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