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

   1 //#include "sysdeps.h"
   2 #include "h264encode.h"
   3
   4 #include <EGL/eglplatform.h>
   5 #include <X11/X.h>
   6 #include <X11/Xlib.h>
   7 #include <assert.h>
   8 #include <epoxy/egl.h>
   9 extern "C" {
  10 #include <libavcodec/avcodec.h>
  11 #include <libavformat/avformat.h>
  12 #include <libavutil/channel_layout.h>
  13 #include <libavutil/frame.h>
  14 #include <libavutil/rational.h>
  15 #include <libavutil/samplefmt.h>
  16 }
  17 #include <libdrm/drm_fourcc.h>
  18 #include <stdio.h>
  19 #include <stdlib.h>
  20 #include <string.h>
  21 #include <fcntl.h>
  22 #include <va/va.h>
  23 #include <va/va_drm.h>
  24 #include <va/va_drmcommon.h>
  25 #include <va/va_enc_h264.h>
  26 #include <va/va_x11.h>
  27 #include <algorithm>
  28 #include <condition_variable>
  29 #include <cstdint>
  30 #include <map>
  31 #include <memory>
  32 #include <mutex>
  33 #include <queue>
  34 #include <string>
  35 #include <thread>
  36 #include <utility>
  37
  38 #include "context.h"
  39 #include "defs.h"
  40 #include "httpd.h"
  41 #include "timebase.h"
  42
  43 using namespace std;
  44
  45 class QOpenGLContext;
  46 class QSurface;
  47
  48 #define CHECK_VASTATUS(va_status, func)                                 \
  49     if (va_status != VA_STATUS_SUCCESS) {                               \
  50         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  51         exit(1);                                                        \
  52     }
  53
  54 //#include "loadsurface.h"
  55
  56 #define NAL_REF_IDC_NONE        0
  57 #define NAL_REF_IDC_LOW         1
  58 #define NAL_REF_IDC_MEDIUM      2
  59 #define NAL_REF_IDC_HIGH        3
  60
  61 #define NAL_NON_IDR             1
  62 #define NAL_IDR                 5
  63 #define NAL_SPS                 7
  64 #define NAL_PPS                 8
  65 #define NAL_SEI                 6
  66
  67 #define SLICE_TYPE_P            0
  68 #define SLICE_TYPE_B            1
  69 #define SLICE_TYPE_I            2
  70 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  71 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  72 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  73
  74
  75 #define ENTROPY_MODE_CAVLC      0
  76 #define ENTROPY_MODE_CABAC      1
  77
  78 #define PROFILE_IDC_BASELINE    66
  79 #define PROFILE_IDC_MAIN        77
  80 #define PROFILE_IDC_HIGH        100
  81
  82 #define BITSTREAM_ALLOCATE_STEPPING     4096
  83 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  84
  85 static constexpr unsigned int MaxFrameNum = (2<<16);
  86 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  87 static constexpr unsigned int Log2MaxFrameNum = 16;
  88 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  89 static constexpr int rc_default_modes[] = {  // Priority list of modes.
  90     VA_RC_VBR,
  91     VA_RC_CQP,
  92     VA_RC_VBR_CONSTRAINED,
  93     VA_RC_CBR,
  94     VA_RC_VCM,
  95     VA_RC_NONE,
  96 };
  97
  98 /* thread to save coded data */
  99 #define SRC_SURFACE_FREE        0
 100 #define SRC_SURFACE_IN_ENCODING 1
 101
 102 struct __bitstream {
 103     unsigned int *buffer;
 104     int bit_offset;
 105     int max_size_in_dword;
 106 };
 107 typedef struct __bitstream bitstream;
 108
 109 using namespace std;
 110
 111 class H264EncoderImpl {
 112 public:
 113         H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
 114         ~H264EncoderImpl();
 115         void add_audio(int64_t pts, vector<float> audio);
 116         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 117         void end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames);
 118         void shutdown();
 119
 120 private:
 121         struct storage_task {
 122                 unsigned long long display_order;
 123                 int frame_type;
 124                 vector<float> audio;
 125                 int64_t pts, dts;
 126         };
 127         struct PendingFrame {
 128                 RefCountedGLsync fence;
 129                 vector<RefCountedFrame> input_frames;
 130                 int64_t pts;
 131         };
 132
 133         void encode_thread_func();
 134         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 135         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 136                           int frame_type, int64_t pts, int64_t dts);
 137         void storage_task_thread();
 138         void storage_task_enqueue(storage_task task);
 139         void save_codeddata(storage_task task);
 140         int render_packedsequence();
 141         int render_packedpicture();
 142         void render_packedslice();
 143         int render_sequence();
 144         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 145         void sps_rbsp(bitstream *bs);
 146         void pps_rbsp(bitstream *bs);
 147         int build_packed_pic_buffer(unsigned char **header_buffer);
 148         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 149         void slice_header(bitstream *bs);
 150         int build_packed_seq_buffer(unsigned char **header_buffer);
 151         int build_packed_slice_buffer(unsigned char **header_buffer);
 152         int init_va(const string &va_display);
 153         int deinit_va();
 154         VADisplay va_open_display(const string &va_display);
 155         void va_close_display(VADisplay va_dpy);
 156         int setup_encode();
 157         int release_encode();
 158         void update_ReferenceFrames(int frame_type);
 159         int update_RefPicList(int frame_type);
 160
 161         bool is_shutdown = false;
 162         bool use_drm;
 163         int drm_fd = -1;
 164
 165         thread encode_thread, storage_thread;
 166
 167         mutex storage_task_queue_mutex;
 168         condition_variable storage_task_queue_changed;
 169         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 170         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 171         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 172
 173         mutex frame_queue_mutex;
 174         condition_variable frame_queue_nonempty;
 175         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 176
 177         int current_storage_frame;
 178
 179         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 180         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 181         QSurface *surface;
 182
 183         AVCodecContext *context_audio;
 184         HTTPD *httpd;
 185
 186         Display *x11_display;
 187         Window x11_window;
 188
 189         // Encoder parameters
 190         VADisplay va_dpy;
 191         VAProfile h264_profile = (VAProfile)~0;
 192         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 193         int config_attrib_num = 0, enc_packed_header_idx;
 194
 195         struct GLSurface {
 196                 VASurfaceID src_surface, ref_surface;
 197                 VABufferID coded_buf;
 198
 199                 VAImage surface_image;
 200                 GLuint y_tex, cbcr_tex;
 201                 EGLImage y_egl_image, cbcr_egl_image;
 202         };
 203         GLSurface gl_surfaces[SURFACE_NUM];
 204
 205         VAConfigID config_id;
 206         VAContextID context_id;
 207         VAEncSequenceParameterBufferH264 seq_param;
 208         VAEncPictureParameterBufferH264 pic_param;
 209         VAEncSliceParameterBufferH264 slice_param;
 210         VAPictureH264 CurrentCurrPic;
 211         VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
 212
 213         // Static quality settings.
 214         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 215         static constexpr unsigned int num_ref_frames = 2;
 216         static constexpr int initial_qp = 15;
 217         static constexpr int minimal_qp = 0;
 218         static constexpr int intra_period = 30;
 219         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 220
 221         // Quality settings that are meant to be static, but might be overridden
 222         // by the profile.
 223         int constraint_set_flag = 0;
 224         int h264_packedheader = 0; /* support pack header? */
 225         int h264_maxref = (1<<16|1);
 226         int h264_entropy_mode = 1; /* cabac */
 227         int ip_period = 3;
 228
 229         int rc_mode = -1;
 230         unsigned int current_frame_num = 0;
 231         unsigned int numShortTerm = 0;
 232
 233         int frame_width;
 234         int frame_height;
 235         int frame_width_mbaligned;
 236         int frame_height_mbaligned;
 237 };
 238
 239 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 240 // but if we don't delete it here, we get leaks. The GStreamer implementation
 241 // does the same.
 242 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 243 {
 244     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 245     CHECK_VASTATUS(va_status, "vaRenderPicture");
 246
 247     for (int i = 0; i < num_buffers; ++i) {
 248         va_status = vaDestroyBuffer(dpy, buffers[i]);
 249         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 250     }
 251 }
 252
 253 static unsigned int
 254 va_swap32(unsigned int val)
 255 {
 256     unsigned char *pval = (unsigned char *)&val;
 257
 258     return ((pval[0] << 24)     |
 259             (pval[1] << 16)     |
 260             (pval[2] << 8)      |
 261             (pval[3] << 0));
 262 }
 263
 264 static void
 265 bitstream_start(bitstream *bs)
 266 {
 267     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 268     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 269     bs->bit_offset = 0;
 270 }
 271
 272 static void
 273 bitstream_end(bitstream *bs)
 274 {
 275     int pos = (bs->bit_offset >> 5);
 276     int bit_offset = (bs->bit_offset & 0x1f);
 277     int bit_left = 32 - bit_offset;
 278
 279     if (bit_offset) {
 280         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 281     }
 282 }
 283
 284 static void
 285 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 286 {
 287     int pos = (bs->bit_offset >> 5);
 288     int bit_offset = (bs->bit_offset & 0x1f);
 289     int bit_left = 32 - bit_offset;
 290
 291     if (!size_in_bits)
 292         return;
 293
 294     bs->bit_offset += size_in_bits;
 295
 296     if (bit_left > size_in_bits) {
 297         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 298     } else {
 299         size_in_bits -= bit_left;
 300         if (bit_left >= 32) {
 301             bs->buffer[pos] = (val >> size_in_bits);
 302         } else {
 303             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 304         }
 305         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 306
 307         if (pos + 1 == bs->max_size_in_dword) {
 308             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 309             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 310         }
 311
 312         bs->buffer[pos + 1] = val;
 313     }
 314 }
 315
 316 static void
 317 bitstream_put_ue(bitstream *bs, unsigned int val)
 318 {
 319     int size_in_bits = 0;
 320     int tmp_val = ++val;
 321
 322     while (tmp_val) {
 323         tmp_val >>= 1;
 324         size_in_bits++;
 325     }
 326
 327     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 328     bitstream_put_ui(bs, val, size_in_bits);
 329 }
 330
 331 static void
 332 bitstream_put_se(bitstream *bs, int val)
 333 {
 334     unsigned int new_val;
 335
 336     if (val <= 0)
 337         new_val = -2 * val;
 338     else
 339         new_val = 2 * val - 1;
 340
 341     bitstream_put_ue(bs, new_val);
 342 }
 343
 344 static void
 345 bitstream_byte_aligning(bitstream *bs, int bit)
 346 {
 347     int bit_offset = (bs->bit_offset & 0x7);
 348     int bit_left = 8 - bit_offset;
 349     int new_val;
 350
 351     if (!bit_offset)
 352         return;
 353
 354     assert(bit == 0 || bit == 1);
 355
 356     if (bit)
 357         new_val = (1 << bit_left) - 1;
 358     else
 359         new_val = 0;
 360
 361     bitstream_put_ui(bs, new_val, bit_left);
 362 }
 363
 364 static void
 365 rbsp_trailing_bits(bitstream *bs)
 366 {
 367     bitstream_put_ui(bs, 1, 1);
 368     bitstream_byte_aligning(bs, 0);
 369 }
 370
 371 static void nal_start_code_prefix(bitstream *bs)
 372 {
 373     bitstream_put_ui(bs, 0x00000001, 32);
 374 }
 375
 376 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 377 {
 378     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 379     bitstream_put_ui(bs, nal_ref_idc, 2);
 380     bitstream_put_ui(bs, nal_unit_type, 5);
 381 }
 382
 383 void H264EncoderImpl::sps_rbsp(bitstream *bs)
 384 {
 385     int profile_idc = PROFILE_IDC_BASELINE;
 386
 387     if (h264_profile  == VAProfileH264High)
 388         profile_idc = PROFILE_IDC_HIGH;
 389     else if (h264_profile  == VAProfileH264Main)
 390         profile_idc = PROFILE_IDC_MAIN;
 391
 392     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 393     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 394     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 395     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 396     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 397     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 398     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 399     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 400
 401     if ( profile_idc == PROFILE_IDC_HIGH) {
 402         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 403         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 404         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 405         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 406         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 407     }
 408
 409     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 410     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 411
 412     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 413         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 414     else {
 415         assert(0);
 416     }
 417
 418     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 419     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 420
 421     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 422     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 423     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 424
 425     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 426         assert(0);
 427     }
 428
 429     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 430     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 431
 432     if (seq_param.frame_cropping_flag) {
 433         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 434         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 435         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 436         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 437     }
 438
 439     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 440     if ( false ) {
 441         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 442     } else {
 443         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 444         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 445         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 446         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 447         {
 448             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 449             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 450             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 451             {
 452                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 453                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 454                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 455             }
 456         }
 457         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 458         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 459         {
 460             bitstream_put_ui(bs, 1, 32);  // FPS
 461             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 462             bitstream_put_ui(bs, 1, 1);
 463         }
 464         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 465         {
 466             // hrd_parameters
 467             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 468             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 469             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 470
 471             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 472             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 473             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 474
 475             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 476             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 477             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 478             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 479         }
 480         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 481         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 482
 483         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 484         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 485     }
 486
 487     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 488 }
 489
 490
 491 void H264EncoderImpl::pps_rbsp(bitstream *bs)
 492 {
 493     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 494     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 495
 496     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 497
 498     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 499
 500     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 501
 502     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 503     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 504
 505     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 506     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 507
 508     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 509     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 510     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 511
 512     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 513     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 514     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 515
 516     /* more_rbsp_data */
 517     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 518     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 519     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 520
 521     rbsp_trailing_bits(bs);
 522 }
 523
 524 void H264EncoderImpl::slice_header(bitstream *bs)
 525 {
 526     int first_mb_in_slice = slice_param.macroblock_address;
 527
 528     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 529     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 530     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 531     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 532
 533     /* frame_mbs_only_flag == 1 */
 534     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 535         /* FIXME: */
 536         assert(0);
 537     }
 538
 539     if (pic_param.pic_fields.bits.idr_pic_flag)
 540         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 541
 542     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 543         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 544         /* pic_order_present_flag == 0 */
 545     } else {
 546         /* FIXME: */
 547         assert(0);
 548     }
 549
 550     /* redundant_pic_cnt_present_flag == 0 */
 551     /* slice type */
 552     if (IS_P_SLICE(slice_param.slice_type)) {
 553         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 554
 555         if (slice_param.num_ref_idx_active_override_flag)
 556             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 557
 558         /* ref_pic_list_reordering */
 559         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 560     } else if (IS_B_SLICE(slice_param.slice_type)) {
 561         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 562
 563         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 564
 565         if (slice_param.num_ref_idx_active_override_flag) {
 566             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 567             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 568         }
 569
 570         /* ref_pic_list_reordering */
 571         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 572         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 573     }
 574
 575     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 576          IS_P_SLICE(slice_param.slice_type)) ||
 577         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 578          IS_B_SLICE(slice_param.slice_type))) {
 579         /* FIXME: fill weight/offset table */
 580         assert(0);
 581     }
 582
 583     /* dec_ref_pic_marking */
 584     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 585         unsigned char no_output_of_prior_pics_flag = 0;
 586         unsigned char long_term_reference_flag = 0;
 587         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 588
 589         if (pic_param.pic_fields.bits.idr_pic_flag) {
 590             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 591             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 592         } else {
 593             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 594         }
 595     }
 596
 597     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 598         !IS_I_SLICE(slice_param.slice_type))
 599         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 600
 601     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 602
 603     /* ignore for SP/SI */
 604
 605     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 606         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 607
 608         if (slice_param.disable_deblocking_filter_idc != 1) {
 609             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 610             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 611         }
 612     }
 613
 614     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 615         bitstream_byte_aligning(bs, 1);
 616     }
 617 }
 618
 619 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 620 {
 621     bitstream bs;
 622
 623     bitstream_start(&bs);
 624     nal_start_code_prefix(&bs);
 625     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 626     pps_rbsp(&bs);
 627     bitstream_end(&bs);
 628
 629     *header_buffer = (unsigned char *)bs.buffer;
 630     return bs.bit_offset;
 631 }
 632
 633 int
 634 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 635 {
 636     bitstream bs;
 637
 638     bitstream_start(&bs);
 639     nal_start_code_prefix(&bs);
 640     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 641     sps_rbsp(&bs);
 642     bitstream_end(&bs);
 643
 644     *header_buffer = (unsigned char *)bs.buffer;
 645     return bs.bit_offset;
 646 }
 647
 648 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 649 {
 650     bitstream bs;
 651     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 652     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 653
 654     bitstream_start(&bs);
 655     nal_start_code_prefix(&bs);
 656
 657     if (IS_I_SLICE(slice_param.slice_type)) {
 658         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 659     } else if (IS_P_SLICE(slice_param.slice_type)) {
 660         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 661     } else {
 662         assert(IS_B_SLICE(slice_param.slice_type));
 663         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 664     }
 665
 666     slice_header(&bs);
 667     bitstream_end(&bs);
 668
 669     *header_buffer = (unsigned char *)bs.buffer;
 670     return bs.bit_offset;
 671 }
 672
 673
 674 /*
 675   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 676   1) period between Frame #X and Frame #N = #X - #N
 677   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 678   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 679   4) intra_period and intra_idr_period take precedence over ip_period
 680   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 681      of I/IDR frames, see bellow examples
 682   -------------------------------------------------------------------
 683   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 684   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 685   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 686   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 687   1            1                ignored    IDR IDR IDR IDR...
 688   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 689   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 690   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 691                                               (PBB)(IBB)(PBB)(IBB)...
 692   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 693                                            IDRPPPPP IPPPPP IPPPPP...
 694   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 695                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 696                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 697                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 698                                            {IDR(PBB)(PBB)}                     (6/6/3)
 699                                            {IDR(PBB)(PBB)}.
 700 */
 701
 702 // General pts/dts strategy:
 703 //
 704 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 705 // bit tricky. We assume first of all that the frame rate never goes _above_
 706 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 707 // in at least this speed, as long at dts <= pts (the frame is not attempted
 708 // presented before it is decoded). Furthermore, we never have longer chains of
 709 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 710 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 711 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 712 // frame, with an order O <= C.)
 713 //
 714 // Many strategies are possible, but we establish these rules:
 715 //
 716 //  - Tip frames have dts = pts - (C-O)*N.
 717 //  - Non-tip frames have dts = dts_last + N.
 718 //
 719 // An example, with C=2 and N=10 and the data flow showed with arrows:
 720 //
 721 //        I  B  P  B  B  P
 722 //   pts: 30 40 50 60 70 80
 723 //        ↓  ↓     ↓
 724 //   dts: 10 30 20 60 50←40
 725 //         |  |  ↑        ↑
 726 //         `--|--'        |
 727 //             `----------'
 728 //
 729 // To show that this works fine also with irregular spacings, let's say that
 730 // the third frame is delayed a bit (something earlier was dropped). Now the
 731 // situation looks like this:
 732 //
 733 //        I  B  P  B  B   P
 734 //   pts: 30 40 80 90 100 110
 735 //        ↓  ↓     ↓
 736 //   dts: 10 30 20 90 50←40
 737 //         |  |  ↑        ↑
 738 //         `--|--'        |
 739 //             `----------'
 740 //
 741 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 742 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 743 //
 744 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 745 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 746 // a tip frame and should be given a dts based on the previous one.
 747 #define FRAME_P 0
 748 #define FRAME_B 1
 749 #define FRAME_I 2
 750 #define FRAME_IDR 7
 751 void encoding2display_order(
 752     int encoding_order, int intra_period,
 753     int intra_idr_period, int ip_period,
 754     int *displaying_order,
 755     int *frame_type, int *pts_lag)
 756 {
 757     int encoding_order_gop = 0;
 758
 759     *pts_lag = 0;
 760
 761     if (intra_period == 1) { /* all are I/IDR frames */
 762         *displaying_order = encoding_order;
 763         if (intra_idr_period == 0)
 764             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 765         else
 766             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 767         return;
 768     }
 769
 770     if (intra_period == 0)
 771         intra_idr_period = 0;
 772
 773     if (ip_period == 1) {
 774         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 775         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 776         *displaying_order = encoding_order;
 777
 778         if (encoding_order_gop == 0) { /* the first frame */
 779             *frame_type = FRAME_IDR;
 780         } else if (intra_period != 0 && /* have I frames */
 781                    encoding_order_gop >= 2 &&
 782                    (encoding_order_gop % intra_period == 0)) {
 783             *frame_type = FRAME_I;
 784         } else {
 785             *frame_type = FRAME_P;
 786         }
 787         return;
 788     }
 789
 790     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 791     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 792     *pts_lag = -1;  // Most frames are not tip frames.
 793
 794     if (encoding_order_gop == 0) { /* the first frame */
 795         *frame_type = FRAME_IDR;
 796         *displaying_order = encoding_order;
 797         // IDR frames are a special case; I honestly can't find the logic behind
 798         // why this is the right thing, but it seems to line up nicely in practice :-)
 799         *pts_lag = TIMEBASE / MAX_FPS;
 800     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 801         *frame_type = FRAME_B;
 802         *displaying_order = encoding_order - 1;
 803         if ((encoding_order_gop % ip_period) == 0) {
 804             *pts_lag = 0;  // Last B-frame.
 805         }
 806     } else if (intra_period != 0 && /* have I frames */
 807                encoding_order_gop >= 2 &&
 808                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 809         *frame_type = FRAME_I;
 810         *displaying_order = encoding_order + ip_period - 1;
 811     } else {
 812         *frame_type = FRAME_P;
 813         *displaying_order = encoding_order + ip_period - 1;
 814     }
 815 }
 816
 817
 818 static const char *rc_to_string(int rc_mode)
 819 {
 820     switch (rc_mode) {
 821     case VA_RC_NONE:
 822         return "NONE";
 823     case VA_RC_CBR:
 824         return "CBR";
 825     case VA_RC_VBR:
 826         return "VBR";
 827     case VA_RC_VCM:
 828         return "VCM";
 829     case VA_RC_CQP:
 830         return "CQP";
 831     case VA_RC_VBR_CONSTRAINED:
 832         return "VBR_CONSTRAINED";
 833     default:
 834         return "Unknown";
 835     }
 836 }
 837
 838 VADisplay H264EncoderImpl::va_open_display(const string &va_display)
 839 {
 840         if (va_display.empty()) {
 841                 x11_display = XOpenDisplay(NULL);
 842                 if (!x11_display) {
 843                         fprintf(stderr, "error: can't connect to X server!\n");
 844                         return NULL;
 845                 }
 846                 use_drm = false;
 847                 return vaGetDisplay(x11_display);
 848         } else if (va_display[0] != '/') {
 849                 x11_display = XOpenDisplay(va_display.c_str());
 850                 if (!x11_display) {
 851                         fprintf(stderr, "error: can't connect to X server!\n");
 852                         return NULL;
 853                 }
 854                 use_drm = false;
 855                 return vaGetDisplay(x11_display);
 856         } else {
 857                 drm_fd = open(va_display.c_str(), O_RDWR);
 858                 if (drm_fd == -1) {
 859                         perror(va_display.c_str());
 860                         return NULL;
 861                 }
 862                 use_drm = true;
 863                 return vaGetDisplayDRM(drm_fd);
 864         }
 865 }
 866
 867 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
 868 {
 869     if (!x11_display)
 870         return;
 871
 872     if (x11_window) {
 873         XUnmapWindow(x11_display, x11_window);
 874         XDestroyWindow(x11_display, x11_window);
 875         x11_window = None;
 876     }
 877     XCloseDisplay(x11_display);
 878     x11_display = NULL;
 879 }
 880
 881 int H264EncoderImpl::init_va(const string &va_display)
 882 {
 883     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 884     VAEntrypoint *entrypoints;
 885     int num_entrypoints, slice_entrypoint;
 886     int support_encode = 0;
 887     int major_ver, minor_ver;
 888     VAStatus va_status;
 889     unsigned int i;
 890
 891     va_dpy = va_open_display(va_display);
 892     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 893     CHECK_VASTATUS(va_status, "vaInitialize");
 894
 895     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 896     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 897     if (!entrypoints) {
 898         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 899         exit(1);
 900     }
 901
 902     /* use the highest profile */
 903     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 904         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 905             continue;
 906
 907         h264_profile = profile_list[i];
 908         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 909         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 910             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 911                 support_encode = 1;
 912                 break;
 913             }
 914         }
 915         if (support_encode == 1)
 916             break;
 917     }
 918
 919     if (support_encode == 0) {
 920         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 921         exit(1);
 922     } else {
 923         switch (h264_profile) {
 924             case VAProfileH264Baseline:
 925                 ip_period = 1;
 926                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 927                 h264_entropy_mode = 0;
 928                 break;
 929             case VAProfileH264ConstrainedBaseline:
 930                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 931                 ip_period = 1;
 932                 break;
 933
 934             case VAProfileH264Main:
 935                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 936                 break;
 937
 938             case VAProfileH264High:
 939                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 940                 break;
 941             default:
 942                 h264_profile = VAProfileH264Baseline;
 943                 ip_period = 1;
 944                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 945                 break;
 946         }
 947     }
 948
 949     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 950
 951     /* find out the format for the render target, and rate control mode */
 952     for (i = 0; i < VAConfigAttribTypeMax; i++)
 953         attrib[i].type = (VAConfigAttribType)i;
 954
 955     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 956                                       &attrib[0], VAConfigAttribTypeMax);
 957     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 958     /* check the interested configattrib */
 959     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 960         printf("Not find desired YUV420 RT format\n");
 961         exit(1);
 962     } else {
 963         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 964         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 965         config_attrib_num++;
 966     }
 967
 968     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 969         int tmp = attrib[VAConfigAttribRateControl].value;
 970
 971         if (rc_mode == -1 || !(rc_mode & tmp))  {
 972             if (rc_mode != -1) {
 973                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
 974             }
 975
 976             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
 977                 if (rc_default_modes[i] & tmp) {
 978                     rc_mode = rc_default_modes[i];
 979                     break;
 980                 }
 981             }
 982         }
 983
 984         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 985         config_attrib[config_attrib_num].value = rc_mode;
 986         config_attrib_num++;
 987     }
 988
 989
 990     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 991         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 992
 993         h264_packedheader = 1;
 994         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 995         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 996
 997         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 998             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 999         }
1000
1001         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1002             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1003         }
1004
1005         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1006             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1007         }
1008
1009         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1010             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1011         }
1012
1013         enc_packed_header_idx = config_attrib_num;
1014         config_attrib_num++;
1015     }
1016
1017     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1018         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1019         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1020         config_attrib_num++;
1021     }
1022
1023     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1024         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1025     }
1026
1027     free(entrypoints);
1028     return 0;
1029 }
1030
1031 int H264EncoderImpl::setup_encode()
1032 {
1033     VAStatus va_status;
1034     VASurfaceID *tmp_surfaceid;
1035     int codedbuf_size, i;
1036     static VASurfaceID src_surface[SURFACE_NUM];
1037     static VASurfaceID ref_surface[SURFACE_NUM];
1038
1039     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1040             &config_attrib[0], config_attrib_num, &config_id);
1041     CHECK_VASTATUS(va_status, "vaCreateConfig");
1042
1043     /* create source surfaces */
1044     va_status = vaCreateSurfaces(va_dpy,
1045                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1046                                  &src_surface[0], SURFACE_NUM,
1047                                  NULL, 0);
1048     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1049
1050     /* create reference surfaces */
1051     va_status = vaCreateSurfaces(va_dpy,
1052                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1053                                  &ref_surface[0], SURFACE_NUM,
1054                                  NULL, 0);
1055     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1056
1057     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1058     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1059     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1060
1061     /* Create a context for this encode pipe */
1062     va_status = vaCreateContext(va_dpy, config_id,
1063                                 frame_width_mbaligned, frame_height_mbaligned,
1064                                 VA_PROGRESSIVE,
1065                                 tmp_surfaceid, 2 * SURFACE_NUM,
1066                                 &context_id);
1067     CHECK_VASTATUS(va_status, "vaCreateContext");
1068     free(tmp_surfaceid);
1069
1070     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1071
1072     for (i = 0; i < SURFACE_NUM; i++) {
1073         /* create coded buffer once for all
1074          * other VA buffers which won't be used again after vaRenderPicture.
1075          * so APP can always vaCreateBuffer for every frame
1076          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1077          * so VA won't maintain the coded buffer
1078          */
1079         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1080                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1081         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1082     }
1083
1084     /* create OpenGL objects */
1085     //glGenFramebuffers(SURFACE_NUM, fbos);
1086
1087     for (i = 0; i < SURFACE_NUM; i++) {
1088         glGenTextures(1, &gl_surfaces[i].y_tex);
1089         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1090     }
1091
1092     for (i = 0; i < SURFACE_NUM; i++) {
1093         gl_surfaces[i].src_surface = src_surface[i];
1094         gl_surfaces[i].ref_surface = ref_surface[i];
1095     }
1096
1097     return 0;
1098 }
1099
1100 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1101 //
1102 //   2 1 0 [3] 4 8 9
1103 template<class T, class C>
1104 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1105 {
1106         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1107         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1108         sort(middle, end, less_than);
1109 }
1110
1111 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1112 {
1113     int i;
1114
1115     if (frame_type == FRAME_B)
1116         return;
1117
1118     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1119     numShortTerm++;
1120     if (numShortTerm > num_ref_frames)
1121         numShortTerm = num_ref_frames;
1122     for (i=numShortTerm-1; i>0; i--)
1123         ReferenceFrames[i] = ReferenceFrames[i-1];
1124     ReferenceFrames[0] = CurrentCurrPic;
1125
1126     current_frame_num++;
1127     if (current_frame_num > MaxFrameNum)
1128         current_frame_num = 0;
1129 }
1130
1131
1132 int H264EncoderImpl::update_RefPicList(int frame_type)
1133 {
1134     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1135         return a.frame_idx > b.frame_idx;
1136     };
1137     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1138         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1139     };
1140     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1141         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1142     };
1143
1144     if (frame_type == FRAME_P) {
1145         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1146         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1147     } else if (frame_type == FRAME_B) {
1148         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1149         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1150
1151         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1152         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1153     }
1154
1155     return 0;
1156 }
1157
1158
1159 int H264EncoderImpl::render_sequence()
1160 {
1161     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1162     VAStatus va_status;
1163     VAEncMiscParameterBuffer *misc_param;
1164     VAEncMiscParameterRateControl *misc_rate_ctrl;
1165
1166     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1167     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1168     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1169     seq_param.bits_per_second = frame_bitrate;
1170
1171     seq_param.intra_period = intra_period;
1172     seq_param.intra_idr_period = intra_idr_period;
1173     seq_param.ip_period = ip_period;
1174
1175     seq_param.max_num_ref_frames = num_ref_frames;
1176     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1177     seq_param.time_scale = TIMEBASE * 2;
1178     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1179     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1180     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1181     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1182     seq_param.seq_fields.bits.chroma_format_idc = 1;
1183     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1184
1185     if (frame_width != frame_width_mbaligned ||
1186         frame_height != frame_height_mbaligned) {
1187         seq_param.frame_cropping_flag = 1;
1188         seq_param.frame_crop_left_offset = 0;
1189         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1190         seq_param.frame_crop_top_offset = 0;
1191         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1192     }
1193
1194     va_status = vaCreateBuffer(va_dpy, context_id,
1195                                VAEncSequenceParameterBufferType,
1196                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1197     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1198
1199     va_status = vaCreateBuffer(va_dpy, context_id,
1200                                VAEncMiscParameterBufferType,
1201                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1202                                1, NULL, &rc_param_buf);
1203     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1204
1205     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1206     misc_param->type = VAEncMiscParameterTypeRateControl;
1207     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1208     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1209     misc_rate_ctrl->bits_per_second = frame_bitrate;
1210     misc_rate_ctrl->target_percentage = 66;
1211     misc_rate_ctrl->window_size = 1000;
1212     misc_rate_ctrl->initial_qp = initial_qp;
1213     misc_rate_ctrl->min_qp = minimal_qp;
1214     misc_rate_ctrl->basic_unit_size = 0;
1215     vaUnmapBuffer(va_dpy, rc_param_buf);
1216
1217     render_id[0] = seq_param_buf;
1218     render_id[1] = rc_param_buf;
1219
1220     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1221
1222     return 0;
1223 }
1224
1225 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1226 {
1227     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1228     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1229     int PicOrderCntMsb, TopFieldOrderCnt;
1230
1231     if (frame_type == FRAME_IDR)
1232         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1233     else {
1234         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1235         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1236     }
1237
1238     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1239         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1240         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1241     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1242              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1243         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1244     else
1245         PicOrderCntMsb = prevPicOrderCntMsb;
1246
1247     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1248
1249     if (frame_type != FRAME_B) {
1250         PicOrderCntMsb_ref = PicOrderCntMsb;
1251         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1252     }
1253
1254     return TopFieldOrderCnt;
1255 }
1256
1257 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1258 {
1259     VABufferID pic_param_buf;
1260     VAStatus va_status;
1261     int i = 0;
1262
1263     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1264     pic_param.CurrPic.frame_idx = current_frame_num;
1265     pic_param.CurrPic.flags = 0;
1266     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1267     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1268     CurrentCurrPic = pic_param.CurrPic;
1269
1270     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1271     for (i = numShortTerm; i < SURFACE_NUM; i++) {
1272         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1273         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1274     }
1275
1276     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1277     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1278     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1279     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1280     pic_param.frame_num = current_frame_num;
1281     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1282     pic_param.last_picture = false;  // FIXME
1283     pic_param.pic_init_qp = initial_qp;
1284
1285     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1286                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1287     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1288
1289     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1290
1291     return 0;
1292 }
1293
1294 int H264EncoderImpl::render_packedsequence()
1295 {
1296     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1297     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1298     unsigned int length_in_bits;
1299     unsigned char *packedseq_buffer = NULL;
1300     VAStatus va_status;
1301
1302     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1303
1304     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1305
1306     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1307     packedheader_param_buffer.has_emulation_bytes = 0;
1308     va_status = vaCreateBuffer(va_dpy,
1309                                context_id,
1310                                VAEncPackedHeaderParameterBufferType,
1311                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1312                                &packedseq_para_bufid);
1313     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1314
1315     va_status = vaCreateBuffer(va_dpy,
1316                                context_id,
1317                                VAEncPackedHeaderDataBufferType,
1318                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1319                                &packedseq_data_bufid);
1320     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1321
1322     render_id[0] = packedseq_para_bufid;
1323     render_id[1] = packedseq_data_bufid;
1324     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1325
1326     free(packedseq_buffer);
1327
1328     return 0;
1329 }
1330
1331
1332 int H264EncoderImpl::render_packedpicture()
1333 {
1334     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1335     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1336     unsigned int length_in_bits;
1337     unsigned char *packedpic_buffer = NULL;
1338     VAStatus va_status;
1339
1340     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1341     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1342     packedheader_param_buffer.bit_length = length_in_bits;
1343     packedheader_param_buffer.has_emulation_bytes = 0;
1344
1345     va_status = vaCreateBuffer(va_dpy,
1346                                context_id,
1347                                VAEncPackedHeaderParameterBufferType,
1348                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1349                                &packedpic_para_bufid);
1350     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1351
1352     va_status = vaCreateBuffer(va_dpy,
1353                                context_id,
1354                                VAEncPackedHeaderDataBufferType,
1355                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1356                                &packedpic_data_bufid);
1357     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1358
1359     render_id[0] = packedpic_para_bufid;
1360     render_id[1] = packedpic_data_bufid;
1361     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1362
1363     free(packedpic_buffer);
1364
1365     return 0;
1366 }
1367
1368 void H264EncoderImpl::render_packedslice()
1369 {
1370     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1371     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1372     unsigned int length_in_bits;
1373     unsigned char *packedslice_buffer = NULL;
1374     VAStatus va_status;
1375
1376     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1377     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1378     packedheader_param_buffer.bit_length = length_in_bits;
1379     packedheader_param_buffer.has_emulation_bytes = 0;
1380
1381     va_status = vaCreateBuffer(va_dpy,
1382                                context_id,
1383                                VAEncPackedHeaderParameterBufferType,
1384                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1385                                &packedslice_para_bufid);
1386     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1387
1388     va_status = vaCreateBuffer(va_dpy,
1389                                context_id,
1390                                VAEncPackedHeaderDataBufferType,
1391                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1392                                &packedslice_data_bufid);
1393     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1394
1395     render_id[0] = packedslice_para_bufid;
1396     render_id[1] = packedslice_data_bufid;
1397     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1398
1399     free(packedslice_buffer);
1400 }
1401
1402 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1403 {
1404     VABufferID slice_param_buf;
1405     VAStatus va_status;
1406     int i;
1407
1408     update_RefPicList(frame_type);
1409
1410     /* one frame, one slice */
1411     slice_param.macroblock_address = 0;
1412     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1413     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1414     if (frame_type == FRAME_IDR) {
1415         if (encoding_frame_num != 0)
1416             ++slice_param.idr_pic_id;
1417     } else if (frame_type == FRAME_P) {
1418         int refpiclist0_max = h264_maxref & 0xffff;
1419         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1420
1421         for (i = refpiclist0_max; i < 32; i++) {
1422             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1423             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1424         }
1425     } else if (frame_type == FRAME_B) {
1426         int refpiclist0_max = h264_maxref & 0xffff;
1427         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1428
1429         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1430         for (i = refpiclist0_max; i < 32; i++) {
1431             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1432             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1433         }
1434
1435         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1436         for (i = refpiclist1_max; i < 32; i++) {
1437             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1438             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1439         }
1440     }
1441
1442     slice_param.slice_alpha_c0_offset_div2 = 0;
1443     slice_param.slice_beta_offset_div2 = 0;
1444     slice_param.direct_spatial_mv_pred_flag = 1;
1445     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1446
1447
1448     if (h264_packedheader &&
1449         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1450         render_packedslice();
1451
1452     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1453                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1454     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1455
1456     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1457
1458     return 0;
1459 }
1460
1461
1462
1463 void H264EncoderImpl::save_codeddata(storage_task task)
1464 {
1465     VACodedBufferSegment *buf_list = NULL;
1466     VAStatus va_status;
1467
1468     string data;
1469
1470     const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1471
1472     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1473     CHECK_VASTATUS(va_status, "vaMapBuffer");
1474     while (buf_list != NULL) {
1475         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1476         buf_list = (VACodedBufferSegment *) buf_list->next;
1477     }
1478     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1479
1480     {
1481         // Add video.
1482         AVPacket pkt;
1483         memset(&pkt, 0, sizeof(pkt));
1484         pkt.buf = nullptr;
1485         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1486         pkt.size = data.size();
1487         pkt.stream_index = 0;
1488         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1489             pkt.flags = AV_PKT_FLAG_KEY;
1490         } else {
1491             pkt.flags = 0;
1492         }
1493         //pkt.duration = 1;
1494         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1495     }
1496     // Encode and add all audio frames up to and including the pts of this video frame.
1497     for ( ;; ) {
1498         int64_t audio_pts;
1499         vector<float> audio;
1500         {
1501              unique_lock<mutex> lock(frame_queue_mutex);
1502              frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1503              if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1504              auto it = pending_audio_frames.begin();
1505              if (it->first > task.pts) break;
1506              audio_pts = it->first;
1507              audio = move(it->second);
1508              pending_audio_frames.erase(it);
1509         }
1510
1511         AVFrame *frame = avcodec_alloc_frame();
1512         frame->nb_samples = audio.size() / 2;
1513         frame->format = AV_SAMPLE_FMT_S32;
1514         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1515
1516         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1517         int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1518         if (ret < 0) {
1519             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1520             exit(1);
1521         }
1522         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1523             if (audio[i] >= 1.0f) {
1524                 int_samples[i] = 2147483647;
1525             } else if (audio[i] <= -1.0f) {
1526                 int_samples[i] = -2147483647;
1527             } else {
1528                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1529             }
1530         }
1531
1532         AVPacket pkt;
1533         av_init_packet(&pkt);
1534         pkt.data = nullptr;
1535         pkt.size = 0;
1536         int got_output;
1537         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1538         if (got_output) {
1539             pkt.stream_index = 1;
1540             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1541         }
1542         // TODO: Delayed frames.
1543         avcodec_free_frame(&frame);
1544         av_free_packet(&pkt);
1545         if (audio_pts == task.pts) break;
1546     }
1547
1548 #if 0
1549     printf("\r      "); /* return back to startpoint */
1550     switch (encode_order % 4) {
1551         case 0:
1552             printf("|");
1553             break;
1554         case 1:
1555             printf("/");
1556             break;
1557         case 2:
1558             printf("-");
1559             break;
1560         case 3:
1561             printf("\\");
1562             break;
1563     }
1564     printf("%08lld", encode_order);
1565 #endif
1566 }
1567
1568
1569 // this is weird. but it seems to put a new frame onto the queue
1570 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1571 {
1572         unique_lock<mutex> lock(storage_task_queue_mutex);
1573         storage_task_queue.push(move(task));
1574         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1575         storage_task_queue_changed.notify_all();
1576 }
1577
1578 void H264EncoderImpl::storage_task_thread()
1579 {
1580         for ( ;; ) {
1581                 storage_task current;
1582                 {
1583                         // wait until there's an encoded frame
1584                         unique_lock<mutex> lock(storage_task_queue_mutex);
1585                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1586                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1587                         current = move(storage_task_queue.front());
1588                         storage_task_queue.pop();
1589                 }
1590
1591                 VAStatus va_status;
1592
1593                 // waits for data, then saves it to disk.
1594                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1595                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1596                 save_codeddata(move(current));
1597
1598                 {
1599                         unique_lock<mutex> lock(storage_task_queue_mutex);
1600                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1601                         storage_task_queue_changed.notify_all();
1602                 }
1603         }
1604 }
1605
1606 int H264EncoderImpl::release_encode()
1607 {
1608     int i;
1609
1610     for (i = 0; i < SURFACE_NUM; i++) {
1611         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1612         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1613         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1614     }
1615
1616     vaDestroyContext(va_dpy, context_id);
1617     vaDestroyConfig(va_dpy, config_id);
1618
1619     return 0;
1620 }
1621
1622 int H264EncoderImpl::deinit_va()
1623 {
1624     vaTerminate(va_dpy);
1625
1626     va_close_display(va_dpy);
1627
1628     return 0;
1629 }
1630
1631
1632 H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1633         : current_storage_frame(0), surface(surface), httpd(httpd)
1634 {
1635         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1636         context_audio = avcodec_alloc_context3(codec_audio);
1637         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1638         context_audio->sample_rate = OUTPUT_FREQUENCY;
1639         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1640         context_audio->channels = 2;
1641         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1642         context_audio->time_base = AVRational{1, TIMEBASE};
1643         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1644                 fprintf(stderr, "Could not open codec\n");
1645                 exit(1);
1646         }
1647
1648         frame_width = width;
1649         frame_height = height;
1650         frame_width_mbaligned = (frame_width + 15) & (~15);
1651         frame_height_mbaligned = (frame_height + 15) & (~15);
1652
1653         //print_input();
1654
1655         init_va(va_display);
1656         setup_encode();
1657
1658         // No frames are ready yet.
1659         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1660
1661         memset(&seq_param, 0, sizeof(seq_param));
1662         memset(&pic_param, 0, sizeof(pic_param));
1663         memset(&slice_param, 0, sizeof(slice_param));
1664
1665         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1666
1667         encode_thread = thread([this]{
1668                 //SDL_GL_MakeCurrent(window, context);
1669                 QOpenGLContext *context = create_context(this->surface);
1670                 eglBindAPI(EGL_OPENGL_API);
1671                 if (!make_current(context, this->surface)) {
1672                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1673                                 eglGetError());
1674                         exit(1);
1675                 }
1676                 encode_thread_func();
1677         });
1678 }
1679
1680 H264EncoderImpl::~H264EncoderImpl()
1681 {
1682         shutdown();
1683 }
1684
1685 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1686 {
1687         assert(!is_shutdown);
1688         {
1689                 // Wait until this frame slot is done encoding.
1690                 unique_lock<mutex> lock(storage_task_queue_mutex);
1691                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1692                 if (storage_thread_should_quit) return false;
1693         }
1694
1695         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1696         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1697         *y_tex = surf->y_tex;
1698         *cbcr_tex = surf->cbcr_tex;
1699
1700         VASurfaceID surface = surf->src_surface;
1701         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1702         CHECK_VASTATUS(va_status, "vaDeriveImage");
1703
1704         VABufferInfo buf_info;
1705         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1706         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1707         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1708
1709         // Create Y image.
1710         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1711         EGLint y_attribs[] = {
1712                 EGL_WIDTH, frame_width,
1713                 EGL_HEIGHT, frame_height,
1714                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1715                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1716                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1717                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1718                 EGL_NONE
1719         };
1720
1721         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1722         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1723
1724         // Associate Y image to a texture.
1725         glBindTexture(GL_TEXTURE_2D, *y_tex);
1726         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1727
1728         // Create CbCr image.
1729         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1730         EGLint cbcr_attribs[] = {
1731                 EGL_WIDTH, frame_width,
1732                 EGL_HEIGHT, frame_height,
1733                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1734                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1735                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1736                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1737                 EGL_NONE
1738         };
1739
1740         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1741         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1742
1743         // Associate CbCr image to a texture.
1744         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1745         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1746
1747         return true;
1748 }
1749
1750 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
1751 {
1752         assert(!is_shutdown);
1753         {
1754                 unique_lock<mutex> lock(frame_queue_mutex);
1755                 pending_audio_frames[pts] = move(audio);
1756         }
1757         frame_queue_nonempty.notify_all();
1758 }
1759
1760 void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1761 {
1762         assert(!is_shutdown);
1763         {
1764                 unique_lock<mutex> lock(frame_queue_mutex);
1765                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1766                 ++current_storage_frame;
1767         }
1768         frame_queue_nonempty.notify_all();
1769 }
1770
1771 void H264EncoderImpl::shutdown()
1772 {
1773         if (is_shutdown) {
1774                 return;
1775         }
1776
1777         {
1778                 unique_lock<mutex> lock(frame_queue_mutex);
1779                 encode_thread_should_quit = true;
1780                 frame_queue_nonempty.notify_all();
1781         }
1782         encode_thread.join();
1783         {
1784                 unique_lock<mutex> lock(storage_task_queue_mutex);
1785                 storage_thread_should_quit = true;
1786                 frame_queue_nonempty.notify_all();
1787                 storage_task_queue_changed.notify_all();
1788         }
1789         storage_thread.join();
1790
1791         release_encode();
1792         deinit_va();
1793         is_shutdown = true;
1794 }
1795
1796 void H264EncoderImpl::encode_thread_func()
1797 {
1798         int64_t last_dts = -1;
1799         int gop_start_display_frame_num = 0;
1800         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1801                 PendingFrame frame;
1802                 int pts_lag;
1803                 int frame_type, display_frame_num;
1804                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1805                                        &display_frame_num, &frame_type, &pts_lag);
1806                 if (frame_type == FRAME_IDR) {
1807                         numShortTerm = 0;
1808                         current_frame_num = 0;
1809                         gop_start_display_frame_num = display_frame_num;
1810                 }
1811
1812                 {
1813                         unique_lock<mutex> lock(frame_queue_mutex);
1814                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1815                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1816                         });
1817                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1818                                 // We have queued frames that were supposed to be B-frames,
1819                                 // but will be no P-frame to encode them against. Encode them all
1820                                 // as P-frames instead. Note that this happens under the mutex,
1821                                 // but nobody else uses it at this point, since we're shutting down,
1822                                 // so there's no contention.
1823                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1824                                 return;
1825                         } else {
1826                                 frame = move(pending_video_frames[display_frame_num]);
1827                                 pending_video_frames.erase(display_frame_num);
1828                         }
1829                 }
1830
1831                 // Determine the dts of this frame.
1832                 int64_t dts;
1833                 if (pts_lag == -1) {
1834                         assert(last_dts != -1);
1835                         dts = last_dts + (TIMEBASE / MAX_FPS);
1836                 } else {
1837                         dts = frame.pts - pts_lag;
1838                 }
1839                 last_dts = dts;
1840
1841                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
1842         }
1843 }
1844
1845 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1846 {
1847         if (pending_video_frames.empty()) {
1848                 return;
1849         }
1850
1851         for (auto &pending_frame : pending_video_frames) {
1852                 int display_frame_num = pending_frame.first;
1853                 assert(display_frame_num > 0);
1854                 PendingFrame frame = move(pending_frame.second);
1855                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1856                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1857                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
1858                 last_dts = dts;
1859         }
1860 }
1861
1862 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1863                                    int frame_type, int64_t pts, int64_t dts)
1864 {
1865         // Wait for the GPU to be done with the frame.
1866         glClientWaitSync(frame.fence.get(), 0, 0);
1867
1868         // Release back any input frames we needed to render this frame.
1869         frame.input_frames.clear();
1870
1871         // Unmap the image.
1872         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1873         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1874         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1875         VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1876         CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1877         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1878         CHECK_VASTATUS(va_status, "vaDestroyImage");
1879
1880         VASurfaceID surface = surf->src_surface;
1881
1882         // Schedule the frame for encoding.
1883         va_status = vaBeginPicture(va_dpy, context_id, surface);
1884         CHECK_VASTATUS(va_status, "vaBeginPicture");
1885
1886         if (frame_type == FRAME_IDR) {
1887                 render_sequence();
1888                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1889                 if (h264_packedheader) {
1890                         render_packedsequence();
1891                         render_packedpicture();
1892                 }
1893         } else {
1894                 //render_sequence();
1895                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1896         }
1897         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1898
1899         va_status = vaEndPicture(va_dpy, context_id);
1900         CHECK_VASTATUS(va_status, "vaEndPicture");
1901
1902         // so now the data is done encoding (well, async job kicked off)...
1903         // we send that to the storage thread
1904         storage_task tmp;
1905         tmp.display_order = display_frame_num;
1906         tmp.frame_type = frame_type;
1907         tmp.pts = pts;
1908         tmp.dts = dts;
1909         storage_task_enqueue(move(tmp));
1910
1911         update_ReferenceFrames(frame_type);
1912 }
1913
1914 // Proxy object.
1915 H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1916         : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
1917
1918 // Must be defined here because unique_ptr<> destructor needs to know the impl.
1919 H264Encoder::~H264Encoder() {}
1920
1921 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
1922 {
1923         impl->add_audio(pts, audio);
1924 }
1925
1926 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1927 {
1928         return impl->begin_frame(y_tex, cbcr_tex);
1929 }
1930
1931 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1932 {
1933         impl->end_frame(fence, pts, input_frames);
1934 }
1935
1936 void H264Encoder::shutdown()
1937 {
1938         impl->shutdown();
1939 }
1940
1941 // Real class.