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 <condition_variable>
  26 #include <cstdint>
  27 #include <map>
  28 #include <memory>
  29 #include <mutex>
  30 #include <queue>
  31 #include <string>
  32 #include <thread>
  33 #include <utility>
  34
  35 #include "context.h"
  36 #include "defs.h"
  37 #include "httpd.h"
  38 #include "timebase.h"
  39
  40 using namespace std;
  41
  42 class QOpenGLContext;
  43 class QSurface;
  44
  45 #define CHECK_VASTATUS(va_status, func)                                 \
  46     if (va_status != VA_STATUS_SUCCESS) {                               \
  47         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  48         exit(1);                                                        \
  49     }
  50
  51 //#include "loadsurface.h"
  52
  53 #define NAL_REF_IDC_NONE        0
  54 #define NAL_REF_IDC_LOW         1
  55 #define NAL_REF_IDC_MEDIUM      2
  56 #define NAL_REF_IDC_HIGH        3
  57
  58 #define NAL_NON_IDR             1
  59 #define NAL_IDR                 5
  60 #define NAL_SPS                 7
  61 #define NAL_PPS                 8
  62 #define NAL_SEI                 6
  63
  64 #define SLICE_TYPE_P            0
  65 #define SLICE_TYPE_B            1
  66 #define SLICE_TYPE_I            2
  67 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  68 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  69 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  70
  71
  72 #define ENTROPY_MODE_CAVLC      0
  73 #define ENTROPY_MODE_CABAC      1
  74
  75 #define PROFILE_IDC_BASELINE    66
  76 #define PROFILE_IDC_MAIN        77
  77 #define PROFILE_IDC_HIGH        100
  78
  79 #define BITSTREAM_ALLOCATE_STEPPING     4096
  80
  81 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  82 static  VADisplay va_dpy;
  83 static  VAProfile h264_profile = (VAProfile)~0;
  84 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  85 static  int config_attrib_num = 0, enc_packed_header_idx;
  86
  87 struct GLSurface {
  88         VASurfaceID src_surface, ref_surface;
  89         VABufferID coded_buf;
  90
  91         VAImage surface_image;
  92         GLuint y_tex, cbcr_tex;
  93         EGLImage y_egl_image, cbcr_egl_image;
  94 };
  95 GLSurface gl_surfaces[SURFACE_NUM];
  96
  97 static  VAConfigID config_id;
  98 static  VAContextID context_id;
  99 static  VAEncSequenceParameterBufferH264 seq_param;
 100 static  VAEncPictureParameterBufferH264 pic_param;
 101 static  VAEncSliceParameterBufferH264 slice_param;
 102 static  VAPictureH264 CurrentCurrPic;
 103 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
 104
 105 static  unsigned int MaxFrameNum = (2<<16);
 106 static  unsigned int MaxPicOrderCntLsb = (2<<8);
 107 static  unsigned int Log2MaxFrameNum = 16;
 108 static  unsigned int Log2MaxPicOrderCntLsb = 8;
 109
 110 static  unsigned int num_ref_frames = 2;
 111 static  unsigned int numShortTerm = 0;
 112 static  int constraint_set_flag = 0;
 113 static  int h264_packedheader = 0; /* support pack header? */
 114 static  int h264_maxref = (1<<16|1);
 115 static  int h264_entropy_mode = 1; /* cabac */
 116
 117 static  int frame_width = 176;
 118 static  int frame_height = 144;
 119 static  int frame_width_mbaligned;
 120 static  int frame_height_mbaligned;
 121 static  unsigned int frame_bitrate = 0;
 122 static  double frame_size = 0;
 123 static  int initial_qp = 15;
 124 //static  int initial_qp = 28;
 125 static  int minimal_qp = 0;
 126 static  int intra_period = 30;
 127 static  int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 128 static  int ip_period = 3;
 129 static  int rc_mode = -1;
 130 static  int rc_default_modes[] = {
 131     VA_RC_VBR,
 132     VA_RC_CQP,
 133     VA_RC_VBR_CONSTRAINED,
 134     VA_RC_CBR,
 135     VA_RC_VCM,
 136     VA_RC_NONE,
 137 };
 138 static  unsigned int current_frame_num = 0;
 139
 140 static  int misc_priv_type = 0;
 141 static  int misc_priv_value = 0;
 142
 143 /* thread to save coded data */
 144 #define SRC_SURFACE_FREE        0
 145 #define SRC_SURFACE_IN_ENCODING 1
 146
 147 struct __bitstream {
 148     unsigned int *buffer;
 149     int bit_offset;
 150     int max_size_in_dword;
 151 };
 152 typedef struct __bitstream bitstream;
 153
 154 using namespace std;
 155
 156 class H264EncoderImpl {
 157 public:
 158         H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd);
 159         ~H264EncoderImpl();
 160         void add_audio(int64_t pts, std::vector<float> audio);  // Needs to come before end_frame() of same pts.
 161         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 162         void end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames);
 163
 164 private:
 165         struct storage_task {
 166                 unsigned long long display_order;
 167                 int frame_type;
 168                 std::vector<float> audio;
 169                 int64_t pts, dts;
 170         };
 171         struct PendingFrame {
 172                 RefCountedGLsync fence;
 173                 std::vector<RefCountedFrame> input_frames;
 174                 int64_t pts;
 175         };
 176
 177         void encode_thread_func();
 178         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 179         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 180                           int frame_type, int64_t pts, int64_t dts);
 181         void storage_task_thread();
 182         void storage_task_enqueue(storage_task task);
 183         void save_codeddata(storage_task task);
 184
 185         std::thread encode_thread, storage_thread;
 186
 187         std::mutex storage_task_queue_mutex;
 188         std::condition_variable storage_task_queue_changed;
 189         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 190         std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 191         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 192
 193         std::mutex frame_queue_mutex;
 194         std::condition_variable frame_queue_nonempty;
 195         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 196
 197         //int frame_width, frame_height;
 198         //int ;
 199         int current_storage_frame;
 200
 201         std::map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 202         std::map<int64_t, std::vector<float>> pending_audio_frames;  // under frame_queue_mutex
 203         QSurface *surface;
 204
 205         AVCodecContext *context_audio;
 206         HTTPD *httpd;
 207 };
 208
 209
 210 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 211 // but if we don't delete it here, we get leaks. The GStreamer implementation
 212 // does the same.
 213 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 214 {
 215     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 216     CHECK_VASTATUS(va_status, "vaRenderPicture");
 217
 218     for (int i = 0; i < num_buffers; ++i) {
 219         va_status = vaDestroyBuffer(va_dpy, buffers[i]);
 220         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 221     }
 222 }
 223
 224 static unsigned int
 225 va_swap32(unsigned int val)
 226 {
 227     unsigned char *pval = (unsigned char *)&val;
 228
 229     return ((pval[0] << 24)     |
 230             (pval[1] << 16)     |
 231             (pval[2] << 8)      |
 232             (pval[3] << 0));
 233 }
 234
 235 static void
 236 bitstream_start(bitstream *bs)
 237 {
 238     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 239     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 240     bs->bit_offset = 0;
 241 }
 242
 243 static void
 244 bitstream_end(bitstream *bs)
 245 {
 246     int pos = (bs->bit_offset >> 5);
 247     int bit_offset = (bs->bit_offset & 0x1f);
 248     int bit_left = 32 - bit_offset;
 249
 250     if (bit_offset) {
 251         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 252     }
 253 }
 254
 255 static void
 256 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 257 {
 258     int pos = (bs->bit_offset >> 5);
 259     int bit_offset = (bs->bit_offset & 0x1f);
 260     int bit_left = 32 - bit_offset;
 261
 262     if (!size_in_bits)
 263         return;
 264
 265     bs->bit_offset += size_in_bits;
 266
 267     if (bit_left > size_in_bits) {
 268         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 269     } else {
 270         size_in_bits -= bit_left;
 271         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 272         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 273
 274         if (pos + 1 == bs->max_size_in_dword) {
 275             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 276             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 277         }
 278
 279         bs->buffer[pos + 1] = val;
 280     }
 281 }
 282
 283 static void
 284 bitstream_put_ue(bitstream *bs, unsigned int val)
 285 {
 286     int size_in_bits = 0;
 287     int tmp_val = ++val;
 288
 289     while (tmp_val) {
 290         tmp_val >>= 1;
 291         size_in_bits++;
 292     }
 293
 294     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 295     bitstream_put_ui(bs, val, size_in_bits);
 296 }
 297
 298 static void
 299 bitstream_put_se(bitstream *bs, int val)
 300 {
 301     unsigned int new_val;
 302
 303     if (val <= 0)
 304         new_val = -2 * val;
 305     else
 306         new_val = 2 * val - 1;
 307
 308     bitstream_put_ue(bs, new_val);
 309 }
 310
 311 static void
 312 bitstream_byte_aligning(bitstream *bs, int bit)
 313 {
 314     int bit_offset = (bs->bit_offset & 0x7);
 315     int bit_left = 8 - bit_offset;
 316     int new_val;
 317
 318     if (!bit_offset)
 319         return;
 320
 321     assert(bit == 0 || bit == 1);
 322
 323     if (bit)
 324         new_val = (1 << bit_left) - 1;
 325     else
 326         new_val = 0;
 327
 328     bitstream_put_ui(bs, new_val, bit_left);
 329 }
 330
 331 static void
 332 rbsp_trailing_bits(bitstream *bs)
 333 {
 334     bitstream_put_ui(bs, 1, 1);
 335     bitstream_byte_aligning(bs, 0);
 336 }
 337
 338 static void nal_start_code_prefix(bitstream *bs)
 339 {
 340     bitstream_put_ui(bs, 0x00000001, 32);
 341 }
 342
 343 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 344 {
 345     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 346     bitstream_put_ui(bs, nal_ref_idc, 2);
 347     bitstream_put_ui(bs, nal_unit_type, 5);
 348 }
 349
 350 static void sps_rbsp(bitstream *bs)
 351 {
 352     int profile_idc = PROFILE_IDC_BASELINE;
 353
 354     if (h264_profile  == VAProfileH264High)
 355         profile_idc = PROFILE_IDC_HIGH;
 356     else if (h264_profile  == VAProfileH264Main)
 357         profile_idc = PROFILE_IDC_MAIN;
 358
 359     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 360     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 361     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 362     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 363     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 364     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 365     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 366     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 367
 368     if ( profile_idc == PROFILE_IDC_HIGH) {
 369         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 370         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 371         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 372         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 373         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 374     }
 375
 376     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 377     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 378
 379     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 380         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 381     else {
 382         assert(0);
 383     }
 384
 385     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 386     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 387
 388     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 389     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 390     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 391
 392     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 393         assert(0);
 394     }
 395
 396     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 397     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 398
 399     if (seq_param.frame_cropping_flag) {
 400         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 401         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 402         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 403         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 404     }
 405
 406     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 407     if ( false ) {
 408         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 409     } else {
 410         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 411         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 412         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 413         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 414         {
 415             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 416             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 417             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 418             {
 419                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 420                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 421                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 422             }
 423         }
 424         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 425         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 426         {
 427             bitstream_put_ui(bs, 1, 32);  // FPS
 428             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 429             bitstream_put_ui(bs, 1, 1);
 430         }
 431         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 432         {
 433             // hrd_parameters
 434             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 435             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 436             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 437
 438             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 439             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 440             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 441
 442             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 443             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 444             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 445             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 446         }
 447         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 448         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 449
 450         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 451         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 452     }
 453
 454     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 455 }
 456
 457
 458 static void pps_rbsp(bitstream *bs)
 459 {
 460     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 461     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 462
 463     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 464
 465     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 466
 467     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 468
 469     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 470     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 471
 472     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 473     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 474
 475     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 476     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 477     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 478
 479     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 480     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 481     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 482
 483     /* more_rbsp_data */
 484     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 485     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 486     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 487
 488     rbsp_trailing_bits(bs);
 489 }
 490
 491 static void slice_header(bitstream *bs)
 492 {
 493     int first_mb_in_slice = slice_param.macroblock_address;
 494
 495     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 496     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 497     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 498     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 499
 500     /* frame_mbs_only_flag == 1 */
 501     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 502         /* FIXME: */
 503         assert(0);
 504     }
 505
 506     if (pic_param.pic_fields.bits.idr_pic_flag)
 507         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 508
 509     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 510         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 511         /* pic_order_present_flag == 0 */
 512     } else {
 513         /* FIXME: */
 514         assert(0);
 515     }
 516
 517     /* redundant_pic_cnt_present_flag == 0 */
 518     /* slice type */
 519     if (IS_P_SLICE(slice_param.slice_type)) {
 520         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 521
 522         if (slice_param.num_ref_idx_active_override_flag)
 523             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 524
 525         /* ref_pic_list_reordering */
 526         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 527     } else if (IS_B_SLICE(slice_param.slice_type)) {
 528         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 529
 530         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 531
 532         if (slice_param.num_ref_idx_active_override_flag) {
 533             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 534             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 535         }
 536
 537         /* ref_pic_list_reordering */
 538         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 539         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 540     }
 541
 542     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 543          IS_P_SLICE(slice_param.slice_type)) ||
 544         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 545          IS_B_SLICE(slice_param.slice_type))) {
 546         /* FIXME: fill weight/offset table */
 547         assert(0);
 548     }
 549
 550     /* dec_ref_pic_marking */
 551     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 552         unsigned char no_output_of_prior_pics_flag = 0;
 553         unsigned char long_term_reference_flag = 0;
 554         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 555
 556         if (pic_param.pic_fields.bits.idr_pic_flag) {
 557             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 558             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 559         } else {
 560             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 561         }
 562     }
 563
 564     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 565         !IS_I_SLICE(slice_param.slice_type))
 566         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 567
 568     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 569
 570     /* ignore for SP/SI */
 571
 572     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 573         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 574
 575         if (slice_param.disable_deblocking_filter_idc != 1) {
 576             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 577             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 578         }
 579     }
 580
 581     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 582         bitstream_byte_aligning(bs, 1);
 583     }
 584 }
 585
 586 static int
 587 build_packed_pic_buffer(unsigned char **header_buffer)
 588 {
 589     bitstream bs;
 590
 591     bitstream_start(&bs);
 592     nal_start_code_prefix(&bs);
 593     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 594     pps_rbsp(&bs);
 595     bitstream_end(&bs);
 596
 597     *header_buffer = (unsigned char *)bs.buffer;
 598     return bs.bit_offset;
 599 }
 600
 601 static int
 602 build_packed_seq_buffer(unsigned char **header_buffer)
 603 {
 604     bitstream bs;
 605
 606     bitstream_start(&bs);
 607     nal_start_code_prefix(&bs);
 608     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 609     sps_rbsp(&bs);
 610     bitstream_end(&bs);
 611
 612     *header_buffer = (unsigned char *)bs.buffer;
 613     return bs.bit_offset;
 614 }
 615
 616 static int build_packed_slice_buffer(unsigned char **header_buffer)
 617 {
 618     bitstream bs;
 619     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 620     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 621
 622     bitstream_start(&bs);
 623     nal_start_code_prefix(&bs);
 624
 625     if (IS_I_SLICE(slice_param.slice_type)) {
 626         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 627     } else if (IS_P_SLICE(slice_param.slice_type)) {
 628         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 629     } else {
 630         assert(IS_B_SLICE(slice_param.slice_type));
 631         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 632     }
 633
 634     slice_header(&bs);
 635     bitstream_end(&bs);
 636
 637     *header_buffer = (unsigned char *)bs.buffer;
 638     return bs.bit_offset;
 639 }
 640
 641
 642 /*
 643   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 644   1) period between Frame #X and Frame #N = #X - #N
 645   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 646   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 647   4) intra_period and intra_idr_period take precedence over ip_period
 648   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 649      of I/IDR frames, see bellow examples
 650   -------------------------------------------------------------------
 651   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 652   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 653   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 654   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 655   1            1                ignored    IDR IDR IDR IDR...
 656   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 657   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 658   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 659                                               (PBB)(IBB)(PBB)(IBB)...
 660   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 661                                            IDRPPPPP IPPPPP IPPPPP...
 662   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 663                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 664                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 665                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 666                                            {IDR(PBB)(PBB)}                     (6/6/3)
 667                                            {IDR(PBB)(PBB)}.
 668 */
 669
 670 // General pts/dts strategy:
 671 //
 672 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 673 // bit tricky. We assume first of all that the frame rate never goes _above_
 674 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 675 // in at least this speed, as long at dts <= pts (the frame is not attempted
 676 // presented before it is decoded). Furthermore, we never have longer chains of
 677 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 678 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 679 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 680 // frame, with an order O <= C.)
 681 //
 682 // Many strategies are possible, but we establish these rules:
 683 //
 684 //  - Tip frames have dts = pts - (C-O)*N.
 685 //  - Non-tip frames have dts = dts_last + N.
 686 //
 687 // An example, with C=2 and N=10 and the data flow showed with arrows:
 688 //
 689 //        I  B  P  B  B  P
 690 //   pts: 30 40 50 60 70 80
 691 //        ↓  ↓     ↓
 692 //   dts: 10 30 20 60 50←40
 693 //         |  |  ↑        ↑
 694 //         `--|--'        |
 695 //             `----------'
 696 //
 697 // To show that this works fine also with irregular spacings, let's say that
 698 // the third frame is delayed a bit (something earlier was dropped). Now the
 699 // situation looks like this:
 700 //
 701 //        I  B  P  B  B   P
 702 //   pts: 30 40 80 90 100 110
 703 //        ↓  ↓     ↓
 704 //   dts: 10 30 20 90 50←40
 705 //         |  |  ↑        ↑
 706 //         `--|--'        |
 707 //             `----------'
 708 //
 709 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 710 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 711 //
 712 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 713 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 714 // a tip frame and should be given a dts based on the previous one.
 715 #define FRAME_P 0
 716 #define FRAME_B 1
 717 #define FRAME_I 2
 718 #define FRAME_IDR 7
 719 void encoding2display_order(
 720     int encoding_order, int intra_period,
 721     int intra_idr_period, int ip_period,
 722     int *displaying_order,
 723     int *frame_type, int *pts_lag)
 724 {
 725     int encoding_order_gop = 0;
 726
 727     *pts_lag = 0;
 728
 729     if (intra_period == 1) { /* all are I/IDR frames */
 730         *displaying_order = encoding_order;
 731         if (intra_idr_period == 0)
 732             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 733         else
 734             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 735         return;
 736     }
 737
 738     if (intra_period == 0)
 739         intra_idr_period = 0;
 740
 741     if (ip_period == 1) {
 742         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 743         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 744         *displaying_order = encoding_order;
 745
 746         if (encoding_order_gop == 0) { /* the first frame */
 747             *frame_type = FRAME_IDR;
 748         } else if (intra_period != 0 && /* have I frames */
 749                    encoding_order_gop >= 2 &&
 750                    (encoding_order_gop % intra_period == 0)) {
 751             *frame_type = FRAME_I;
 752         } else {
 753             *frame_type = FRAME_P;
 754         }
 755         return;
 756     }
 757
 758     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 759     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 760     *pts_lag = -1;  // Most frames are not tip frames.
 761
 762     if (encoding_order_gop == 0) { /* the first frame */
 763         *frame_type = FRAME_IDR;
 764         *displaying_order = encoding_order;
 765         // IDR frames are a special case; I honestly can't find the logic behind
 766         // why this is the right thing, but it seems to line up nicely in practice :-)
 767         *pts_lag = TIMEBASE / MAX_FPS;
 768     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 769         *frame_type = FRAME_B;
 770         *displaying_order = encoding_order - 1;
 771         if ((encoding_order_gop % ip_period) == 0) {
 772             *pts_lag = 0;  // Last B-frame.
 773         }
 774     } else if (intra_period != 0 && /* have I frames */
 775                encoding_order_gop >= 2 &&
 776                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 777         *frame_type = FRAME_I;
 778         *displaying_order = encoding_order + ip_period - 1;
 779     } else {
 780         *frame_type = FRAME_P;
 781         *displaying_order = encoding_order + ip_period - 1;
 782     }
 783 }
 784
 785
 786 static const char *rc_to_string(int rcmode)
 787 {
 788     switch (rc_mode) {
 789     case VA_RC_NONE:
 790         return "NONE";
 791     case VA_RC_CBR:
 792         return "CBR";
 793     case VA_RC_VBR:
 794         return "VBR";
 795     case VA_RC_VCM:
 796         return "VCM";
 797     case VA_RC_CQP:
 798         return "CQP";
 799     case VA_RC_VBR_CONSTRAINED:
 800         return "VBR_CONSTRAINED";
 801     default:
 802         return "Unknown";
 803     }
 804 }
 805
 806 #if 0
 807 static int process_cmdline(int argc, char *argv[])
 808 {
 809     char c;
 810     const struct option long_opts[] = {
 811         {"help", no_argument, NULL, 0 },
 812         {"bitrate", required_argument, NULL, 1 },
 813         {"minqp", required_argument, NULL, 2 },
 814         {"initialqp", required_argument, NULL, 3 },
 815         {"intra_period", required_argument, NULL, 4 },
 816         {"idr_period", required_argument, NULL, 5 },
 817         {"ip_period", required_argument, NULL, 6 },
 818         {"rcmode", required_argument, NULL, 7 },
 819         {"srcyuv", required_argument, NULL, 9 },
 820         {"recyuv", required_argument, NULL, 10 },
 821         {"fourcc", required_argument, NULL, 11 },
 822         {"syncmode", no_argument, NULL, 12 },
 823         {"enablePSNR", no_argument, NULL, 13 },
 824         {"prit", required_argument, NULL, 14 },
 825         {"priv", required_argument, NULL, 15 },
 826         {"framecount", required_argument, NULL, 16 },
 827         {"entropy", required_argument, NULL, 17 },
 828         {"profile", required_argument, NULL, 18 },
 829         {NULL, no_argument, NULL, 0 }};
 830     int long_index;
 831
 832     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 833         switch (c) {
 834         case 'w':
 835             frame_width = atoi(optarg);
 836             break;
 837         case 'h':
 838             frame_height = atoi(optarg);
 839             break;
 840         case 'n':
 841         case 'f':
 842             frame_rate = atoi(optarg);
 843             break;
 844         case 'o':
 845             coded_fn = strdup(optarg);
 846             break;
 847         case 0:
 848             print_help();
 849             exit(0);
 850         case 1:
 851             frame_bitrate = atoi(optarg);
 852             break;
 853         case 2:
 854             minimal_qp = atoi(optarg);
 855             break;
 856         case 3:
 857             initial_qp = atoi(optarg);
 858             break;
 859         case 4:
 860             intra_period = atoi(optarg);
 861             break;
 862         case 5:
 863             intra_idr_period = atoi(optarg);
 864             break;
 865         case 6:
 866             ip_period = atoi(optarg);
 867             break;
 868         case 7:
 869             rc_mode = string_to_rc(optarg);
 870             if (rc_mode < 0) {
 871                 print_help();
 872                 exit(1);
 873             }
 874             break;
 875         case 9:
 876             srcyuv_fn = strdup(optarg);
 877             break;
 878         case 11:
 879             srcyuv_fourcc = string_to_fourcc(optarg);
 880             if (srcyuv_fourcc <= 0) {
 881                 print_help();
 882                 exit(1);
 883             }
 884             break;
 885         case 13:
 886             calc_psnr = 1;
 887             break;
 888         case 14:
 889             misc_priv_type = strtol(optarg, NULL, 0);
 890             break;
 891         case 15:
 892             misc_priv_value = strtol(optarg, NULL, 0);
 893             break;
 894         case 17:
 895             h264_entropy_mode = atoi(optarg) ? 1: 0;
 896             break;
 897         case 18:
 898             if (strncmp(optarg, "BP", 2) == 0)
 899                 h264_profile = VAProfileH264Baseline;
 900             else if (strncmp(optarg, "MP", 2) == 0)
 901                 h264_profile = VAProfileH264Main;
 902             else if (strncmp(optarg, "HP", 2) == 0)
 903                 h264_profile = VAProfileH264High;
 904             else
 905                 h264_profile = (VAProfile)0;
 906             break;
 907         case ':':
 908         case '?':
 909             print_help();
 910             exit(0);
 911         }
 912     }
 913
 914     if (ip_period < 1) {
 915         printf(" ip_period must be greater than 0\n");
 916         exit(0);
 917     }
 918     if (intra_period != 1 && intra_period % ip_period != 0) {
 919         printf(" intra_period must be a multiplier of ip_period\n");
 920         exit(0);
 921     }
 922     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 923         printf(" intra_idr_period must be a multiplier of intra_period\n");
 924         exit(0);
 925     }
 926
 927     if (frame_bitrate == 0)
 928         frame_bitrate = frame_width * frame_height * 12 * MAX_FPS / 50;
 929
 930     if (coded_fn == NULL) {
 931         struct stat buf;
 932         if (stat("/tmp", &buf) == 0)
 933             coded_fn = strdup("/tmp/test.264");
 934         else if (stat("/sdcard", &buf) == 0)
 935             coded_fn = strdup("/sdcard/test.264");
 936         else
 937             coded_fn = strdup("./test.264");
 938     }
 939
 940
 941     frame_width_mbaligned = (frame_width + 15) & (~15);
 942     frame_height_mbaligned = (frame_height + 15) & (~15);
 943     if (frame_width != frame_width_mbaligned ||
 944         frame_height != frame_height_mbaligned) {
 945         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 946                frame_width, frame_height,
 947                frame_width_mbaligned, frame_height_mbaligned
 948                );
 949     }
 950
 951     return 0;
 952 }
 953 #endif
 954
 955 static Display *x11_display;
 956 static Window   x11_window;
 957
 958 VADisplay
 959 va_open_display(void)
 960 {
 961     x11_display = XOpenDisplay(NULL);
 962     if (!x11_display) {
 963         fprintf(stderr, "error: can't connect to X server!\n");
 964         return NULL;
 965     }
 966     return vaGetDisplay(x11_display);
 967 }
 968
 969 void
 970 va_close_display(VADisplay va_dpy)
 971 {
 972     if (!x11_display)
 973         return;
 974
 975     if (x11_window) {
 976         XUnmapWindow(x11_display, x11_window);
 977         XDestroyWindow(x11_display, x11_window);
 978         x11_window = None;
 979     }
 980     XCloseDisplay(x11_display);
 981     x11_display = NULL;
 982 }
 983
 984 static int init_va(void)
 985 {
 986     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 987     VAEntrypoint *entrypoints;
 988     int num_entrypoints, slice_entrypoint;
 989     int support_encode = 0;
 990     int major_ver, minor_ver;
 991     VAStatus va_status;
 992     unsigned int i;
 993
 994     va_dpy = va_open_display();
 995     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 996     CHECK_VASTATUS(va_status, "vaInitialize");
 997
 998     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 999     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1000     if (!entrypoints) {
1001         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1002         exit(1);
1003     }
1004
1005     /* use the highest profile */
1006     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1007         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1008             continue;
1009
1010         h264_profile = profile_list[i];
1011         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1012         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1013             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1014                 support_encode = 1;
1015                 break;
1016             }
1017         }
1018         if (support_encode == 1)
1019             break;
1020     }
1021
1022     if (support_encode == 0) {
1023         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
1024         exit(1);
1025     } else {
1026         switch (h264_profile) {
1027             case VAProfileH264Baseline:
1028                 ip_period = 1;
1029                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1030                 h264_entropy_mode = 0;
1031                 break;
1032             case VAProfileH264ConstrainedBaseline:
1033                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1034                 ip_period = 1;
1035                 break;
1036
1037             case VAProfileH264Main:
1038                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1039                 break;
1040
1041             case VAProfileH264High:
1042                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1043                 break;
1044             default:
1045                 h264_profile = VAProfileH264Baseline;
1046                 ip_period = 1;
1047                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1048                 break;
1049         }
1050     }
1051
1052     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1053
1054     /* find out the format for the render target, and rate control mode */
1055     for (i = 0; i < VAConfigAttribTypeMax; i++)
1056         attrib[i].type = (VAConfigAttribType)i;
1057
1058     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1059                                       &attrib[0], VAConfigAttribTypeMax);
1060     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1061     /* check the interested configattrib */
1062     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1063         printf("Not find desired YUV420 RT format\n");
1064         exit(1);
1065     } else {
1066         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1067         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1068         config_attrib_num++;
1069     }
1070
1071     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1072         int tmp = attrib[VAConfigAttribRateControl].value;
1073
1074         if (rc_mode == -1 || !(rc_mode & tmp))  {
1075             if (rc_mode != -1) {
1076                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1077             }
1078
1079             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1080                 if (rc_default_modes[i] & tmp) {
1081                     rc_mode = rc_default_modes[i];
1082                     break;
1083                 }
1084             }
1085         }
1086
1087         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1088         config_attrib[config_attrib_num].value = rc_mode;
1089         config_attrib_num++;
1090     }
1091
1092
1093     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1094         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1095
1096         h264_packedheader = 1;
1097         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1098         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1099
1100         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1101             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1102         }
1103
1104         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1105             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1106         }
1107
1108         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1109             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1110         }
1111
1112         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1113             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1114         }
1115
1116         enc_packed_header_idx = config_attrib_num;
1117         config_attrib_num++;
1118     }
1119
1120     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1121         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1122         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1123         config_attrib_num++;
1124     }
1125
1126     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1127         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1128     }
1129
1130     free(entrypoints);
1131     return 0;
1132 }
1133
1134 static int setup_encode()
1135 {
1136     VAStatus va_status;
1137     VASurfaceID *tmp_surfaceid;
1138     int codedbuf_size, i;
1139     static VASurfaceID src_surface[SURFACE_NUM];
1140     static VASurfaceID ref_surface[SURFACE_NUM];
1141
1142     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1143             &config_attrib[0], config_attrib_num, &config_id);
1144     CHECK_VASTATUS(va_status, "vaCreateConfig");
1145
1146     /* create source surfaces */
1147     va_status = vaCreateSurfaces(va_dpy,
1148                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1149                                  &src_surface[0], SURFACE_NUM,
1150                                  NULL, 0);
1151     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1152
1153     /* create reference surfaces */
1154     va_status = vaCreateSurfaces(va_dpy,
1155                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1156                                  &ref_surface[0], SURFACE_NUM,
1157                                  NULL, 0);
1158     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1159
1160     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1161     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1162     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1163
1164     /* Create a context for this encode pipe */
1165     va_status = vaCreateContext(va_dpy, config_id,
1166                                 frame_width_mbaligned, frame_height_mbaligned,
1167                                 VA_PROGRESSIVE,
1168                                 tmp_surfaceid, 2 * SURFACE_NUM,
1169                                 &context_id);
1170     CHECK_VASTATUS(va_status, "vaCreateContext");
1171     free(tmp_surfaceid);
1172
1173     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1174
1175     for (i = 0; i < SURFACE_NUM; i++) {
1176         /* create coded buffer once for all
1177          * other VA buffers which won't be used again after vaRenderPicture.
1178          * so APP can always vaCreateBuffer for every frame
1179          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1180          * so VA won't maintain the coded buffer
1181          */
1182         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1183                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1184         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1185     }
1186
1187     /* create OpenGL objects */
1188     //glGenFramebuffers(SURFACE_NUM, fbos);
1189
1190     for (i = 0; i < SURFACE_NUM; i++) {
1191         glGenTextures(1, &gl_surfaces[i].y_tex);
1192         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1193     }
1194
1195     for (i = 0; i < SURFACE_NUM; i++) {
1196         gl_surfaces[i].src_surface = src_surface[i];
1197         gl_surfaces[i].ref_surface = ref_surface[i];
1198     }
1199
1200     return 0;
1201 }
1202
1203
1204
1205 #define partition(ref, field, key, ascending)   \
1206     while (i <= j) {                            \
1207         if (ascending) {                        \
1208             while (ref[i].field < key)          \
1209                 i++;                            \
1210             while (ref[j].field > key)          \
1211                 j--;                            \
1212         } else {                                \
1213             while (ref[i].field > key)          \
1214                 i++;                            \
1215             while (ref[j].field < key)          \
1216                 j--;                            \
1217         }                                       \
1218         if (i <= j) {                           \
1219             tmp = ref[i];                       \
1220             ref[i] = ref[j];                    \
1221             ref[j] = tmp;                       \
1222             i++;                                \
1223             j--;                                \
1224         }                                       \
1225     }                                           \
1226
1227 static void sort_one(VAPictureH264 ref[], int left, int right,
1228                      int ascending, int frame_idx)
1229 {
1230     int i = left, j = right;
1231     unsigned int key;
1232     VAPictureH264 tmp;
1233
1234     if (frame_idx) {
1235         key = ref[(left + right) / 2].frame_idx;
1236         partition(ref, frame_idx, key, ascending);
1237     } else {
1238         key = ref[(left + right) / 2].TopFieldOrderCnt;
1239         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1240     }
1241
1242     /* recursion */
1243     if (left < j)
1244         sort_one(ref, left, j, ascending, frame_idx);
1245
1246     if (i < right)
1247         sort_one(ref, i, right, ascending, frame_idx);
1248 }
1249
1250 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1251                      int partition_ascending, int list0_ascending, int list1_ascending)
1252 {
1253     int i = left, j = right;
1254     VAPictureH264 tmp;
1255
1256     if (frame_idx) {
1257         partition(ref, frame_idx, key, partition_ascending);
1258     } else {
1259         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1260     }
1261
1262
1263     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1264     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1265 }
1266
1267 static void update_ReferenceFrames(int frame_type)
1268 {
1269     int i;
1270
1271     if (frame_type == FRAME_B)
1272         return;
1273
1274     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1275     numShortTerm++;
1276     if (numShortTerm > num_ref_frames)
1277         numShortTerm = num_ref_frames;
1278     for (i=numShortTerm-1; i>0; i--)
1279         ReferenceFrames[i] = ReferenceFrames[i-1];
1280     ReferenceFrames[0] = CurrentCurrPic;
1281
1282     current_frame_num++;
1283     if (current_frame_num > MaxFrameNum)
1284         current_frame_num = 0;
1285 }
1286
1287
1288 static int update_RefPicList(int frame_type)
1289 {
1290     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1291
1292     if (frame_type == FRAME_P) {
1293         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1294         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1295     }
1296
1297     if (frame_type == FRAME_B) {
1298         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1299         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1300                  1, 0, 1);
1301
1302         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1303         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1304                  0, 1, 0);
1305     }
1306
1307     return 0;
1308 }
1309
1310
1311 static int render_sequence(void)
1312 {
1313     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1314     VAStatus va_status;
1315     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1316     VAEncMiscParameterRateControl *misc_rate_ctrl;
1317
1318     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1319     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1320     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1321     seq_param.bits_per_second = frame_bitrate;
1322
1323     seq_param.intra_period = intra_period;
1324     seq_param.intra_idr_period = intra_idr_period;
1325     seq_param.ip_period = ip_period;
1326
1327     seq_param.max_num_ref_frames = num_ref_frames;
1328     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1329     seq_param.time_scale = TIMEBASE * 2;
1330     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1331     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1332     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1333     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1334     seq_param.seq_fields.bits.chroma_format_idc = 1;
1335     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1336
1337     if (frame_width != frame_width_mbaligned ||
1338         frame_height != frame_height_mbaligned) {
1339         seq_param.frame_cropping_flag = 1;
1340         seq_param.frame_crop_left_offset = 0;
1341         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1342         seq_param.frame_crop_top_offset = 0;
1343         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1344     }
1345
1346     va_status = vaCreateBuffer(va_dpy, context_id,
1347                                VAEncSequenceParameterBufferType,
1348                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1349     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1350
1351     va_status = vaCreateBuffer(va_dpy, context_id,
1352                                VAEncMiscParameterBufferType,
1353                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1354                                1, NULL, &rc_param_buf);
1355     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1356
1357     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1358     misc_param->type = VAEncMiscParameterTypeRateControl;
1359     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1360     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1361     misc_rate_ctrl->bits_per_second = frame_bitrate;
1362     misc_rate_ctrl->target_percentage = 66;
1363     misc_rate_ctrl->window_size = 1000;
1364     misc_rate_ctrl->initial_qp = initial_qp;
1365     misc_rate_ctrl->min_qp = minimal_qp;
1366     misc_rate_ctrl->basic_unit_size = 0;
1367     vaUnmapBuffer(va_dpy, rc_param_buf);
1368
1369     render_id[0] = seq_param_buf;
1370     render_id[1] = rc_param_buf;
1371
1372     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1373
1374     if (misc_priv_type != 0) {
1375         va_status = vaCreateBuffer(va_dpy, context_id,
1376                                    VAEncMiscParameterBufferType,
1377                                    sizeof(VAEncMiscParameterBuffer),
1378                                    1, NULL, &misc_param_tmpbuf);
1379         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1380         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1381         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1382         misc_param_tmp->data[0] = misc_priv_value;
1383         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1384
1385         render_picture_and_delete(va_dpy, context_id, &misc_param_tmpbuf, 1);
1386     }
1387
1388     return 0;
1389 }
1390
1391 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1392 {
1393     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1394     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1395     int PicOrderCntMsb, TopFieldOrderCnt;
1396
1397     if (frame_type == FRAME_IDR)
1398         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1399     else {
1400         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1401         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1402     }
1403
1404     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1405         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1406         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1407     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1408              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1409         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1410     else
1411         PicOrderCntMsb = prevPicOrderCntMsb;
1412
1413     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1414
1415     if (frame_type != FRAME_B) {
1416         PicOrderCntMsb_ref = PicOrderCntMsb;
1417         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1418     }
1419
1420     return TopFieldOrderCnt;
1421 }
1422
1423 static int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1424 {
1425     VABufferID pic_param_buf;
1426     VAStatus va_status;
1427     int i = 0;
1428
1429     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1430     pic_param.CurrPic.frame_idx = current_frame_num;
1431     pic_param.CurrPic.flags = 0;
1432     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1433     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1434     CurrentCurrPic = pic_param.CurrPic;
1435
1436     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1437     for (i = numShortTerm; i < SURFACE_NUM; i++) {
1438         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1439         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1440     }
1441
1442     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1443     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1444     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1445     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1446     pic_param.frame_num = current_frame_num;
1447     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1448     pic_param.last_picture = false;  // FIXME
1449     pic_param.pic_init_qp = initial_qp;
1450
1451     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1452                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1453     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1454
1455     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1456
1457     return 0;
1458 }
1459
1460 static int render_packedsequence(void)
1461 {
1462     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1463     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1464     unsigned int length_in_bits;
1465     unsigned char *packedseq_buffer = NULL;
1466     VAStatus va_status;
1467
1468     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1469
1470     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1471
1472     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1473     packedheader_param_buffer.has_emulation_bytes = 0;
1474     va_status = vaCreateBuffer(va_dpy,
1475                                context_id,
1476                                VAEncPackedHeaderParameterBufferType,
1477                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1478                                &packedseq_para_bufid);
1479     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1480
1481     va_status = vaCreateBuffer(va_dpy,
1482                                context_id,
1483                                VAEncPackedHeaderDataBufferType,
1484                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1485                                &packedseq_data_bufid);
1486     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1487
1488     render_id[0] = packedseq_para_bufid;
1489     render_id[1] = packedseq_data_bufid;
1490     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1491
1492     free(packedseq_buffer);
1493
1494     return 0;
1495 }
1496
1497
1498 static int render_packedpicture(void)
1499 {
1500     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1501     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1502     unsigned int length_in_bits;
1503     unsigned char *packedpic_buffer = NULL;
1504     VAStatus va_status;
1505
1506     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1507     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1508     packedheader_param_buffer.bit_length = length_in_bits;
1509     packedheader_param_buffer.has_emulation_bytes = 0;
1510
1511     va_status = vaCreateBuffer(va_dpy,
1512                                context_id,
1513                                VAEncPackedHeaderParameterBufferType,
1514                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1515                                &packedpic_para_bufid);
1516     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1517
1518     va_status = vaCreateBuffer(va_dpy,
1519                                context_id,
1520                                VAEncPackedHeaderDataBufferType,
1521                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1522                                &packedpic_data_bufid);
1523     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1524
1525     render_id[0] = packedpic_para_bufid;
1526     render_id[1] = packedpic_data_bufid;
1527     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1528
1529     free(packedpic_buffer);
1530
1531     return 0;
1532 }
1533
1534 static void render_packedslice()
1535 {
1536     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1537     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1538     unsigned int length_in_bits;
1539     unsigned char *packedslice_buffer = NULL;
1540     VAStatus va_status;
1541
1542     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1543     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1544     packedheader_param_buffer.bit_length = length_in_bits;
1545     packedheader_param_buffer.has_emulation_bytes = 0;
1546
1547     va_status = vaCreateBuffer(va_dpy,
1548                                context_id,
1549                                VAEncPackedHeaderParameterBufferType,
1550                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1551                                &packedslice_para_bufid);
1552     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1553
1554     va_status = vaCreateBuffer(va_dpy,
1555                                context_id,
1556                                VAEncPackedHeaderDataBufferType,
1557                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1558                                &packedslice_data_bufid);
1559     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1560
1561     render_id[0] = packedslice_para_bufid;
1562     render_id[1] = packedslice_data_bufid;
1563     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1564
1565     free(packedslice_buffer);
1566 }
1567
1568 static int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1569 {
1570     VABufferID slice_param_buf;
1571     VAStatus va_status;
1572     int i;
1573
1574     update_RefPicList(frame_type);
1575
1576     /* one frame, one slice */
1577     slice_param.macroblock_address = 0;
1578     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1579     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1580     if (frame_type == FRAME_IDR) {
1581         if (encoding_frame_num != 0)
1582             ++slice_param.idr_pic_id;
1583     } else if (frame_type == FRAME_P) {
1584         int refpiclist0_max = h264_maxref & 0xffff;
1585         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1586
1587         for (i = refpiclist0_max; i < 32; i++) {
1588             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1589             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1590         }
1591     } else if (frame_type == FRAME_B) {
1592         int refpiclist0_max = h264_maxref & 0xffff;
1593         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1594
1595         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1596         for (i = refpiclist0_max; i < 32; i++) {
1597             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1598             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1599         }
1600
1601         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1602         for (i = refpiclist1_max; i < 32; i++) {
1603             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1604             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1605         }
1606     }
1607
1608     slice_param.slice_alpha_c0_offset_div2 = 0;
1609     slice_param.slice_beta_offset_div2 = 0;
1610     slice_param.direct_spatial_mv_pred_flag = 1;
1611     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1612
1613
1614     if (h264_packedheader &&
1615         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1616         render_packedslice();
1617
1618     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1619                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1620     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1621
1622     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1623
1624     return 0;
1625 }
1626
1627
1628
1629 void H264EncoderImpl::save_codeddata(storage_task task)
1630 {
1631     VACodedBufferSegment *buf_list = NULL;
1632     VAStatus va_status;
1633     unsigned int coded_size = 0;
1634
1635     string data;
1636
1637     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1638
1639     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1640     CHECK_VASTATUS(va_status, "vaMapBuffer");
1641     while (buf_list != NULL) {
1642         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1643         buf_list = (VACodedBufferSegment *) buf_list->next;
1644
1645         frame_size += coded_size;
1646     }
1647     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1648
1649     {
1650         // Add video.
1651         AVPacket pkt;
1652         memset(&pkt, 0, sizeof(pkt));
1653         pkt.buf = nullptr;
1654         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1655         pkt.size = data.size();
1656         pkt.stream_index = 0;
1657         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1658             pkt.flags = AV_PKT_FLAG_KEY;
1659         } else {
1660             pkt.flags = 0;
1661         }
1662         //pkt.duration = 1;
1663         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1664     }
1665     // Encode and add all audio frames up to and including the pts of this video frame.
1666     for ( ;; ) {
1667         int64_t audio_pts;
1668         vector<float> audio;
1669         {
1670              unique_lock<mutex> lock(frame_queue_mutex);
1671              frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1672              if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1673              auto it = pending_audio_frames.begin();
1674              if (it->first > task.pts) break;
1675              audio_pts = it->first;
1676              audio = move(it->second);
1677              pending_audio_frames.erase(it);
1678         }
1679
1680         AVFrame *frame = avcodec_alloc_frame();
1681         frame->nb_samples = audio.size() / 2;
1682         frame->format = AV_SAMPLE_FMT_S32;
1683         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1684
1685         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1686         int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1687         if (ret < 0) {
1688             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1689             exit(1);
1690         }
1691         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1692             if (audio[i] >= 1.0f) {
1693                 int_samples[i] = 2147483647;
1694             } else if (audio[i] <= -1.0f) {
1695                 int_samples[i] = -2147483647;
1696             } else {
1697                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1698             }
1699         }
1700
1701         AVPacket pkt;
1702         av_init_packet(&pkt);
1703         pkt.data = nullptr;
1704         pkt.size = 0;
1705         int got_output;
1706         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1707         if (got_output) {
1708             pkt.stream_index = 1;
1709             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1710         }
1711         // TODO: Delayed frames.
1712         avcodec_free_frame(&frame);
1713         av_free_packet(&pkt);
1714         if (audio_pts == task.pts) break;
1715     }
1716
1717 #if 0
1718     printf("\r      "); /* return back to startpoint */
1719     switch (encode_order % 4) {
1720         case 0:
1721             printf("|");
1722             break;
1723         case 1:
1724             printf("/");
1725             break;
1726         case 2:
1727             printf("-");
1728             break;
1729         case 3:
1730             printf("\\");
1731             break;
1732     }
1733     printf("%08lld", encode_order);
1734     printf("(%06d bytes coded)", coded_size);
1735 #endif
1736 }
1737
1738
1739 // this is weird. but it seems to put a new frame onto the queue
1740 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1741 {
1742         unique_lock<mutex> lock(storage_task_queue_mutex);
1743         storage_task_queue.push(move(task));
1744         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1745         storage_task_queue_changed.notify_all();
1746 }
1747
1748 void H264EncoderImpl::storage_task_thread()
1749 {
1750         for ( ;; ) {
1751                 storage_task current;
1752                 {
1753                         // wait until there's an encoded frame
1754                         unique_lock<mutex> lock(storage_task_queue_mutex);
1755                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1756                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1757                         current = move(storage_task_queue.front());
1758                         storage_task_queue.pop();
1759                 }
1760
1761                 VAStatus va_status;
1762
1763                 // waits for data, then saves it to disk.
1764                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1765                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1766                 save_codeddata(move(current));
1767
1768                 {
1769                         unique_lock<mutex> lock(storage_task_queue_mutex);
1770                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1771                         storage_task_queue_changed.notify_all();
1772                 }
1773         }
1774 }
1775
1776 static int release_encode()
1777 {
1778     int i;
1779
1780     for (i = 0; i < SURFACE_NUM; i++) {
1781         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1782         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1783         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1784     }
1785
1786     vaDestroyContext(va_dpy, context_id);
1787     vaDestroyConfig(va_dpy, config_id);
1788
1789     return 0;
1790 }
1791
1792 static int deinit_va()
1793 {
1794     vaTerminate(va_dpy);
1795
1796     va_close_display(va_dpy);
1797
1798     return 0;
1799 }
1800
1801
1802 H264EncoderImpl::H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd)
1803         : current_storage_frame(0), surface(surface), httpd(httpd)
1804 {
1805         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1806         context_audio = avcodec_alloc_context3(codec_audio);
1807         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1808         context_audio->sample_rate = OUTPUT_FREQUENCY;
1809         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1810         context_audio->channels = 2;
1811         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1812         context_audio->time_base = AVRational{1, TIMEBASE};
1813         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1814                 fprintf(stderr, "Could not open codec\n");
1815                 exit(1);
1816         }
1817
1818         frame_width = width;
1819         frame_height = height;
1820         frame_width_mbaligned = (frame_width + 15) & (~15);
1821         frame_height_mbaligned = (frame_height + 15) & (~15);
1822         frame_bitrate = 15000000;  // / 60;
1823
1824         //print_input();
1825
1826         init_va();
1827         setup_encode();
1828
1829         // No frames are ready yet.
1830         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1831
1832         memset(&seq_param, 0, sizeof(seq_param));
1833         memset(&pic_param, 0, sizeof(pic_param));
1834         memset(&slice_param, 0, sizeof(slice_param));
1835
1836         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1837
1838         encode_thread = thread([this]{
1839                 //SDL_GL_MakeCurrent(window, context);
1840                 QOpenGLContext *context = create_context(this->surface);
1841                 eglBindAPI(EGL_OPENGL_API);
1842                 if (!make_current(context, this->surface)) {
1843                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1844                                 eglGetError());
1845                         exit(1);
1846                 }
1847                 encode_thread_func();
1848         });
1849 }
1850
1851 H264EncoderImpl::~H264EncoderImpl()
1852 {
1853         {
1854                 unique_lock<mutex> lock(frame_queue_mutex);
1855                 encode_thread_should_quit = true;
1856                 frame_queue_nonempty.notify_all();
1857         }
1858         encode_thread.join();
1859         {
1860                 unique_lock<mutex> lock(storage_task_queue_mutex);
1861                 storage_thread_should_quit = true;
1862                 frame_queue_nonempty.notify_all();
1863                 storage_task_queue_changed.notify_all();
1864         }
1865         storage_thread.join();
1866
1867         release_encode();
1868         deinit_va();
1869 }
1870
1871 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1872 {
1873         {
1874                 // Wait until this frame slot is done encoding.
1875                 unique_lock<mutex> lock(storage_task_queue_mutex);
1876                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1877                 if (storage_thread_should_quit) return false;
1878         }
1879
1880         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1881         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1882         *y_tex = surf->y_tex;
1883         *cbcr_tex = surf->cbcr_tex;
1884
1885         VASurfaceID surface = surf->src_surface;
1886         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1887         CHECK_VASTATUS(va_status, "vaDeriveImage");
1888
1889         VABufferInfo buf_info;
1890         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1891         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1892         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1893
1894         // Create Y image.
1895         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1896         EGLint y_attribs[] = {
1897                 EGL_WIDTH, frame_width,
1898                 EGL_HEIGHT, frame_height,
1899                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1900                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1901                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1902                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1903                 EGL_NONE
1904         };
1905
1906         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1907         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1908
1909         // Associate Y image to a texture.
1910         glBindTexture(GL_TEXTURE_2D, *y_tex);
1911         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1912
1913         // Create CbCr image.
1914         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1915         EGLint cbcr_attribs[] = {
1916                 EGL_WIDTH, frame_width,
1917                 EGL_HEIGHT, frame_height,
1918                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1919                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1920                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1921                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1922                 EGL_NONE
1923         };
1924
1925         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1926         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1927
1928         // Associate CbCr image to a texture.
1929         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1930         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1931
1932         return true;
1933 }
1934
1935 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
1936 {
1937         {
1938                 unique_lock<mutex> lock(frame_queue_mutex);
1939                 pending_audio_frames[pts] = move(audio);
1940         }
1941         frame_queue_nonempty.notify_all();
1942 }
1943
1944 void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1945 {
1946         {
1947                 unique_lock<mutex> lock(frame_queue_mutex);
1948                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1949                 ++current_storage_frame;
1950         }
1951         frame_queue_nonempty.notify_all();
1952 }
1953
1954 void H264EncoderImpl::encode_thread_func()
1955 {
1956         int64_t last_dts = -1;
1957         int gop_start_display_frame_num = 0;
1958         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1959                 PendingFrame frame;
1960                 int pts_lag;
1961                 int frame_type, display_frame_num;
1962                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1963                                        &display_frame_num, &frame_type, &pts_lag);
1964                 if (frame_type == FRAME_IDR) {
1965                         numShortTerm = 0;
1966                         current_frame_num = 0;
1967                         gop_start_display_frame_num = display_frame_num;
1968                 }
1969
1970                 {
1971                         unique_lock<mutex> lock(frame_queue_mutex);
1972                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1973                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1974                         });
1975                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1976                                 // We have queued frames that were supposed to be B-frames,
1977                                 // but will be no P-frame to encode them against. Encode them all
1978                                 // as P-frames instead. Note that this happens under the mutex,
1979                                 // but nobody else uses it at this point, since we're shutting down,
1980                                 // so there's no contention.
1981                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1982                                 return;
1983                         } else {
1984                                 frame = move(pending_video_frames[display_frame_num]);
1985                                 pending_video_frames.erase(display_frame_num);
1986                         }
1987                 }
1988
1989                 // Determine the dts of this frame.
1990                 int64_t dts;
1991                 if (pts_lag == -1) {
1992                         assert(last_dts != -1);
1993                         dts = last_dts + (TIMEBASE / MAX_FPS);
1994                 } else {
1995                         dts = frame.pts - pts_lag;
1996                 }
1997                 last_dts = dts;
1998
1999                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
2000         }
2001 }
2002
2003 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2004 {
2005         if (pending_video_frames.empty()) {
2006                 return;
2007         }
2008
2009         for (auto &pending_frame : pending_video_frames) {
2010                 int display_frame_num = pending_frame.first;
2011                 assert(display_frame_num > 0);
2012                 PendingFrame frame = move(pending_frame.second);
2013                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2014                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2015                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
2016                 last_dts = dts;
2017         }
2018 }
2019
2020 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2021                                    int frame_type, int64_t pts, int64_t dts)
2022 {
2023         // Wait for the GPU to be done with the frame.
2024         glClientWaitSync(frame.fence.get(), 0, 0);
2025
2026         // Release back any input frames we needed to render this frame.
2027         frame.input_frames.clear();
2028
2029         // Unmap the image.
2030         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2031         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2032         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2033         VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2034         CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2035         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2036         CHECK_VASTATUS(va_status, "vaDestroyImage");
2037
2038         VASurfaceID surface = surf->src_surface;
2039
2040         // Schedule the frame for encoding.
2041         va_status = vaBeginPicture(va_dpy, context_id, surface);
2042         CHECK_VASTATUS(va_status, "vaBeginPicture");
2043
2044         if (frame_type == FRAME_IDR) {
2045                 render_sequence();
2046                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2047                 if (h264_packedheader) {
2048                         render_packedsequence();
2049                         render_packedpicture();
2050                 }
2051         } else {
2052                 //render_sequence();
2053                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2054         }
2055         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2056
2057         va_status = vaEndPicture(va_dpy, context_id);
2058         CHECK_VASTATUS(va_status, "vaEndPicture");
2059
2060         // so now the data is done encoding (well, async job kicked off)...
2061         // we send that to the storage thread
2062         storage_task tmp;
2063         tmp.display_order = display_frame_num;
2064         tmp.frame_type = frame_type;
2065         tmp.pts = pts;
2066         tmp.dts = dts;
2067         storage_task_enqueue(move(tmp));
2068
2069         update_ReferenceFrames(frame_type);
2070 }
2071
2072 // Proxy object.
2073 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
2074         : impl(new H264EncoderImpl(surface, width, height, httpd)) {}
2075
2076 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2077 H264Encoder::~H264Encoder() {}
2078
2079 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
2080 {
2081         impl->add_audio(pts, audio);
2082 }
2083
2084 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2085 {
2086         return impl->begin_frame(y_tex, cbcr_tex);
2087 }
2088
2089 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
2090 {
2091         impl->end_frame(fence, pts, input_frames);
2092 }
2093
2094 // Real class.