893
 894     if (ip_period == 1) {
 895         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 896         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 897         *displaying_order = encoding_order;
 898
 899         if (encoding_order_gop == 0) { /* the first frame */
 900             *frame_type = FRAME_IDR;
 901         } else if (intra_period != 0 && /* have I frames */
 902                    encoding_order_gop >= 2 &&
 903                    (encoding_order_gop % intra_period == 0)) {
 904             *frame_type = FRAME_I;
 905         } else {
 906             *frame_type = FRAME_P;
 907         }
 908         return;
 909     }
 910
 911     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 912     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 913     *pts_lag = -1;  // Most frames are not tip frames.
 914
 915     if (encoding_order_gop == 0) { /* the first frame */
 916         *frame_type = FRAME_IDR;
 917         *displaying_order = encoding_order;
 918         // IDR frames are a special case; I honestly can't find the logic behind
 919         // why this is the right thing, but it seems to line up nicely in practice :-)
 920         *pts_lag = TIMEBASE / MAX_FPS;
 921     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 922         *frame_type = FRAME_B;
 923         *displaying_order = encoding_order - 1;
 924         if ((encoding_order_gop % ip_period) == 0) {
 925             *pts_lag = 0;  // Last B-frame.
 926         }
 927     } else if (intra_period != 0 && /* have I frames */
 928                encoding_order_gop >= 2 &&
 929                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 930         *frame_type = FRAME_I;
 931         *displaying_order = encoding_order + ip_period - 1;
 932     } else {
 933         *frame_type = FRAME_P;
 934         *displaying_order = encoding_order + ip_period - 1;
 935     }
 936 }
 937
 938
 939 static const char *rc_to_string(int rc_mode)
 940 {
 941     switch (rc_mode) {
 942     case VA_RC_NONE:
 943         return "NONE";
 944     case VA_RC_CBR:
 945         return "CBR";
 946     case VA_RC_VBR:
 947         return "VBR";
 948     case VA_RC_VCM:
 949         return "VCM";
 950     case VA_RC_CQP:
 951         return "CQP";
 952     case VA_RC_VBR_CONSTRAINED:
 953         return "VBR_CONSTRAINED";
 954     default:
 955         return "Unknown";
 956     }
 957 }
 958
 959 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
 960 {
 961         if (global_flags.uncompressed_video_to_http) {
 962                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 963                 use_zerocopy = false;
 964         } else if (global_flags.x264_video_to_http) {
 965                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 966                 use_zerocopy = false;
 967         } else {
 968                 use_zerocopy = true;
 969         }
 970 }
 971
 972 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
 973 {
 974         if (va_display.empty()) {
 975                 x11_display = XOpenDisplay(NULL);
 976                 if (!x11_display) {
 977                         fprintf(stderr, "error: can't connect to X server!\n");
 978                         return NULL;
 979                 }
 980                 enable_zerocopy_if_possible();
 981                 return vaGetDisplay(x11_display);
 982         } else if (va_display[0] != '/') {
 983                 x11_display = XOpenDisplay(va_display.c_str());
 984                 if (!x11_display) {
 985                         fprintf(stderr, "error: can't connect to X server!\n");
 986                         return NULL;
 987                 }
 988                 enable_zerocopy_if_possible();
 989                 return vaGetDisplay(x11_display);
 990         } else {
 991                 drm_fd = open(va_display.c_str(), O_RDWR);
 992                 if (drm_fd == -1) {
 993                         perror(va_display.c_str());
 994                         return NULL;
 995                 }
 996                 use_zerocopy = false;
 997                 return vaGetDisplayDRM(drm_fd);
 998         }
 999 }
1000
1001 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
1002 {
1003         if (x11_display) {
1004                 XCloseDisplay(x11_display);
1005                 x11_display = nullptr;
1006         }
1007         if (drm_fd != -1) {
1008                 close(drm_fd);
1009         }
1010 }
1011
1012 int QuickSyncEncoderImpl::init_va(const string &va_display)
1013 {
1014     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1015     VAEntrypoint *entrypoints;
1016     int num_entrypoints, slice_entrypoint;
1017     int support_encode = 0;
1018     int major_ver, minor_ver;
1019     VAStatus va_status;
1020     unsigned int i;
1021
1022     va_dpy = va_open_display(va_display);
1023     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1024     CHECK_VASTATUS(va_status, "vaInitialize");
1025
1026     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1027     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1028     if (!entrypoints) {
1029         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1030         exit(1);
1031     }
1032
1033     /* use the highest profile */
1034     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1035         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1036             continue;
1037
1038         h264_profile = profile_list[i];
1039         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1040         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1041             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1042                 support_encode = 1;
1043                 break;
1044             }
1045         }
1046         if (support_encode == 1)
1047             break;
1048     }
1049
1050     if (support_encode == 0) {
1051         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1052         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1053         printf("to use VA-API against DRM instead of X11.\n");
1054         exit(1);
1055     } else {
1056         switch (h264_profile) {
1057             case VAProfileH264Baseline:
1058                 ip_period = 1;
1059                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1060                 h264_entropy_mode = 0;
1061                 break;
1062             case VAProfileH264ConstrainedBaseline:
1063                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1064                 ip_period = 1;
1065                 break;
1066
1067             case VAProfileH264Main:
1068                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1069                 break;
1070
1071             case VAProfileH264High:
1072                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1073                 break;
1074             default:
1075                 h264_profile = VAProfileH264Baseline;
1076                 ip_period = 1;
1077                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1078                 break;
1079         }
1080     }
1081
1082     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1083
1084     /* find out the format for the render target, and rate control mode */
1085     for (i = 0; i < VAConfigAttribTypeMax; i++)
1086         attrib[i].type = (VAConfigAttribType)i;
1087
1088     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1089                                       &attrib[0], VAConfigAttribTypeMax);
1090     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1091     /* check the interested configattrib */
1092     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1093         printf("Not find desired YUV420 RT format\n");
1094         exit(1);
1095     } else {
1096         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1097         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1098         config_attrib_num++;
1099     }
1100
1101     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1102         int tmp = attrib[VAConfigAttribRateControl].value;
1103
1104         if (rc_mode == -1 || !(rc_mode & tmp))  {
1105             if (rc_mode != -1) {
1106                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1107             }
1108
1109             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1110                 if (rc_default_modes[i] & tmp) {
1111                     rc_mode = rc_default_modes[i];
1112                     break;
1113                 }
1114             }
1115         }
1116
1117         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1118         config_attrib[config_attrib_num].value = rc_mode;
1119         config_attrib_num++;
1120     }
1121
1122
1123     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1124         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1125
1126         h264_packedheader = 1;
1127         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1128         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1129
1130         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1131             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1132         }
1133
1134         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1135             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1136         }
1137
1138         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1139             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1140         }
1141
1142         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1143             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1144         }
1145
1146         enc_packed_header_idx = config_attrib_num;
1147         config_attrib_num++;
1148     }
1149
1150     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1151         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1152         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1153         config_attrib_num++;
1154     }
1155
1156     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1157         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1158     }
1159
1160     free(entrypoints);
1161     return 0;
1162 }
1163
1164 int QuickSyncEncoderImpl::setup_encode()
1165 {
1166     VAStatus va_status;
1167     VASurfaceID *tmp_surfaceid;
1168     int codedbuf_size, i;
1169     VASurfaceID src_surface[SURFACE_NUM];
1170     VASurfaceID ref_surface[SURFACE_NUM];
1171
1172     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1173             &config_attrib[0], config_attrib_num, &config_id);
1174     CHECK_VASTATUS(va_status, "vaCreateConfig");
1175
1176     /* create source surfaces */
1177     va_status = vaCreateSurfaces(va_dpy,
1178                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1179                                  &src_surface[0], SURFACE_NUM,
1180                                  NULL, 0);
1181     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1182
1183     /* create reference surfaces */
1184     va_status = vaCreateSurfaces(va_dpy,
1185                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1186                                  &ref_surface[0], SURFACE_NUM,
1187                                  NULL, 0);
1188     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1189
1190     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1191     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1192     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1193
1194     /* Create a context for this encode pipe */
1195     va_status = vaCreateContext(va_dpy, config_id,
1196                                 frame_width_mbaligned, frame_height_mbaligned,
1197                                 VA_PROGRESSIVE,
1198                                 tmp_surfaceid, 2 * SURFACE_NUM,
1199                                 &context_id);
1200     CHECK_VASTATUS(va_status, "vaCreateContext");
1201     free(tmp_surfaceid);
1202
1203     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1204
1205     for (i = 0; i < SURFACE_NUM; i++) {
1206         /* create coded buffer once for all
1207          * other VA buffers which won't be used again after vaRenderPicture.
1208          * so APP can always vaCreateBuffer for every frame
1209          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1210          * so VA won't maintain the coded buffer
1211          */
1212         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1213                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1214         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1215     }
1216
1217     /* create OpenGL objects */
1218     //glGenFramebuffers(SURFACE_NUM, fbos);
1219
1220     for (i = 0; i < SURFACE_NUM; i++) {
1221         if (use_zerocopy) {
1222             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
1223             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
1224         } else {
1225             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
1226             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
1227
1228             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1229             // buffers, due to potentially differing pitch.
1230             glGenBuffers(1, &gl_surfaces[i].pbo);
1231             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1232             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1233             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1234             gl_surfaces[i].y_offset = 0;
1235             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1236             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1237             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1238             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1239         }
1240     }
1241
1242     for (i = 0; i < SURFACE_NUM; i++) {
1243         gl_surfaces[i].src_surface = src_surface[i];
1244         gl_surfaces[i].ref_surface = ref_surface[i];
1245     }
1246
1247     return 0;
1248 }
1249
1250 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1251 //
1252 //   2 1 0 [3] 4 8 9
1253 template<class T, class C>
1254 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1255 {
1256         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1257         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1258         sort(middle, end, less_than);
1259 }
1260
1261 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1262 {
1263     int i;
1264
1265     if (frame_type == FRAME_B)
1266         return;
1267
1268     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1269     numShortTerm++;
1270     if (numShortTerm > num_ref_frames)
1271         numShortTerm = num_ref_frames;
1272     for (i=numShortTerm-1; i>0; i--)
1273         ReferenceFrames[i] = ReferenceFrames[i-1];
1274     ReferenceFrames[0] = CurrentCurrPic;
1275
1276     current_frame_num++;
1277     if (current_frame_num > MaxFrameNum)
1278         current_frame_num = 0;
1279 }
1280
1281
1282 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1283 {
1284     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1285         return a.frame_idx > b.frame_idx;
1286     };
1287     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1288         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1289     };
1290     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1291         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1292     };
1293
1294     if (frame_type == FRAME_P) {
1295         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1296         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1297     } else if (frame_type == FRAME_B) {
1298         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1299         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1300
1301         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1302         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1303     }
1304
1305     return 0;
1306 }
1307
1308
1309 int QuickSyncEncoderImpl::render_sequence()
1310 {
1311     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1312     VAStatus va_status;
1313     VAEncMiscParameterBuffer *misc_param;
1314     VAEncMiscParameterRateControl *misc_rate_ctrl;
1315
1316     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1317     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1318     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1319     seq_param.bits_per_second = frame_bitrate;
1320
1321     seq_param.intra_period = intra_period;
1322     seq_param.intra_idr_period = intra_idr_period;
1323     seq_param.ip_period = ip_period;
1324
1325     seq_param.max_num_ref_frames = num_ref_frames;
1326     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1327     seq_param.time_scale = TIMEBASE * 2;
1328     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1329     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1330     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1331     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1332     seq_param.seq_fields.bits.chroma_format_idc = 1;
1333     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1334
1335     if (frame_width != frame_width_mbaligned ||
1336         frame_height != frame_height_mbaligned) {
1337         seq_param.frame_cropping_flag = 1;
1338         seq_param.frame_crop_left_offset = 0;
1339         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1340         seq_param.frame_crop_top_offset = 0;
1341         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1342     }
1343
1344     va_status = vaCreateBuffer(va_dpy, context_id,
1345                                VAEncSequenceParameterBufferType,
1346                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1347     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1348
1349     va_status = vaCreateBuffer(va_dpy, context_id,
1350                                VAEncMiscParameterBufferType,
1351                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1352                                1, NULL, &rc_param_buf);
1353     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1354
1355     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1356     misc_param->type = VAEncMiscParameterTypeRateControl;
1357     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1358     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1359     misc_rate_ctrl->bits_per_second = frame_bitrate;
1360     misc_rate_ctrl->target_percentage = 66;
1361     misc_rate_ctrl->window_size = 1000;
1362     misc_rate_ctrl->initial_qp = initial_qp;
1363     misc_rate_ctrl->min_qp = minimal_qp;
1364     misc_rate_ctrl->basic_unit_size = 0;
1365     vaUnmapBuffer(va_dpy, rc_param_buf);
1366
1367     render_id[0] = seq_param_buf;
1368     render_id[1] = rc_param_buf;
1369
1370     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1371
1372     return 0;
1373 }
1374
1375 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1376 {
1377     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1378     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1379     int PicOrderCntMsb, TopFieldOrderCnt;
1380
1381     if (frame_type == FRAME_IDR)
1382         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1383     else {
1384         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1385         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1386     }
1387
1388     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1389         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1390         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1391     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1392              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1393         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1394     else
1395         PicOrderCntMsb = prevPicOrderCntMsb;
1396
1397     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1398
1399     if (frame_type != FRAME_B) {
1400         PicOrderCntMsb_ref = PicOrderCntMsb;
1401         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1402     }
1403
1404     return TopFieldOrderCnt;
1405 }
1406
1407 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1408 {
1409     VABufferID pic_param_buf;
1410     VAStatus va_status;
1411     int i = 0;
1412
1413     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1414     pic_param.CurrPic.frame_idx = current_frame_num;
1415     pic_param.CurrPic.flags = 0;
1416     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1417     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1418     CurrentCurrPic = pic_param.CurrPic;
1419
1420     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1421     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1422         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1423         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1424     }
1425
1426     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1427     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1428     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1429     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1430     pic_param.frame_num = current_frame_num;
1431     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1432     pic_param.last_picture = false;  // FIXME
1433     pic_param.pic_init_qp = initial_qp;
1434
1435     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1436                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1437     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1438
1439     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1440
1441     return 0;
1442 }
1443
1444 int QuickSyncEncoderImpl::render_packedsequence()
1445 {
1446     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1447     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1448     unsigned int length_in_bits;
1449     unsigned char *packedseq_buffer = NULL;
1450     VAStatus va_status;
1451
1452     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1453
1454     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1455
1456     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1457     packedheader_param_buffer.has_emulation_bytes = 0;
1458     va_status = vaCreateBuffer(va_dpy,
1459                                context_id,
1460                                VAEncPackedHeaderParameterBufferType,
1461                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1462                                &packedseq_para_bufid);
1463     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1464
1465     va_status = vaCreateBuffer(va_dpy,
1466                                context_id,
1467                                VAEncPackedHeaderDataBufferType,
1468                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1469                                &packedseq_data_bufid);
1470     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1471
1472     render_id[0] = packedseq_para_bufid;
1473     render_id[1] = packedseq_data_bufid;
1474     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1475
1476     free(packedseq_buffer);
1477
1478     return 0;
1479 }
1480
1481
1482 int QuickSyncEncoderImpl::render_packedpicture()
1483 {
1484     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1485     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1486     unsigned int length_in_bits;
1487     unsigned char *packedpic_buffer = NULL;
1488     VAStatus va_status;
1489
1490     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1491     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1492     packedheader_param_buffer.bit_length = length_in_bits;
1493     packedheader_param_buffer.has_emulation_bytes = 0;
1494
1495     va_status = vaCreateBuffer(va_dpy,
1496                                context_id,
1497                                VAEncPackedHeaderParameterBufferType,
1498                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1499                                &packedpic_para_bufid);
1500     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1501
1502     va_status = vaCreateBuffer(va_dpy,
1503                                context_id,
1504                                VAEncPackedHeaderDataBufferType,
1505                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1506                                &packedpic_data_bufid);
1507     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1508
1509     render_id[0] = packedpic_para_bufid;
1510     render_id[1] = packedpic_data_bufid;
1511     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1512
1513     free(packedpic_buffer);
1514
1515     return 0;
1516 }
1517
1518 void QuickSyncEncoderImpl::render_packedslice()
1519 {
1520     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1521     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1522     unsigned int length_in_bits;
1523     unsigned char *packedslice_buffer = NULL;
1524     VAStatus va_status;
1525
1526     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1527     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1528     packedheader_param_buffer.bit_length = length_in_bits;
1529     packedheader_param_buffer.has_emulation_bytes = 0;
1530
1531     va_status = vaCreateBuffer(va_dpy,
1532                                context_id,
1533                                VAEncPackedHeaderParameterBufferType,
1534                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1535                                &packedslice_para_bufid);
1536     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1537
1538     va_status = vaCreateBuffer(va_dpy,
1539                                context_id,
1540                                VAEncPackedHeaderDataBufferType,
1541                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1542                                &packedslice_data_bufid);
1543     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1544
1545     render_id[0] = packedslice_para_bufid;
1546     render_id[1] = packedslice_data_bufid;
1547     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1548
1549     free(packedslice_buffer);
1550 }
1551
1552 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1553 {
1554     VABufferID slice_param_buf;
1555     VAStatus va_status;
1556     int i;
1557
1558     update_RefPicList(frame_type);
1559
1560     /* one frame, one slice */
1561     slice_param.macroblock_address = 0;
1562     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1563     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1564     if (frame_type == FRAME_IDR) {
1565         if (encoding_frame_num != 0)
1566             ++slice_param.idr_pic_id;
1567     } else if (frame_type == FRAME_P) {
1568         int refpiclist0_max = h264_maxref & 0xffff;
1569         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1570
1571         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1572             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1573             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1574         }
1575     } else if (frame_type == FRAME_B) {
1576         int refpiclist0_max = h264_maxref & 0xffff;
1577         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1578
1579         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1580         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1581             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1582             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1583         }
1584
1585         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1586         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1587             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1588             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1589         }
1590     }
1591
1592     slice_param.slice_alpha_c0_offset_div2 = 0;
1593     slice_param.slice_beta_offset_div2 = 0;
1594     slice_param.direct_spatial_mv_pred_flag = 1;
1595     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1596
1597
1598     if (h264_packedheader &&
1599         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1600         render_packedslice();
1601
1602     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1603                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1604     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1605
1606     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1607
1608     return 0;
1609 }
1610
1611
1612
1613 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1614 {
1615         VACodedBufferSegment *buf_list = NULL;
1616         VAStatus va_status;
1617
1618         string data;
1619
1620         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1621         CHECK_VASTATUS(va_status, "vaMapBuffer");
1622         while (buf_list != NULL) {
1623                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1624                 buf_list = (VACodedBufferSegment *) buf_list->next;
1625         }
1626         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1627
1628         {
1629                 // Add video.
1630                 AVPacket pkt;
1631                 memset(&pkt, 0, sizeof(pkt));
1632                 pkt.buf = nullptr;
1633                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1634                 pkt.size = data.size();
1635                 pkt.stream_index = 0;
1636                 if (task.frame_type == FRAME_IDR) {
1637                         pkt.flags = AV_PKT_FLAG_KEY;
1638                 } else {
1639                         pkt.flags = 0;
1640                 }
1641                 pkt.duration = task.duration;
1642                 if (file_mux) {
1643                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1644                 }
1645                 if (!global_flags.uncompressed_video_to_http &&
1646                     !global_flags.x264_video_to_http) {
1647                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1648                 }
1649         }
1650 }
1651
1652
1653 // this is weird. but it seems to put a new frame onto the queue
1654 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1655 {
1656         unique_lock<mutex> lock(storage_task_queue_mutex);
1657         storage_task_queue.push(move(task));
1658         storage_task_queue_changed.notify_all();
1659 }
1660
1661 void QuickSyncEncoderImpl::storage_task_thread()
1662 {
1663         for ( ;; ) {
1664                 storage_task current;
1665                 {
1666                         // wait until there's an encoded frame
1667                         unique_lock<mutex> lock(storage_task_queue_mutex);
1668                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1669                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1670                         current = move(storage_task_queue.front());
1671                         storage_task_queue.pop();
1672                 }
1673
1674                 VAStatus va_status;
1675
1676                 // waits for data, then saves it to disk.
1677                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1678                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1679                 save_codeddata(move(current));
1680
1681                 {
1682                         unique_lock<mutex> lock(storage_task_queue_mutex);
1683                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1684                         storage_task_queue_changed.notify_all();
1685                 }
1686         }
1687 }
1688
1689 void QuickSyncEncoderImpl::release_encode()
1690 {
1691         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1692                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1693                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1694                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1695         }
1696
1697         vaDestroyContext(va_dpy, context_id);
1698         vaDestroyConfig(va_dpy, config_id);
1699 }
1700
1701 void QuickSyncEncoderImpl::release_gl_resources()
1702 {
1703         assert(is_shutdown);
1704         if (has_released_gl_resources) {
1705                 return;
1706         }
1707
1708         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1709                 if (!use_zerocopy) {
1710                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1711                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1712                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1713                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1714                 }
1715                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1716                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1717         }
1718
1719         has_released_gl_resources = true;
1720 }
1721
1722 int QuickSyncEncoderImpl::deinit_va()
1723 {
1724     vaTerminate(va_dpy);
1725
1726     va_close_display(va_dpy);
1727
1728     return 0;
1729 }
1730
1731 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)
1732         : current_storage_frame(0), resource_pool(resource_pool), surface(surface), x264_encoder(x264_encoder), frame_width(width), frame_height(height)
1733 {
1734         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1735         open_output_file(filename);
1736         file_audio_encoder->add_mux(file_mux.get());
1737
1738         frame_width_mbaligned = (frame_width + 15) & (~15);
1739         frame_height_mbaligned = (frame_height + 15) & (~15);
1740
1741         //print_input();
1742
1743         if (global_flags.uncompressed_video_to_http ||
1744             global_flags.x264_video_to_http) {
1745                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1746         }
1747         if (global_flags.x264_video_to_http) {
1748                 assert(x264_encoder != nullptr);
1749         } else {
1750                 assert(x264_encoder == nullptr);
1751         }
1752
1753         init_va(va_display);
1754         setup_encode();
1755
1756         // No frames are ready yet.
1757         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1758
1759         memset(&seq_param, 0, sizeof(seq_param));
1760         memset(&pic_param, 0, sizeof(pic_param));
1761         memset(&slice_param, 0, sizeof(slice_param));
1762
1763         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1764
1765         encode_thread = thread([this]{
1766                 //SDL_GL_MakeCurrent(window, context);
1767                 QOpenGLContext *context = create_context(this->surface);
1768                 eglBindAPI(EGL_OPENGL_API);
1769                 if (!make_current(context, this->surface)) {
1770                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1771                                 eglGetError());
1772                         exit(1);
1773                 }
1774                 encode_thread_func();
1775                 delete_context(context);
1776         });
1777 }
1778
1779 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1780 {
1781         shutdown();
1782         release_gl_resources();
1783 }
1784
1785 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1786 {
1787         assert(!is_shutdown);
1788         {
1789                 // Wait until this frame slot is done encoding.
1790                 unique_lock<mutex> lock(storage_task_queue_mutex);
1791                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1792                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1793                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1794                 }
1795                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1796                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1797                 if (storage_thread_should_quit) return false;
1798         }
1799
1800         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1801         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1802         *y_tex = surf->y_tex;
1803         *cbcr_tex = surf->cbcr_tex;
1804
1805         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1806         CHECK_VASTATUS(va_status, "vaDeriveImage");
1807
1808         if (use_zerocopy) {
1809                 VABufferInfo buf_info;
1810                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1811                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1812                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1813
1814                 // Create Y image.
1815                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1816                 EGLint y_attribs[] = {
1817                         EGL_WIDTH, frame_width,
1818                         EGL_HEIGHT, frame_height,
1819                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1820                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1821                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1822                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1823                         EGL_NONE
1824                 };
1825
1826                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1827                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1828
1829                 // Associate Y image to a texture.
1830                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1831                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1832
1833                 // Create CbCr image.
1834                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1835                 EGLint cbcr_attribs[] = {
1836                         EGL_WIDTH, frame_width,
1837                         EGL_HEIGHT, frame_height,
1838                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1839                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1840                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1841                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1842                         EGL_NONE
1843                 };
1844
1845                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1846                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1847
1848                 // Associate CbCr image to a texture.
1849                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1850                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1851         }
1852
1853         return true;
1854 }
1855
1856 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1857 {
1858         assert(!is_shutdown);
1859         file_audio_encoder->encode_audio(audio, pts + global_delay());
1860 }
1861
1862 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1863 {
1864         assert(!is_shutdown);
1865
1866         if (!use_zerocopy) {
1867                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1868
1869                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1870                 check_error();
1871
1872                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1873                 check_error();
1874
1875                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1876                 check_error();
1877                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1878                 check_error();
1879
1880                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1881                 check_error();
1882                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1883                 check_error();
1884
1885                 glBindTexture(GL_TEXTURE_2D, 0);
1886                 check_error();
1887                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1888                 check_error();
1889
1890                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1891                 check_error();
1892         }
1893
1894         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1895         check_error();
1896         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1897         check_error();
1898
1899         {
1900                 unique_lock<mutex> lock(frame_queue_mutex);
1901                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1902                 ++current_storage_frame;
1903         }
1904         frame_queue_nonempty.notify_all();
1905         return fence;
1906 }
1907
1908 void QuickSyncEncoderImpl::shutdown()
1909 {
1910         if (is_shutdown) {
1911                 return;
1912         }
1913
1914         {
1915                 unique_lock<mutex> lock(frame_queue_mutex);
1916                 encode_thread_should_quit = true;
1917                 frame_queue_nonempty.notify_all();
1918         }
1919         encode_thread.join();
1920         {
1921                 unique_lock<mutex> lock(storage_task_queue_mutex);
1922                 storage_thread_should_quit = true;
1923                 frame_queue_nonempty.notify_all();
1924                 storage_task_queue_changed.notify_all();
1925         }
1926         storage_thread.join();
1927
1928         // Encode any leftover audio in the queues, and also any delayed frames.
1929         file_audio_encoder->encode_last_audio();
1930
1931         release_encode();
1932         deinit_va();
1933         file_mux.reset();
1934         is_shutdown = true;
1935 }
1936
1937 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1938 {
1939         AVFormatContext *avctx = avformat_alloc_context();
1940         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1941         assert(filename.size() < sizeof(avctx->filename) - 1);
1942         strcpy(avctx->filename, filename.c_str());
1943
1944         string url = "file:" + filename;
1945         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1946         if (ret < 0) {
1947                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1948                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1949                 exit(1);
1950         }
1951
1952         string video_extradata = "";  // FIXME: See other comment about global headers.
1953         AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1954         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE));
1955 }
1956
1957 void QuickSyncEncoderImpl::encode_thread_func()
1958 {
1959         int64_t last_dts = -1;
1960         int gop_start_display_frame_num = 0;
1961         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1962                 PendingFrame frame;
1963                 int pts_lag;
1964                 int frame_type, display_frame_num;
1965                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1966                                        &display_frame_num, &frame_type, &pts_lag);
1967                 if (frame_type == FRAME_IDR) {
1968                         numShortTerm = 0;
1969                         current_frame_num = 0;
1970                         gop_start_display_frame_num = display_frame_num;
1971                 }
1972
1973                 {
1974                         unique_lock<mutex> lock(frame_queue_mutex);
1975                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1976                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1977                         });
1978                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1979                                 // We have queued frames that were supposed to be B-frames,
1980                                 // but will be no P-frame to encode them against. Encode them all
1981                                 // as P-frames instead. Note that this happens under the mutex,
1982                                 // but nobody else uses it at this point, since we're shutting down,
1983                                 // so there's no contention.
1984                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1985                                 return;
1986                         } else {
1987                                 frame = move(pending_video_frames[display_frame_num]);
1988                                 pending_video_frames.erase(display_frame_num);
1989                         }
1990                 }
1991
1992                 // Determine the dts of this frame.
1993                 int64_t dts;
1994                 if (pts_lag == -1) {
1995                         assert(last_dts != -1);
1996                         dts = last_dts + (TIMEBASE / MAX_FPS);
1997                 } else {
1998                         dts = frame.pts - pts_lag;
1999                 }
2000                 last_dts = dts;
2001
2002                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2003         }
2004 }
2005
2006 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2007 {
2008         if (pending_video_frames.empty()) {
2009                 return;
2010         }
2011
2012         for (auto &pending_frame : pending_video_frames) {
2013                 int display_frame_num = pending_frame.first;
2014                 assert(display_frame_num > 0);
2015                 PendingFrame frame = move(pending_frame.second);
2016                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2017                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2018                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2019                 last_dts = dts;
2020         }
2021
2022         if (global_flags.uncompressed_video_to_http ||
2023             global_flags.x264_video_to_http) {
2024                 // Add frames left in reorderer.
2025                 while (!reorderer->empty()) {
2026                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2027                         if (global_flags.uncompressed_video_to_http) {
2028                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2029                         } else {
2030                                 assert(global_flags.x264_video_to_http);
2031                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2032                         }
2033                 }
2034         }
2035 }
2036
2037 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2038 {
2039         AVPacket pkt;
2040         memset(&pkt, 0, sizeof(pkt));
2041         pkt.buf = nullptr;
2042         pkt.data = const_cast<uint8_t *>(data);
2043         pkt.size = frame_width * frame_height * 2;
2044         pkt.stream_index = 0;
2045         pkt.flags = AV_PKT_FLAG_KEY;
2046         pkt.duration = duration;
2047         stream_mux->add_packet(pkt, pts, pts);
2048 }
2049
2050 namespace {
2051
2052 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2053 {
2054         if (src_width == dst_pitch) {
2055                 memcpy(dst, src, src_width * height);
2056         } else {
2057                 for (size_t y = 0; y < height; ++y) {
2058                         const uint8_t *sptr = src + y * src_width;
2059                         uint8_t *dptr = dst + y * dst_pitch;
2060                         memcpy(dptr, sptr, src_width);
2061                 }
2062         }
2063 }
2064
2065 }  // namespace
2066
2067 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2068                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2069 {
2070         // Wait for the GPU to be done with the frame.
2071         GLenum sync_status;
2072         do {
2073                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2074                 check_error();
2075         } while (sync_status == GL_TIMEOUT_EXPIRED);
2076         assert(sync_status != GL_WAIT_FAILED);
2077
2078         // Release back any input frames we needed to render this frame.
2079         frame.input_frames.clear();
2080
2081         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2082         VAStatus va_status;
2083
2084         if (use_zerocopy) {
2085                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2086                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2087                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2088                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2089         } else {
2090                 unsigned char *surface_p = nullptr;
2091                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2092
2093                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2094                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2095
2096                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2097                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2098
2099                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2100                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2101
2102                 if (global_flags.uncompressed_video_to_http ||
2103                     global_flags.x264_video_to_http) {
2104                         // Add uncompressed video. (Note that pts == dts here.)
2105                         // Delay needs to match audio.
2106                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2107                         if (output_frame.data != nullptr) {
2108                                 if (global_flags.uncompressed_video_to_http) {
2109                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2110                                 } else {
2111                                         assert(global_flags.x264_video_to_http);
2112                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2113                                 }
2114                         }
2115                 }
2116         }
2117
2118         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2119         CHECK_VASTATUS(va_status, "vaDestroyImage");
2120
2121         // Schedule the frame for encoding.
2122         VASurfaceID va_surface = surf->src_surface;
2123         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2124         CHECK_VASTATUS(va_status, "vaBeginPicture");
2125
2126         if (frame_type == FRAME_IDR) {
2127                 // FIXME: If the mux wants global headers, we should not put the
2128                 // SPS/PPS before each IDR frame, but rather put it into the
2129                 // codec extradata (formatted differently?).
2130                 render_sequence();
2131                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2132                 if (h264_packedheader) {
2133                         render_packedsequence();
2134                         render_packedpicture();
2135                 }
2136         } else {
2137                 //render_sequence();
2138                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2139         }
2140         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2141
2142         va_status = vaEndPicture(va_dpy, context_id);
2143         CHECK_VASTATUS(va_status, "vaEndPicture");
2144
2145         // so now the data is done encoding (well, async job kicked off)...
2146         // we send that to the storage thread
2147         storage_task tmp;
2148         tmp.display_order = display_frame_num;
2149         tmp.frame_type = frame_type;
2150         tmp.pts = pts;
2151         tmp.dts = dts;
2152         tmp.duration = duration;
2153         storage_task_enqueue(move(tmp));
2154
2155         update_ReferenceFrames(frame_type);
2156 }
2157
2158 // Proxy object.
2159 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)
2160         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder)) {}
2161
2162 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2163 QuickSyncEncoder::~QuickSyncEncoder() {}
2164
2165 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2166 {
2167         impl->add_audio(pts, audio);
2168 }
2169
2170 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2171 {
2172         return impl->begin_frame(y_tex, cbcr_tex);
2173 }
2174
2175 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2176 {
2177         return impl->end_frame(pts, duration, input_frames);
2178 }
2179
2180 void QuickSyncEncoder::shutdown()
2181 {
2182         impl->shutdown();
2183 }
2184
2185 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2186 {
2187         impl->set_stream_mux(mux);
2188 }
2189
2190 int64_t QuickSyncEncoder::global_delay() const {
2191         return impl->global_delay();
2192 }