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