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