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