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 #include <libavcodec/avcodec.h>
  10 #include <libavutil/channel_layout.h>
  11 #include <libavutil/frame.h>
  12 #include <libavutil/rational.h>
  13 #include <libavutil/samplefmt.h>
  14 #include <libdrm/drm_fourcc.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <string.h>
  18 #include <va/va.h>
  19 #include <va/va_drmcommon.h>
  20 #include <va/va_enc_h264.h>
  21 #include <va/va_x11.h>
  22 #include <condition_variable>
  23 #include <cstdint>
  24 #include <memory>
  25 #include <mutex>
  26 #include <queue>
  27 #include <string>
  28 #include <thread>
  29 #include <utility>
  30
  31 #include "context.h"
  32 #include "defs.h"
  33 #include "httpd.h"
  34 #include "timebase.h"
  35
  36 class QOpenGLContext;
  37 class QSurface;
  38
  39 #define CHECK_VASTATUS(va_status, func)                                 \
  40     if (va_status != VA_STATUS_SUCCESS) {                               \
  41         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  42         exit(1);                                                        \
  43     }
  44
  45 //#include "loadsurface.h"
  46
  47 #define NAL_REF_IDC_NONE        0
  48 #define NAL_REF_IDC_LOW         1
  49 #define NAL_REF_IDC_MEDIUM      2
  50 #define NAL_REF_IDC_HIGH        3
  51
  52 #define NAL_NON_IDR             1
  53 #define NAL_IDR                 5
  54 #define NAL_SPS                 7
  55 #define NAL_PPS                 8
  56 #define NAL_SEI                 6
  57
  58 #define SLICE_TYPE_P            0
  59 #define SLICE_TYPE_B            1
  60 #define SLICE_TYPE_I            2
  61 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  62 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  63 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  64
  65
  66 #define ENTROPY_MODE_CAVLC      0
  67 #define ENTROPY_MODE_CABAC      1
  68
  69 #define PROFILE_IDC_BASELINE    66
  70 #define PROFILE_IDC_MAIN        77
  71 #define PROFILE_IDC_HIGH        100
  72
  73 #define BITSTREAM_ALLOCATE_STEPPING     4096
  74
  75 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  76 static  VADisplay va_dpy;
  77 static  VAProfile h264_profile = (VAProfile)~0;
  78 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  79 static  int config_attrib_num = 0, enc_packed_header_idx;
  80
  81 struct GLSurface {
  82         VASurfaceID src_surface, ref_surface;
  83         VABufferID coded_buf;
  84
  85         VAImage surface_image;
  86         GLuint y_tex, cbcr_tex;
  87         EGLImage y_egl_image, cbcr_egl_image;
  88 };
  89 GLSurface gl_surfaces[SURFACE_NUM];
  90
  91 static  VAConfigID config_id;
  92 static  VAContextID context_id;
  93 static  VAEncSequenceParameterBufferH264 seq_param;
  94 static  VAEncPictureParameterBufferH264 pic_param;
  95 static  VAEncSliceParameterBufferH264 slice_param;
  96 static  VAPictureH264 CurrentCurrPic;
  97 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  98
  99 static  unsigned int MaxFrameNum = (2<<16);
 100 static  unsigned int MaxPicOrderCntLsb = (2<<8);
 101 static  unsigned int Log2MaxFrameNum = 16;
 102 static  unsigned int Log2MaxPicOrderCntLsb = 8;
 103
 104 static  unsigned int num_ref_frames = 2;
 105 static  unsigned int numShortTerm = 0;
 106 static  int constraint_set_flag = 0;
 107 static  int h264_packedheader = 0; /* support pack header? */
 108 static  int h264_maxref = (1<<16|1);
 109 static  int h264_entropy_mode = 1; /* cabac */
 110
 111 static  int frame_width = 176;
 112 static  int frame_height = 144;
 113 static  int frame_width_mbaligned;
 114 static  int frame_height_mbaligned;
 115 static  unsigned int frame_bitrate = 0;
 116 static  double frame_size = 0;
 117 static  int initial_qp = 15;
 118 //static  int initial_qp = 28;
 119 static  int minimal_qp = 0;
 120 static  int intra_period = 30;
 121 static  int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 122 static  int ip_period = 3;
 123 static  int rc_mode = -1;
 124 static  int rc_default_modes[] = {
 125     VA_RC_VBR,
 126     VA_RC_CQP,
 127     VA_RC_VBR_CONSTRAINED,
 128     VA_RC_CBR,
 129     VA_RC_VCM,
 130     VA_RC_NONE,
 131 };
 132 static  unsigned long long current_frame_encoding = 0;
 133 static  unsigned long long current_frame_display = 0;
 134 static  unsigned long long current_IDR_display = 0;
 135 static  unsigned int current_frame_num = 0;
 136 static  int current_frame_type;
 137
 138 static  int misc_priv_type = 0;
 139 static  int misc_priv_value = 0;
 140
 141 /* thread to save coded data */
 142 #define SRC_SURFACE_FREE        0
 143 #define SRC_SURFACE_IN_ENCODING 1
 144
 145 struct __bitstream {
 146     unsigned int *buffer;
 147     int bit_offset;
 148     int max_size_in_dword;
 149 };
 150 typedef struct __bitstream bitstream;
 151
 152 using namespace std;
 153
 154 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 155 // but if we don't delete it here, we get leaks. The GStreamer implementation
 156 // does the same.
 157 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 158 {
 159     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 160     CHECK_VASTATUS(va_status, "vaRenderPicture");
 161
 162     for (int i = 0; i < num_buffers; ++i) {
 163         va_status = vaDestroyBuffer(va_dpy, buffers[i]);
 164         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 165     }
 166 }
 167
 168 static unsigned int
 169 va_swap32(unsigned int val)
 170 {
 171     unsigned char *pval = (unsigned char *)&val;
 172
 173     return ((pval[0] << 24)     |
 174             (pval[1] << 16)     |
 175             (pval[2] << 8)      |
 176             (pval[3] << 0));
 177 }
 178
 179 static void
 180 bitstream_start(bitstream *bs)
 181 {
 182     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 183     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 184     bs->bit_offset = 0;
 185 }
 186
 187 static void
 188 bitstream_end(bitstream *bs)
 189 {
 190     int pos = (bs->bit_offset >> 5);
 191     int bit_offset = (bs->bit_offset & 0x1f);
 192     int bit_left = 32 - bit_offset;
 193
 194     if (bit_offset) {
 195         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 196     }
 197 }
 198
 199 static void
 200 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 201 {
 202     int pos = (bs->bit_offset >> 5);
 203     int bit_offset = (bs->bit_offset & 0x1f);
 204     int bit_left = 32 - bit_offset;
 205
 206     if (!size_in_bits)
 207         return;
 208
 209     bs->bit_offset += size_in_bits;
 210
 211     if (bit_left > size_in_bits) {
 212         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 213     } else {
 214         size_in_bits -= bit_left;
 215         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 216         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 217
 218         if (pos + 1 == bs->max_size_in_dword) {
 219             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 220             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 221         }
 222
 223         bs->buffer[pos + 1] = val;
 224     }
 225 }
 226
 227 static void
 228 bitstream_put_ue(bitstream *bs, unsigned int val)
 229 {
 230     int size_in_bits = 0;
 231     int tmp_val = ++val;
 232
 233     while (tmp_val) {
 234         tmp_val >>= 1;
 235         size_in_bits++;
 236     }
 237
 238     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 239     bitstream_put_ui(bs, val, size_in_bits);
 240 }
 241
 242 static void
 243 bitstream_put_se(bitstream *bs, int val)
 244 {
 245     unsigned int new_val;
 246
 247     if (val <= 0)
 248         new_val = -2 * val;
 249     else
 250         new_val = 2 * val - 1;
 251
 252     bitstream_put_ue(bs, new_val);
 253 }
 254
 255 static void
 256 bitstream_byte_aligning(bitstream *bs, int bit)
 257 {
 258     int bit_offset = (bs->bit_offset & 0x7);
 259     int bit_left = 8 - bit_offset;
 260     int new_val;
 261
 262     if (!bit_offset)
 263         return;
 264
 265     assert(bit == 0 || bit == 1);
 266
 267     if (bit)
 268         new_val = (1 << bit_left) - 1;
 269     else
 270         new_val = 0;
 271
 272     bitstream_put_ui(bs, new_val, bit_left);
 273 }
 274
 275 static void
 276 rbsp_trailing_bits(bitstream *bs)
 277 {
 278     bitstream_put_ui(bs, 1, 1);
 279     bitstream_byte_aligning(bs, 0);
 280 }
 281
 282 static void nal_start_code_prefix(bitstream *bs)
 283 {
 284     bitstream_put_ui(bs, 0x00000001, 32);
 285 }
 286
 287 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 288 {
 289     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 290     bitstream_put_ui(bs, nal_ref_idc, 2);
 291     bitstream_put_ui(bs, nal_unit_type, 5);
 292 }
 293
 294 static void sps_rbsp(bitstream *bs)
 295 {
 296     int profile_idc = PROFILE_IDC_BASELINE;
 297
 298     if (h264_profile  == VAProfileH264High)
 299         profile_idc = PROFILE_IDC_HIGH;
 300     else if (h264_profile  == VAProfileH264Main)
 301         profile_idc = PROFILE_IDC_MAIN;
 302
 303     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 304     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 305     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 306     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 307     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 308     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 309     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 310     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 311
 312     if ( profile_idc == PROFILE_IDC_HIGH) {
 313         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 314         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 315         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 316         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 317         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 318     }
 319
 320     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 321     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 322
 323     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 324         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 325     else {
 326         assert(0);
 327     }
 328
 329     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 330     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 331
 332     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 333     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 334     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 335
 336     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 337         assert(0);
 338     }
 339
 340     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 341     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 342
 343     if (seq_param.frame_cropping_flag) {
 344         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 345         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 346         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 347         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 348     }
 349
 350     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 351     if ( false ) {
 352         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 353     } else {
 354         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 355         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 356         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 357         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 358         {
 359             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 360             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 361             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 362             {
 363                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 364                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 365                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 366             }
 367         }
 368         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 369         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 370         {
 371             bitstream_put_ui(bs, 1, 32);  // FPS
 372             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 373             bitstream_put_ui(bs, 1, 1);
 374         }
 375         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 376         {
 377             // hrd_parameters
 378             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 379             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 380             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 381
 382             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 383             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 384             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 385
 386             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 387             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 388             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 389             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 390         }
 391         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 392         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 393
 394         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 395         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 396     }
 397
 398     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 399 }
 400
 401
 402 static void pps_rbsp(bitstream *bs)
 403 {
 404     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 405     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 406
 407     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 408
 409     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 410
 411     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 412
 413     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 414     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 415
 416     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 417     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 418
 419     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 420     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 421     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 422
 423     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 424     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 425     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 426
 427     /* more_rbsp_data */
 428     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 429     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 430     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 431
 432     rbsp_trailing_bits(bs);
 433 }
 434
 435 static void slice_header(bitstream *bs)
 436 {
 437     int first_mb_in_slice = slice_param.macroblock_address;
 438
 439     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 440     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 441     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 442     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 443
 444     /* frame_mbs_only_flag == 1 */
 445     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 446         /* FIXME: */
 447         assert(0);
 448     }
 449
 450     if (pic_param.pic_fields.bits.idr_pic_flag)
 451         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 452
 453     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 454         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 455         /* pic_order_present_flag == 0 */
 456     } else {
 457         /* FIXME: */
 458         assert(0);
 459     }
 460
 461     /* redundant_pic_cnt_present_flag == 0 */
 462     /* slice type */
 463     if (IS_P_SLICE(slice_param.slice_type)) {
 464         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 465
 466         if (slice_param.num_ref_idx_active_override_flag)
 467             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 468
 469         /* ref_pic_list_reordering */
 470         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 471     } else if (IS_B_SLICE(slice_param.slice_type)) {
 472         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 473
 474         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 475
 476         if (slice_param.num_ref_idx_active_override_flag) {
 477             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 478             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 479         }
 480
 481         /* ref_pic_list_reordering */
 482         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 483         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 484     }
 485
 486     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 487          IS_P_SLICE(slice_param.slice_type)) ||
 488         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 489          IS_B_SLICE(slice_param.slice_type))) {
 490         /* FIXME: fill weight/offset table */
 491         assert(0);
 492     }
 493
 494     /* dec_ref_pic_marking */
 495     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 496         unsigned char no_output_of_prior_pics_flag = 0;
 497         unsigned char long_term_reference_flag = 0;
 498         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 499
 500         if (pic_param.pic_fields.bits.idr_pic_flag) {
 501             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 502             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 503         } else {
 504             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 505         }
 506     }
 507
 508     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 509         !IS_I_SLICE(slice_param.slice_type))
 510         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 511
 512     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 513
 514     /* ignore for SP/SI */
 515
 516     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 517         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 518
 519         if (slice_param.disable_deblocking_filter_idc != 1) {
 520             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 521             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 522         }
 523     }
 524
 525     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 526         bitstream_byte_aligning(bs, 1);
 527     }
 528 }
 529
 530 static int
 531 build_packed_pic_buffer(unsigned char **header_buffer)
 532 {
 533     bitstream bs;
 534
 535     bitstream_start(&bs);
 536     nal_start_code_prefix(&bs);
 537     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 538     pps_rbsp(&bs);
 539     bitstream_end(&bs);
 540
 541     *header_buffer = (unsigned char *)bs.buffer;
 542     return bs.bit_offset;
 543 }
 544
 545 static int
 546 build_packed_seq_buffer(unsigned char **header_buffer)
 547 {
 548     bitstream bs;
 549
 550     bitstream_start(&bs);
 551     nal_start_code_prefix(&bs);
 552     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 553     sps_rbsp(&bs);
 554     bitstream_end(&bs);
 555
 556     *header_buffer = (unsigned char *)bs.buffer;
 557     return bs.bit_offset;
 558 }
 559
 560 static int build_packed_slice_buffer(unsigned char **header_buffer)
 561 {
 562     bitstream bs;
 563     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 564     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 565
 566     bitstream_start(&bs);
 567     nal_start_code_prefix(&bs);
 568
 569     if (IS_I_SLICE(slice_param.slice_type)) {
 570         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 571     } else if (IS_P_SLICE(slice_param.slice_type)) {
 572         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 573     } else {
 574         assert(IS_B_SLICE(slice_param.slice_type));
 575         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 576     }
 577
 578     slice_header(&bs);
 579     bitstream_end(&bs);
 580
 581     *header_buffer = (unsigned char *)bs.buffer;
 582     return bs.bit_offset;
 583 }
 584
 585
 586 /*
 587   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 588   1) period between Frame #X and Frame #N = #X - #N
 589   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 590   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 591   4) intra_period and intra_idr_period take precedence over ip_period
 592   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 593      of I/IDR frames, see bellow examples
 594   -------------------------------------------------------------------
 595   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 596   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 597   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 598   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 599   1            1                ignored    IDR IDR IDR IDR...
 600   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 601   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 602   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 603                                               (PBB)(IBB)(PBB)(IBB)...
 604   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 605                                            IDRPPPPP IPPPPP IPPPPP...
 606   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 607                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 608                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 609                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 610                                            {IDR(PBB)(PBB)}                     (6/6/3)
 611                                            {IDR(PBB)(PBB)}.
 612 */
 613
 614 // General pts/dts strategy:
 615 //
 616 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 617 // bit tricky. We assume first of all that the frame rate never goes _above_
 618 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 619 // in at least this speed, as long at dts <= pts (the frame is not attempted
 620 // presented before it is decoded). Furthermore, we never have longer chains of
 621 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 622 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 623 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 624 // frame, with an order O <= C.)
 625 //
 626 // Many strategies are possible, but we establish these rules:
 627 //
 628 //  - Tip frames have dts = pts - (C-O)*N.
 629 //  - Non-tip frames have dts = dts_last + N.
 630 //
 631 // An example, with C=2 and N=10 and the data flow showed with arrows:
 632 //
 633 //        I  B  P  B  B  P
 634 //   pts: 30 40 50 60 70 80
 635 //        ↓  ↓     ↓
 636 //   dts: 10 30 20 60 50←40
 637 //         |  |  ↑        ↑
 638 //         `--|--'        |
 639 //             `----------'
 640 //
 641 // To show that this works fine also with irregular spacings, let's say that
 642 // the third frame is delayed a bit (something earlier was dropped). Now the
 643 // situation looks like this:
 644 //
 645 //        I  B  P  B  B   P
 646 //   pts: 30 40 80 90 100 110
 647 //        ↓  ↓     ↓
 648 //   dts: 10 30 20 90 50←40
 649 //         |  |  ↑        ↑
 650 //         `--|--'        |
 651 //             `----------'
 652 //
 653 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 654 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 655 //
 656 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 657 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 658 // a tip frame and should be given a dts based on the previous one.
 659 #define FRAME_P 0
 660 #define FRAME_B 1
 661 #define FRAME_I 2
 662 #define FRAME_IDR 7
 663 void encoding2display_order(
 664     unsigned long long encoding_order, int intra_period,
 665     int intra_idr_period, int ip_period,
 666     unsigned long long *displaying_order,
 667     int *frame_type, int *pts_lag)
 668 {
 669     int encoding_order_gop = 0;
 670
 671     *pts_lag = 0;
 672
 673     if (intra_period == 1) { /* all are I/IDR frames */
 674         *displaying_order = encoding_order;
 675         if (intra_idr_period == 0)
 676             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 677         else
 678             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 679         return;
 680     }
 681
 682     if (intra_period == 0)
 683         intra_idr_period = 0;
 684
 685     if (ip_period == 1) {
 686         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 687         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 688         *displaying_order = encoding_order;
 689
 690         if (encoding_order_gop == 0) { /* the first frame */
 691             *frame_type = FRAME_IDR;
 692         } else if (intra_period != 0 && /* have I frames */
 693                    encoding_order_gop >= 2 &&
 694                    (encoding_order_gop % intra_period == 0)) {
 695             *frame_type = FRAME_I;
 696         } else {
 697             *frame_type = FRAME_P;
 698         }
 699         return;
 700     }
 701
 702     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 703     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 704     *pts_lag = -1;  // Most frames are not tip frames.
 705
 706     if (encoding_order_gop == 0) { /* the first frame */
 707         *frame_type = FRAME_IDR;
 708         *displaying_order = encoding_order;
 709         // IDR frames are a special case; I honestly can't find the logic behind
 710         // why this is the right thing, but it seems to line up nicely in practice :-)
 711         *pts_lag = TIMEBASE / MAX_FPS;
 712     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 713         *frame_type = FRAME_B;
 714         *displaying_order = encoding_order - 1;
 715         if ((encoding_order_gop % ip_period) == 0) {
 716             *pts_lag = 0;  // Last B-frame.
 717         }
 718     } else if (intra_period != 0 && /* have I frames */
 719                encoding_order_gop >= 2 &&
 720                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 721         *frame_type = FRAME_I;
 722         *displaying_order = encoding_order + ip_period - 1;
 723     } else {
 724         *frame_type = FRAME_P;
 725         *displaying_order = encoding_order + ip_period - 1;
 726     }
 727 }
 728
 729
 730 static const char *rc_to_string(int rcmode)
 731 {
 732     switch (rc_mode) {
 733     case VA_RC_NONE:
 734         return "NONE";
 735     case VA_RC_CBR:
 736         return "CBR";
 737     case VA_RC_VBR:
 738         return "VBR";
 739     case VA_RC_VCM:
 740         return "VCM";
 741     case VA_RC_CQP:
 742         return "CQP";
 743     case VA_RC_VBR_CONSTRAINED:
 744         return "VBR_CONSTRAINED";
 745     default:
 746         return "Unknown";
 747     }
 748 }
 749
 750 #if 0
 751 static int process_cmdline(int argc, char *argv[])
 752 {
 753     char c;
 754     const struct option long_opts[] = {
 755         {"help", no_argument, NULL, 0 },
 756         {"bitrate", required_argument, NULL, 1 },
 757         {"minqp", required_argument, NULL, 2 },
 758         {"initialqp", required_argument, NULL, 3 },
 759         {"intra_period", required_argument, NULL, 4 },
 760         {"idr_period", required_argument, NULL, 5 },
 761         {"ip_period", required_argument, NULL, 6 },
 762         {"rcmode", required_argument, NULL, 7 },
 763         {"srcyuv", required_argument, NULL, 9 },
 764         {"recyuv", required_argument, NULL, 10 },
 765         {"fourcc", required_argument, NULL, 11 },
 766         {"syncmode", no_argument, NULL, 12 },
 767         {"enablePSNR", no_argument, NULL, 13 },
 768         {"prit", required_argument, NULL, 14 },
 769         {"priv", required_argument, NULL, 15 },
 770         {"framecount", required_argument, NULL, 16 },
 771         {"entropy", required_argument, NULL, 17 },
 772         {"profile", required_argument, NULL, 18 },
 773         {NULL, no_argument, NULL, 0 }};
 774     int long_index;
 775
 776     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 777         switch (c) {
 778         case 'w':
 779             frame_width = atoi(optarg);
 780             break;
 781         case 'h':
 782             frame_height = atoi(optarg);
 783             break;
 784         case 'n':
 785         case 'f':
 786             frame_rate = atoi(optarg);
 787             break;
 788         case 'o':
 789             coded_fn = strdup(optarg);
 790             break;
 791         case 0:
 792             print_help();
 793             exit(0);
 794         case 1:
 795             frame_bitrate = atoi(optarg);
 796             break;
 797         case 2:
 798             minimal_qp = atoi(optarg);
 799             break;
 800         case 3:
 801             initial_qp = atoi(optarg);
 802             break;
 803         case 4:
 804             intra_period = atoi(optarg);
 805             break;
 806         case 5:
 807             intra_idr_period = atoi(optarg);
 808             break;
 809         case 6:
 810             ip_period = atoi(optarg);
 811             break;
 812         case 7:
 813             rc_mode = string_to_rc(optarg);
 814             if (rc_mode < 0) {
 815                 print_help();
 816                 exit(1);
 817             }
 818             break;
 819         case 9:
 820             srcyuv_fn = strdup(optarg);
 821             break;
 822         case 11:
 823             srcyuv_fourcc = string_to_fourcc(optarg);
 824             if (srcyuv_fourcc <= 0) {
 825                 print_help();
 826                 exit(1);
 827             }
 828             break;
 829         case 13:
 830             calc_psnr = 1;
 831             break;
 832         case 14:
 833             misc_priv_type = strtol(optarg, NULL, 0);
 834             break;
 835         case 15:
 836             misc_priv_value = strtol(optarg, NULL, 0);
 837             break;
 838         case 17:
 839             h264_entropy_mode = atoi(optarg) ? 1: 0;
 840             break;
 841         case 18:
 842             if (strncmp(optarg, "BP", 2) == 0)
 843                 h264_profile = VAProfileH264Baseline;
 844             else if (strncmp(optarg, "MP", 2) == 0)
 845                 h264_profile = VAProfileH264Main;
 846             else if (strncmp(optarg, "HP", 2) == 0)
 847                 h264_profile = VAProfileH264High;
 848             else
 849                 h264_profile = (VAProfile)0;
 850             break;
 851         case ':':
 852         case '?':
 853             print_help();
 854             exit(0);
 855         }
 856     }
 857
 858     if (ip_period < 1) {
 859         printf(" ip_period must be greater than 0\n");
 860         exit(0);
 861     }
 862     if (intra_period != 1 && intra_period % ip_period != 0) {
 863         printf(" intra_period must be a multiplier of ip_period\n");
 864         exit(0);
 865     }
 866     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 867         printf(" intra_idr_period must be a multiplier of intra_period\n");
 868         exit(0);
 869     }
 870
 871     if (frame_bitrate == 0)
 872         frame_bitrate = frame_width * frame_height * 12 * MAX_FPS / 50;
 873
 874     if (coded_fn == NULL) {
 875         struct stat buf;
 876         if (stat("/tmp", &buf) == 0)
 877             coded_fn = strdup("/tmp/test.264");
 878         else if (stat("/sdcard", &buf) == 0)
 879             coded_fn = strdup("/sdcard/test.264");
 880         else
 881             coded_fn = strdup("./test.264");
 882     }
 883
 884
 885     frame_width_mbaligned = (frame_width + 15) & (~15);
 886     frame_height_mbaligned = (frame_height + 15) & (~15);
 887     if (frame_width != frame_width_mbaligned ||
 888         frame_height != frame_height_mbaligned) {
 889         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 890                frame_width, frame_height,
 891                frame_width_mbaligned, frame_height_mbaligned
 892                );
 893     }
 894
 895     return 0;
 896 }
 897 #endif
 898
 899 static Display *x11_display;
 900 static Window   x11_window;
 901
 902 VADisplay
 903 va_open_display(void)
 904 {
 905     x11_display = XOpenDisplay(NULL);
 906     if (!x11_display) {
 907         fprintf(stderr, "error: can't connect to X server!\n");
 908         return NULL;
 909     }
 910     return vaGetDisplay(x11_display);
 911 }
 912
 913 void
 914 va_close_display(VADisplay va_dpy)
 915 {
 916     if (!x11_display)
 917         return;
 918
 919     if (x11_window) {
 920         XUnmapWindow(x11_display, x11_window);
 921         XDestroyWindow(x11_display, x11_window);
 922         x11_window = None;
 923     }
 924     XCloseDisplay(x11_display);
 925     x11_display = NULL;
 926 }
 927
 928 static int init_va(void)
 929 {
 930     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 931     VAEntrypoint *entrypoints;
 932     int num_entrypoints, slice_entrypoint;
 933     int support_encode = 0;
 934     int major_ver, minor_ver;
 935     VAStatus va_status;
 936     unsigned int i;
 937
 938     va_dpy = va_open_display();
 939     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 940     CHECK_VASTATUS(va_status, "vaInitialize");
 941
 942     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 943     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 944     if (!entrypoints) {
 945         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 946         exit(1);
 947     }
 948
 949     /* use the highest profile */
 950     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 951         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 952             continue;
 953
 954         h264_profile = profile_list[i];
 955         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 956         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 957             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 958                 support_encode = 1;
 959                 break;
 960             }
 961         }
 962         if (support_encode == 1)
 963             break;
 964     }
 965
 966     if (support_encode == 0) {
 967         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 968         exit(1);
 969     } else {
 970         switch (h264_profile) {
 971             case VAProfileH264Baseline:
 972                 ip_period = 1;
 973                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 974                 h264_entropy_mode = 0;
 975                 break;
 976             case VAProfileH264ConstrainedBaseline:
 977                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 978                 ip_period = 1;
 979                 break;
 980
 981             case VAProfileH264Main:
 982                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 983                 break;
 984
 985             case VAProfileH264High:
 986                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 987                 break;
 988             default:
 989                 h264_profile = VAProfileH264Baseline;
 990                 ip_period = 1;
 991                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 992                 break;
 993         }
 994     }
 995
 996     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 997
 998     /* find out the format for the render target, and rate control mode */
 999     for (i = 0; i < VAConfigAttribTypeMax; i++)
1000         attrib[i].type = (VAConfigAttribType)i;
1001
1002     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1003                                       &attrib[0], VAConfigAttribTypeMax);
1004     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1005     /* check the interested configattrib */
1006     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1007         printf("Not find desired YUV420 RT format\n");
1008         exit(1);
1009     } else {
1010         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1011         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1012         config_attrib_num++;
1013     }
1014
1015     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1016         int tmp = attrib[VAConfigAttribRateControl].value;
1017
1018         if (rc_mode == -1 || !(rc_mode & tmp))  {
1019             if (rc_mode != -1) {
1020                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1021             }
1022
1023             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1024                 if (rc_default_modes[i] & tmp) {
1025                     rc_mode = rc_default_modes[i];
1026                     break;
1027                 }
1028             }
1029         }
1030
1031         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1032         config_attrib[config_attrib_num].value = rc_mode;
1033         config_attrib_num++;
1034     }
1035
1036
1037     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1038         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1039
1040         h264_packedheader = 1;
1041         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1042         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1043
1044         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1045             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1046         }
1047
1048         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1049             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1050         }
1051
1052         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1053             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1054         }
1055
1056         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1057             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1058         }
1059
1060         enc_packed_header_idx = config_attrib_num;
1061         config_attrib_num++;
1062     }
1063
1064     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1065         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1066         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1067         config_attrib_num++;
1068     }
1069
1070     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1071         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1072     }
1073
1074     free(entrypoints);
1075     return 0;
1076 }
1077
1078 static int setup_encode()
1079 {
1080     VAStatus va_status;
1081     VASurfaceID *tmp_surfaceid;
1082     int codedbuf_size, i;
1083     static VASurfaceID src_surface[SURFACE_NUM];
1084     static VASurfaceID ref_surface[SURFACE_NUM];
1085
1086     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1087             &config_attrib[0], config_attrib_num, &config_id);
1088     CHECK_VASTATUS(va_status, "vaCreateConfig");
1089
1090     /* create source surfaces */
1091     va_status = vaCreateSurfaces(va_dpy,
1092                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1093                                  &src_surface[0], SURFACE_NUM,
1094                                  NULL, 0);
1095     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1096
1097     /* create reference surfaces */
1098     va_status = vaCreateSurfaces(va_dpy,
1099                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1100                                  &ref_surface[0], SURFACE_NUM,
1101                                  NULL, 0);
1102     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1103
1104     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1105     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1106     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1107
1108     /* Create a context for this encode pipe */
1109     va_status = vaCreateContext(va_dpy, config_id,
1110                                 frame_width_mbaligned, frame_height_mbaligned,
1111                                 VA_PROGRESSIVE,
1112                                 tmp_surfaceid, 2 * SURFACE_NUM,
1113                                 &context_id);
1114     CHECK_VASTATUS(va_status, "vaCreateContext");
1115     free(tmp_surfaceid);
1116
1117     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1118
1119     for (i = 0; i < SURFACE_NUM; i++) {
1120         /* create coded buffer once for all
1121          * other VA buffers which won't be used again after vaRenderPicture.
1122          * so APP can always vaCreateBuffer for every frame
1123          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1124          * so VA won't maintain the coded buffer
1125          */
1126         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1127                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1128         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1129     }
1130
1131     /* create OpenGL objects */
1132     //glGenFramebuffers(SURFACE_NUM, fbos);
1133
1134     for (i = 0; i < SURFACE_NUM; i++) {
1135         glGenTextures(1, &gl_surfaces[i].y_tex);
1136         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1137     }
1138
1139     for (i = 0; i < SURFACE_NUM; i++) {
1140         gl_surfaces[i].src_surface = src_surface[i];
1141         gl_surfaces[i].ref_surface = ref_surface[i];
1142     }
1143
1144     return 0;
1145 }
1146
1147
1148
1149 #define partition(ref, field, key, ascending)   \
1150     while (i <= j) {                            \
1151         if (ascending) {                        \
1152             while (ref[i].field < key)          \
1153                 i++;                            \
1154             while (ref[j].field > key)          \
1155                 j--;                            \
1156         } else {                                \
1157             while (ref[i].field > key)          \
1158                 i++;                            \
1159             while (ref[j].field < key)          \
1160                 j--;                            \
1161         }                                       \
1162         if (i <= j) {                           \
1163             tmp = ref[i];                       \
1164             ref[i] = ref[j];                    \
1165             ref[j] = tmp;                       \
1166             i++;                                \
1167             j--;                                \
1168         }                                       \
1169     }                                           \
1170
1171 static void sort_one(VAPictureH264 ref[], int left, int right,
1172                      int ascending, int frame_idx)
1173 {
1174     int i = left, j = right;
1175     unsigned int key;
1176     VAPictureH264 tmp;
1177
1178     if (frame_idx) {
1179         key = ref[(left + right) / 2].frame_idx;
1180         partition(ref, frame_idx, key, ascending);
1181     } else {
1182         key = ref[(left + right) / 2].TopFieldOrderCnt;
1183         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1184     }
1185
1186     /* recursion */
1187     if (left < j)
1188         sort_one(ref, left, j, ascending, frame_idx);
1189
1190     if (i < right)
1191         sort_one(ref, i, right, ascending, frame_idx);
1192 }
1193
1194 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1195                      int partition_ascending, int list0_ascending, int list1_ascending)
1196 {
1197     int i = left, j = right;
1198     VAPictureH264 tmp;
1199
1200     if (frame_idx) {
1201         partition(ref, frame_idx, key, partition_ascending);
1202     } else {
1203         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1204     }
1205
1206
1207     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1208     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1209 }
1210
1211 static int update_ReferenceFrames(void)
1212 {
1213     int i;
1214
1215     if (current_frame_type == FRAME_B)
1216         return 0;
1217
1218     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1219     numShortTerm++;
1220     if (numShortTerm > num_ref_frames)
1221         numShortTerm = num_ref_frames;
1222     for (i=numShortTerm-1; i>0; i--)
1223         ReferenceFrames[i] = ReferenceFrames[i-1];
1224     ReferenceFrames[0] = CurrentCurrPic;
1225
1226     if (current_frame_type != FRAME_B)
1227         current_frame_num++;
1228     if (current_frame_num > MaxFrameNum)
1229         current_frame_num = 0;
1230
1231     return 0;
1232 }
1233
1234
1235 static int update_RefPicList(void)
1236 {
1237     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1238
1239     if (current_frame_type == FRAME_P) {
1240         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1241         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1242     }
1243
1244     if (current_frame_type == FRAME_B) {
1245         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1246         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1247                  1, 0, 1);
1248
1249         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1250         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1251                  0, 1, 0);
1252     }
1253
1254     return 0;
1255 }
1256
1257
1258 static int render_sequence(void)
1259 {
1260     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1261     VAStatus va_status;
1262     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1263     VAEncMiscParameterRateControl *misc_rate_ctrl;
1264
1265     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1266     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1267     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1268     seq_param.bits_per_second = frame_bitrate;
1269
1270     seq_param.intra_period = intra_period;
1271     seq_param.intra_idr_period = intra_idr_period;
1272     seq_param.ip_period = ip_period;
1273
1274     seq_param.max_num_ref_frames = num_ref_frames;
1275     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1276     seq_param.time_scale = TIMEBASE * 2;
1277     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1278     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1279     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1280     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1281     seq_param.seq_fields.bits.chroma_format_idc = 1;
1282     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1283
1284     if (frame_width != frame_width_mbaligned ||
1285         frame_height != frame_height_mbaligned) {
1286         seq_param.frame_cropping_flag = 1;
1287         seq_param.frame_crop_left_offset = 0;
1288         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1289         seq_param.frame_crop_top_offset = 0;
1290         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1291     }
1292
1293     va_status = vaCreateBuffer(va_dpy, context_id,
1294                                VAEncSequenceParameterBufferType,
1295                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1296     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1297
1298     va_status = vaCreateBuffer(va_dpy, context_id,
1299                                VAEncMiscParameterBufferType,
1300                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1301                                1, NULL, &rc_param_buf);
1302     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1303
1304     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1305     misc_param->type = VAEncMiscParameterTypeRateControl;
1306     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1307     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1308     misc_rate_ctrl->bits_per_second = frame_bitrate;
1309     misc_rate_ctrl->target_percentage = 66;
1310     misc_rate_ctrl->window_size = 1000;
1311     misc_rate_ctrl->initial_qp = initial_qp;
1312     misc_rate_ctrl->min_qp = minimal_qp;
1313     misc_rate_ctrl->basic_unit_size = 0;
1314     vaUnmapBuffer(va_dpy, rc_param_buf);
1315
1316     render_id[0] = seq_param_buf;
1317     render_id[1] = rc_param_buf;
1318
1319     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1320
1321     if (misc_priv_type != 0) {
1322         va_status = vaCreateBuffer(va_dpy, context_id,
1323                                    VAEncMiscParameterBufferType,
1324                                    sizeof(VAEncMiscParameterBuffer),
1325                                    1, NULL, &misc_param_tmpbuf);
1326         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1327         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1328         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1329         misc_param_tmp->data[0] = misc_priv_value;
1330         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1331
1332         render_picture_and_delete(va_dpy, context_id, &misc_param_tmpbuf, 1);
1333     }
1334
1335     return 0;
1336 }
1337
1338 static int calc_poc(int pic_order_cnt_lsb)
1339 {
1340     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1341     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1342     int PicOrderCntMsb, TopFieldOrderCnt;
1343
1344     if (current_frame_type == FRAME_IDR)
1345         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1346     else {
1347         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1348         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1349     }
1350
1351     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1352         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1353         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1354     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1355              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1356         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1357     else
1358         PicOrderCntMsb = prevPicOrderCntMsb;
1359
1360     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1361
1362     if (current_frame_type != FRAME_B) {
1363         PicOrderCntMsb_ref = PicOrderCntMsb;
1364         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1365     }
1366
1367     return TopFieldOrderCnt;
1368 }
1369
1370 static int render_picture(void)
1371 {
1372     VABufferID pic_param_buf;
1373     VAStatus va_status;
1374     int i = 0;
1375
1376     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1377     pic_param.CurrPic.frame_idx = current_frame_num;
1378     pic_param.CurrPic.flags = 0;
1379     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1380     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1381     CurrentCurrPic = pic_param.CurrPic;
1382
1383     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1384         update_RefPicList(); /* calc RefPicList */
1385         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1386         if (current_frame_type == FRAME_P) {
1387             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1388         } else if (current_frame_type == FRAME_B) {
1389             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1390             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1391         }
1392     } else {
1393         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1394         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1395             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1396             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1397         }
1398     }
1399
1400     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1401     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1402     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1403     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1404     pic_param.frame_num = current_frame_num;
1405     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1406     pic_param.last_picture = false;  // FIXME
1407     pic_param.pic_init_qp = initial_qp;
1408
1409     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1410                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1411     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1412
1413     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1414
1415     return 0;
1416 }
1417
1418 static int render_packedsequence(void)
1419 {
1420     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1421     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1422     unsigned int length_in_bits;
1423     unsigned char *packedseq_buffer = NULL;
1424     VAStatus va_status;
1425
1426     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1427
1428     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1429
1430     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1431     packedheader_param_buffer.has_emulation_bytes = 0;
1432     va_status = vaCreateBuffer(va_dpy,
1433                                context_id,
1434                                VAEncPackedHeaderParameterBufferType,
1435                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1436                                &packedseq_para_bufid);
1437     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1438
1439     va_status = vaCreateBuffer(va_dpy,
1440                                context_id,
1441                                VAEncPackedHeaderDataBufferType,
1442                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1443                                &packedseq_data_bufid);
1444     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1445
1446     render_id[0] = packedseq_para_bufid;
1447     render_id[1] = packedseq_data_bufid;
1448     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1449
1450     free(packedseq_buffer);
1451
1452     return 0;
1453 }
1454
1455
1456 static int render_packedpicture(void)
1457 {
1458     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1459     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1460     unsigned int length_in_bits;
1461     unsigned char *packedpic_buffer = NULL;
1462     VAStatus va_status;
1463
1464     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1465     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1466     packedheader_param_buffer.bit_length = length_in_bits;
1467     packedheader_param_buffer.has_emulation_bytes = 0;
1468
1469     va_status = vaCreateBuffer(va_dpy,
1470                                context_id,
1471                                VAEncPackedHeaderParameterBufferType,
1472                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1473                                &packedpic_para_bufid);
1474     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1475
1476     va_status = vaCreateBuffer(va_dpy,
1477                                context_id,
1478                                VAEncPackedHeaderDataBufferType,
1479                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1480                                &packedpic_data_bufid);
1481     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1482
1483     render_id[0] = packedpic_para_bufid;
1484     render_id[1] = packedpic_data_bufid;
1485     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1486
1487     free(packedpic_buffer);
1488
1489     return 0;
1490 }
1491
1492 static void render_packedslice()
1493 {
1494     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1495     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1496     unsigned int length_in_bits;
1497     unsigned char *packedslice_buffer = NULL;
1498     VAStatus va_status;
1499
1500     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1501     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1502     packedheader_param_buffer.bit_length = length_in_bits;
1503     packedheader_param_buffer.has_emulation_bytes = 0;
1504
1505     va_status = vaCreateBuffer(va_dpy,
1506                                context_id,
1507                                VAEncPackedHeaderParameterBufferType,
1508                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1509                                &packedslice_para_bufid);
1510     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1511
1512     va_status = vaCreateBuffer(va_dpy,
1513                                context_id,
1514                                VAEncPackedHeaderDataBufferType,
1515                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1516                                &packedslice_data_bufid);
1517     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1518
1519     render_id[0] = packedslice_para_bufid;
1520     render_id[1] = packedslice_data_bufid;
1521     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1522
1523     free(packedslice_buffer);
1524 }
1525
1526 static int render_slice(void)
1527 {
1528     VABufferID slice_param_buf;
1529     VAStatus va_status;
1530     int i;
1531
1532     update_RefPicList();
1533
1534     /* one frame, one slice */
1535     slice_param.macroblock_address = 0;
1536     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1537     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1538     if (current_frame_type == FRAME_IDR) {
1539         if (current_frame_encoding != 0)
1540             ++slice_param.idr_pic_id;
1541     } else if (current_frame_type == FRAME_P) {
1542         int refpiclist0_max = h264_maxref & 0xffff;
1543         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1544
1545         for (i = refpiclist0_max; i < 32; i++) {
1546             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1547             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1548         }
1549     } else if (current_frame_type == FRAME_B) {
1550         int refpiclist0_max = h264_maxref & 0xffff;
1551         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1552
1553         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1554         for (i = refpiclist0_max; i < 32; i++) {
1555             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1556             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1557         }
1558
1559         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1560         for (i = refpiclist1_max; i < 32; i++) {
1561             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1562             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1563         }
1564     }
1565
1566     slice_param.slice_alpha_c0_offset_div2 = 0;
1567     slice_param.slice_beta_offset_div2 = 0;
1568     slice_param.direct_spatial_mv_pred_flag = 1;
1569     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1570
1571
1572     if (h264_packedheader &&
1573         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1574         render_packedslice();
1575
1576     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1577                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1578     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1579
1580     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1581
1582     return 0;
1583 }
1584
1585
1586
1587 void H264Encoder::save_codeddata(storage_task task)
1588 {
1589     VACodedBufferSegment *buf_list = NULL;
1590     VAStatus va_status;
1591     unsigned int coded_size = 0;
1592
1593     string data;
1594
1595     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1596
1597     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1598     CHECK_VASTATUS(va_status, "vaMapBuffer");
1599     while (buf_list != NULL) {
1600         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1601         buf_list = (VACodedBufferSegment *) buf_list->next;
1602
1603         frame_size += coded_size;
1604     }
1605     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1606
1607     {
1608         // Add video.
1609         AVPacket pkt;
1610         memset(&pkt, 0, sizeof(pkt));
1611         pkt.buf = nullptr;
1612         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1613         pkt.size = data.size();
1614         pkt.stream_index = 0;
1615         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1616             pkt.flags = AV_PKT_FLAG_KEY;
1617         } else {
1618             pkt.flags = 0;
1619         }
1620         //pkt.duration = 1;
1621         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1622     }
1623     // Encode and add all audio frames up to and including the pts of this video frame.
1624     for ( ;; ) {
1625         int64_t audio_pts;
1626         std::vector<float> audio;
1627         {
1628              unique_lock<mutex> lock(frame_queue_mutex);
1629              frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || !pending_audio_frames.empty(); });
1630              if (copy_thread_should_quit) return;
1631              auto it = pending_audio_frames.begin();
1632              if (it->first > task.pts) break;
1633              audio_pts = it->first;
1634              audio = move(it->second);
1635              pending_audio_frames.erase(it);
1636         }
1637
1638         AVFrame *frame = avcodec_alloc_frame();
1639         frame->nb_samples = audio.size() / 2;
1640         frame->format = AV_SAMPLE_FMT_S32;
1641         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1642
1643         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1644         int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1645         if (ret < 0) {
1646             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1647             exit(1);
1648         }
1649         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1650             if (audio[i] >= 1.0f) {
1651                 int_samples[i] = 2147483647;
1652             } else if (audio[i] <= -1.0f) {
1653                 int_samples[i] = -2147483647;
1654             } else {
1655                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1656             }
1657         }
1658
1659         AVPacket pkt;
1660         av_init_packet(&pkt);
1661         pkt.data = nullptr;
1662         pkt.size = 0;
1663         int got_output;
1664         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1665         if (got_output) {
1666             pkt.stream_index = 1;
1667             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1668         }
1669         // TODO: Delayed frames.
1670         avcodec_free_frame(&frame);
1671         av_free_packet(&pkt);
1672         if (audio_pts == task.pts) break;
1673     }
1674
1675 #if 0
1676     printf("\r      "); /* return back to startpoint */
1677     switch (encode_order % 4) {
1678         case 0:
1679             printf("|");
1680             break;
1681         case 1:
1682             printf("/");
1683             break;
1684         case 2:
1685             printf("-");
1686             break;
1687         case 3:
1688             printf("\\");
1689             break;
1690     }
1691     printf("%08lld", encode_order);
1692     printf("(%06d bytes coded)", coded_size);
1693 #endif
1694 }
1695
1696
1697 // this is weird. but it seems to put a new frame onto the queue
1698 void H264Encoder::storage_task_enqueue(storage_task task)
1699 {
1700         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1701         storage_task_queue.push(move(task));
1702         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1703         storage_task_queue_changed.notify_all();
1704 }
1705
1706 void H264Encoder::storage_task_thread()
1707 {
1708         for ( ;; ) {
1709                 storage_task current;
1710                 {
1711                         // wait until there's an encoded frame
1712                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1713                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1714                         if (storage_thread_should_quit) return;
1715                         current = move(storage_task_queue.front());
1716                         storage_task_queue.pop();
1717                 }
1718
1719                 VAStatus va_status;
1720
1721                 // waits for data, then saves it to disk.
1722                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1723                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1724                 save_codeddata(move(current));
1725
1726                 {
1727                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1728                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1729                         storage_task_queue_changed.notify_all();
1730                 }
1731         }
1732 }
1733
1734 static int release_encode()
1735 {
1736     int i;
1737
1738     for (i = 0; i < SURFACE_NUM; i++) {
1739         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1740         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1741         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1742     }
1743
1744     vaDestroyContext(va_dpy, context_id);
1745     vaDestroyConfig(va_dpy, config_id);
1746
1747     return 0;
1748 }
1749
1750 static int deinit_va()
1751 {
1752     vaTerminate(va_dpy);
1753
1754     va_close_display(va_dpy);
1755
1756     return 0;
1757 }
1758
1759
1760 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1761         : current_storage_frame(0), surface(surface), httpd(httpd)
1762 {
1763         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1764         context_audio = avcodec_alloc_context3(codec_audio);
1765         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1766         context_audio->sample_rate = OUTPUT_FREQUENCY;
1767         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1768         context_audio->channels = 2;
1769         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1770         context_audio->time_base = AVRational{1, TIMEBASE};
1771         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1772                 fprintf(stderr, "Could not open codec\n");
1773                 exit(1);
1774         }
1775
1776         frame_width = width;
1777         frame_height = height;
1778         frame_width_mbaligned = (frame_width + 15) & (~15);
1779         frame_height_mbaligned = (frame_height + 15) & (~15);
1780         frame_bitrate = 15000000;  // / 60;
1781         current_frame_encoding = 0;
1782
1783         //print_input();
1784
1785         init_va();
1786         setup_encode();
1787
1788         // No frames are ready yet.
1789         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1790
1791         memset(&seq_param, 0, sizeof(seq_param));
1792         memset(&pic_param, 0, sizeof(pic_param));
1793         memset(&slice_param, 0, sizeof(slice_param));
1794
1795         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1796
1797         copy_thread = std::thread([this]{
1798                 //SDL_GL_MakeCurrent(window, context);
1799                 QOpenGLContext *context = create_context(this->surface);
1800                 eglBindAPI(EGL_OPENGL_API);
1801                 if (!make_current(context, this->surface)) {
1802                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1803                                 eglGetError());
1804                         exit(1);
1805                 }
1806                 copy_thread_func();
1807         });
1808 }
1809
1810 H264Encoder::~H264Encoder()
1811 {
1812         {
1813                 unique_lock<mutex> lock(storage_task_queue_mutex);
1814                 storage_thread_should_quit = true;
1815                 storage_task_queue_changed.notify_all();
1816         }
1817         {
1818                 unique_lock<mutex> lock(frame_queue_mutex);
1819                 copy_thread_should_quit = true;
1820                 frame_queue_nonempty.notify_all();
1821         }
1822         storage_thread.join();
1823         copy_thread.join();
1824
1825         release_encode();
1826         deinit_va();
1827 }
1828
1829 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1830 {
1831         {
1832                 // Wait until this frame slot is done encoding.
1833                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1834                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1835                 if (storage_thread_should_quit) return false;
1836         }
1837
1838         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1839         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1840         *y_tex = surf->y_tex;
1841         *cbcr_tex = surf->cbcr_tex;
1842
1843         VASurfaceID surface = surf->src_surface;
1844         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1845         CHECK_VASTATUS(va_status, "vaDeriveImage");
1846
1847         VABufferInfo buf_info;
1848         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1849         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1850         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1851
1852         // Create Y image.
1853         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1854         EGLint y_attribs[] = {
1855                 EGL_WIDTH, frame_width,
1856                 EGL_HEIGHT, frame_height,
1857                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1858                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1859                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1860                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1861                 EGL_NONE
1862         };
1863
1864         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1865         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1866
1867         // Associate Y image to a texture.
1868         glBindTexture(GL_TEXTURE_2D, *y_tex);
1869         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1870
1871         // Create CbCr image.
1872         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1873         EGLint cbcr_attribs[] = {
1874                 EGL_WIDTH, frame_width,
1875                 EGL_HEIGHT, frame_height,
1876                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1877                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1878                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1879                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1880                 EGL_NONE
1881         };
1882
1883         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1884         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1885
1886         // Associate CbCr image to a texture.
1887         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1888         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1889
1890         return true;
1891 }
1892
1893 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1894 {
1895         {
1896                 unique_lock<mutex> lock(frame_queue_mutex);
1897                 pending_audio_frames[pts] = move(audio);
1898         }
1899         frame_queue_nonempty.notify_all();
1900 }
1901
1902 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
1903 {
1904         {
1905                 unique_lock<mutex> lock(frame_queue_mutex);
1906                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1907                 ++current_storage_frame;
1908         }
1909         frame_queue_nonempty.notify_all();
1910 }
1911
1912 void H264Encoder::copy_thread_func()
1913 {
1914         int64_t last_dts = -1;
1915         for ( ;; ) {
1916                 PendingFrame frame;
1917                 int pts_lag;
1918                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1919                                        &current_frame_display, &current_frame_type, &pts_lag);
1920                 if (current_frame_type == FRAME_IDR) {
1921                         numShortTerm = 0;
1922                         current_frame_num = 0;
1923                         current_IDR_display = current_frame_display;
1924                 }
1925
1926                 {
1927                         unique_lock<mutex> lock(frame_queue_mutex);
1928                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
1929                         if (copy_thread_should_quit) return;
1930                         frame = move(pending_video_frames[current_frame_display]);
1931                         pending_video_frames.erase(current_frame_display);
1932                 }
1933
1934                 // Wait for the GPU to be done with the frame.
1935                 glClientWaitSync(frame.fence.get(), 0, 0);
1936
1937                 // Release back any input frames we needed to render this frame.
1938                 frame.input_frames.clear();
1939
1940                 // Unmap the image.
1941                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
1942                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1943                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1944                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1945                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1946                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1947                 CHECK_VASTATUS(va_status, "vaDestroyImage");
1948
1949                 VASurfaceID surface = surf->src_surface;
1950
1951                 // Schedule the frame for encoding.
1952                 va_status = vaBeginPicture(va_dpy, context_id, surface);
1953                 CHECK_VASTATUS(va_status, "vaBeginPicture");
1954
1955                 if (current_frame_type == FRAME_IDR) {
1956                         render_sequence();
1957                         render_picture();
1958                         if (h264_packedheader) {
1959                                 render_packedsequence();
1960                                 render_packedpicture();
1961                         }
1962                 } else {
1963                         //render_sequence();
1964                         render_picture();
1965                 }
1966                 render_slice();
1967
1968                 va_status = vaEndPicture(va_dpy, context_id);
1969                 CHECK_VASTATUS(va_status, "vaEndPicture");
1970
1971                 // Determine the pts and dts of this frame.
1972                 int64_t pts = frame.pts;
1973                 int64_t dts;
1974                 if (pts_lag == -1) {
1975                         assert(last_dts != -1);
1976                         dts = last_dts + (TIMEBASE / MAX_FPS);
1977                 } else {
1978                         dts = pts - pts_lag;
1979                 }
1980                 last_dts = dts;
1981
1982                 // so now the data is done encoding (well, async job kicked off)...
1983                 // we send that to the storage thread
1984                 storage_task tmp;
1985                 tmp.display_order = current_frame_display;
1986                 tmp.frame_type = current_frame_type;
1987                 tmp.pts = pts;
1988                 tmp.dts = dts;
1989                 storage_task_enqueue(move(tmp));
1990
1991                 update_ReferenceFrames();
1992                 ++current_frame_encoding;
1993         }
1994 }