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 <libavresample/avresample.h>
  14 #include <libavutil/channel_layout.h>
  15 #include <libavutil/frame.h>
  16 #include <libavutil/rational.h>
  17 #include <libavutil/samplefmt.h>
  18 #include <libavutil/opt.h>
  19 }
  20 #include <libdrm/drm_fourcc.h>
  21 #include <stdio.h>
  22 #include <stdlib.h>
  23 #include <string.h>
  24 #include <fcntl.h>
  25 #include <va/va.h>
  26 #include <va/va_drm.h>
  27 #include <va/va_drmcommon.h>
  28 #include <va/va_enc_h264.h>
  29 #include <va/va_x11.h>
  30 #include <algorithm>
  31 #include <condition_variable>
  32 #include <cstdint>
  33 #include <map>
  34 #include <memory>
  35 #include <mutex>
  36 #include <queue>
  37 #include <string>
  38 #include <thread>
  39 #include <utility>
  40
  41 #include "context.h"
  42 #include "defs.h"
  43 #include "flags.h"
  44 #include "httpd.h"
  45 #include "mux.h"
  46 #include "timebase.h"
  47 #include "x264encode.h"
  48
  49 using namespace std;
  50
  51 class QOpenGLContext;
  52 class QSurface;
  53
  54 #define CHECK_VASTATUS(va_status, func)                                 \
  55     if (va_status != VA_STATUS_SUCCESS) {                               \
  56         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  57         exit(1);                                                        \
  58     }
  59
  60 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  61
  62 //#include "loadsurface.h"
  63
  64 #define NAL_REF_IDC_NONE        0
  65 #define NAL_REF_IDC_LOW         1
  66 #define NAL_REF_IDC_MEDIUM      2
  67 #define NAL_REF_IDC_HIGH        3
  68
  69 #define NAL_NON_IDR             1
  70 #define NAL_IDR                 5
  71 #define NAL_SPS                 7
  72 #define NAL_PPS                 8
  73 #define NAL_SEI                 6
  74
  75 #define SLICE_TYPE_P            0
  76 #define SLICE_TYPE_B            1
  77 #define SLICE_TYPE_I            2
  78 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  79 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  80 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  81
  82
  83 #define ENTROPY_MODE_CAVLC      0
  84 #define ENTROPY_MODE_CABAC      1
  85
  86 #define PROFILE_IDC_BASELINE    66
  87 #define PROFILE_IDC_MAIN        77
  88 #define PROFILE_IDC_HIGH        100
  89
  90 #define BITSTREAM_ALLOCATE_STEPPING     4096
  91 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  92 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  93 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  94
  95 static constexpr unsigned int MaxFrameNum = (2<<16);
  96 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  97 static constexpr unsigned int Log2MaxFrameNum = 16;
  98 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  99 static constexpr int rc_default_modes[] = {  // Priority list of modes.
 100     VA_RC_VBR,
 101     VA_RC_CQP,
 102     VA_RC_VBR_CONSTRAINED,
 103     VA_RC_CBR,
 104     VA_RC_VCM,
 105     VA_RC_NONE,
 106 };
 107
 108 /* thread to save coded data */
 109 #define SRC_SURFACE_FREE        0
 110 #define SRC_SURFACE_IN_ENCODING 1
 111
 112 struct __bitstream {
 113     unsigned int *buffer;
 114     int bit_offset;
 115     int max_size_in_dword;
 116 };
 117 typedef struct __bitstream bitstream;
 118
 119 using namespace std;
 120
 121 // H.264 video comes out in encoding order (e.g. with two B-frames:
 122 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 123 // come in the right order. Since we do everything, including waiting
 124 // for the frames to come out of OpenGL, in encoding order, we need
 125 // a reordering buffer for uncompressed frames so that they come out
 126 // correctly. We go the super-lazy way of not making it understand
 127 // anything about the true order (which introduces some extra latency,
 128 // though); we know that for N B-frames we need at most (N-1) frames
 129 // in the reorder buffer, and can just sort on that.
 130 //
 131 // The class also deals with keeping a freelist as needed.
 132 class FrameReorderer {
 133 public:
 134         FrameReorderer(unsigned queue_length, int width, int height);
 135
 136         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 137         // Does _not_ take ownership of data; a copy is taken if needed.
 138         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 139         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 140         pair<int64_t, const uint8_t *> reorder_frame(int64_t pts, const uint8_t *data);
 141
 142         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 143         pair<int64_t, const uint8_t *> get_first_frame();
 144
 145         bool empty() const { return frames.empty(); }
 146
 147 private:
 148         unsigned queue_length;
 149         int width, height;
 150
 151         priority_queue<pair<int64_t, uint8_t *>> frames;
 152         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 153
 154         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 155         // except priority_queue doesn't work well with movable-only types.
 156         vector<unique_ptr<uint8_t[]>> owner;
 157 };
 158
 159 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 160     : queue_length(queue_length), width(width), height(height)
 161 {
 162         for (unsigned i = 0; i < queue_length; ++i) {
 163                 owner.emplace_back(new uint8_t[width * height * 2]);
 164                 freelist.push(owner.back().get());
 165         }
 166 }
 167
 168 pair<int64_t, const uint8_t *> FrameReorderer::reorder_frame(int64_t pts, const uint8_t *data)
 169 {
 170         if (queue_length == 0) {
 171                 return make_pair(pts, data);
 172         }
 173
 174         assert(!freelist.empty());
 175         uint8_t *storage = freelist.top();
 176         freelist.pop();
 177         memcpy(storage, data, width * height * 2);
 178         frames.emplace(-pts, storage);  // Invert pts to get smallest first.
 179
 180         if (frames.size() >= queue_length) {
 181                 return get_first_frame();
 182         } else {
 183                 return make_pair(-1, nullptr);
 184         }
 185 }
 186
 187 pair<int64_t, const uint8_t *> FrameReorderer::get_first_frame()
 188 {
 189         assert(!frames.empty());
 190         pair<int64_t, uint8_t *> storage = frames.top();
 191         frames.pop();
 192         int64_t pts = storage.first;
 193         freelist.push(storage.second);
 194         return make_pair(-pts, storage.second);  // Re-invert pts (see reorder_frame()).
 195 }
 196
 197 class H264EncoderImpl : public KeyFrameSignalReceiver {
 198 public:
 199         H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
 200         ~H264EncoderImpl();
 201         void add_audio(int64_t pts, vector<float> audio);
 202         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 203         RefCountedGLsync end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames);
 204         void shutdown();
 205         void open_output_file(const std::string &filename);
 206         void close_output_file();
 207
 208         virtual void signal_keyframe() override {
 209                 stream_mux_writing_keyframes = true;
 210         }
 211
 212 private:
 213         struct storage_task {
 214                 unsigned long long display_order;
 215                 int frame_type;
 216                 vector<float> audio;
 217                 int64_t pts, dts;
 218         };
 219         struct PendingFrame {
 220                 RefCountedGLsync fence;
 221                 vector<RefCountedFrame> input_frames;
 222                 int64_t pts;
 223         };
 224
 225         // So we never get negative dts.
 226         int64_t global_delay() const {
 227                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 228         }
 229
 230         void encode_thread_func();
 231         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 232         void add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data);
 233         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 234                           int frame_type, int64_t pts, int64_t dts);
 235         void storage_task_thread();
 236         void encode_audio(const vector<float> &audio,
 237                           vector<float> *audio_queue,
 238                           int64_t audio_pts,
 239                           AVCodecContext *ctx,
 240                           AVAudioResampleContext *resampler,
 241                           const vector<Mux *> &muxes);
 242         void encode_audio_one_frame(const float *audio,
 243                                     size_t num_samples,  // In each channel.
 244                                     int64_t audio_pts,
 245                                     AVCodecContext *ctx,
 246                                     AVAudioResampleContext *resampler,
 247                                     const vector<Mux *> &muxes);
 248         void encode_last_audio(vector<float> *audio_queue,
 249                                int64_t audio_pts,
 250                                AVCodecContext *ctx,
 251                                AVAudioResampleContext *resampler,
 252                                const vector<Mux *> &muxes);
 253         void encode_remaining_audio();
 254         void storage_task_enqueue(storage_task task);
 255         void save_codeddata(storage_task task);
 256         int render_packedsequence();
 257         int render_packedpicture();
 258         void render_packedslice();
 259         int render_sequence();
 260         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 261         void sps_rbsp(bitstream *bs);
 262         void pps_rbsp(bitstream *bs);
 263         int build_packed_pic_buffer(unsigned char **header_buffer);
 264         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 265         void slice_header(bitstream *bs);
 266         int build_packed_seq_buffer(unsigned char **header_buffer);
 267         int build_packed_slice_buffer(unsigned char **header_buffer);
 268         int init_va(const string &va_display);
 269         int deinit_va();
 270         void enable_zerocopy_if_possible();
 271         VADisplay va_open_display(const string &va_display);
 272         void va_close_display(VADisplay va_dpy);
 273         int setup_encode();
 274         int release_encode();
 275         void update_ReferenceFrames(int frame_type);
 276         int update_RefPicList(int frame_type);
 277         void open_output_stream();
 278         void close_output_stream();
 279         static int write_packet_thunk(void *opaque, uint8_t *buf, int buf_size);
 280         int write_packet(uint8_t *buf, int buf_size);
 281
 282         bool is_shutdown = false;
 283         bool use_zerocopy;
 284         int drm_fd = -1;
 285
 286         thread encode_thread, storage_thread;
 287
 288         mutex storage_task_queue_mutex;
 289         condition_variable storage_task_queue_changed;
 290         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 291         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 292         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 293
 294         mutex frame_queue_mutex;
 295         condition_variable frame_queue_nonempty;
 296         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 297
 298         int current_storage_frame;
 299
 300         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 301         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 302         int64_t last_audio_pts = 0;  // The first pts after all audio we've encoded.
 303         QSurface *surface;
 304
 305         AVCodecContext *context_audio_file;
 306         AVCodecContext *context_audio_stream = nullptr;  // nullptr = don't code separate audio for stream.
 307
 308         AVAudioResampleContext *resampler_audio_file = nullptr;
 309         AVAudioResampleContext *resampler_audio_stream = nullptr;
 310
 311         vector<float> audio_queue_file;
 312         vector<float> audio_queue_stream;
 313
 314         unique_ptr<Mux> stream_mux;  // To HTTP.
 315         unique_ptr<Mux> file_mux;  // To local disk.
 316
 317         // While Mux object is constructing, <stream_mux_writing_header> is true,
 318         // and the header is being collected into stream_mux_header.
 319         bool stream_mux_writing_header;
 320         string stream_mux_header;
 321
 322         bool stream_mux_writing_keyframes = false;
 323
 324         AVFrame *audio_frame = nullptr;
 325         HTTPD *httpd;
 326         unique_ptr<FrameReorderer> reorderer;
 327         unique_ptr<X264Encoder> x264_encoder;  // nullptr if not using x264.
 328
 329         Display *x11_display = nullptr;
 330
 331         // Encoder parameters
 332         VADisplay va_dpy;
 333         VAProfile h264_profile = (VAProfile)~0;
 334         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 335         int config_attrib_num = 0, enc_packed_header_idx;
 336
 337         struct GLSurface {
 338                 VASurfaceID src_surface, ref_surface;
 339                 VABufferID coded_buf;
 340
 341                 VAImage surface_image;
 342                 GLuint y_tex, cbcr_tex;
 343
 344                 // Only if use_zerocopy == true.
 345                 EGLImage y_egl_image, cbcr_egl_image;
 346
 347                 // Only if use_zerocopy == false.
 348                 GLuint pbo;
 349                 uint8_t *y_ptr, *cbcr_ptr;
 350                 size_t y_offset, cbcr_offset;
 351         };
 352         GLSurface gl_surfaces[SURFACE_NUM];
 353
 354         VAConfigID config_id;
 355         VAContextID context_id;
 356         VAEncSequenceParameterBufferH264 seq_param;
 357         VAEncPictureParameterBufferH264 pic_param;
 358         VAEncSliceParameterBufferH264 slice_param;
 359         VAPictureH264 CurrentCurrPic;
 360         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 361
 362         // Static quality settings.
 363         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 364         static constexpr unsigned int num_ref_frames = 2;
 365         static constexpr int initial_qp = 15;
 366         static constexpr int minimal_qp = 0;
 367         static constexpr int intra_period = 30;
 368         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 369
 370         // Quality settings that are meant to be static, but might be overridden
 371         // by the profile.
 372         int constraint_set_flag = 0;
 373         int h264_packedheader = 0; /* support pack header? */
 374         int h264_maxref = (1<<16|1);
 375         int h264_entropy_mode = 1; /* cabac */
 376         int ip_period = 3;
 377
 378         int rc_mode = -1;
 379         unsigned int current_frame_num = 0;
 380         unsigned int numShortTerm = 0;
 381
 382         int frame_width;
 383         int frame_height;
 384         int frame_width_mbaligned;
 385         int frame_height_mbaligned;
 386 };
 387
 388 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 389 // but if we don't delete it here, we get leaks. The GStreamer implementation
 390 // does the same.
 391 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 392 {
 393     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 394     CHECK_VASTATUS(va_status, "vaRenderPicture");
 395
 396     for (int i = 0; i < num_buffers; ++i) {
 397         va_status = vaDestroyBuffer(dpy, buffers[i]);
 398         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 399     }
 400 }
 401
 402 static unsigned int
 403 va_swap32(unsigned int val)
 404 {
 405     unsigned char *pval = (unsigned char *)&val;
 406
 407     return ((pval[0] << 24)     |
 408             (pval[1] << 16)     |
 409             (pval[2] << 8)      |
 410             (pval[3] << 0));
 411 }
 412
 413 static void
 414 bitstream_start(bitstream *bs)
 415 {
 416     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 417     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 418     bs->bit_offset = 0;
 419 }
 420
 421 static void
 422 bitstream_end(bitstream *bs)
 423 {
 424     int pos = (bs->bit_offset >> 5);
 425     int bit_offset = (bs->bit_offset & 0x1f);
 426     int bit_left = 32 - bit_offset;
 427
 428     if (bit_offset) {
 429         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 430     }
 431 }
 432
 433 static void
 434 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 435 {
 436     int pos = (bs->bit_offset >> 5);
 437     int bit_offset = (bs->bit_offset & 0x1f);
 438     int bit_left = 32 - bit_offset;
 439
 440     if (!size_in_bits)
 441         return;
 442
 443     bs->bit_offset += size_in_bits;
 444
 445     if (bit_left > size_in_bits) {
 446         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 447     } else {
 448         size_in_bits -= bit_left;
 449         if (bit_left >= 32) {
 450             bs->buffer[pos] = (val >> size_in_bits);
 451         } else {
 452             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 453         }
 454         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 455
 456         if (pos + 1 == bs->max_size_in_dword) {
 457             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 458             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 459         }
 460
 461         bs->buffer[pos + 1] = val;
 462     }
 463 }
 464
 465 static void
 466 bitstream_put_ue(bitstream *bs, unsigned int val)
 467 {
 468     int size_in_bits = 0;
 469     int tmp_val = ++val;
 470
 471     while (tmp_val) {
 472         tmp_val >>= 1;
 473         size_in_bits++;
 474     }
 475
 476     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 477     bitstream_put_ui(bs, val, size_in_bits);
 478 }
 479
 480 static void
 481 bitstream_put_se(bitstream *bs, int val)
 482 {
 483     unsigned int new_val;
 484
 485     if (val <= 0)
 486         new_val = -2 * val;
 487     else
 488         new_val = 2 * val - 1;
 489
 490     bitstream_put_ue(bs, new_val);
 491 }
 492
 493 static void
 494 bitstream_byte_aligning(bitstream *bs, int bit)
 495 {
 496     int bit_offset = (bs->bit_offset & 0x7);
 497     int bit_left = 8 - bit_offset;
 498     int new_val;
 499
 500     if (!bit_offset)
 501         return;
 502
 503     assert(bit == 0 || bit == 1);
 504
 505     if (bit)
 506         new_val = (1 << bit_left) - 1;
 507     else
 508         new_val = 0;
 509
 510     bitstream_put_ui(bs, new_val, bit_left);
 511 }
 512
 513 static void
 514 rbsp_trailing_bits(bitstream *bs)
 515 {
 516     bitstream_put_ui(bs, 1, 1);
 517     bitstream_byte_aligning(bs, 0);
 518 }
 519
 520 static void nal_start_code_prefix(bitstream *bs)
 521 {
 522     bitstream_put_ui(bs, 0x00000001, 32);
 523 }
 524
 525 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 526 {
 527     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 528     bitstream_put_ui(bs, nal_ref_idc, 2);
 529     bitstream_put_ui(bs, nal_unit_type, 5);
 530 }
 531
 532 void H264EncoderImpl::sps_rbsp(bitstream *bs)
 533 {
 534     int profile_idc = PROFILE_IDC_BASELINE;
 535
 536     if (h264_profile  == VAProfileH264High)
 537         profile_idc = PROFILE_IDC_HIGH;
 538     else if (h264_profile  == VAProfileH264Main)
 539         profile_idc = PROFILE_IDC_MAIN;
 540
 541     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 542     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 543     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 544     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 545     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 546     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 547     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 548     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 549
 550     if ( profile_idc == PROFILE_IDC_HIGH) {
 551         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 552         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 553         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 554         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 555         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 556     }
 557
 558     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 559     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 560
 561     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 562         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 563     else {
 564         assert(0);
 565     }
 566
 567     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 568     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 569
 570     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 571     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 572     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 573
 574     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 575         assert(0);
 576     }
 577
 578     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 579     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 580
 581     if (seq_param.frame_cropping_flag) {
 582         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 583         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 584         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 585         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 586     }
 587
 588     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 589     if ( false ) {
 590         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 591     } else {
 592         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 593         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 594         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 595         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 596         {
 597             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 598             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 599             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 600             {
 601                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 602                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 603                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 604             }
 605         }
 606         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 607         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 608         {
 609             bitstream_put_ui(bs, 1, 32);  // FPS
 610             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 611             bitstream_put_ui(bs, 1, 1);
 612         }
 613         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 614         {
 615             // hrd_parameters
 616             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 617             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 618             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 619
 620             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 621             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 622             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 623
 624             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 625             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 626             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 627             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 628         }
 629         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 630         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 631
 632         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 633         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 634     }
 635
 636     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 637 }
 638
 639
 640 void H264EncoderImpl::pps_rbsp(bitstream *bs)
 641 {
 642     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 643     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 644
 645     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 646
 647     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 648
 649     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 650
 651     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 652     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 653
 654     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 655     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 656
 657     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 658     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 659     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 660
 661     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 662     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 663     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 664
 665     /* more_rbsp_data */
 666     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 667     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 668     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 669
 670     rbsp_trailing_bits(bs);
 671 }
 672
 673 void H264EncoderImpl::slice_header(bitstream *bs)
 674 {
 675     int first_mb_in_slice = slice_param.macroblock_address;
 676
 677     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 678     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 679     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 680     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 681
 682     /* frame_mbs_only_flag == 1 */
 683     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 684         /* FIXME: */
 685         assert(0);
 686     }
 687
 688     if (pic_param.pic_fields.bits.idr_pic_flag)
 689         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 690
 691     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 692         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 693         /* pic_order_present_flag == 0 */
 694     } else {
 695         /* FIXME: */
 696         assert(0);
 697     }
 698
 699     /* redundant_pic_cnt_present_flag == 0 */
 700     /* slice type */
 701     if (IS_P_SLICE(slice_param.slice_type)) {
 702         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 703
 704         if (slice_param.num_ref_idx_active_override_flag)
 705             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 706
 707         /* ref_pic_list_reordering */
 708         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 709     } else if (IS_B_SLICE(slice_param.slice_type)) {
 710         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 711
 712         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 713
 714         if (slice_param.num_ref_idx_active_override_flag) {
 715             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 716             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 717         }
 718
 719         /* ref_pic_list_reordering */
 720         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 721         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 722     }
 723
 724     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 725          IS_P_SLICE(slice_param.slice_type)) ||
 726         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 727          IS_B_SLICE(slice_param.slice_type))) {
 728         /* FIXME: fill weight/offset table */
 729         assert(0);
 730     }
 731
 732     /* dec_ref_pic_marking */
 733     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 734         unsigned char no_output_of_prior_pics_flag = 0;
 735         unsigned char long_term_reference_flag = 0;
 736         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 737
 738         if (pic_param.pic_fields.bits.idr_pic_flag) {
 739             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 740             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 741         } else {
 742             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 743         }
 744     }
 745
 746     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 747         !IS_I_SLICE(slice_param.slice_type))
 748         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 749
 750     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 751
 752     /* ignore for SP/SI */
 753
 754     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 755         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 756
 757         if (slice_param.disable_deblocking_filter_idc != 1) {
 758             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 759             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 760         }
 761     }
 762
 763     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 764         bitstream_byte_aligning(bs, 1);
 765     }
 766 }
 767
 768 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 769 {
 770     bitstream bs;
 771
 772     bitstream_start(&bs);
 773     nal_start_code_prefix(&bs);
 774     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 775     pps_rbsp(&bs);
 776     bitstream_end(&bs);
 777
 778     *header_buffer = (unsigned char *)bs.buffer;
 779     return bs.bit_offset;
 780 }
 781
 782 int
 783 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 784 {
 785     bitstream bs;
 786
 787     bitstream_start(&bs);
 788     nal_start_code_prefix(&bs);
 789     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 790     sps_rbsp(&bs);
 791     bitstream_end(&bs);
 792
 793     *header_buffer = (unsigned char *)bs.buffer;
 794     return bs.bit_offset;
 795 }
 796
 797 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 798 {
 799     bitstream bs;
 800     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 801     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 802
 803     bitstream_start(&bs);
 804     nal_start_code_prefix(&bs);
 805
 806     if (IS_I_SLICE(slice_param.slice_type)) {
 807         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 808     } else if (IS_P_SLICE(slice_param.slice_type)) {
 809         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 810     } else {
 811         assert(IS_B_SLICE(slice_param.slice_type));
 812         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 813     }
 814
 815     slice_header(&bs);
 816     bitstream_end(&bs);
 817
 818     *header_buffer = (unsigned char *)bs.buffer;
 819     return bs.bit_offset;
 820 }
 821
 822
 823 /*
 824   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 825   1) period between Frame #X and Frame #N = #X - #N
 826   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 827   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 828   4) intra_period and intra_idr_period take precedence over ip_period
 829   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 830      of I/IDR frames, see bellow examples
 831   -------------------------------------------------------------------
 832   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 833   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 834   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 835   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 836   1            1                ignored    IDR IDR IDR IDR...
 837   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 838   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 839   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 840                                               (PBB)(IBB)(PBB)(IBB)...
 841   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 842                                            IDRPPPPP IPPPPP IPPPPP...
 843   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 844                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 845                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 846                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 847                                            {IDR(PBB)(PBB)}                     (6/6/3)
 848                                            {IDR(PBB)(PBB)}.
 849 */
 850
 851 // General pts/dts strategy:
 852 //
 853 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 854 // bit tricky. We assume first of all that the frame rate never goes _above_
 855 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 856 // in at least this speed, as long at dts <= pts (the frame is not attempted
 857 // presented before it is decoded). Furthermore, we never have longer chains of
 858 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 859 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 860 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 861 // frame, with an order O <= C.)
 862 //
 863 // Many strategies are possible, but we establish these rules:
 864 //
 865 //  - Tip frames have dts = pts - (C-O)*N.
 866 //  - Non-tip frames have dts = dts_last + N.
 867 //
 868 // An example, with C=2 and N=10 and the data flow showed with arrows:
 869 //
 870 //        I  B  P  B  B  P
 871 //   pts: 30 40 50 60 70 80
 872 //        ↓  ↓     ↓
 873 //   dts: 10 30 20 60 50←40
 874 //         |  |  ↑        ↑
 875 //         `--|--'        |
 876 //             `----------'
 877 //
 878 // To show that this works fine also with irregular spacings, let's say that
 879 // the third frame is delayed a bit (something earlier was dropped). Now the
 880 // situation looks like this:
 881 //
 882 //        I  B  P  B  B   P
 883 //   pts: 30 40 80 90 100 110
 884 //        ↓  ↓     ↓
 885 //   dts: 10 30 20 90 50←40
 886 //         |  |  ↑        ↑
 887 //         `--|--'        |
 888 //             `----------'
 889 //
 890 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 891 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 892 //
 893 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 894 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 895 // a tip frame and should be given a dts based on the previous one.
 896 #define FRAME_P 0
 897 #define FRAME_B 1
 898 #define FRAME_I 2
 899 #define FRAME_IDR 7
 900 void encoding2display_order(
 901     int encoding_order, int intra_period,
 902     int intra_idr_period, int ip_period,
 903     int *displaying_order,
 904     int *frame_type, int *pts_lag)
 905 {
 906     int encoding_order_gop = 0;
 907
 908     *pts_lag = 0;
 909
 910     if (intra_period == 1) { /* all are I/IDR frames */
 911         *displaying_order = encoding_order;
 912         if (intra_idr_period == 0)
 913             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 914         else
 915             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 916         return;
 917     }
 918
 919     if (intra_period == 0)
 920         intra_idr_period = 0;
 921
 922     if (ip_period == 1) {
 923         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 924         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 925         *displaying_order = encoding_order;
 926
 927         if (encoding_order_gop == 0) { /* the first frame */
 928             *frame_type = FRAME_IDR;
 929         } else if (intra_period != 0 && /* have I frames */
 930                    encoding_order_gop >= 2 &&
 931                    (encoding_order_gop % intra_period == 0)) {
 932             *frame_type = FRAME_I;
 933         } else {
 934             *frame_type = FRAME_P;
 935         }
 936         return;
 937     }
 938
 939     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 940     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 941     *pts_lag = -1;  // Most frames are not tip frames.
 942
 943     if (encoding_order_gop == 0) { /* the first frame */
 944         *frame_type = FRAME_IDR;
 945         *displaying_order = encoding_order;
 946         // IDR frames are a special case; I honestly can't find the logic behind
 947         // why this is the right thing, but it seems to line up nicely in practice :-)
 948         *pts_lag = TIMEBASE / MAX_FPS;
 949     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 950         *frame_type = FRAME_B;
 951         *displaying_order = encoding_order - 1;
 952         if ((encoding_order_gop % ip_period) == 0) {
 953             *pts_lag = 0;  // Last B-frame.
 954         }
 955     } else if (intra_period != 0 && /* have I frames */
 956                encoding_order_gop >= 2 &&
 957                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 958         *frame_type = FRAME_I;
 959         *displaying_order = encoding_order + ip_period - 1;
 960     } else {
 961         *frame_type = FRAME_P;
 962         *displaying_order = encoding_order + ip_period - 1;
 963     }
 964 }
 965
 966
 967 static const char *rc_to_string(int rc_mode)
 968 {
 969     switch (rc_mode) {
 970     case VA_RC_NONE:
 971         return "NONE";
 972     case VA_RC_CBR:
 973         return "CBR";
 974     case VA_RC_VBR:
 975         return "VBR";
 976     case VA_RC_VCM:
 977         return "VCM";
 978     case VA_RC_CQP:
 979         return "CQP";
 980     case VA_RC_VBR_CONSTRAINED:
 981         return "VBR_CONSTRAINED";
 982     default:
 983         return "Unknown";
 984     }
 985 }
 986
 987 void H264EncoderImpl::enable_zerocopy_if_possible()
 988 {
 989         if (global_flags.uncompressed_video_to_http) {
 990                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
 991                 use_zerocopy = false;
 992         } else if (global_flags.x264_video_to_http) {
 993                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
 994                 use_zerocopy = false;
 995         } else {
 996                 use_zerocopy = true;
 997         }
 998 }
 999
1000 VADisplay H264EncoderImpl::va_open_display(const string &va_display)
1001 {
1002         if (va_display.empty()) {
1003                 x11_display = XOpenDisplay(NULL);
1004                 if (!x11_display) {
1005                         fprintf(stderr, "error: can't connect to X server!\n");
1006                         return NULL;
1007                 }
1008                 enable_zerocopy_if_possible();
1009                 return vaGetDisplay(x11_display);
1010         } else if (va_display[0] != '/') {
1011                 x11_display = XOpenDisplay(va_display.c_str());
1012                 if (!x11_display) {
1013                         fprintf(stderr, "error: can't connect to X server!\n");
1014                         return NULL;
1015                 }
1016                 enable_zerocopy_if_possible();
1017                 return vaGetDisplay(x11_display);
1018         } else {
1019                 drm_fd = open(va_display.c_str(), O_RDWR);
1020                 if (drm_fd == -1) {
1021                         perror(va_display.c_str());
1022                         return NULL;
1023                 }
1024                 use_zerocopy = false;
1025                 return vaGetDisplayDRM(drm_fd);
1026         }
1027 }
1028
1029 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
1030 {
1031         if (x11_display) {
1032                 XCloseDisplay(x11_display);
1033                 x11_display = nullptr;
1034         }
1035         if (drm_fd != -1) {
1036                 close(drm_fd);
1037         }
1038 }
1039
1040 int H264EncoderImpl::init_va(const string &va_display)
1041 {
1042     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1043     VAEntrypoint *entrypoints;
1044     int num_entrypoints, slice_entrypoint;
1045     int support_encode = 0;
1046     int major_ver, minor_ver;
1047     VAStatus va_status;
1048     unsigned int i;
1049
1050     va_dpy = va_open_display(va_display);
1051     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1052     CHECK_VASTATUS(va_status, "vaInitialize");
1053
1054     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1055     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1056     if (!entrypoints) {
1057         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1058         exit(1);
1059     }
1060
1061     /* use the highest profile */
1062     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1063         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1064             continue;
1065
1066         h264_profile = profile_list[i];
1067         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1068         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1069             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1070                 support_encode = 1;
1071                 break;
1072             }
1073         }
1074         if (support_encode == 1)
1075             break;
1076     }
1077
1078     if (support_encode == 0) {
1079         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1080         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1081         printf("to use VA-API against DRM instead of X11.\n");
1082         exit(1);
1083     } else {
1084         switch (h264_profile) {
1085             case VAProfileH264Baseline:
1086                 ip_period = 1;
1087                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1088                 h264_entropy_mode = 0;
1089                 break;
1090             case VAProfileH264ConstrainedBaseline:
1091                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1092                 ip_period = 1;
1093                 break;
1094
1095             case VAProfileH264Main:
1096                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1097                 break;
1098
1099             case VAProfileH264High:
1100                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1101                 break;
1102             default:
1103                 h264_profile = VAProfileH264Baseline;
1104                 ip_period = 1;
1105                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1106                 break;
1107         }
1108     }
1109
1110     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1111
1112     /* find out the format for the render target, and rate control mode */
1113     for (i = 0; i < VAConfigAttribTypeMax; i++)
1114         attrib[i].type = (VAConfigAttribType)i;
1115
1116     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1117                                       &attrib[0], VAConfigAttribTypeMax);
1118     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1119     /* check the interested configattrib */
1120     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1121         printf("Not find desired YUV420 RT format\n");
1122         exit(1);
1123     } else {
1124         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1125         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1126         config_attrib_num++;
1127     }
1128
1129     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1130         int tmp = attrib[VAConfigAttribRateControl].value;
1131
1132         if (rc_mode == -1 || !(rc_mode & tmp))  {
1133             if (rc_mode != -1) {
1134                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1135             }
1136
1137             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1138                 if (rc_default_modes[i] & tmp) {
1139                     rc_mode = rc_default_modes[i];
1140                     break;
1141                 }
1142             }
1143         }
1144
1145         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1146         config_attrib[config_attrib_num].value = rc_mode;
1147         config_attrib_num++;
1148     }
1149
1150
1151     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1152         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1153
1154         h264_packedheader = 1;
1155         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1156         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1157
1158         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1159             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1160         }
1161
1162         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1163             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1164         }
1165
1166         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1167             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1168         }
1169
1170         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1171             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1172         }
1173
1174         enc_packed_header_idx = config_attrib_num;
1175         config_attrib_num++;
1176     }
1177
1178     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1179         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1180         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1181         config_attrib_num++;
1182     }
1183
1184     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1185         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1186     }
1187
1188     free(entrypoints);
1189     return 0;
1190 }
1191
1192 int H264EncoderImpl::setup_encode()
1193 {
1194     VAStatus va_status;
1195     VASurfaceID *tmp_surfaceid;
1196     int codedbuf_size, i;
1197     static VASurfaceID src_surface[SURFACE_NUM];
1198     static VASurfaceID ref_surface[SURFACE_NUM];
1199
1200     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1201             &config_attrib[0], config_attrib_num, &config_id);
1202     CHECK_VASTATUS(va_status, "vaCreateConfig");
1203
1204     /* create source surfaces */
1205     va_status = vaCreateSurfaces(va_dpy,
1206                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1207                                  &src_surface[0], SURFACE_NUM,
1208                                  NULL, 0);
1209     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1210
1211     /* create reference surfaces */
1212     va_status = vaCreateSurfaces(va_dpy,
1213                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1214                                  &ref_surface[0], SURFACE_NUM,
1215                                  NULL, 0);
1216     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1217
1218     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1219     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1220     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1221
1222     /* Create a context for this encode pipe */
1223     va_status = vaCreateContext(va_dpy, config_id,
1224                                 frame_width_mbaligned, frame_height_mbaligned,
1225                                 VA_PROGRESSIVE,
1226                                 tmp_surfaceid, 2 * SURFACE_NUM,
1227                                 &context_id);
1228     CHECK_VASTATUS(va_status, "vaCreateContext");
1229     free(tmp_surfaceid);
1230
1231     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1232
1233     for (i = 0; i < SURFACE_NUM; i++) {
1234         /* create coded buffer once for all
1235          * other VA buffers which won't be used again after vaRenderPicture.
1236          * so APP can always vaCreateBuffer for every frame
1237          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1238          * so VA won't maintain the coded buffer
1239          */
1240         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1241                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1242         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1243     }
1244
1245     /* create OpenGL objects */
1246     //glGenFramebuffers(SURFACE_NUM, fbos);
1247
1248     for (i = 0; i < SURFACE_NUM; i++) {
1249         glGenTextures(1, &gl_surfaces[i].y_tex);
1250         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1251
1252         if (!use_zerocopy) {
1253             // Create Y image.
1254             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1255             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1256
1257             // Create CbCr image.
1258             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1259             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1260
1261             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1262             // buffers, due to potentially differing pitch.
1263             glGenBuffers(1, &gl_surfaces[i].pbo);
1264             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1265             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1266             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1267             gl_surfaces[i].y_offset = 0;
1268             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1269             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1270             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1271             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1272         }
1273     }
1274
1275     for (i = 0; i < SURFACE_NUM; i++) {
1276         gl_surfaces[i].src_surface = src_surface[i];
1277         gl_surfaces[i].ref_surface = ref_surface[i];
1278     }
1279
1280     return 0;
1281 }
1282
1283 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1284 //
1285 //   2 1 0 [3] 4 8 9
1286 template<class T, class C>
1287 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1288 {
1289         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1290         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1291         sort(middle, end, less_than);
1292 }
1293
1294 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1295 {
1296     int i;
1297
1298     if (frame_type == FRAME_B)
1299         return;
1300
1301     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1302     numShortTerm++;
1303     if (numShortTerm > num_ref_frames)
1304         numShortTerm = num_ref_frames;
1305     for (i=numShortTerm-1; i>0; i--)
1306         ReferenceFrames[i] = ReferenceFrames[i-1];
1307     ReferenceFrames[0] = CurrentCurrPic;
1308
1309     current_frame_num++;
1310     if (current_frame_num > MaxFrameNum)
1311         current_frame_num = 0;
1312 }
1313
1314
1315 int H264EncoderImpl::update_RefPicList(int frame_type)
1316 {
1317     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1318         return a.frame_idx > b.frame_idx;
1319     };
1320     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1321         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1322     };
1323     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1324         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1325     };
1326
1327     if (frame_type == FRAME_P) {
1328         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1329         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1330     } else if (frame_type == FRAME_B) {
1331         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1332         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1333
1334         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1335         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1336     }
1337
1338     return 0;
1339 }
1340
1341
1342 int H264EncoderImpl::render_sequence()
1343 {
1344     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1345     VAStatus va_status;
1346     VAEncMiscParameterBuffer *misc_param;
1347     VAEncMiscParameterRateControl *misc_rate_ctrl;
1348
1349     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1350     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1351     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1352     seq_param.bits_per_second = frame_bitrate;
1353
1354     seq_param.intra_period = intra_period;
1355     seq_param.intra_idr_period = intra_idr_period;
1356     seq_param.ip_period = ip_period;
1357
1358     seq_param.max_num_ref_frames = num_ref_frames;
1359     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1360     seq_param.time_scale = TIMEBASE * 2;
1361     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1362     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1363     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1364     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1365     seq_param.seq_fields.bits.chroma_format_idc = 1;
1366     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1367
1368     if (frame_width != frame_width_mbaligned ||
1369         frame_height != frame_height_mbaligned) {
1370         seq_param.frame_cropping_flag = 1;
1371         seq_param.frame_crop_left_offset = 0;
1372         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1373         seq_param.frame_crop_top_offset = 0;
1374         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1375     }
1376
1377     va_status = vaCreateBuffer(va_dpy, context_id,
1378                                VAEncSequenceParameterBufferType,
1379                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1380     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1381
1382     va_status = vaCreateBuffer(va_dpy, context_id,
1383                                VAEncMiscParameterBufferType,
1384                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1385                                1, NULL, &rc_param_buf);
1386     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1387
1388     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1389     misc_param->type = VAEncMiscParameterTypeRateControl;
1390     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1391     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1392     misc_rate_ctrl->bits_per_second = frame_bitrate;
1393     misc_rate_ctrl->target_percentage = 66;
1394     misc_rate_ctrl->window_size = 1000;
1395     misc_rate_ctrl->initial_qp = initial_qp;
1396     misc_rate_ctrl->min_qp = minimal_qp;
1397     misc_rate_ctrl->basic_unit_size = 0;
1398     vaUnmapBuffer(va_dpy, rc_param_buf);
1399
1400     render_id[0] = seq_param_buf;
1401     render_id[1] = rc_param_buf;
1402
1403     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1404
1405     return 0;
1406 }
1407
1408 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1409 {
1410     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1411     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1412     int PicOrderCntMsb, TopFieldOrderCnt;
1413
1414     if (frame_type == FRAME_IDR)
1415         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1416     else {
1417         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1418         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1419     }
1420
1421     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1422         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1423         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1424     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1425              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1426         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1427     else
1428         PicOrderCntMsb = prevPicOrderCntMsb;
1429
1430     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1431
1432     if (frame_type != FRAME_B) {
1433         PicOrderCntMsb_ref = PicOrderCntMsb;
1434         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1435     }
1436
1437     return TopFieldOrderCnt;
1438 }
1439
1440 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1441 {
1442     VABufferID pic_param_buf;
1443     VAStatus va_status;
1444     int i = 0;
1445
1446     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1447     pic_param.CurrPic.frame_idx = current_frame_num;
1448     pic_param.CurrPic.flags = 0;
1449     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1450     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1451     CurrentCurrPic = pic_param.CurrPic;
1452
1453     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1454     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1455         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1456         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1457     }
1458
1459     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1460     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1461     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1462     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1463     pic_param.frame_num = current_frame_num;
1464     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1465     pic_param.last_picture = false;  // FIXME
1466     pic_param.pic_init_qp = initial_qp;
1467
1468     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1469                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1470     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1471
1472     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1473
1474     return 0;
1475 }
1476
1477 int H264EncoderImpl::render_packedsequence()
1478 {
1479     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1480     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1481     unsigned int length_in_bits;
1482     unsigned char *packedseq_buffer = NULL;
1483     VAStatus va_status;
1484
1485     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1486
1487     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1488
1489     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1490     packedheader_param_buffer.has_emulation_bytes = 0;
1491     va_status = vaCreateBuffer(va_dpy,
1492                                context_id,
1493                                VAEncPackedHeaderParameterBufferType,
1494                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1495                                &packedseq_para_bufid);
1496     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1497
1498     va_status = vaCreateBuffer(va_dpy,
1499                                context_id,
1500                                VAEncPackedHeaderDataBufferType,
1501                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1502                                &packedseq_data_bufid);
1503     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1504
1505     render_id[0] = packedseq_para_bufid;
1506     render_id[1] = packedseq_data_bufid;
1507     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1508
1509     free(packedseq_buffer);
1510
1511     return 0;
1512 }
1513
1514
1515 int H264EncoderImpl::render_packedpicture()
1516 {
1517     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1518     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1519     unsigned int length_in_bits;
1520     unsigned char *packedpic_buffer = NULL;
1521     VAStatus va_status;
1522
1523     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1524     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1525     packedheader_param_buffer.bit_length = length_in_bits;
1526     packedheader_param_buffer.has_emulation_bytes = 0;
1527
1528     va_status = vaCreateBuffer(va_dpy,
1529                                context_id,
1530                                VAEncPackedHeaderParameterBufferType,
1531                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1532                                &packedpic_para_bufid);
1533     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1534
1535     va_status = vaCreateBuffer(va_dpy,
1536                                context_id,
1537                                VAEncPackedHeaderDataBufferType,
1538                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1539                                &packedpic_data_bufid);
1540     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1541
1542     render_id[0] = packedpic_para_bufid;
1543     render_id[1] = packedpic_data_bufid;
1544     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1545
1546     free(packedpic_buffer);
1547
1548     return 0;
1549 }
1550
1551 void H264EncoderImpl::render_packedslice()
1552 {
1553     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1554     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1555     unsigned int length_in_bits;
1556     unsigned char *packedslice_buffer = NULL;
1557     VAStatus va_status;
1558
1559     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1560     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1561     packedheader_param_buffer.bit_length = length_in_bits;
1562     packedheader_param_buffer.has_emulation_bytes = 0;
1563
1564     va_status = vaCreateBuffer(va_dpy,
1565                                context_id,
1566                                VAEncPackedHeaderParameterBufferType,
1567                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1568                                &packedslice_para_bufid);
1569     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1570
1571     va_status = vaCreateBuffer(va_dpy,
1572                                context_id,
1573                                VAEncPackedHeaderDataBufferType,
1574                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1575                                &packedslice_data_bufid);
1576     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1577
1578     render_id[0] = packedslice_para_bufid;
1579     render_id[1] = packedslice_data_bufid;
1580     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1581
1582     free(packedslice_buffer);
1583 }
1584
1585 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1586 {
1587     VABufferID slice_param_buf;
1588     VAStatus va_status;
1589     int i;
1590
1591     update_RefPicList(frame_type);
1592
1593     /* one frame, one slice */
1594     slice_param.macroblock_address = 0;
1595     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1596     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1597     if (frame_type == FRAME_IDR) {
1598         if (encoding_frame_num != 0)
1599             ++slice_param.idr_pic_id;
1600     } else if (frame_type == FRAME_P) {
1601         int refpiclist0_max = h264_maxref & 0xffff;
1602         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1603
1604         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1605             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1606             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1607         }
1608     } else if (frame_type == FRAME_B) {
1609         int refpiclist0_max = h264_maxref & 0xffff;
1610         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1611
1612         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1613         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1614             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1615             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1616         }
1617
1618         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1619         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1620             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1621             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1622         }
1623     }
1624
1625     slice_param.slice_alpha_c0_offset_div2 = 0;
1626     slice_param.slice_beta_offset_div2 = 0;
1627     slice_param.direct_spatial_mv_pred_flag = 1;
1628     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1629
1630
1631     if (h264_packedheader &&
1632         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1633         render_packedslice();
1634
1635     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1636                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1637     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1638
1639     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1640
1641     return 0;
1642 }
1643
1644
1645
1646 void H264EncoderImpl::save_codeddata(storage_task task)
1647 {
1648         VACodedBufferSegment *buf_list = NULL;
1649         VAStatus va_status;
1650
1651         string data;
1652
1653         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1654         CHECK_VASTATUS(va_status, "vaMapBuffer");
1655         while (buf_list != NULL) {
1656                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1657                 buf_list = (VACodedBufferSegment *) buf_list->next;
1658         }
1659         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1660
1661         {
1662                 // Add video.
1663                 AVPacket pkt;
1664                 memset(&pkt, 0, sizeof(pkt));
1665                 pkt.buf = nullptr;
1666                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1667                 pkt.size = data.size();
1668                 pkt.stream_index = 0;
1669                 if (task.frame_type == FRAME_IDR) {
1670                         pkt.flags = AV_PKT_FLAG_KEY;
1671                 } else {
1672                         pkt.flags = 0;
1673                 }
1674                 //pkt.duration = 1;
1675                 if (file_mux) {
1676                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1677                 }
1678                 if (!global_flags.uncompressed_video_to_http &&
1679                     !global_flags.x264_video_to_http) {
1680                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1681                 }
1682         }
1683         // Encode and add all audio frames up to and including the pts of this video frame.
1684         for ( ;; ) {
1685                 int64_t audio_pts;
1686                 vector<float> audio;
1687                 {
1688                         unique_lock<mutex> lock(frame_queue_mutex);
1689                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1690                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1691                         auto it = pending_audio_frames.begin();
1692                         if (it->first > task.pts) break;
1693                         audio_pts = it->first;
1694                         audio = move(it->second);
1695                         pending_audio_frames.erase(it);
1696                 }
1697
1698                 if (context_audio_stream) {
1699                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { file_mux.get() });
1700                         encode_audio(audio, &audio_queue_stream, audio_pts, context_audio_stream, resampler_audio_stream, { stream_mux.get() });
1701                 } else {
1702                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { stream_mux.get(), file_mux.get() });
1703                 }
1704                 last_audio_pts = audio_pts + audio.size() * TIMEBASE / (OUTPUT_FREQUENCY * 2);
1705
1706                 if (audio_pts == task.pts) break;
1707         }
1708 }
1709
1710 void H264EncoderImpl::encode_audio(
1711         const vector<float> &audio,
1712         vector<float> *audio_queue,
1713         int64_t audio_pts,
1714         AVCodecContext *ctx,
1715         AVAudioResampleContext *resampler,
1716         const vector<Mux *> &muxes)
1717 {
1718         if (ctx->frame_size == 0) {
1719                 // No queueing needed.
1720                 assert(audio_queue->empty());
1721                 assert(audio.size() % 2 == 0);
1722                 encode_audio_one_frame(&audio[0], audio.size() / 2, audio_pts, ctx, resampler, muxes);
1723                 return;
1724         }
1725
1726         int64_t sample_offset = audio_queue->size();
1727
1728         audio_queue->insert(audio_queue->end(), audio.begin(), audio.end());
1729         size_t sample_num;
1730         for (sample_num = 0;
1731              sample_num + ctx->frame_size * 2 <= audio_queue->size();
1732              sample_num += ctx->frame_size * 2) {
1733                 int64_t adjusted_audio_pts = audio_pts + (int64_t(sample_num) - sample_offset) * TIMEBASE / (OUTPUT_FREQUENCY * 2);
1734                 encode_audio_one_frame(&(*audio_queue)[sample_num],
1735                                        ctx->frame_size,
1736                                        adjusted_audio_pts,
1737                                        ctx,
1738                                        resampler,
1739                                        muxes);
1740         }
1741         audio_queue->erase(audio_queue->begin(), audio_queue->begin() + sample_num);
1742 }
1743
1744 void H264EncoderImpl::encode_audio_one_frame(
1745         const float *audio,
1746         size_t num_samples,
1747         int64_t audio_pts,
1748         AVCodecContext *ctx,
1749         AVAudioResampleContext *resampler,
1750         const vector<Mux *> &muxes)
1751 {
1752         audio_frame->pts = audio_pts + global_delay();
1753         audio_frame->nb_samples = num_samples;
1754         audio_frame->channel_layout = AV_CH_LAYOUT_STEREO;
1755         audio_frame->format = ctx->sample_fmt;
1756         audio_frame->sample_rate = OUTPUT_FREQUENCY;
1757
1758         if (av_samples_alloc(audio_frame->data, nullptr, 2, num_samples, ctx->sample_fmt, 0) < 0) {
1759                 fprintf(stderr, "Could not allocate %ld samples.\n", num_samples);
1760                 exit(1);
1761         }
1762
1763         if (avresample_convert(resampler, audio_frame->data, 0, num_samples,
1764                                (uint8_t **)&audio, 0, num_samples) < 0) {
1765                 fprintf(stderr, "Audio conversion failed.\n");
1766                 exit(1);
1767         }
1768
1769         AVPacket pkt;
1770         av_init_packet(&pkt);
1771         pkt.data = nullptr;
1772         pkt.size = 0;
1773         int got_output = 0;
1774         avcodec_encode_audio2(ctx, &pkt, audio_frame, &got_output);
1775         if (got_output) {
1776                 pkt.stream_index = 1;
1777                 pkt.flags = 0;
1778                 for (Mux *mux : muxes) {
1779                         mux->add_packet(pkt, pkt.pts, pkt.dts);
1780                 }
1781         }
1782
1783         av_freep(&audio_frame->data[0]);
1784
1785         av_frame_unref(audio_frame);
1786         av_free_packet(&pkt);
1787 }
1788
1789 void H264EncoderImpl::encode_last_audio(
1790         vector<float> *audio_queue,
1791         int64_t audio_pts,
1792         AVCodecContext *ctx,
1793         AVAudioResampleContext *resampler,
1794         const vector<Mux *> &muxes)
1795 {
1796         if (!audio_queue->empty()) {
1797                 // Last frame can be whatever size we want.
1798                 assert(audio_queue->size() % 2 == 0);
1799                 encode_audio_one_frame(&(*audio_queue)[0], audio_queue->size() / 2, audio_pts, ctx, resampler, muxes);
1800                 audio_queue->clear();
1801         }
1802
1803         if (ctx->codec->capabilities & AV_CODEC_CAP_DELAY) {
1804                 // Collect any delayed frames.
1805                 for ( ;; ) {
1806                         int got_output = 0;
1807                         AVPacket pkt;
1808                         av_init_packet(&pkt);
1809                         pkt.data = nullptr;
1810                         pkt.size = 0;
1811                         avcodec_encode_audio2(ctx, &pkt, nullptr, &got_output);
1812                         if (!got_output) break;
1813
1814                         pkt.stream_index = 1;
1815                         pkt.flags = 0;
1816                         for (Mux *mux : muxes) {
1817                                 mux->add_packet(pkt, pkt.pts, pkt.dts);
1818                         }
1819                         av_free_packet(&pkt);
1820                 }
1821         }
1822 }
1823
1824 // this is weird. but it seems to put a new frame onto the queue
1825 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1826 {
1827         unique_lock<mutex> lock(storage_task_queue_mutex);
1828         storage_task_queue.push(move(task));
1829         storage_task_queue_changed.notify_all();
1830 }
1831
1832 void H264EncoderImpl::storage_task_thread()
1833 {
1834         for ( ;; ) {
1835                 storage_task current;
1836                 {
1837                         // wait until there's an encoded frame
1838                         unique_lock<mutex> lock(storage_task_queue_mutex);
1839                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1840                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1841                         current = move(storage_task_queue.front());
1842                         storage_task_queue.pop();
1843                 }
1844
1845                 VAStatus va_status;
1846
1847                 // waits for data, then saves it to disk.
1848                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1849                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1850                 save_codeddata(move(current));
1851
1852                 {
1853                         unique_lock<mutex> lock(storage_task_queue_mutex);
1854                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1855                         storage_task_queue_changed.notify_all();
1856                 }
1857         }
1858 }
1859
1860 int H264EncoderImpl::release_encode()
1861 {
1862         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1863                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1864                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1865                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1866
1867                 if (!use_zerocopy) {
1868                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1869                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1870                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1871                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1872                 }
1873                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1874                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1875         }
1876
1877         vaDestroyContext(va_dpy, context_id);
1878         vaDestroyConfig(va_dpy, config_id);
1879
1880         return 0;
1881 }
1882
1883 int H264EncoderImpl::deinit_va()
1884 {
1885     vaTerminate(va_dpy);
1886
1887     va_close_display(va_dpy);
1888
1889     return 0;
1890 }
1891
1892 namespace {
1893
1894 void init_audio_encoder(const string &codec_name, int bit_rate, AVCodecContext **ctx, AVAudioResampleContext **resampler)
1895 {
1896         AVCodec *codec_audio = avcodec_find_encoder_by_name(codec_name.c_str());
1897         if (codec_audio == nullptr) {
1898                 fprintf(stderr, "ERROR: Could not find codec '%s'\n", codec_name.c_str());
1899                 exit(1);
1900         }
1901
1902         AVCodecContext *context_audio = avcodec_alloc_context3(codec_audio);
1903         context_audio->bit_rate = bit_rate;
1904         context_audio->sample_rate = OUTPUT_FREQUENCY;
1905         context_audio->sample_fmt = codec_audio->sample_fmts[0];
1906         context_audio->channels = 2;
1907         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1908         context_audio->time_base = AVRational{1, TIMEBASE};
1909         context_audio->flags |= CODEC_FLAG_GLOBAL_HEADER;
1910         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1911                 fprintf(stderr, "Could not open codec '%s'\n", codec_name.c_str());
1912                 exit(1);
1913         }
1914
1915         *ctx = context_audio;
1916
1917         *resampler = avresample_alloc_context();
1918         if (*resampler == nullptr) {
1919                 fprintf(stderr, "Allocating resampler failed.\n");
1920                 exit(1);
1921         }
1922
1923         av_opt_set_int(*resampler, "in_channel_layout",  AV_CH_LAYOUT_STEREO,       0);
1924         av_opt_set_int(*resampler, "out_channel_layout", AV_CH_LAYOUT_STEREO,       0);
1925         av_opt_set_int(*resampler, "in_sample_rate",     OUTPUT_FREQUENCY,          0);
1926         av_opt_set_int(*resampler, "out_sample_rate",    OUTPUT_FREQUENCY,          0);
1927         av_opt_set_int(*resampler, "in_sample_fmt",      AV_SAMPLE_FMT_FLT,         0);
1928         av_opt_set_int(*resampler, "out_sample_fmt",     context_audio->sample_fmt, 0);
1929
1930         if (avresample_open(*resampler) < 0) {
1931                 fprintf(stderr, "Could not open resample context.\n");
1932                 exit(1);
1933         }
1934 }
1935
1936 }  // namespace
1937
1938 H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1939         : current_storage_frame(0), surface(surface), httpd(httpd), frame_width(width), frame_height(height)
1940 {
1941         init_audio_encoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, &context_audio_file, &resampler_audio_file);
1942
1943         if (!global_flags.stream_audio_codec_name.empty()) {
1944                 init_audio_encoder(global_flags.stream_audio_codec_name,
1945                         global_flags.stream_audio_codec_bitrate, &context_audio_stream, &resampler_audio_stream);
1946         }
1947
1948         frame_width_mbaligned = (frame_width + 15) & (~15);
1949         frame_height_mbaligned = (frame_height + 15) & (~15);
1950
1951         open_output_stream();
1952
1953         audio_frame = av_frame_alloc();
1954
1955         //print_input();
1956
1957         if (global_flags.uncompressed_video_to_http ||
1958             global_flags.x264_video_to_http) {
1959                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1960         }
1961         if (global_flags.x264_video_to_http) {
1962                 x264_encoder.reset(new X264Encoder(stream_mux.get()));
1963         }
1964
1965         init_va(va_display);
1966         setup_encode();
1967
1968         // No frames are ready yet.
1969         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1970
1971         memset(&seq_param, 0, sizeof(seq_param));
1972         memset(&pic_param, 0, sizeof(pic_param));
1973         memset(&slice_param, 0, sizeof(slice_param));
1974
1975         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1976
1977         encode_thread = thread([this]{
1978                 //SDL_GL_MakeCurrent(window, context);
1979                 QOpenGLContext *context = create_context(this->surface);
1980                 eglBindAPI(EGL_OPENGL_API);
1981                 if (!make_current(context, this->surface)) {
1982                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1983                                 eglGetError());
1984                         exit(1);
1985                 }
1986                 encode_thread_func();
1987         });
1988 }
1989
1990 H264EncoderImpl::~H264EncoderImpl()
1991 {
1992         shutdown();
1993         av_frame_free(&audio_frame);
1994         avresample_free(&resampler_audio_file);
1995         avresample_free(&resampler_audio_stream);
1996         avcodec_free_context(&context_audio_file);
1997         avcodec_free_context(&context_audio_stream);
1998         close_output_stream();
1999 }
2000
2001 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2002 {
2003         assert(!is_shutdown);
2004         {
2005                 // Wait until this frame slot is done encoding.
2006                 unique_lock<mutex> lock(storage_task_queue_mutex);
2007                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
2008                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
2009                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
2010                 }
2011                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
2012                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
2013                 if (storage_thread_should_quit) return false;
2014         }
2015
2016         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
2017         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
2018         *y_tex = surf->y_tex;
2019         *cbcr_tex = surf->cbcr_tex;
2020
2021         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
2022         CHECK_VASTATUS(va_status, "vaDeriveImage");
2023
2024         if (use_zerocopy) {
2025                 VABufferInfo buf_info;
2026                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
2027                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
2028                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
2029
2030                 // Create Y image.
2031                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
2032                 EGLint y_attribs[] = {
2033                         EGL_WIDTH, frame_width,
2034                         EGL_HEIGHT, frame_height,
2035                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
2036                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
2037                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
2038                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
2039                         EGL_NONE
2040                 };
2041
2042                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
2043                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
2044
2045                 // Associate Y image to a texture.
2046                 glBindTexture(GL_TEXTURE_2D, *y_tex);
2047                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
2048
2049                 // Create CbCr image.
2050                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
2051                 EGLint cbcr_attribs[] = {
2052                         EGL_WIDTH, frame_width,
2053                         EGL_HEIGHT, frame_height,
2054                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
2055                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
2056                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
2057                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
2058                         EGL_NONE
2059                 };
2060
2061                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
2062                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
2063
2064                 // Associate CbCr image to a texture.
2065                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
2066                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
2067         }
2068
2069         return true;
2070 }
2071
2072 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
2073 {
2074         assert(!is_shutdown);
2075         {
2076                 unique_lock<mutex> lock(frame_queue_mutex);
2077                 pending_audio_frames[pts] = move(audio);
2078         }
2079         frame_queue_nonempty.notify_all();
2080 }
2081
2082 RefCountedGLsync H264EncoderImpl::end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames)
2083 {
2084         assert(!is_shutdown);
2085
2086         if (!use_zerocopy) {
2087                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
2088
2089                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
2090                 check_error();
2091
2092                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
2093                 check_error();
2094
2095                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
2096                 check_error();
2097                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
2098                 check_error();
2099
2100                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
2101                 check_error();
2102                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
2103                 check_error();
2104
2105                 glBindTexture(GL_TEXTURE_2D, 0);
2106                 check_error();
2107                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
2108                 check_error();
2109
2110                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
2111                 check_error();
2112         }
2113
2114         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
2115         check_error();
2116         glFlush();  // Make the H.264 thread see the fence as soon as possible.
2117         check_error();
2118
2119         {
2120                 unique_lock<mutex> lock(frame_queue_mutex);
2121                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
2122                 ++current_storage_frame;
2123         }
2124         frame_queue_nonempty.notify_all();
2125         return fence;
2126 }
2127
2128 void H264EncoderImpl::shutdown()
2129 {
2130         if (is_shutdown) {
2131                 return;
2132         }
2133
2134         {
2135                 unique_lock<mutex> lock(frame_queue_mutex);
2136                 encode_thread_should_quit = true;
2137                 frame_queue_nonempty.notify_all();
2138         }
2139         encode_thread.join();
2140         x264_encoder.reset();
2141         {
2142                 unique_lock<mutex> lock(storage_task_queue_mutex);
2143                 storage_thread_should_quit = true;
2144                 frame_queue_nonempty.notify_all();
2145                 storage_task_queue_changed.notify_all();
2146         }
2147         storage_thread.join();
2148         encode_remaining_audio();
2149
2150         release_encode();
2151         deinit_va();
2152         is_shutdown = true;
2153 }
2154
2155 void H264EncoderImpl::open_output_file(const std::string &filename)
2156 {
2157         AVFormatContext *avctx = avformat_alloc_context();
2158         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
2159         assert(filename.size() < sizeof(avctx->filename) - 1);
2160         strcpy(avctx->filename, filename.c_str());
2161
2162         string url = "file:" + filename;
2163         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
2164         if (ret < 0) {
2165                 char tmp[AV_ERROR_MAX_STRING_SIZE];
2166                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
2167                 exit(1);
2168         }
2169
2170         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, context_audio_file->codec, TIMEBASE, DEFAULT_AUDIO_OUTPUT_BIT_RATE, nullptr));
2171 }
2172
2173 void H264EncoderImpl::close_output_file()
2174 {
2175         file_mux.reset();
2176 }
2177
2178 void H264EncoderImpl::open_output_stream()
2179 {
2180         AVFormatContext *avctx = avformat_alloc_context();
2181         AVOutputFormat *oformat = av_guess_format(global_flags.stream_mux_name.c_str(), nullptr, nullptr);
2182         assert(oformat != nullptr);
2183         avctx->oformat = oformat;
2184
2185         string codec_name;
2186         int bit_rate;
2187
2188         if (global_flags.stream_audio_codec_name.empty()) {
2189                 codec_name = AUDIO_OUTPUT_CODEC_NAME;
2190                 bit_rate = DEFAULT_AUDIO_OUTPUT_BIT_RATE;
2191         } else {
2192                 codec_name = global_flags.stream_audio_codec_name;
2193                 bit_rate = global_flags.stream_audio_codec_bitrate;
2194         }
2195
2196         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
2197         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, &H264EncoderImpl::write_packet_thunk, nullptr);
2198
2199         Mux::Codec video_codec;
2200         if (global_flags.uncompressed_video_to_http) {
2201                 video_codec = Mux::CODEC_NV12;
2202         } else {
2203                 video_codec = Mux::CODEC_H264;
2204         }
2205
2206         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
2207         AVCodec *codec_audio = avcodec_find_encoder_by_name(codec_name.c_str());
2208         if (codec_audio == nullptr) {
2209                 fprintf(stderr, "ERROR: Could not find codec '%s'\n", codec_name.c_str());
2210                 exit(1);
2211         }
2212
2213         int time_base = global_flags.stream_coarse_timebase ? COARSE_TIMEBASE : TIMEBASE;
2214         stream_mux_writing_header = true;
2215         stream_mux.reset(new Mux(avctx, frame_width, frame_height, video_codec, codec_audio, time_base, bit_rate, this));
2216         stream_mux_writing_header = false;
2217         httpd->set_header(stream_mux_header);
2218         stream_mux_header.clear();
2219 }
2220
2221 void H264EncoderImpl::close_output_stream()
2222 {
2223         stream_mux.reset();
2224 }
2225
2226 int H264EncoderImpl::write_packet_thunk(void *opaque, uint8_t *buf, int buf_size)
2227 {
2228         H264EncoderImpl *h264_encoder = (H264EncoderImpl *)opaque;
2229         return h264_encoder->write_packet(buf, buf_size);
2230 }
2231
2232 int H264EncoderImpl::write_packet(uint8_t *buf, int buf_size)
2233 {
2234         if (stream_mux_writing_header) {
2235                 stream_mux_header.append((char *)buf, buf_size);
2236         } else {
2237                 httpd->add_data((char *)buf, buf_size, stream_mux_writing_keyframes);
2238                 stream_mux_writing_keyframes = false;
2239         }
2240         return buf_size;
2241 }
2242
2243 void H264EncoderImpl::encode_thread_func()
2244 {
2245         int64_t last_dts = -1;
2246         int gop_start_display_frame_num = 0;
2247         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
2248                 PendingFrame frame;
2249                 int pts_lag;
2250                 int frame_type, display_frame_num;
2251                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
2252                                        &display_frame_num, &frame_type, &pts_lag);
2253                 if (frame_type == FRAME_IDR) {
2254                         numShortTerm = 0;
2255                         current_frame_num = 0;
2256                         gop_start_display_frame_num = display_frame_num;
2257                 }
2258
2259                 {
2260                         unique_lock<mutex> lock(frame_queue_mutex);
2261                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
2262                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
2263                         });
2264                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
2265                                 // We have queued frames that were supposed to be B-frames,
2266                                 // but will be no P-frame to encode them against. Encode them all
2267                                 // as P-frames instead. Note that this happens under the mutex,
2268                                 // but nobody else uses it at this point, since we're shutting down,
2269                                 // so there's no contention.
2270                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
2271                                 return;
2272                         } else {
2273                                 frame = move(pending_video_frames[display_frame_num]);
2274                                 pending_video_frames.erase(display_frame_num);
2275                         }
2276                 }
2277
2278                 // Determine the dts of this frame.
2279                 int64_t dts;
2280                 if (pts_lag == -1) {
2281                         assert(last_dts != -1);
2282                         dts = last_dts + (TIMEBASE / MAX_FPS);
2283                 } else {
2284                         dts = frame.pts - pts_lag;
2285                 }
2286                 last_dts = dts;
2287
2288                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
2289         }
2290 }
2291
2292 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2293 {
2294         if (pending_video_frames.empty()) {
2295                 return;
2296         }
2297
2298         for (auto &pending_frame : pending_video_frames) {
2299                 int display_frame_num = pending_frame.first;
2300                 assert(display_frame_num > 0);
2301                 PendingFrame frame = move(pending_frame.second);
2302                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2303                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2304                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
2305                 last_dts = dts;
2306         }
2307
2308         if (global_flags.uncompressed_video_to_http ||
2309             global_flags.x264_video_to_http) {
2310                 // Add frames left in reorderer.
2311                 while (!reorderer->empty()) {
2312                         pair<int64_t, const uint8_t *> output_frame = reorderer->get_first_frame();
2313                         if (global_flags.uncompressed_video_to_http) {
2314                                 add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
2315                         } else {
2316                                 assert(global_flags.x264_video_to_http);
2317                                 x264_encoder->add_frame(output_frame.first, output_frame.second);
2318                         }
2319                 }
2320         }
2321 }
2322
2323 void H264EncoderImpl::encode_remaining_audio()
2324 {
2325         // This really ought to be empty by now, but just to be sure...
2326         for (auto &pending_frame : pending_audio_frames) {
2327                 int64_t audio_pts = pending_frame.first;
2328                 vector<float> audio = move(pending_frame.second);
2329
2330                 if (context_audio_stream) {
2331                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { file_mux.get() });
2332                         encode_audio(audio, &audio_queue_stream, audio_pts, context_audio_stream, resampler_audio_stream, { stream_mux.get() });
2333                 } else {
2334                         encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { stream_mux.get(), file_mux.get() });
2335                 }
2336                 last_audio_pts = audio_pts + audio.size() * TIMEBASE / (OUTPUT_FREQUENCY * 2);
2337         }
2338         pending_audio_frames.clear();
2339
2340         // Encode any leftover audio in the queues, and also any delayed frames.
2341         if (context_audio_stream) {
2342                 encode_last_audio(&audio_queue_file, last_audio_pts, context_audio_file, resampler_audio_file, { file_mux.get() });
2343                 encode_last_audio(&audio_queue_stream, last_audio_pts, context_audio_stream, resampler_audio_stream, { stream_mux.get() });
2344         } else {
2345                 encode_last_audio(&audio_queue_file, last_audio_pts, context_audio_file, resampler_audio_file, { stream_mux.get(), file_mux.get() });
2346         }
2347 }
2348
2349 void H264EncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data)
2350 {
2351         AVPacket pkt;
2352         memset(&pkt, 0, sizeof(pkt));
2353         pkt.buf = nullptr;
2354         pkt.data = const_cast<uint8_t *>(data);
2355         pkt.size = frame_width * frame_height * 2;
2356         pkt.stream_index = 0;
2357         pkt.flags = AV_PKT_FLAG_KEY;
2358         stream_mux->add_packet(pkt, pts, pts);
2359 }
2360
2361 namespace {
2362
2363 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2364 {
2365         if (src_width == dst_pitch) {
2366                 memcpy(dst, src, src_width * height);
2367         } else {
2368                 for (size_t y = 0; y < height; ++y) {
2369                         const uint8_t *sptr = src + y * src_width;
2370                         uint8_t *dptr = dst + y * dst_pitch;
2371                         memcpy(dptr, sptr, src_width);
2372                 }
2373         }
2374 }
2375
2376 }  // namespace
2377
2378 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2379                                    int frame_type, int64_t pts, int64_t dts)
2380 {
2381         // Wait for the GPU to be done with the frame.
2382         GLenum sync_status;
2383         do {
2384                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2385                 check_error();
2386         } while (sync_status == GL_TIMEOUT_EXPIRED);
2387         assert(sync_status != GL_WAIT_FAILED);
2388
2389         // Release back any input frames we needed to render this frame.
2390         frame.input_frames.clear();
2391
2392         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2393         VAStatus va_status;
2394
2395         if (use_zerocopy) {
2396                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2397                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2398                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2399                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2400         } else {
2401                 unsigned char *surface_p = nullptr;
2402                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2403
2404                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2405                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2406
2407                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2408                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2409
2410                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2411                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2412
2413                 if (global_flags.uncompressed_video_to_http ||
2414                     global_flags.x264_video_to_http) {
2415                         // Add uncompressed video. (Note that pts == dts here.)
2416                         // Delay needs to match audio.
2417                         pair<int64_t, const uint8_t *> output_frame = reorderer->reorder_frame(pts + global_delay(), reinterpret_cast<uint8_t *>(surf->y_ptr));
2418                         if (output_frame.second != nullptr) {
2419                                 if (global_flags.uncompressed_video_to_http) {
2420                                         add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
2421                                 } else {
2422                                         assert(global_flags.x264_video_to_http);
2423                                         x264_encoder->add_frame(output_frame.first, output_frame.second);
2424                                 }
2425                         }
2426                 }
2427         }
2428
2429         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2430         CHECK_VASTATUS(va_status, "vaDestroyImage");
2431
2432         // Schedule the frame for encoding.
2433         VASurfaceID va_surface = surf->src_surface;
2434         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2435         CHECK_VASTATUS(va_status, "vaBeginPicture");
2436
2437         if (frame_type == FRAME_IDR) {
2438                 render_sequence();
2439                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2440                 if (h264_packedheader) {
2441                         render_packedsequence();
2442                         render_packedpicture();
2443                 }
2444         } else {
2445                 //render_sequence();
2446                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2447         }
2448         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2449
2450         va_status = vaEndPicture(va_dpy, context_id);
2451         CHECK_VASTATUS(va_status, "vaEndPicture");
2452
2453         // so now the data is done encoding (well, async job kicked off)...
2454         // we send that to the storage thread
2455         storage_task tmp;
2456         tmp.display_order = display_frame_num;
2457         tmp.frame_type = frame_type;
2458         tmp.pts = pts;
2459         tmp.dts = dts;
2460         storage_task_enqueue(move(tmp));
2461
2462         update_ReferenceFrames(frame_type);
2463 }
2464
2465 // Proxy object.
2466 H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
2467         : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
2468
2469 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2470 H264Encoder::~H264Encoder() {}
2471
2472 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
2473 {
2474         impl->add_audio(pts, audio);
2475 }
2476
2477 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2478 {
2479         return impl->begin_frame(y_tex, cbcr_tex);
2480 }
2481
2482 RefCountedGLsync H264Encoder::end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames)
2483 {
2484         return impl->end_frame(pts, input_frames);
2485 }
2486
2487 void H264Encoder::shutdown()
2488 {
2489         impl->shutdown();
2490 }
2491
2492 void H264Encoder::open_output_file(const std::string &filename)
2493 {
2494         impl->open_output_file(filename);
2495 }
2496
2497 void H264Encoder::close_output_file()
2498 {
2499         impl->close_output_file();
2500 }