git.sesse.net Git - nageru/blob - h264encode.cpp

   1 //#include "sysdeps.h"
   2 #include "h264encode.h"
   3
   4 #include <movit/util.h>
   5 #include <EGL/eglplatform.h>
   6 #include <X11/X.h>
   7 #include <X11/Xlib.h>
   8 #include <assert.h>
   9 #include <epoxy/egl.h>
  10 extern "C" {
  11 #include <libavcodec/avcodec.h>
  12 #include <libavformat/avformat.h>
  13 #include <libavutil/channel_layout.h>
  14 #include <libavutil/frame.h>
  15 #include <libavutil/rational.h>
  16 #include <libavutil/samplefmt.h>
  17 }
  18 #include <libdrm/drm_fourcc.h>
  19 #include <stdio.h>
  20 #include <stdlib.h>
  21 #include <string.h>
  22 #include <fcntl.h>
  23 #include <va/va.h>
  24 #include <va/va_drm.h>
  25 #include <va/va_drmcommon.h>
  26 #include <va/va_enc_h264.h>
  27 #include <va/va_x11.h>
  28 #include <algorithm>
  29 #include <condition_variable>
  30 #include <cstdint>
  31 #include <map>
  32 #include <memory>
  33 #include <mutex>
  34 #include <queue>
  35 #include <string>
  36 #include <thread>
  37 #include <utility>
  38
  39 #include "context.h"
  40 #include "defs.h"
  41 #include "flags.h"
  42 #include "httpd.h"
  43 #include "timebase.h"
  44 #include "x264encode.h"
  45
  46 using namespace std;
  47
  48 class QOpenGLContext;
  49 class QSurface;
  50
  51 #define CHECK_VASTATUS(va_status, func)                                 \
  52     if (va_status != VA_STATUS_SUCCESS) {                               \
  53         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  54         exit(1);                                                        \
  55     }
  56
  57 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  58
  59 //#include "loadsurface.h"
  60
  61 #define NAL_REF_IDC_NONE        0
  62 #define NAL_REF_IDC_LOW         1
  63 #define NAL_REF_IDC_MEDIUM      2
  64 #define NAL_REF_IDC_HIGH        3
  65
  66 #define NAL_NON_IDR             1
  67 #define NAL_IDR                 5
  68 #define NAL_SPS                 7
  69 #define NAL_PPS                 8
  70 #define NAL_SEI                 6
  71
  72 #define SLICE_TYPE_P            0
  73 #define SLICE_TYPE_B            1
  74 #define SLICE_TYPE_I            2
  75 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  76 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  77 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  78
  79
  80 #define ENTROPY_MODE_CAVLC      0
  81 #define ENTROPY_MODE_CABAC      1
  82
  83 #define PROFILE_IDC_BASELINE    66
  84 #define PROFILE_IDC_MAIN        77
  85 #define PROFILE_IDC_HIGH        100
  86
  87 #define BITSTREAM_ALLOCATE_STEPPING     4096
  88 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  89 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  90 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  91
  92 static constexpr unsigned int MaxFrameNum = (2<<16);
  93 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  94 static constexpr unsigned int Log2MaxFrameNum = 16;
  95 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  96 static constexpr int rc_default_modes[] = {  // Priority list of modes.
  97     VA_RC_VBR,
  98     VA_RC_CQP,
  99     VA_RC_VBR_CONSTRAINED,
 100     VA_RC_CBR,
 101     VA_RC_VCM,
 102     VA_RC_NONE,
 103 };
 104
 105 /* thread to save coded data */
 106 #define SRC_SURFACE_FREE        0
 107 #define SRC_SURFACE_IN_ENCODING 1
 108
 109 struct __bitstream {
 110     unsigned int *buffer;
 111     int bit_offset;
 112     int max_size_in_dword;
 113 };
 114 typedef struct __bitstream bitstream;
 115
 116 using namespace std;
 117
 118 // H.264 video comes out in encoding order (e.g. with two B-frames:
 119 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 120 // come in the right order. Since we do everything, including waiting
 121 // for the frames to come out of OpenGL, in encoding order, we need
 122 // a reordering buffer for uncompressed frames so that they come out
 123 // correctly. We go the super-lazy way of not making it understand
 124 // anything about the true order (which introduces some extra latency,
 125 // though); we know that for N B-frames we need at most (N-1) frames
 126 // in the reorder buffer, and can just sort on that.
 127 //
 128 // The class also deals with keeping a freelist as needed.
 129 class FrameReorderer {
 130 public:
 131         FrameReorderer(unsigned queue_length, int width, int height);
 132
 133         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 134         // Does _not_ take ownership of data; a copy is taken if needed.
 135         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 136         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 137         pair<int64_t, const uint8_t *> reorder_frame(int64_t pts, const uint8_t *data);
 138
 139         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 140         pair<int64_t, const uint8_t *> get_first_frame();
 141
 142         bool empty() const { return frames.empty(); }
 143
 144 private:
 145         unsigned queue_length;
 146         int width, height;
 147
 148         priority_queue<pair<int64_t, uint8_t *>> frames;
 149         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 150
 151         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 152         // except priority_queue doesn't work well with movable-only types.
 153         vector<unique_ptr<uint8_t[]>> owner;
 154 };
 155
 156 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 157     : queue_length(queue_length), width(width), height(height)
 158 {
 159         for (unsigned i = 0; i < queue_length; ++i) {
 160                 owner.emplace_back(new uint8_t[width * height * 2]);
 161                 freelist.push(owner.back().get());
 162         }
 163 }
 164
 165 pair<int64_t, const uint8_t *> FrameReorderer::reorder_frame(int64_t pts, const uint8_t *data)
 166 {
 167         if (queue_length == 0) {
 168                 return make_pair(pts, data);
 169         }
 170
 171         assert(!freelist.empty());
 172         uint8_t *storage = freelist.top();
 173         freelist.pop();
 174         memcpy(storage, data, width * height * 2);
 175         frames.emplace(-pts, storage);  // Invert pts to get smallest first.
 176
 177         if (frames.size() >= queue_length) {
 178                 return get_first_frame();
 179         } else {
 180                 return make_pair(-1, nullptr);
 181         }
 182 }
 183
 184 pair<int64_t, const uint8_t *> FrameReorderer::get_first_frame()
 185 {
 186         assert(!frames.empty());
 187         pair<int64_t, uint8_t *> storage = frames.top();
 188         frames.pop();
 189         int64_t pts = storage.first;
 190         freelist.push(storage.second);
 191         return make_pair(-pts, storage.second);  // Re-invert pts (see reorder_frame()).
 192 }
 193
 194 class H264EncoderImpl {
 195 public:
 196         H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
 197         ~H264EncoderImpl();
 198         void add_audio(int64_t pts, vector<float> audio);
 199         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 200         RefCountedGLsync end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames);
 201         void shutdown();
 202         void open_output_file(const std::string &filename);
 203         void close_output_file();
 204
 205 private:
 206         struct storage_task {
 207                 unsigned long long display_order;
 208                 int frame_type;
 209                 vector<float> audio;
 210                 int64_t pts, dts;
 211         };
 212         struct PendingFrame {
 213                 RefCountedGLsync fence;
 214                 vector<RefCountedFrame> input_frames;
 215                 int64_t pts;
 216         };
 217
 218         // So we never get negative dts.
 219         int64_t global_delay() const {
 220                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 221         }
 222
 223         void encode_thread_func();
 224         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 225         void add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data);
 226         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 227                           int frame_type, int64_t pts, int64_t dts);
 228         void storage_task_thread();
 229         void encode_audio(const vector<float> &audio,
 230                           vector<float> *audio_queue,
 231                           int64_t audio_pts,
 232                           AVCodecContext *ctx,
 233                           const vector<PacketDestination *> &destinations);
 234         void encode_audio_one_frame(const float *audio,
 235                                     size_t num_samples,  // In each channel.
 236                                     int64_t audio_pts,
 237                                     AVCodecContext *ctx,
 238                                     const vector<PacketDestination *> &destinations);
 239         void storage_task_enqueue(storage_task task);
 240         void save_codeddata(storage_task task);
 241         int render_packedsequence();
 242         int render_packedpicture();
 243         void render_packedslice();
 244         int render_sequence();
 245         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 246         void sps_rbsp(bitstream *bs);
 247         void pps_rbsp(bitstream *bs);
 248         int build_packed_pic_buffer(unsigned char **header_buffer);
 249         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 250         void slice_header(bitstream *bs);
 251         int build_packed_seq_buffer(unsigned char **header_buffer);
 252         int build_packed_slice_buffer(unsigned char **header_buffer);
 253         int init_va(const string &va_display);
 254         int deinit_va();
 255         void enable_zerocopy_if_possible();
 256         VADisplay va_open_display(const string &va_display);
 257         void va_close_display(VADisplay va_dpy);
 258         int setup_encode();
 259         int release_encode();
 260         void update_ReferenceFrames(int frame_type);
 261         int update_RefPicList(int frame_type);
 262
 263         bool is_shutdown = false;
 264         bool use_zerocopy;
 265         int drm_fd = -1;
 266
 267         thread encode_thread, storage_thread;
 268
 269         mutex storage_task_queue_mutex;
 270         condition_variable storage_task_queue_changed;
 271         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 272         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 273         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 274
 275         mutex frame_queue_mutex;
 276         condition_variable frame_queue_nonempty;
 277         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 278
 279         int current_storage_frame;
 280
 281         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 282         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 283         QSurface *surface;
 284
 285         AVCodecContext *context_audio_file;
 286         AVCodecContext *context_audio_stream = nullptr;  // nullptr = don't code separate audio for stream.
 287
 288         vector<float> audio_queue_file;
 289         vector<float> audio_queue_stream;
 290
 291         AVFrame *audio_frame = nullptr;
 292         HTTPD *httpd;
 293         unique_ptr<FrameReorderer> reorderer;
 294         unique_ptr<X264Encoder> x264_encoder;  // nullptr if not using x264.
 295
 296         Display *x11_display = nullptr;
 297
 298         // Encoder parameters
 299         VADisplay va_dpy;
 300         VAProfile h264_profile = (VAProfile)~0;
 301         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 302         int config_attrib_num = 0, enc_packed_header_idx;
 303
 304         struct GLSurface {
 305                 VASurfaceID src_surface, ref_surface;
 306                 VABufferID coded_buf;
 307
 308                 VAImage surface_image;
 309                 GLuint y_tex, cbcr_tex;
 310
 311                 // Only if use_zerocopy == true.
 312                 EGLImage y_egl_image, cbcr_egl_image;
 313
 314                 // Only if use_zerocopy == false.
 315                 GLuint pbo;
 316                 uint8_t *y_ptr, *cbcr_ptr;
 317                 size_t y_offset, cbcr_offset;
 318         };
 319         GLSurface gl_surfaces[SURFACE_NUM];
 320
 321         VAConfigID config_id;
 322         VAContextID context_id;
 323         VAEncSequenceParameterBufferH264 seq_param;
 324         VAEncPictureParameterBufferH264 pic_param;
 325         VAEncSliceParameterBufferH264 slice_param;
 326         VAPictureH264 CurrentCurrPic;
 327         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 328
 329         // Static quality settings.
 330         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 331         static constexpr unsigned int num_ref_frames = 2;
 332         static constexpr int initial_qp = 15;
 333         static constexpr int minimal_qp = 0;
 334         static constexpr int intra_period = 30;
 335         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 336
 337         // Quality settings that are meant to be static, but might be overridden
 338         // by the profile.
 339         int constraint_set_flag = 0;
 340         int h264_packedheader = 0; /* support pack header? */
 341         int h264_maxref = (1<<16|1);
 342         int h264_entropy_mode = 1; /* cabac */
 343         int ip_period = 3;
 344
 345         int rc_mode = -1;
 346         unsigned int current_frame_num = 0;
 347         unsigned int numShortTerm = 0;
 348
 349         int frame_width;
 350         int frame_height;
 351         int frame_width_mbaligned;
 352         int frame_height_mbaligned;
 353
 354         unique_ptr<Mux> file_mux;  // To local disk.
 355 };
 356
 357 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 358 // but if we don't delete it here, we get leaks. The GStreamer implementation
 359 // does the same.
 360 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 361 {
 362     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 363     CHECK_VASTATUS(va_status, "vaRenderPicture");
 364
 365     for (int i = 0; i < num_buffers; ++i) {
 366         va_status = vaDestroyBuffer(dpy, buffers[i]);
 367         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 368     }
 369 }
 370
 371 static unsigned int
 372 va_swap32(unsigned int val)
 373 {
 374     unsigned char *pval = (unsigned char *)&val;
 375
 376     return ((pval[0] << 24)     |
 377             (pval[1] << 16)     |
 378             (pval[2] << 8)      |
 379             (pval[3] << 0));
 380 }
 381
 382 static void
 383 bitstream_start(bitstream *bs)
 384 {
 385     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 386     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 387     bs->bit_offset = 0;
 388 }
 389
 390 static void
 391 bitstream_end(bitstream *bs)
 392 {
 393     int pos = (bs->bit_offset >> 5);
 394     int bit_offset = (bs->bit_offset & 0x1f);
 395     int bit_left = 32 - bit_offset;
 396
 397     if (bit_offset) {
 398         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 399     }
 400 }
 401
 402 static void
 403 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 404 {
 405     int pos = (bs->bit_offset >> 5);
 406     int bit_offset = (bs->bit_offset & 0x1f);
 407     int bit_left = 32 - bit_offset;
 408
 409     if (!size_in_bits)
 410         return;
 411
 412     bs->bit_offset += size_in_bits;
 413
 414     if (bit_left > size_in_bits) {
 415         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 416     } else {
 417         size_in_bits -= bit_left;
 418         if (bit_left >= 32) {
 419             bs->buffer[pos] = (val >> size_in_bits);
 420         } else {
 421             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 422         }
 423         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 424
 425         if (pos + 1 == bs->max_size_in_dword) {
 426             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 427             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 428         }
 429
 430         bs->buffer[pos + 1] = val;
 431     }
 432 }
 433
 434 static void
 435 bitstream_put_ue(bitstream *bs, unsigned int val)
 436 {
 437     int size_in_bits = 0;
 438     int tmp_val = ++val;
 439
 440     while (tmp_val) {
 441         tmp_val >>= 1;
 442         size_in_bits++;
 443     }
 444
 445     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 446     bitstream_put_ui(bs, val, size_in_bits);
 447 }
 448
 449 static void
 450 bitstream_put_se(bitstream *bs, int val)
 451 {
 452     unsigned int new_val;
 453
 454     if (val <= 0)
 455         new_val = -2 * val;
 456     else
 457         new_val = 2 * val - 1;
 458
 459     bitstream_put_ue(bs, new_val);
 460 }
 461
 462 static void
 463 bitstream_byte_aligning(bitstream *bs, int bit)
 464 {
 465     int bit_offset = (bs->bit_offset & 0x7);
 466     int bit_left = 8 - bit_offset;
 467     int new_val;
 468
 469     if (!bit_offset)
 470         return;
 471
 472     assert(bit == 0 || bit == 1);
 473
 474     if (bit)
 475         new_val = (1 << bit_left) - 1;
 476     else
 477         new_val = 0;
 478
 479     bitstream_put_ui(bs, new_val, bit_left);
 480 }
 481
 482 static void
 483 rbsp_trailing_bits(bitstream *bs)
 484 {
 485     bitstream_put_ui(bs, 1, 1);
 486     bitstream_byte_aligning(bs, 0);
 487 }
 488
 489 static void nal_start_code_prefix(bitstream *bs)
 490 {
 491     bitstream_put_ui(bs, 0x00000001, 32);
 492 }
 493
 494 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 495 {
 496     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 497     bitstream_put_ui(bs, nal_ref_idc, 2);
 498     bitstream_put_ui(bs, nal_unit_type, 5);
 499 }
 500
 501 void H264EncoderImpl::sps_rbsp(bitstream *bs)
 502 {
 503     int profile_idc = PROFILE_IDC_BASELINE;
 504
 505     if (h264_profile  == VAProfileH264High)
 506         profile_idc = PROFILE_IDC_HIGH;
 507     else if (h264_profile  == VAProfileH264Main)
 508         profile_idc = PROFILE_IDC_MAIN;
 509
 510     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 511     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 512     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 513     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 514     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 515     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 516     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 517     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 518
 519     if ( profile_idc == PROFILE_IDC_HIGH) {
 520         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 521         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 522         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 523         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 524         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 525     }
 526
 527     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 528     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 529
 530     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 531         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 532     else {
 533         assert(0);
 534     }
 535
 536     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 537     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 538
 539     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 540     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 541     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 542
 543     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 544         assert(0);
 545     }
 546
 547     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 548     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 549
 550     if (seq_param.frame_cropping_flag) {
 551         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 552         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 553         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 554         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 555     }
 556
 557     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 558     if ( false ) {
 559         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 560     } else {
 561         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 562         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 563         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 564         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 565         {
 566             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 567             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 568             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 569             {
 570                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 571                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 572                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 573             }
 574         }
 575         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 576         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 577         {
 578             bitstream_put_ui(bs, 1, 32);  // FPS
 579             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 580             bitstream_put_ui(bs, 1, 1);
 581         }
 582         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 583         {
 584             // hrd_parameters
 585             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 586             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 587             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 588
 589             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 590             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 591             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 592
 593             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 594             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 595             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 596             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 597         }
 598         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 599         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 600
 601         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 602         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 603     }
 604
 605     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 606 }
 607
 608
 609 void H264EncoderImpl::pps_rbsp(bitstream *bs)
 610 {
 611     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 612     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 613
 614     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 615
 616     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 617
 618     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 619
 620     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 621     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 622
 623     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 624     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 625
 626     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 627     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 628     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 629
 630     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 631     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 632     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 633
 634     /* more_rbsp_data */
 635     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 636     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 637     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 638
 639     rbsp_trailing_bits(bs);
 640 }
 641
 642 void H264EncoderImpl::slice_header(bitstream *bs)
 643 {
 644     int first_mb_in_slice = slice_param.macroblock_address;
 645
 646     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 647     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 648     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 649     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 650
 651     /* frame_mbs_only_flag == 1 */
 652     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 653         /* FIXME: */
 654         assert(0);
 655     }
 656
 657     if (pic_param.pic_fields.bits.idr_pic_flag)
 658         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 659
 660     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 661         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 662         /* pic_order_present_flag == 0 */
 663     } else {
 664         /* FIXME: */
 665         assert(0);
 666     }
 667
 668     /* redundant_pic_cnt_present_flag == 0 */
 669     /* slice type */
 670     if (IS_P_SLICE(slice_param.slice_type)) {
 671         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 672
 673         if (slice_param.num_ref_idx_active_override_flag)
 674             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 675
 676         /* ref_pic_list_reordering */
 677         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 678     } else if (IS_B_SLICE(slice_param.slice_type)) {
 679         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 680
 681         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 682
 683         if (slice_param.num_ref_idx_active_override_flag) {
 684             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 685             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 686         }
 687
 688         /* ref_pic_list_reordering */
 689         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 690         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 691     }
 692
 693     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 694          IS_P_SLICE(slice_param.slice_type)) ||
 695         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 696          IS_B_SLICE(slice_param.slice_type))) {
 697         /* FIXME: fill weight/offset table */
 698         assert(0);
 699     }
 700
 701     /* dec_ref_pic_marking */
 702     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 703         unsigned char no_output_of_prior_pics_flag = 0;
 704         unsigned char long_term_reference_flag = 0;
 705         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 706
 707         if (pic_param.pic_fields.bits.idr_pic_flag) {
 708             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 709             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 710         } else {
 711             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 712         }
 713     }
 714
 715     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 716         !IS_I_SLICE(slice_param.slice_type))
 717         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 718
 719     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 720
 721     /* ignore for SP/SI */
 722
 723     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 724         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 725
 726         if (slice_param.disable_deblocking_filter_idc != 1) {
 727             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 728             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 729         }
 730     }
 731
 732     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 733         bitstream_byte_aligning(bs, 1);
 734     }
 735 }
 736
 737 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 738 {
 739     bitstream bs;
 740
 741     bitstream_start(&bs);
 742     nal_start_code_prefix(&bs);
 743     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 744     pps_rbsp(&bs);
 745     bitstream_end(&bs);
 746
 747     *header_buffer = (unsigned char *)bs.buffer;
 748     return bs.bit_offset;
 749 }
 750
 751 int
 752 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 753 {
 754     bitstream bs;
 755
 756     bitstream_start(&bs);
 757     nal_start_code_prefix(&bs);
 758     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 759     sps_rbsp(&bs);
 760     bitstream_end(&bs);
 761
 762     *header_buffer = (unsigned char *)bs.buffer;
 763     return bs.bit_offset;
 764 }
 765
 766 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 767 {
 768     bitstream bs;
 769     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 770     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 771
 772     bitstream_start(&bs);
 773     nal_start_code_prefix(&bs);
 774
 775     if (IS_I_SLICE(slice_param.slice_type)) {
 776         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 777     } else if (IS_P_SLICE(slice_param.slice_type)) {
 778         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 779     } else {
 780         assert(IS_B_SLICE(slice_param.slice_type));
 781         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 782     }
 783
 784     slice_header(&bs);
 785     bitstream_end(&bs);
 786
 787     *header_buffer = (unsigned char *)bs.buffer;
 788     return bs.bit_offset;
 789 }
 790
 791
 792 /*
 793   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 794   1) period between Frame #X and Frame #N = #X - #N
 795   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 796   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 797   4) intra_period and intra_idr_period take precedence over ip_period
 798   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 799      of I/IDR frames, see bellow examples
 800   -------------------------------------------------------------------
 801   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 802   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 803   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 804   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 805   1            1                ignored    IDR IDR IDR IDR...
 806   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 807   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 808   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 809                                               (PBB)(IBB)(PBB)(IBB)...
 810   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 811                                            IDRPPPPP IPPPPP IPPPPP...
 812   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 813                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 814                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 815                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 816                                            {IDR(PBB)(PBB)}                     (6/6/3)
 817                                            {IDR(PBB)(PBB)}.
 818 */
 819
 820 // General pts/dts strategy:
 821 //
 822 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 823 // bit tricky. We assume first of all that the frame rate never goes _above_
 824 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 825 // in at least this speed, as long at dts <= pts (the frame is not attempted
 826 // presented before it is decoded). Furthermore, we never have longer chains of
 827 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 828 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 829 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 830 // frame, with an order O <= C.)
 831 //
 832 // Many strategies are possible, but we establish these rules:
 833 //
 834 //  - Tip frames have dts = pts - (C-O)*N.
 835 //  - Non-tip frames have dts = dts_last + N.
 836 //
 837 // An example, with C=2 and N=10 and the data flow showed with arrows:
 838 //
 839 //        I  B  P  B  B  P
 840 //   pts: 30 40 50 60 70 80
 841 //        ↓  ↓     ↓
 842 //   dts: 10 30 20 60 50←40
 843 //         |  |  ↑        ↑
 844 //         `--|--'        |
 845 //             `----------'
 846 //
 847 // To show that this works fine also with irregular spacings, let's say that
 848 // the third frame is delayed a bit (something earlier was dropped). Now the
 849 // situation looks like this:
 850 //
 851 //        I  B  P  B  B   P
 852 //   pts: 30 40 80 90 100 110
 853 //        ↓  ↓     ↓
 854 //   dts: 10 30 20 90 50←40
 855 //         |  |  ↑        ↑
 856 //         `--|--'        |
 857 //             `----------'
 858 //
 859 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 860 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 861 //
 862 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 863 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 864 // a tip frame and should be given a dts based on the previous one.
 865 #define FRAME_P 0
 866 #define FRAME_B 1
 867 #define FRAME_I 2
 868 #define FRAME_IDR 7
 869 void encoding2display_order(
 870     int encoding_order, int intra_period,
 871     int intra_idr_period, int ip_period,
 872     int *displaying_order,
 873     int *frame_type, int *pts_lag)
 874 {
 875     int encoding_order_gop = 0;
 876
 877     *pts_lag = 0;
 878
 879     if (intra_period == 1) { /* all are I/IDR frames */
 880         *displaying_order = encoding_order;
 881         if (intra_idr_period == 0)
 882             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 883         else
 884             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 885         return;
 886     }
 887
 888     if (intra_period == 0)
 889         intra_idr_period = 0;
 890
 891     if (ip_period == 1) {
 892         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 893         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 894         *displaying_order = encoding_order;
 895
 896         if (encoding_order_gop == 0) { /* the first frame */
 897             *frame_type = FRAME_IDR;
 898         } else if (intra_period != 0 && /* have I frames */
 899                    encoding_order_gop >= 2 &&
 900                    (encoding_order_gop % intra_period == 0)) {
 901             *frame_type = FRAME_I;
 902         } else {
 903             *frame_type = FRAME_P;
 904         }
 905         return;
 906     }
 907
 908     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 909     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 910     *pts_lag = -1;  // Most frames are not tip frames.
 911
 912     if (encoding_order_gop == 0) { /* the first frame */
 913         *frame_type = FRAME_IDR;
 914         *displaying_order = encoding_order;
 915         // IDR frames are a special case; I honestly can't find the logic behind
 916         // why this is the right thing, but it seems to line up nicely in practice :-)
 917         *pts_lag = TIMEBASE / MAX_FPS;
 918     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 919         *frame_type = FRAME_B;
 920         *displaying_order = encoding_order - 1;
 921         if ((encoding_order_gop % ip_period) == 0) {
 922             *pts_lag = 0;  // Last B-frame.
 923         }
 924     } else if (intra_period != 0 && /* have I frames */
 925                encoding_order_gop >= 2 &&
 926                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 927         *frame_type = FRAME_I;
 928         *displaying_order = encoding_order + ip_period - 1;
 929     } else {
 930         *frame_type = FRAME_P;
 931         *displaying_order = encoding_order + ip_period - 1;
 932     }
 933 }
 934
 935
 936 static const char *rc_to_string(int rc_mode)
 937 {
 938     switch (rc_mode) {
 939     case VA_RC_NONE:
 940         return "NONE";
 941     case VA_RC_CBR:
 942         return "CBR";
 943     case VA_RC_VBR:
 944         return "VBR";
 945     case VA_RC_VCM:
 946         return "VCM";
 947     case VA_RC_CQP:
 948         return "CQP";
 949     case VA_RC_VBR_CONSTRAINED:
 950         return "VBR_CONSTRAINED";
 951     default:
 952         return "Unknown";
 953     }
 954 }
 955
 956 void H264EncoderImpl::enable_zerocopy_if_possible()
 957 {
 958         if (global_flags.uncompressed_video_to_http) {
 959                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n");
 960                 use_zerocopy = false;
 961         } else if (global_flags.x264_video_to_http) {
 962                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --x264_video_to_http.\n");
 963                 use_zerocopy = false;
 964         } else {
 965                 use_zerocopy = true;
 966         }
 967 }
 968
 969 VADisplay H264EncoderImpl::va_open_display(const string &va_display)
 970 {
 971         if (va_display.empty()) {
 972                 x11_display = XOpenDisplay(NULL);
 973                 if (!x11_display) {
 974                         fprintf(stderr, "error: can't connect to X server!\n");
 975                         return NULL;
 976                 }
 977                 enable_zerocopy_if_possible();
 978                 return vaGetDisplay(x11_display);
 979         } else if (va_display[0] != '/') {
 980                 x11_display = XOpenDisplay(va_display.c_str());
 981                 if (!x11_display) {
 982                         fprintf(stderr, "error: can't connect to X server!\n");
 983                         return NULL;
 984                 }
 985                 enable_zerocopy_if_possible();
 986                 return vaGetDisplay(x11_display);
 987         } else {
 988                 drm_fd = open(va_display.c_str(), O_RDWR);
 989                 if (drm_fd == -1) {
 990                         perror(va_display.c_str());
 991                         return NULL;
 992                 }
 993                 use_zerocopy = false;
 994                 return vaGetDisplayDRM(drm_fd);
 995         }
 996 }
 997
 998 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
 999 {
1000         if (x11_display) {
1001                 XCloseDisplay(x11_display);
1002                 x11_display = nullptr;
1003         }
1004         if (drm_fd != -1) {
1005                 close(drm_fd);
1006         }
1007 }
1008
1009 int H264EncoderImpl::init_va(const string &va_display)
1010 {
1011     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1012     VAEntrypoint *entrypoints;
1013     int num_entrypoints, slice_entrypoint;
1014     int support_encode = 0;
1015     int major_ver, minor_ver;
1016     VAStatus va_status;
1017     unsigned int i;
1018
1019     va_dpy = va_open_display(va_display);
1020     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1021     CHECK_VASTATUS(va_status, "vaInitialize");
1022
1023     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1024     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1025     if (!entrypoints) {
1026         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1027         exit(1);
1028     }
1029
1030     /* use the highest profile */
1031     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1032         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1033             continue;
1034
1035         h264_profile = profile_list[i];
1036         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1037         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1038             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1039                 support_encode = 1;
1040                 break;
1041             }
1042         }
1043         if (support_encode == 1)
1044             break;
1045     }
1046
1047     if (support_encode == 0) {
1048         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1049         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1050         printf("to use VA-API against DRM instead of X11.\n");
1051         exit(1);
1052     } else {
1053         switch (h264_profile) {
1054             case VAProfileH264Baseline:
1055                 ip_period = 1;
1056                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1057                 h264_entropy_mode = 0;
1058                 break;
1059             case VAProfileH264ConstrainedBaseline:
1060                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1061                 ip_period = 1;
1062                 break;
1063
1064             case VAProfileH264Main:
1065                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1066                 break;
1067
1068             case VAProfileH264High:
1069                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1070                 break;
1071             default:
1072                 h264_profile = VAProfileH264Baseline;
1073                 ip_period = 1;
1074                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1075                 break;
1076         }
1077     }
1078
1079     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1080
1081     /* find out the format for the render target, and rate control mode */
1082     for (i = 0; i < VAConfigAttribTypeMax; i++)
1083         attrib[i].type = (VAConfigAttribType)i;
1084
1085     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1086                                       &attrib[0], VAConfigAttribTypeMax);
1087     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1088     /* check the interested configattrib */
1089     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1090         printf("Not find desired YUV420 RT format\n");
1091         exit(1);
1092     } else {
1093         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1094         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1095         config_attrib_num++;
1096     }
1097
1098     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1099         int tmp = attrib[VAConfigAttribRateControl].value;
1100
1101         if (rc_mode == -1 || !(rc_mode & tmp))  {
1102             if (rc_mode != -1) {
1103                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1104             }
1105
1106             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1107                 if (rc_default_modes[i] & tmp) {
1108                     rc_mode = rc_default_modes[i];
1109                     break;
1110                 }
1111             }
1112         }
1113
1114         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1115         config_attrib[config_attrib_num].value = rc_mode;
1116         config_attrib_num++;
1117     }
1118
1119
1120     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1121         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1122
1123         h264_packedheader = 1;
1124         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1125         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1126
1127         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1128             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1129         }
1130
1131         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1132             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1133         }
1134
1135         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1136             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1137         }
1138
1139         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1140             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1141         }
1142
1143         enc_packed_header_idx = config_attrib_num;
1144         config_attrib_num++;
1145     }
1146
1147     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1148         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1149         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1150         config_attrib_num++;
1151     }
1152
1153     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1154         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1155     }
1156
1157     free(entrypoints);
1158     return 0;
1159 }
1160
1161 int H264EncoderImpl::setup_encode()
1162 {
1163     VAStatus va_status;
1164     VASurfaceID *tmp_surfaceid;
1165     int codedbuf_size, i;
1166     static VASurfaceID src_surface[SURFACE_NUM];
1167     static VASurfaceID ref_surface[SURFACE_NUM];
1168
1169     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1170             &config_attrib[0], config_attrib_num, &config_id);
1171     CHECK_VASTATUS(va_status, "vaCreateConfig");
1172
1173     /* create source surfaces */
1174     va_status = vaCreateSurfaces(va_dpy,
1175                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1176                                  &src_surface[0], SURFACE_NUM,
1177                                  NULL, 0);
1178     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1179
1180     /* create reference surfaces */
1181     va_status = vaCreateSurfaces(va_dpy,
1182                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1183                                  &ref_surface[0], SURFACE_NUM,
1184                                  NULL, 0);
1185     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1186
1187     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1188     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1189     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1190
1191     /* Create a context for this encode pipe */
1192     va_status = vaCreateContext(va_dpy, config_id,
1193                                 frame_width_mbaligned, frame_height_mbaligned,
1194                                 VA_PROGRESSIVE,
1195                                 tmp_surfaceid, 2 * SURFACE_NUM,
1196                                 &context_id);
1197     CHECK_VASTATUS(va_status, "vaCreateContext");
1198     free(tmp_surfaceid);
1199
1200     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1201
1202     for (i = 0; i < SURFACE_NUM; i++) {
1203         /* create coded buffer once for all
1204          * other VA buffers which won't be used again after vaRenderPicture.
1205          * so APP can always vaCreateBuffer for every frame
1206          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1207          * so VA won't maintain the coded buffer
1208          */
1209         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1210                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1211         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1212     }
1213
1214     /* create OpenGL objects */
1215     //glGenFramebuffers(SURFACE_NUM, fbos);
1216
1217     for (i = 0; i < SURFACE_NUM; i++) {
1218         glGenTextures(1, &gl_surfaces[i].y_tex);
1219         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1220
1221         if (!use_zerocopy) {
1222             // Create Y image.
1223             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1224             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1225
1226             // Create CbCr image.
1227             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1228             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1229
1230             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1231             // buffers, due to potentially differing pitch.
1232             glGenBuffers(1, &gl_surfaces[i].pbo);
1233             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1234             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1235             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1236             gl_surfaces[i].y_offset = 0;
1237             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1238             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1239             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1240             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1241         }
1242     }
1243
1244     for (i = 0; i < SURFACE_NUM; i++) {
1245         gl_surfaces[i].src_surface = src_surface[i];
1246         gl_surfaces[i].ref_surface = ref_surface[i];
1247     }
1248
1249     return 0;
1250 }
1251
1252 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1253 //
1254 //   2 1 0 [3] 4 8 9
1255 template<class T, class C>
1256 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1257 {
1258         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1259         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1260         sort(middle, end, less_than);
1261 }
1262
1263 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1264 {
1265     int i;
1266
1267     if (frame_type == FRAME_B)
1268         return;
1269
1270     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1271     numShortTerm++;
1272     if (numShortTerm > num_ref_frames)
1273         numShortTerm = num_ref_frames;
1274     for (i=numShortTerm-1; i>0; i--)
1275         ReferenceFrames[i] = ReferenceFrames[i-1];
1276     ReferenceFrames[0] = CurrentCurrPic;
1277
1278     current_frame_num++;
1279     if (current_frame_num > MaxFrameNum)
1280         current_frame_num = 0;
1281 }
1282
1283
1284 int H264EncoderImpl::update_RefPicList(int frame_type)
1285 {
1286     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1287         return a.frame_idx > b.frame_idx;
1288     };
1289     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1290         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1291     };
1292     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1293         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1294     };
1295
1296     if (frame_type == FRAME_P) {
1297         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1298         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1299     } else if (frame_type == FRAME_B) {
1300         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1301         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1302
1303         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1304         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1305     }
1306
1307     return 0;
1308 }
1309
1310
1311 int H264EncoderImpl::render_sequence()
1312 {
1313     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1314     VAStatus va_status;
1315     VAEncMiscParameterBuffer *misc_param;
1316     VAEncMiscParameterRateControl *misc_rate_ctrl;
1317
1318     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1319     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1320     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1321     seq_param.bits_per_second = frame_bitrate;
1322
1323     seq_param.intra_period = intra_period;
1324     seq_param.intra_idr_period = intra_idr_period;
1325     seq_param.ip_period = ip_period;
1326
1327     seq_param.max_num_ref_frames = num_ref_frames;
1328     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1329     seq_param.time_scale = TIMEBASE * 2;
1330     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1331     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1332     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1333     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1334     seq_param.seq_fields.bits.chroma_format_idc = 1;
1335     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1336
1337     if (frame_width != frame_width_mbaligned ||
1338         frame_height != frame_height_mbaligned) {
1339         seq_param.frame_cropping_flag = 1;
1340         seq_param.frame_crop_left_offset = 0;
1341         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1342         seq_param.frame_crop_top_offset = 0;
1343         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1344     }
1345
1346     va_status = vaCreateBuffer(va_dpy, context_id,
1347                                VAEncSequenceParameterBufferType,
1348                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1349     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1350
1351     va_status = vaCreateBuffer(va_dpy, context_id,
1352                                VAEncMiscParameterBufferType,
1353                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1354                                1, NULL, &rc_param_buf);
1355     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1356
1357     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1358     misc_param->type = VAEncMiscParameterTypeRateControl;
1359     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1360     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1361     misc_rate_ctrl->bits_per_second = frame_bitrate;
1362     misc_rate_ctrl->target_percentage = 66;
1363     misc_rate_ctrl->window_size = 1000;
1364     misc_rate_ctrl->initial_qp = initial_qp;
1365     misc_rate_ctrl->min_qp = minimal_qp;
1366     misc_rate_ctrl->basic_unit_size = 0;
1367     vaUnmapBuffer(va_dpy, rc_param_buf);
1368
1369     render_id[0] = seq_param_buf;
1370     render_id[1] = rc_param_buf;
1371
1372     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1373
1374     return 0;
1375 }
1376
1377 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1378 {
1379     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1380     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1381     int PicOrderCntMsb, TopFieldOrderCnt;
1382
1383     if (frame_type == FRAME_IDR)
1384         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1385     else {
1386         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1387         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1388     }
1389
1390     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1391         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1392         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1393     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1394              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1395         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1396     else
1397         PicOrderCntMsb = prevPicOrderCntMsb;
1398
1399     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1400
1401     if (frame_type != FRAME_B) {
1402         PicOrderCntMsb_ref = PicOrderCntMsb;
1403         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1404     }
1405
1406     return TopFieldOrderCnt;
1407 }
1408
1409 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1410 {
1411     VABufferID pic_param_buf;
1412     VAStatus va_status;
1413     int i = 0;
1414
1415     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1416     pic_param.CurrPic.frame_idx = current_frame_num;
1417     pic_param.CurrPic.flags = 0;
1418     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1419     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1420     CurrentCurrPic = pic_param.CurrPic;
1421
1422     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1423     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1424         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1425         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1426     }
1427
1428     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1429     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1430     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1431     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1432     pic_param.frame_num = current_frame_num;
1433     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1434     pic_param.last_picture = false;  // FIXME
1435     pic_param.pic_init_qp = initial_qp;
1436
1437     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1438                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1439     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1440
1441     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1442
1443     return 0;
1444 }
1445
1446 int H264EncoderImpl::render_packedsequence()
1447 {
1448     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1449     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1450     unsigned int length_in_bits;
1451     unsigned char *packedseq_buffer = NULL;
1452     VAStatus va_status;
1453
1454     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1455
1456     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1457
1458     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1459     packedheader_param_buffer.has_emulation_bytes = 0;
1460     va_status = vaCreateBuffer(va_dpy,
1461                                context_id,
1462                                VAEncPackedHeaderParameterBufferType,
1463                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1464                                &packedseq_para_bufid);
1465     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1466
1467     va_status = vaCreateBuffer(va_dpy,
1468                                context_id,
1469                                VAEncPackedHeaderDataBufferType,
1470                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1471                                &packedseq_data_bufid);
1472     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1473
1474     render_id[0] = packedseq_para_bufid;
1475     render_id[1] = packedseq_data_bufid;
1476     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1477
1478     free(packedseq_buffer);
1479
1480     return 0;
1481 }
1482
1483
1484 int H264EncoderImpl::render_packedpicture()
1485 {
1486     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1487     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1488     unsigned int length_in_bits;
1489     unsigned char *packedpic_buffer = NULL;
1490     VAStatus va_status;
1491
1492     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1493     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1494     packedheader_param_buffer.bit_length = length_in_bits;
1495     packedheader_param_buffer.has_emulation_bytes = 0;
1496
1497     va_status = vaCreateBuffer(va_dpy,
1498                                context_id,
1499                                VAEncPackedHeaderParameterBufferType,
1500                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1501                                &packedpic_para_bufid);
1502     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1503
1504     va_status = vaCreateBuffer(va_dpy,
1505                                context_id,
1506                                VAEncPackedHeaderDataBufferType,
1507                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1508                                &packedpic_data_bufid);
1509     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1510
1511     render_id[0] = packedpic_para_bufid;
1512     render_id[1] = packedpic_data_bufid;
1513     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1514
1515     free(packedpic_buffer);
1516
1517     return 0;
1518 }
1519
1520 void H264EncoderImpl::render_packedslice()
1521 {
1522     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1523     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1524     unsigned int length_in_bits;
1525     unsigned char *packedslice_buffer = NULL;
1526     VAStatus va_status;
1527
1528     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1529     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1530     packedheader_param_buffer.bit_length = length_in_bits;
1531     packedheader_param_buffer.has_emulation_bytes = 0;
1532
1533     va_status = vaCreateBuffer(va_dpy,
1534                                context_id,
1535                                VAEncPackedHeaderParameterBufferType,
1536                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1537                                &packedslice_para_bufid);
1538     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1539
1540     va_status = vaCreateBuffer(va_dpy,
1541                                context_id,
1542                                VAEncPackedHeaderDataBufferType,
1543                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1544                                &packedslice_data_bufid);
1545     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1546
1547     render_id[0] = packedslice_para_bufid;
1548     render_id[1] = packedslice_data_bufid;
1549     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1550
1551     free(packedslice_buffer);
1552 }
1553
1554 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1555 {
1556     VABufferID slice_param_buf;
1557     VAStatus va_status;
1558     int i;
1559
1560     update_RefPicList(frame_type);
1561
1562     /* one frame, one slice */
1563     slice_param.macroblock_address = 0;
1564     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1565     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1566     if (frame_type == FRAME_IDR) {
1567         if (encoding_frame_num != 0)
1568             ++slice_param.idr_pic_id;
1569     } else if (frame_type == FRAME_P) {
1570         int refpiclist0_max = h264_maxref & 0xffff;
1571         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1572
1573         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1574             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1575             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1576         }
1577     } else if (frame_type == FRAME_B) {
1578         int refpiclist0_max = h264_maxref & 0xffff;
1579         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1580
1581         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1582         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1583             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1584             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1585         }
1586
1587         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1588         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1589             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1590             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1591         }
1592     }
1593
1594     slice_param.slice_alpha_c0_offset_div2 = 0;
1595     slice_param.slice_beta_offset_div2 = 0;
1596     slice_param.direct_spatial_mv_pred_flag = 1;
1597     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1598
1599
1600     if (h264_packedheader &&
1601         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1602         render_packedslice();
1603
1604     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1605                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1606     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1607
1608     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1609
1610     return 0;
1611 }
1612
1613
1614
1615 void H264EncoderImpl::save_codeddata(storage_task task)
1616 {
1617         VACodedBufferSegment *buf_list = NULL;
1618         VAStatus va_status;
1619
1620         string data;
1621
1622         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1623         CHECK_VASTATUS(va_status, "vaMapBuffer");
1624         while (buf_list != NULL) {
1625                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1626                 buf_list = (VACodedBufferSegment *) buf_list->next;
1627         }
1628         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1629
1630         {
1631                 // Add video.
1632                 AVPacket pkt;
1633                 memset(&pkt, 0, sizeof(pkt));
1634                 pkt.buf = nullptr;
1635                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1636                 pkt.size = data.size();
1637                 pkt.stream_index = 0;
1638                 if (task.frame_type == FRAME_IDR) {
1639                         pkt.flags = AV_PKT_FLAG_KEY;
1640                 } else {
1641                         pkt.flags = 0;
1642                 }
1643                 //pkt.duration = 1;
1644                 if (file_mux) {
1645                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1646                 }
1647                 if (!global_flags.uncompressed_video_to_http &&
1648                     !global_flags.x264_video_to_http) {
1649                         httpd->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1650                 }
1651         }
1652         // Encode and add all audio frames up to and including the pts of this video frame.
1653         for ( ;; ) {
1654                 int64_t audio_pts;
1655                 vector<float> audio;
1656                 {
1657                         unique_lock<mutex> lock(frame_queue_mutex);
1658                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1659                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1660                         auto it = pending_audio_frames.begin();
1661                         if (it->first > task.pts) break;
1662                         audio_pts = it->first;
1663                         audio = move(it->second);
1664                         pending_audio_frames.erase(it);
1665                 }
1666
1667                 if (context_audio_stream) {
1668                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, { file_mux.get() });
1669                         encode_audio(audio, &audio_queue_stream, audio_pts, context_audio_stream, { httpd });
1670                 } else {
1671                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, { httpd, file_mux.get() });
1672                 }
1673
1674                 if (audio_pts == task.pts) break;
1675         }
1676 }
1677
1678 void H264EncoderImpl::encode_audio(
1679         const vector<float> &audio,
1680         vector<float> *audio_queue,
1681         int64_t audio_pts,
1682         AVCodecContext *ctx,
1683         const vector<PacketDestination *> &destinations)
1684 {
1685         if (ctx->frame_size == 0) {
1686                 // No queueing needed.
1687                 assert(audio_queue->empty());
1688                 assert(audio.size() % 2 == 0);
1689                 encode_audio_one_frame(&audio[0], audio.size() / 2, audio_pts, ctx, destinations);
1690                 return;
1691         }
1692
1693         audio_queue->insert(audio_queue->end(), audio.begin(), audio.end());
1694         size_t sample_num;
1695         for (sample_num = 0;
1696              sample_num + ctx->frame_size * 2 <= audio_queue->size();
1697              sample_num += ctx->frame_size * 2) {
1698                 encode_audio_one_frame(&(*audio_queue)[sample_num],
1699                                        ctx->frame_size,
1700                                        audio_pts,
1701                                        ctx,
1702                                        destinations);
1703         }
1704         audio_queue->erase(audio_queue->begin(), audio_queue->begin() + sample_num);
1705 }
1706
1707 void H264EncoderImpl::encode_audio_one_frame(
1708         const float *audio,
1709         size_t num_samples,
1710         int64_t audio_pts,
1711         AVCodecContext *ctx,
1712         const vector<PacketDestination *> &destinations)
1713 {
1714         audio_frame->nb_samples = num_samples;
1715         audio_frame->channel_layout = AV_CH_LAYOUT_STEREO;
1716
1717         unique_ptr<float[]> planar_samples;
1718         unique_ptr<int32_t[]> int_samples;
1719
1720         if (ctx->sample_fmt == AV_SAMPLE_FMT_FLTP) {
1721                 audio_frame->format = AV_SAMPLE_FMT_FLTP;
1722                 planar_samples.reset(new float[num_samples * 2]);
1723                 avcodec_fill_audio_frame(audio_frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), num_samples * 2 * sizeof(float), 0);
1724                 for (size_t i = 0; i < num_samples; ++i) {
1725                         planar_samples[i] = audio[i * 2 + 0];
1726                         planar_samples[i + num_samples] = audio[i * 2 + 1];
1727                 }
1728         } else {
1729                 assert(ctx->sample_fmt == AV_SAMPLE_FMT_S32);
1730                 int_samples.reset(new int32_t[num_samples * 2]);
1731                 int ret = avcodec_fill_audio_frame(audio_frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), num_samples * 2 * sizeof(int32_t), 1);
1732                 if (ret < 0) {
1733                         fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1734                         exit(1);
1735                 }
1736                 for (size_t i = 0; i < num_samples * 2; ++i) {
1737                         if (audio[i] >= 1.0f) {
1738                                 int_samples[i] = 2147483647;
1739                         } else if (audio[i] <= -1.0f) {
1740                                 int_samples[i] = -2147483647;
1741                         } else {
1742                                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1743                         }
1744                 }
1745         }
1746
1747         AVPacket pkt;
1748         av_init_packet(&pkt);
1749         pkt.data = nullptr;
1750         pkt.size = 0;
1751         int got_output = 0;
1752         avcodec_encode_audio2(ctx, &pkt, audio_frame, &got_output);
1753         if (got_output) {
1754                 pkt.stream_index = 1;
1755                 pkt.flags = AV_PKT_FLAG_KEY;
1756                 for (PacketDestination *dest : destinations) {
1757                         dest->add_packet(pkt, audio_pts + global_delay(), audio_pts + global_delay());
1758                 }
1759         }
1760         // TODO: Delayed frames.
1761         av_frame_unref(audio_frame);
1762         av_free_packet(&pkt);
1763 }
1764
1765 // this is weird. but it seems to put a new frame onto the queue
1766 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1767 {
1768         unique_lock<mutex> lock(storage_task_queue_mutex);
1769         storage_task_queue.push(move(task));
1770         storage_task_queue_changed.notify_all();
1771 }
1772
1773 void H264EncoderImpl::storage_task_thread()
1774 {
1775         for ( ;; ) {
1776                 storage_task current;
1777                 {
1778                         // wait until there's an encoded frame
1779                         unique_lock<mutex> lock(storage_task_queue_mutex);
1780                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1781                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1782                         current = move(storage_task_queue.front());
1783                         storage_task_queue.pop();
1784                 }
1785
1786                 VAStatus va_status;
1787
1788                 // waits for data, then saves it to disk.
1789                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1790                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1791                 save_codeddata(move(current));
1792
1793                 {
1794                         unique_lock<mutex> lock(storage_task_queue_mutex);
1795                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1796                         storage_task_queue_changed.notify_all();
1797                 }
1798         }
1799 }
1800
1801 int H264EncoderImpl::release_encode()
1802 {
1803         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1804                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1805                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1806                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1807
1808                 if (!use_zerocopy) {
1809                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1810                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1811                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1812                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1813                 }
1814                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1815                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1816         }
1817
1818         vaDestroyContext(va_dpy, context_id);
1819         vaDestroyConfig(va_dpy, config_id);
1820
1821         return 0;
1822 }
1823
1824 int H264EncoderImpl::deinit_va()
1825 {
1826     vaTerminate(va_dpy);
1827
1828     va_close_display(va_dpy);
1829
1830     return 0;
1831 }
1832
1833 namespace {
1834
1835 void init_audio_encoder(const string &codec_name, int bit_rate, AVCodecContext **ctx)
1836 {
1837         AVCodec *codec_audio = avcodec_find_encoder_by_name(codec_name.c_str());
1838         if (codec_audio == nullptr) {
1839                 fprintf(stderr, "ERROR: Could not find codec '%s'\n", codec_name.c_str());
1840                 exit(1);
1841         }
1842
1843         AVCodecContext *context_audio = avcodec_alloc_context3(codec_audio);
1844         context_audio->bit_rate = bit_rate;
1845         context_audio->sample_rate = OUTPUT_FREQUENCY;
1846
1847         // Choose sample format; we currently only support these two
1848         // (see encode_audio), so we're a bit picky.
1849         const AVSampleFormat *ptr = codec_audio->sample_fmts;
1850         for ( ; *ptr != -1; ++ptr) {
1851                 if (*ptr == AV_SAMPLE_FMT_FLTP || *ptr == AV_SAMPLE_FMT_S32) {
1852                         context_audio->sample_fmt = *ptr;
1853                         break;
1854                 }
1855         }
1856         if (*ptr == -1) {
1857                 fprintf(stderr, "ERROR: Audio codec does not support fltp or s32 sample formats\n");
1858                 exit(1);
1859         }
1860
1861         context_audio->channels = 2;
1862         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1863         context_audio->time_base = AVRational{1, TIMEBASE};
1864         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1865                 fprintf(stderr, "Could not open codec '%s'\n", codec_name.c_str());
1866                 exit(1);
1867         }
1868
1869         *ctx = context_audio;
1870 }
1871
1872 }  // namespace
1873
1874 H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1875         : current_storage_frame(0), surface(surface), httpd(httpd)
1876 {
1877         init_audio_encoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, &context_audio_file);
1878
1879         if (!global_flags.stream_audio_codec_name.empty()) {
1880                 init_audio_encoder(global_flags.stream_audio_codec_name,
1881                         global_flags.stream_audio_codec_bitrate, &context_audio_stream);
1882         }
1883
1884         audio_frame = av_frame_alloc();
1885
1886         frame_width = width;
1887         frame_height = height;
1888         frame_width_mbaligned = (frame_width + 15) & (~15);
1889         frame_height_mbaligned = (frame_height + 15) & (~15);
1890
1891         //print_input();
1892
1893         if (global_flags.uncompressed_video_to_http ||
1894             global_flags.x264_video_to_http) {
1895                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1896         }
1897         if (global_flags.x264_video_to_http) {
1898                 x264_encoder.reset(new X264Encoder(httpd));
1899         }
1900
1901         init_va(va_display);
1902         setup_encode();
1903
1904         // No frames are ready yet.
1905         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1906
1907         memset(&seq_param, 0, sizeof(seq_param));
1908         memset(&pic_param, 0, sizeof(pic_param));
1909         memset(&slice_param, 0, sizeof(slice_param));
1910
1911         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1912
1913         encode_thread = thread([this]{
1914                 //SDL_GL_MakeCurrent(window, context);
1915                 QOpenGLContext *context = create_context(this->surface);
1916                 eglBindAPI(EGL_OPENGL_API);
1917                 if (!make_current(context, this->surface)) {
1918                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1919                                 eglGetError());
1920                         exit(1);
1921                 }
1922                 encode_thread_func();
1923         });
1924 }
1925
1926 H264EncoderImpl::~H264EncoderImpl()
1927 {
1928         shutdown();
1929         av_frame_free(&audio_frame);
1930
1931         // TODO: Destroy context.
1932 }
1933
1934 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1935 {
1936         assert(!is_shutdown);
1937         {
1938                 // Wait until this frame slot is done encoding.
1939                 unique_lock<mutex> lock(storage_task_queue_mutex);
1940                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1941                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1942                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1943                 }
1944                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1945                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1946                 if (storage_thread_should_quit) return false;
1947         }
1948
1949         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1950         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1951         *y_tex = surf->y_tex;
1952         *cbcr_tex = surf->cbcr_tex;
1953
1954         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1955         CHECK_VASTATUS(va_status, "vaDeriveImage");
1956
1957         if (use_zerocopy) {
1958                 VABufferInfo buf_info;
1959                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1960                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1961                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1962
1963                 // Create Y image.
1964                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1965                 EGLint y_attribs[] = {
1966                         EGL_WIDTH, frame_width,
1967                         EGL_HEIGHT, frame_height,
1968                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1969                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1970                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1971                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1972                         EGL_NONE
1973                 };
1974
1975                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1976                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1977
1978                 // Associate Y image to a texture.
1979                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1980                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1981
1982                 // Create CbCr image.
1983                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1984                 EGLint cbcr_attribs[] = {
1985                         EGL_WIDTH, frame_width,
1986                         EGL_HEIGHT, frame_height,
1987                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1988                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1989                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1990                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1991                         EGL_NONE
1992                 };
1993
1994                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1995                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1996
1997                 // Associate CbCr image to a texture.
1998                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1999                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
2000         }
2001
2002         return true;
2003 }
2004
2005 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
2006 {
2007         assert(!is_shutdown);
2008         {
2009                 unique_lock<mutex> lock(frame_queue_mutex);
2010                 pending_audio_frames[pts] = move(audio);
2011         }
2012         frame_queue_nonempty.notify_all();
2013 }
2014
2015 RefCountedGLsync H264EncoderImpl::end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames)
2016 {
2017         assert(!is_shutdown);
2018
2019         if (!use_zerocopy) {
2020                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
2021
2022                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
2023                 check_error();
2024
2025                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
2026                 check_error();
2027
2028                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
2029                 check_error();
2030                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
2031                 check_error();
2032
2033                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
2034                 check_error();
2035                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
2036                 check_error();
2037
2038                 glBindTexture(GL_TEXTURE_2D, 0);
2039                 check_error();
2040                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
2041                 check_error();
2042
2043                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
2044                 check_error();
2045         }
2046
2047         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
2048         check_error();
2049         glFlush();  // Make the H.264 thread see the fence as soon as possible.
2050         check_error();
2051
2052         {
2053                 unique_lock<mutex> lock(frame_queue_mutex);
2054                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
2055                 ++current_storage_frame;
2056         }
2057         frame_queue_nonempty.notify_all();
2058         return fence;
2059 }
2060
2061 void H264EncoderImpl::shutdown()
2062 {
2063         if (is_shutdown) {
2064                 return;
2065         }
2066
2067         {
2068                 unique_lock<mutex> lock(frame_queue_mutex);
2069                 encode_thread_should_quit = true;
2070                 frame_queue_nonempty.notify_all();
2071         }
2072         encode_thread.join();
2073         {
2074                 unique_lock<mutex> lock(storage_task_queue_mutex);
2075                 storage_thread_should_quit = true;
2076                 frame_queue_nonempty.notify_all();
2077                 storage_task_queue_changed.notify_all();
2078         }
2079         storage_thread.join();
2080
2081         release_encode();
2082         deinit_va();
2083         is_shutdown = true;
2084 }
2085
2086 void H264EncoderImpl::open_output_file(const std::string &filename)
2087 {
2088         AVFormatContext *avctx = avformat_alloc_context();
2089         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
2090         assert(filename.size() < sizeof(avctx->filename) - 1);
2091         strcpy(avctx->filename, filename.c_str());
2092
2093         string url = "file:" + filename;
2094         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
2095         if (ret < 0) {
2096                 char tmp[AV_ERROR_MAX_STRING_SIZE];
2097                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
2098                 exit(1);
2099         }
2100
2101         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, TIMEBASE, DEFAULT_AUDIO_OUTPUT_BIT_RATE));
2102 }
2103
2104 void H264EncoderImpl::close_output_file()
2105 {
2106         file_mux.reset();
2107 }
2108
2109 void H264EncoderImpl::encode_thread_func()
2110 {
2111         int64_t last_dts = -1;
2112         int gop_start_display_frame_num = 0;
2113         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
2114                 PendingFrame frame;
2115                 int pts_lag;
2116                 int frame_type, display_frame_num;
2117                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
2118                                        &display_frame_num, &frame_type, &pts_lag);
2119                 if (frame_type == FRAME_IDR) {
2120                         numShortTerm = 0;
2121                         current_frame_num = 0;
2122                         gop_start_display_frame_num = display_frame_num;
2123                 }
2124
2125                 {
2126                         unique_lock<mutex> lock(frame_queue_mutex);
2127                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
2128                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
2129                         });
2130                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
2131                                 // We have queued frames that were supposed to be B-frames,
2132                                 // but will be no P-frame to encode them against. Encode them all
2133                                 // as P-frames instead. Note that this happens under the mutex,
2134                                 // but nobody else uses it at this point, since we're shutting down,
2135                                 // so there's no contention.
2136                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
2137                                 return;
2138                         } else {
2139                                 frame = move(pending_video_frames[display_frame_num]);
2140                                 pending_video_frames.erase(display_frame_num);
2141                         }
2142                 }
2143
2144                 // Determine the dts of this frame.
2145                 int64_t dts;
2146                 if (pts_lag == -1) {
2147                         assert(last_dts != -1);
2148                         dts = last_dts + (TIMEBASE / MAX_FPS);
2149                 } else {
2150                         dts = frame.pts - pts_lag;
2151                 }
2152                 last_dts = dts;
2153
2154                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
2155         }
2156 }
2157
2158 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2159 {
2160         if (pending_video_frames.empty()) {
2161                 return;
2162         }
2163
2164         for (auto &pending_frame : pending_video_frames) {
2165                 int display_frame_num = pending_frame.first;
2166                 assert(display_frame_num > 0);
2167                 PendingFrame frame = move(pending_frame.second);
2168                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2169                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2170                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
2171                 last_dts = dts;
2172         }
2173
2174         if (global_flags.uncompressed_video_to_http ||
2175             global_flags.x264_video_to_http) {
2176                 // Add frames left in reorderer.
2177                 while (!reorderer->empty()) {
2178                         pair<int64_t, const uint8_t *> output_frame = reorderer->get_first_frame();
2179                         if (global_flags.uncompressed_video_to_http) {
2180                                 add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
2181                         } else {
2182                                 assert(global_flags.x264_video_to_http);
2183                                 x264_encoder->add_frame(output_frame.first, output_frame.second);
2184                         }
2185                 }
2186         }
2187 }
2188
2189 void H264EncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data)
2190 {
2191         AVPacket pkt;
2192         memset(&pkt, 0, sizeof(pkt));
2193         pkt.buf = nullptr;
2194         pkt.data = const_cast<uint8_t *>(data);
2195         pkt.size = frame_width * frame_height * 2;
2196         pkt.stream_index = 0;
2197         pkt.flags = AV_PKT_FLAG_KEY;
2198         httpd->add_packet(pkt, pts, pts);
2199 }
2200
2201 namespace {
2202
2203 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2204 {
2205         if (src_width == dst_pitch) {
2206                 memcpy(dst, src, src_width * height);
2207         } else {
2208                 for (size_t y = 0; y < height; ++y) {
2209                         const uint8_t *sptr = src + y * src_width;
2210                         uint8_t *dptr = dst + y * dst_pitch;
2211                         memcpy(dptr, sptr, src_width);
2212                 }
2213         }
2214 }
2215
2216 }  // namespace
2217
2218 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2219                                    int frame_type, int64_t pts, int64_t dts)
2220 {
2221         // Wait for the GPU to be done with the frame.
2222         GLenum sync_status;
2223         do {
2224                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2225                 check_error();
2226         } while (sync_status == GL_TIMEOUT_EXPIRED);
2227         assert(sync_status != GL_WAIT_FAILED);
2228
2229         // Release back any input frames we needed to render this frame.
2230         frame.input_frames.clear();
2231
2232         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2233         VAStatus va_status;
2234
2235         if (use_zerocopy) {
2236                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2237                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2238                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2239                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2240         } else {
2241                 unsigned char *surface_p = nullptr;
2242                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2243
2244                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2245                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2246
2247                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2248                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2249
2250                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2251                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2252
2253                 if (global_flags.uncompressed_video_to_http ||
2254                     global_flags.x264_video_to_http) {
2255                         // Add uncompressed video. (Note that pts == dts here.)
2256                         // Delay needs to match audio.
2257                         pair<int64_t, const uint8_t *> output_frame = reorderer->reorder_frame(pts + global_delay(), reinterpret_cast<uint8_t *>(surf->y_ptr));
2258                         if (output_frame.second != nullptr) {
2259                                 if (global_flags.uncompressed_video_to_http) {
2260                                         add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
2261                                 } else {
2262                                         assert(global_flags.x264_video_to_http);
2263                                         x264_encoder->add_frame(output_frame.first, output_frame.second);
2264                                 }
2265                         }
2266                 }
2267         }
2268
2269         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2270         CHECK_VASTATUS(va_status, "vaDestroyImage");
2271
2272         // Schedule the frame for encoding.
2273         VASurfaceID va_surface = surf->src_surface;
2274         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2275         CHECK_VASTATUS(va_status, "vaBeginPicture");
2276
2277         if (frame_type == FRAME_IDR) {
2278                 render_sequence();
2279                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2280                 if (h264_packedheader) {
2281                         render_packedsequence();
2282                         render_packedpicture();
2283                 }
2284         } else {
2285                 //render_sequence();
2286                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2287         }
2288         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2289
2290         va_status = vaEndPicture(va_dpy, context_id);
2291         CHECK_VASTATUS(va_status, "vaEndPicture");
2292
2293         // so now the data is done encoding (well, async job kicked off)...
2294         // we send that to the storage thread
2295         storage_task tmp;
2296         tmp.display_order = display_frame_num;
2297         tmp.frame_type = frame_type;
2298         tmp.pts = pts;
2299         tmp.dts = dts;
2300         storage_task_enqueue(move(tmp));
2301
2302         update_ReferenceFrames(frame_type);
2303 }
2304
2305 // Proxy object.
2306 H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
2307         : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
2308
2309 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2310 H264Encoder::~H264Encoder() {}
2311
2312 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
2313 {
2314         impl->add_audio(pts, audio);
2315 }
2316
2317 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2318 {
2319         return impl->begin_frame(y_tex, cbcr_tex);
2320 }
2321
2322 RefCountedGLsync H264Encoder::end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames)
2323 {
2324         return impl->end_frame(pts, input_frames);
2325 }
2326
2327 void H264Encoder::shutdown()
2328 {
2329         impl->shutdown();
2330 }
2331
2332 void H264Encoder::open_output_file(const std::string &filename)
2333 {
2334         impl->open_output_file(filename);
2335 }
2336
2337 void H264Encoder::close_output_file()
2338 {
2339         impl->close_output_file();
2340 }